Authors: Wei Zhou, Bolei Ma, Annemarie Friedrich, Mohsen Mesgar
Abstract: Table Question Answering (TQA) aims to answer natural language questions about tabular data, often accompanied by additional contexts such as text passages. The task spans diverse settings, varying in table representation, question/answer complexity, modality involved, and domain. While recent advances in large language models (LLMs) have led to substantial progress in TQA, the field still lacks a systematic organization and understanding of task formulations, core challenges, and methodological trends, particularly in light of emerging research directions such as reinforcement learning. This survey addresses this gap by providing a comprehensive and structured overview of TQA research with a focus on LLM-based methods. We provide a comprehensive categorization of existing benchmarks and task setups. We group current modeling strategies according to the challenges they target, and analyze their strengths and limitations. Furthermore, we highlight underexplored but timely topics that have not been systematically covered in prior research. By unifying disparate research threads and identifying open problems, our survey offers a consolidated foundation for the TQA community, enabling a deeper understanding of the state of the art and guiding future developments in this rapidly evolving area.
Authors: Yanran Chen, Lynn Greschner, Roman Klinger, Michael Klenk, Steffen Eger
Abstract: Logical fallacies are common in public communication and can mislead audiences; fallacious arguments may still appear convincing despite lacking soundness, because convincingness is inherently subjective. We present the first computational study of how emotional framing interacts with fallacies and convincingness, using large language models (LLMs) to systematically change emotional appeals in fallacious arguments. We benchmark eight LLMs on injecting emotional appeal into fallacious arguments while preserving their logical structures, then use the best models to generate stimuli for a human study. Our results show that LLM-driven emotional framing reduces human fallacy detection in F1 by 14.5% on average. Humans perform better in fallacy detection when perceiving enjoyment than fear or sadness, and these three emotions also correlate with significantly higher convincingness compared to neutral or other emotion states. Our work has implications for AI-driven emotional manipulation in the context of fallacious argumentation.
Authors: Shin-nosuke Ishikawa, Masato Todo, Taiki Ogihara, Hirotsugu Ohba
Abstract: We present a tendency of large language models (LLMs) to generate absurd patterns despite their clear inappropriateness in a simple task of identifying regularities in number series. Several approaches have been proposed to apply LLMs to complex real-world tasks, such as providing knowledge through retrieval-augmented generation and executing multi-step tasks using AI agent frameworks. However, these approaches rely on the logical consistency and self-coherence of LLMs, making it crucial to evaluate these aspects and consider potential countermeasures. To identify cases where LLMs fail to maintain logical consistency, we conducted an experiment in which LLMs were asked to explain the patterns in various integer sequences, ranging from arithmetic sequences to randomly generated integer series. While the models successfully identified correct patterns in arithmetic and geometric sequences, they frequently over-recognized patterns that were inconsistent with the given numbers when analyzing randomly generated series. This issue was observed even in multi-step reasoning models, including OpenAI o3, o4-mini, and Google Gemini 2.5 Flash Preview Thinking. This tendency to perceive non-existent patterns can be interpreted as the AI model equivalent of Idola Tribus and highlights potential limitations in their capability for applied tasks requiring logical reasoning, even when employing chain-of-thought reasoning mechanisms.
Authors: Xiaonan Si, Meilin Zhu, Simeng Qin, Lijia Yu, Lijun Zhang, Shuaitong Liu, Xinfeng Li, Ranjie Duan, Yang Liu, Xiaojun Jia
Abstract: Retrieval-augmented generation (RAG) systems enhance large language models (LLMs) with external knowledge but are vulnerable to corpus poisoning and contamination attacks, which can compromise output integrity. Existing defenses often apply aggressive filtering, leading to unnecessary loss of valuable information and reduced reliability in generation. To address this problem, we propose a two-stage semantic filtering and conflict-free framework for trustworthy RAG. In the first stage, we perform a joint filter with semantic and cluster-based filtering which is guided by the Entity-intent-relation extractor (EIRE). EIRE extracts entities, latent objectives, and entity relations from both the user query and filtered documents, scores their semantic relevance, and selectively adds valuable documents into the clean retrieval database. In the second stage, we proposed an EIRE-guided conflict-aware filtering module, which analyzes semantic consistency between the query, candidate answers, and retrieved knowledge before final answer generation, filtering out internal and external contradictions that could mislead the model. Through this two-stage process, SeCon-RAG effectively preserves useful knowledge while mitigating conflict contamination, achieving significant improvements in both generation robustness and output trustworthiness. Extensive experiments across various LLMs and datasets demonstrate that the proposed SeCon-RAG markedly outperforms state-of-the-art defense methods.
Authors: Wenbin Guo, Xin Wang, Jiaoyan Chen, Lingbing Guo, Zhao Li, Zirui Chen
Abstract: Large Language Models (LLMs) have recently emerged as a powerful paradigm for Knowledge Graph Completion (KGC), offering strong reasoning and generalization capabilities beyond traditional embedding-based approaches. However, existing LLM-based methods often struggle to fully exploit structured semantic representations, as the continuous embedding space of pretrained KG models is fundamentally misaligned with the discrete token space of LLMs. This discrepancy hinders effective semantic transfer and limits their performance. To address this challenge, we propose ReaLM, a novel and effective framework that bridges the gap between KG embeddings and LLM tokenization through the mechanism of residual vector quantization. ReaLM discretizes pretrained KG embeddings into compact code sequences and integrates them as learnable tokens within the LLM vocabulary, enabling seamless fusion of symbolic and contextual knowledge. Furthermore, we incorporate ontology-guided class constraints to enforce semantic consistency, refining entity predictions based on class-level compatibility. Extensive experiments on two widely used benchmark datasets demonstrate that ReaLM achieves state-of-the-art performance, confirming its effectiveness in aligning structured knowledge with large-scale language models.
Authors: Shiyuan Guo, Henry Sleight, Fabien Roger
Abstract: Detecting harmful AI actions is important as AI agents gain adoption. Chain-of-thought (CoT) monitoring is one method widely used to detect adversarial attacks and AI misalignment. However, attackers and misaligned models might evade CoT monitoring through ciphered reasoning: reasoning hidden in encrypted, translated, or compressed text. To assess this risk, we test whether models can perform ciphered reasoning. For each of 28 different ciphers, we fine-tune and prompt up to 10 models to reason in that cipher. We measure model accuracy on math problems as a proxy for reasoning ability. Across the models we test, we find an asymmetry: model accuracy can drop significantly when reasoning in ciphered text, even though models demonstrate comprehension of ciphered text by being able to translate it accurately to English. Even frontier models struggle with lesser-known ciphers, although they can reason accurately in well-known ciphers like rot13. We show that ciphered reasoning capability correlates with cipher prevalence in pretraining data. We also identify scaling laws showing that ciphered reasoning capability improves slowly with additional fine-tuning data. Our work suggests that evading CoT monitoring using ciphered reasoning may be an ineffective tactic for current models and offers guidance on constraining the development of this capability in future frontier models.
Authors: Haoran Sun, Zekun Zhang, Shaoning Zeng
Abstract: One of the key factors influencing the reasoning capabilities of LLM-based agents is their ability to leverage long-term memory. Integrating long-term memory mechanisms allows agents to make informed decisions grounded in historical interactions. While recent advances have significantly improved the storage and retrieval components, by encoding memory into dense vectors for similarity search or organizing memory as structured knowledge graphs most existing approaches fall short in memory updating. In particular, they lack mechanisms for dynamically refining preference memory representations in response to evolving user behaviors and contexts. To address this gap, we propose a Preference-Aware Memory Update Mechanism (PAMU) that enables dynamic and personalized memory refinement. By integrating sliding window averages (SW) with exponential moving averages (EMA), PAMU constructs a fused preference-aware representation that captures both short-term fluctuations and long-term user tendencies. We conduct experiments on five task scenarios of the LoCoMo dataset, and the results show that our mechanism can significantly improve the output quality of LLM in five baselines, validating its effectiveness in long-term conversations.
Authors: Fanwei Zhu, Jinke Yu, Zulong Chen, Ying Zhou, Junhao Ji, Zhibo Yang, Yuxue Zhang, Haoyuan Hu, Zhenghao Liu
Abstract: Automated resume information extraction is critical for scaling talent acquisition, yet its real-world deployment faces three major challenges: the extreme heterogeneity of resume layouts and content, the high cost and latency of large language models (LLMs), and the lack of standardized datasets and evaluation tools. In this work, we present a layout-aware and efficiency-optimized framework for automated extraction and evaluation that addresses all three challenges. Our system combines a fine-tuned layout parser to normalize diverse document formats, an inference-efficient LLM extractor based on parallel prompting and instruction tuning, and a robust two-stage automated evaluation framework supported by new benchmark datasets. Extensive experiments show that our framework significantly outperforms strong baselines in both accuracy and efficiency. In particular, we demonstrate that a fine-tuned compact 0.6B LLM achieves top-tier accuracy while significantly reducing inference latency and computational cost. The system is fully deployed in Alibaba's intelligent HR platform, supporting real-time applications across its business units.
Authors: Yubo Sun, Chunyi Peng, Yukun Yan, Shi Yu, Zhenghao Liu, Chi Chen, Zhiyuan Liu, Maosong Sun
Abstract: Visual retrieval-augmented generation (VRAG) augments vision-language models (VLMs) with external visual knowledge to ground reasoning and reduce hallucinations. Yet current VRAG systems often fail to reliably perceive and integrate evidence across multiple images, leading to weak grounding and erroneous conclusions. In this paper, we propose EVisRAG, an end-to-end framework that learns to reason with evidence-guided multi-image to address this issue. The model first observes retrieved images and records per-image evidence, then derives the final answer from the aggregated evidence. To train EVisRAG effectively, we introduce Reward-Scoped Group Relative Policy Optimization (RS-GRPO), which binds fine-grained rewards to scope-specific tokens to jointly optimize visual perception and reasoning abilities of VLMs. Experimental results on multiple visual question answering benchmarks demonstrate that EVisRAG delivers substantial end-to-end gains over backbone VLM with 27\% improvements on average. Further analysis shows that, powered by RS-GRPO, EVisRAG improves answer accuracy by precisely perceiving and localizing question-relevant evidence across multiple images and deriving the final answer from that evidence, much like a real detective.
Authors: Steve Han, Gilberto Titericz Junior, Tom Balough, Wenfei Zhou
Abstract: This research introduces the Judge's Verdict Benchmark, a novel two-step methodology to evaluate Large Language Models (LLMs) as judges for response accuracy evaluation tasks. We assess how well 54 LLMs can replicate human judgment when scoring responses from RAG (Retrieval-Augmented Generation) or Agentic pipelines against ground truth answers. Our methodology progresses from traditional correlation analysis to comprehensive Cohen's Kappa analysis that measures actual agreement patterns. The two-step approach includes: (1) a correlation test that filters judges with strong alignment, followed by (2) a human-likeness test using z-scores to identify two distinct judgment patterns: human-like judgment (|z| < 1) that mimics natural human variation, and super-consistent judgment (z > 1) that exceeds typical human-to-human agreement levels. This methodology reveals that 27 out of 54 tested LLMs achieve Tier 1 performance: 23 models exhibit human-like patterns that preserve the nuances of human judgment, while 4 models demonstrate super-consistent behavior, a pattern that could indicate either enhanced reliability or oversimplification of complex judgments. Testing 43 open-source models (1B-405B parameters) and 11 closed models (GPT, Gemini, Claude variants), we demonstrate that judge excellence is not solely dependent on model size but on specific training strategies. Our key contributions include: (1) establishing that correlation alone is insufficient for judge evaluation, (2) introducing a "Turing Test for judges" based on agreement patterns, and (3) providing a standardized benchmark for classifying LLM judges into distinct performance tiers for different evaluation needs.
Authors: Adam Byerly, Daniel Khashabi
Abstract: Large language models exhibit a strong position bias in multi-document contexts, systematically prioritizing information based on location rather than relevance. While existing approaches treat this bias as noise to be mitigated, we introduce Gold Panning Bandits, a framework that leverages position bias as a diagnostic signal: by reordering documents and observing shifts in the model's responses, we can efficiently identify the most relevant content. We frame the problem of choosing reorderings as a bipartite matching problem. While an optimal assignment can be computed at each iteration with the Hungarian algorithm in $O(N^3)$ time, we propose a greedy $O(N \log N)$ strategy that achieves comparable performance by prioritizing the placement of the most uncertain documents in the most informative positions. Our approach identifies relevant documents using up to 65\% fewer language model queries than random permutation baselines on knowledge-intensive NLP tasks, substantially reducing computational cost without model retraining. This work demonstrates that inherent LLM biases can be transformed from liabilities into assets for efficient, inference-time optimization.
Authors: Rakib Hossan, Shubhashis Roy Dipta
Abstract: The BLP-2025 Task 1A requires Bengali hate speech classification into six categories. Traditional supervised approaches need extensive labeled datasets that are expensive for low-resource languages. We developed PromptGuard, a few-shot framework combining chi-square statistical analysis for keyword extraction with adaptive majority voting for decision-making. We explore statistical keyword selection versus random approaches and adaptive voting mechanisms that extend classification based on consensus quality. Chi-square keywords provide consistent improvements across categories, while adaptive voting benefits ambiguous cases requiring extended classification rounds. PromptGuard achieves a micro-F1 of 67.61, outperforming n-gram baselines (60.75) and random approaches (14.65). Ablation studies confirm chi-square-based keywords show the most consistent impact across all categories.
Authors: Michael Freenor, Lauren Alvarez
Abstract: Understanding how language and embedding models encode semantic relationships is fundamental to model interpretability and control. While early word embeddings exhibited intuitive vector arithmetic (''king'' - ''man'' + ''woman'' = ''queen''), modern high-dimensional text representations lack straightforward interpretable geometric properties. We introduce Rotor-Invariant Shift Estimation (RISE), a geometric approach that represents semantic transformations as consistent rotational operations in embedding space, leveraging the manifold structure of modern language representations. RISE operations have the ability to operate across both languages and models with high transfer of performance, suggesting the existence of analogous cross-lingual geometric structure. We evaluate RISE across three embedding models, three datasets, and seven morphologically diverse languages in five major language groups. Our results demonstrate that RISE consistently maps discourse-level semantic transformations with distinct grammatical features (e.g., negation and conditionality) across languages and models. This work provides the first systematic demonstration that discourse-level semantic transformations correspond to consistent geometric operations in multilingual embedding spaces, empirically supporting the Linear Representation Hypothesis at the sentence level.
Authors: Ruizhe Zhu
Abstract: The widespread application of large vision language models has significantly raised safety concerns. In this project, we investigate text prompt injection, a simple yet effective method to mislead these models. We developed an algorithm for this type of attack and demonstrated its effectiveness and efficiency through experiments. Compared to other attack methods, our approach is particularly effective for large models without high demand for computational resources.
Authors: Kaiwen Shi, Zheyuan Zhang, Zhengqing Yuan, Keerthiram Murugesan, Vincent Galass, Chuxu Zhang, Yanfang Ye
Abstract: Diet plays a central role in human health, and Nutrition Question Answering (QA) offers a promising path toward personalized dietary guidance and the prevention of diet-related chronic diseases. However, existing methods face two fundamental challenges: the limited reasoning capacity of single-agent systems and the complexity of designing effective multi-agent architectures, as well as contextual overload that hinders accurate decision-making. We introduce Nutritional-Graph Router (NG-Router), a novel framework that formulates nutritional QA as a supervised, knowledge-graph-guided multi-agent collaboration problem. NG-Router integrates agent nodes into heterogeneous knowledge graphs and employs a graph neural network to learn task-aware routing distributions over agents, leveraging soft supervision derived from empirical agent performance. To further address contextual overload, we propose a gradient-based subgraph retrieval mechanism that identifies salient evidence during training, thereby enhancing multi-hop and relational reasoning. Extensive experiments across multiple benchmarks and backbone models demonstrate that NG-Router consistently outperforms both single-agent and ensemble baselines, offering a principled approach to domain-aware multi-agent reasoning for complex nutritional health tasks.
Authors: Sil Hamilton, Matthew Wilkens, Andrew Piper
Abstract: We present NarraBench, a theory-informed taxonomy of narrative-understanding tasks, as well as an associated survey of 78 existing benchmarks in the area. We find significant need for new evaluations covering aspects of narrative understanding that are either overlooked in current work or are poorly aligned with existing metrics. Specifically, we estimate that only 27% of narrative tasks are well captured by existing benchmarks, and we note that some areas -- including narrative events, style, perspective, and revelation -- are nearly absent from current evaluations. We also note the need for increased development of benchmarks capable of assessing constitutively subjective and perspectival aspects of narrative, that is, aspects for which there is generally no single correct answer. Our taxonomy, survey, and methodology are of value to NLP researchers seeking to test LLM narrative understanding.
Authors: Nafiseh Nikeghbal, Amir Hossein Kargaran, Jana Diesner
Abstract: Improvements in model construction, including fortified safety guardrails, allow Large language models (LLMs) to increasingly pass standard safety checks. However, LLMs sometimes slip into revealing harmful behavior, such as expressing racist viewpoints, during conversations. To analyze this systematically, we introduce CoBia, a suite of lightweight adversarial attacks that allow us to refine the scope of conditions under which LLMs depart from normative or ethical behavior in conversations. CoBia creates a constructed conversation where the model utters a biased claim about a social group. We then evaluate whether the model can recover from the fabricated bias claim and reject biased follow-up questions. We evaluate 11 open-source as well as proprietary LLMs for their outputs related to six socio-demographic categories that are relevant to individual safety and fair treatment, i.e., gender, race, religion, nationality, sex orientation, and others. Our evaluation is based on established LLM-based bias metrics, and we compare the results against human judgments to scope out the LLMs' reliability and alignment. The results suggest that purposefully constructed conversations reliably reveal bias amplification and that LLMs often fail to reject biased follow-up questions during dialogue. This form of stress-testing highlights deeply embedded biases that can be surfaced through interaction. Code and artifacts are available at https://github.com/nafisenik/CoBia.
Authors: Vishal Anand, Milad Alshomary, Kathleen McKeown
Abstract: We present iBERT (interpretable-BERT), an encoder to produce inherently interpretable and controllable embeddings - designed to modularize and expose the discriminative cues present in language, such as stylistic and semantic structure. Each input token is represented as a sparse, non-negative mixture over k context-independent sense vectors, which can be pooled into sentence embeddings or used directly at the token level. This enables modular control over representation, before any decoding or downstream use. To demonstrate our model's interpretability, we evaluate it on a suite of style-focused tasks. On the STEL benchmark, it improves style representation effectiveness by ~8 points over SBERT-style baselines, while maintaining competitive performance on authorship verification. Because each embedding is a structured composition of interpretable senses, we highlight how specific style attributes - such as emoji use, formality, or misspelling can be assigned to specific sense vectors. While our experiments center on style, iBERT is not limited to stylistic modeling. Its structural modularity is designed to interpretably decompose whichever discriminative signals are present in the data - enabling generalization even when supervision blends stylistic and semantic factors.
Authors: Hossein Entezari Zarch, Lei Gao, Chaoyi Jiang, Murali Annavarm
Abstract: Large reasoning models (LRMs) achieve state-of-the-art performance on challenging benchmarks by generating long chains of intermediate steps, but their inference cost is dominated by decoding, where each new token must attend to the entire growing sequence. Existing sparse attention methods reduce computation by pruning the key-value (KV) cache, yet they suffer from severe accuracy degradation on reasoning tasks due to cumulative selection errors and the dynamic importance of tokens over long derivations. We present \textbf{DELTA}, a training-free sparse attention mechanism that achieves computational efficiency without sacrificing model accuracy. DELTA partitions transformer layers into three groups: initial layers that use full attention, a small set of \emph{selection layers} that identify salient tokens via aggregated head-level attention scores, and subsequent \emph{sparse-attention layers} that attend only to the selected subset. This design preserves the full KV cache in GPU memory for accuracy, while avoiding expensive full-attention computation over many layers. On reasoning benchmarks such as AIME and GPQA-Diamond, DELTA matches or surpasses full attention in accuracy, while reducing the number of attended tokens by up to $5\times$ and delivering $1.5\times$ end-to-end speedup. Our results show that selective reuse of intermediate attention maps offers a robust path toward efficient long-context reasoning.
Authors: Xu Pan, Ely Hahami, Jingxuan Fan, Ziqian Xie, Haim Sompolinsky
Abstract: Despite autoregressive large language models (arLLMs) being the current dominant paradigm in language modeling, they resist knowledge injection via fine-tuning due to inherent shortcomings such as the "reversal curse" -- the challenge of answering questions that reverse the original information order in the training sample. Masked diffusion large language models (dLLMs) are rapidly emerging as a powerful alternative to the arLLM paradigm, with evidence of better data efficiency and free of the "reversal curse" in pre-training. However, it is unknown whether these advantages extend to the post-training phase, i.e. whether pre-trained dLLMs can easily acquire new knowledge through fine-tuning. On three diverse datasets, we fine-tune arLLMs and dLLMs, evaluating them with forward and backward style Question Answering (QA) to probe knowledge generalization and the reversal curse. Our results confirm that arLLMs critically rely on extensive data augmentation via paraphrases for QA generalization, and paraphrases are only effective when their information order matches the QA style. Conversely, dLLMs achieve high accuracies on both forward and backward QAs without paraphrases; adding paraphrases yields only marginal gains. Lastly, inspired by the dLLM's performance, we introduce a novel masked fine-tuning paradigm for knowledge injection into pre-trained arLLMs. This proposed method successfully and drastically improves the data efficiency of arLLM fine-tuning, effectively closing the performance gap with dLLMs.
Authors: Yijin Ni, Peng Qi
Abstract: Frontier large language models such as GPT-5 and Claude Sonnet remain prone to overconfidence even after alignment through Reinforcement Learning with Human Feedback (RLHF) and Direct Preference Optimization (DPO). For instance, they tend to offer the same conservative answer "No" to both questions "Can I eat the [food / potato chips] that has been left out overnight?" despite the latter requiring no refridgeration for safe consumption. We find that this failure is potentially attributed to a limitation of existing preference learning: it emphasizes selecting the correct response for a given prompt, while neglecting counterfactual prompts that should alter the response. To address this limitation, we propose abductive preference learning, a fine-tuning paradigm that reverses the conventional conditioning by learning preferences over prompts given a response. To validate this idea, we construct an abductive dataset derived from the HaluEval QA benchmark with 1,001 entries, implementing abductive DPO and its variant DPOP. Experiments reveal complementary strengths: standard methods improve response selection, abductive methods improve prompt discrimination, while a multitask objective unifies both. On the abductive dataset, multitask DPOP boosts accuracy from $90.0\%$ to $99.5\%$ in response selection and $54.7\%$ to $85.0\%$ in prompt discrimination, with qualitative evidence highlighting improved sensitivity to prompt differences. Finally, evaluation on AlpacaEval shows multitask DPOP improves win rate (from $5.26\%$ to $6.17\%$), confirming that abductive preference learning preserves the benefits of conventional preference optimization while addressing the overlooked challenge of counterfactual prompts.
Authors: Guanming Chen, Lingzhi Shen, Xiaohao Cai, Imran Razzak, Shoaib Jameel
Abstract: Personality detection from text aims to infer an individual's personality traits based on linguistic patterns. However, existing machine learning approaches often struggle to capture contextual information spanning multiple posts and tend to fall short in extracting representative and robust features in semantically sparse environments. This paper presents HIPPD, a brain-inspired framework for personality detection that emulates the hierarchical information processing of the human brain. HIPPD utilises a large language model to simulate the cerebral cortex, enabling global semantic reasoning and deep feature abstraction. A dynamic memory module, modelled after the prefrontal cortex, performs adaptive gating and selective retention of critical features, with all adjustments driven by dopaminergic prediction error feedback. Subsequently, a set of specialised lightweight models, emulating the basal ganglia, are dynamically routed via a strict winner-takes-all mechanism to capture the personality-related patterns they are most proficient at recognising. Extensive experiments on the Kaggle and Pandora datasets demonstrate that HIPPD consistently outperforms state-of-the-art baselines.
Authors: Shangbin Feng, Wenhao Yu, Yike Wang, Hongming Zhang, Yulia Tsvetkov, Dong Yu
Abstract: Alignment training has tradeoffs: it helps language models (LMs) gain in reasoning and instruction following but might lose out on skills such as creativity and calibration, where unaligned base models are better at. We aim to make the best of both worlds through model collaboration, where different models in the training pipeline collaborate and complement each other. Since LM responses feature interleaving skills that favor different models, we propose Switch Generation, where pretrained and aligned model versions take turns to ``speak'' in a response sequence. Specifically, we train a switcher LM by learning from outcomes of choosing different models to generate the next segment across diverse queries and contexts. At inference time, the switcher LM guides different model checkpoints to dynamically generate the next segment where their strengths are most needed. Extensive experiments with 8 model collaboration baselines and 18 datasets show that 1) model collaboration consistently outperforms individual models on 16 out of 18 tasks, and 2) Switch Generation further outperforms baselines by 12.9% on average. Further analysis reveals that Switch Generation discovers compositional skills to solve problems where individual models struggle and generalizes to unseen models and tasks, reusing and repurposing by-products in expensive model training pipelines that are otherwise discarded.
Authors: Sicong Huang, Qianqi Yan, Shengze Wang, Ian Lane
Abstract: Abstractive summarization using large language models (LLMs) has become an essential tool for condensing information. However, despite their ability to generate fluent summaries, these models sometimes produce unfaithful summaries, introducing hallucinations at the word, phrase, or concept level. Existing mitigation strategies, such as post-processing corrections or contrastive learning with synthetically generated negative samples, fail to fully address the diverse errors that can occur in LLM-generated summaries. In this paper, we investigate fine-tuning strategies to reduce the occurrence of unfaithful spans in generated summaries. First, we automatically generate summaries for the set of source documents in the training set with a variety of LLMs and then use GPT-4o to annotate any hallucinations it detects at the span-level. Leveraging these annotations, we fine-tune LLMs with both hallucination-free summaries and annotated unfaithful spans to enhance model faithfulness. In this paper, we introduce a new dataset that contains both faithful and unfaithful summaries with span-level labels and we evaluate three techniques to fine-tuning a LLM to improve the faithfulness of the resulting summarization: gradient ascent, unlikelihood training, and task vector negation. Experimental results show that all three approaches successfully leverage span-level annotations to improve faithfulness, with unlikelihood training being the most effective.
Authors: Weibin Cai, Jiayu Li, Reza Zafarani
Abstract: While memes are often humorous, they are frequently used to disseminate hate, causing serious harm to individuals and society. Current approaches to hateful meme detection mainly rely on pre-trained language models. However, less focus has been dedicated to \textit{what make a meme hateful}. Drawing on insights from philosophy and psychology, we argue that hateful memes are characterized by two essential features: a \textbf{presupposed context} and the expression of \textbf{false claims}. To capture presupposed context, we develop \textbf{PCM} for modeling contextual information across modalities. To detect false claims, we introduce the \textbf{FACT} module, which integrates external knowledge and harnesses cross-modal reference graphs. By combining PCM and FACT, we introduce \textbf{\textsf{SHIELD}}, a hateful meme detection framework designed to capture the fundamental nature of hate. Extensive experiments show that SHIELD outperforms state-of-the-art methods across datasets and metrics, while demonstrating versatility on other tasks, such as fake news detection.
Authors: Mir Tafseer Nayeem, Sawsan Alqahtani, Md Tahmid Rahman Laskar, Tasnim Mohiuddin, M Saiful Bari
Abstract: Tokenization is a crucial but under-evaluated step in large language models (LLMs). The standard metric, fertility (the average number of tokens per word), captures compression efficiency but obscures how vocabularies are allocated across languages and domains. We analyze six widely used tokenizers across seven languages and two domains, finding stable fertility for English, high fertility for Chinese, and little domain sensitivity. To address fertility's blind spots, we propose the Single Token Retention Rate (STRR), which measures the proportion of words preserved as single tokens. STRR reveals systematic prioritization of English, strong support for Chinese, and fragmentation in Hindi, offering an interpretable view of cross-lingual fairness. Our results show that STRR complements fertility and provides practical guidance for designing more equitable multilingual tokenizers.
Authors: Jiaqi Wei, Xiang Zhang, Yuejin Yang, Wenxuan Huang, Juntai Cao, Sheng Xu, Xiang Zhuang, Zhangyang Gao, Muhammad Abdul-Mageed, Laks V. S. Lakshmanan, Chenyu You, Wanli Ouyang, Siqi Sun
Abstract: Deliberative tree search is a cornerstone of modern Large Language Model (LLM) research, driving the pivot from brute-force scaling toward algorithmic efficiency. This single paradigm unifies two critical frontiers: \textbf{Test-Time Scaling (TTS)}, which deploys on-demand computation to solve hard problems, and \textbf{Self-Improvement}, which uses search-generated data to durably enhance model parameters. However, this burgeoning field is fragmented and lacks a common formalism, particularly concerning the ambiguous role of the reward signal -- is it a transient heuristic or a durable learning target? This paper resolves this ambiguity by introducing a unified framework that deconstructs search algorithms into three core components: the \emph{Search Mechanism}, \emph{Reward Formulation}, and \emph{Transition Function}. We establish a formal distinction between transient \textbf{Search Guidance} for TTS and durable \textbf{Parametric Reward Modeling} for Self-Improvement. Building on this formalism, we introduce a component-centric taxonomy, synthesize the state-of-the-art, and chart a research roadmap toward more systematic progress in creating autonomous, self-improving agents.
Authors: Yimin Xiao, Yongle Zhang, Dayeon Ki, Calvin Bao, Marianna J. Martindale, Charlotte Vaughn, Ge Gao, Marine Carpuat
Abstract: As Machine Translation (MT) becomes increasingly commonplace, understanding how the general public perceives and relies on imperfect MT is crucial for contextualizing MT research in real-world applications. We present a human study conducted in a public museum (n=452), investigating how fluency and adequacy errors impact bilingual and non-bilingual users' reliance on MT during casual use. Our findings reveal that non-bilingual users often over-rely on MT due to a lack of evaluation strategies and alternatives, while experiencing the impact of errors can prompt users to reassess future reliance. This highlights the need for MT evaluation and NLP explanation techniques to promote not only MT quality, but also MT literacy among its users.
Authors: Jianjin Wang, Runsong Zhao, Xiaoqian Liu, Yuan Ge, Ziqiang Xu, Tong Xiao, Shengxiang Gao, Zhengtao Yu, Jingbo Zhu
Abstract: Current direct speech-to-speech translation methods predominantly employ speech tokens as intermediate representations. However, a single speech token is not dense in semantics, so we generally need multiple tokens to express a complete semantic unit. To address this limitation, we introduce multi-token prediction (MTP) loss into speech-to-unit translation (S2UT) models, enabling models to predict multiple subsequent tokens at each position, thereby capturing more complete semantics and enhancing information density per position. Initial MTP implementations apply the loss at the final layer, which improves output representation but initiates information enrichment too late. We hypothesize that advancing the information enrichment process to intermediate layers can achieve earlier and more effective enhancement of hidden representation. Consequently, we propose MTP-S2UT loss, applying MTP loss to hidden representation where CTC loss is computed. Experiments demonstrate that all MTP loss variants consistently improve the quality of S2UT translation, with MTP-S2UT achieving the best performance.
Authors: Shu Zhao, Tan Yu, Anbang Xu
Abstract: Reasoning-augmented search agents, such as Search-R1, are trained to reason, search, and generate the final answer iteratively. Nevertheless, due to their limited capabilities in reasoning and search, their performance on multi-hop QA benchmarks remains far from satisfactory. To handle complex or compound queries, we train an LLM-based search agent with the native capability of query expansion through reinforcement learning. In each turn, our search agent proposes several query variants, which are searched simultaneously to cover more relevant information. Meanwhile, given limited post-training data and computing resources, it is very challenging for a search agent to master multiple tasks, including query generation, retrieved information understanding, and answer generation. Therefore, we propose incorporating a pre-trained squeezer model that helps the search agent understand the retrieved documents, allowing the search agent to focus on query generation for high retrieval recall. With the assistance of the squeezer model, we discover that even a small-scale 3B LLM can demonstrate a strong capability of query expansion and achieve state-of-the-art accuracy on the multi-hop QA benchmarks. To be specific, our experiments across seven question-answering benchmarks demonstrate that our method, named ExpandSearch, achieves an average improvement of 4.4% compared to state-of-the-art baselines, with strong gains on multi-hop reasoning tasks requiring diverse evidence aggregation.
Authors: Yuyi Huang, Runzhe Zhan, Lidia S. Chao, Ailin Tao, Derek F. Wong
Abstract: As large language models (LLMs) are increasingly deployed for complex reasoning tasks, Long Chain-of-Thought (Long-CoT) prompting has emerged as a key paradigm for structured inference. Despite early-stage safeguards enabled by alignment techniques such as RLHF, we identify a previously underexplored vulnerability: reasoning trajectories in Long-CoT models can drift from aligned paths, resulting in content that violates safety constraints. We term this phenomenon Path Drift. Through empirical analysis, we uncover three behavioral triggers of Path Drift: (1) first-person commitments that induce goal-driven reasoning that delays refusal signals; (2) ethical evaporation, where surface-level disclaimers bypass alignment checkpoints; (3) condition chain escalation, where layered cues progressively steer models toward unsafe completions. Building on these insights, we introduce a three-stage Path Drift Induction Framework comprising cognitive load amplification, self-role priming, and condition chain hijacking. Each stage independently reduces refusal rates, while their combination further compounds the effect. To mitigate these risks, we propose a path-level defense strategy incorporating role attribution correction and metacognitive reflection (reflective safety cues). Our findings highlight the need for trajectory-level alignment oversight in long-form reasoning beyond token-level alignment.
Authors: Jiaqi Liu, Lanruo Wang, Su Liu, Xin Hu
Abstract: Large language models work well for many NLP tasks, but they are hard to deploy in health settings with strict cost, latency, and privacy limits. We revisit a lightweight recipe for medical abstract classification and ask how far compact encoders can go under a controlled budget. Using the public medical abstracts corpus, we finetune BERT base and DistilBERT with three objectives standard cross-entropy, class weighted cross entropy, and focal loss keeping tokenizer, sequence length, optimizer, and schedule fixed. DistilBERT with plain cross-entropy gives the best balance on the test set while using far fewer parameters than BERT base. We report accuracy, Macro F1, and Weighted F1, release the evaluation code, and include confusion analyses to make error patterns clear. Our results suggest a practical default: start with a compact encoder and cross-entropy, then add calibration and task-specific checks before moving to heavier models.
Authors: Adnan El Assadi, Isaac Chung, Roman Solomatin, Niklas Muennighoff, Kenneth Enevoldsen
Abstract: Comparing human and model performance offers a valuable perspective for understanding the strengths and limitations of embedding models, highlighting where they succeed and where they fail to capture meaning and nuance. However, such comparisons are rarely made, as human performance on embedding tasks is difficult to measure. To fill this gap, we introduce HUME: Human Evaluation Framework for Text Embeddings. While frameworks like MTEB provide broad model evaluation, they lack reliable estimates of human performance, limiting the interpretability of model scores. We measure human performance across 16 MTEB datasets spanning reranking, classification, clustering, and semantic textual similarity across linguistically diverse high- and low-resource languages. Humans achieve an average performance of 77.6% compared to 80.1% for the best embedding model, although variation is substantial: models reach near-ceiling performance on some datasets while struggling on others, suggesting dataset issues and revealing shortcomings in low-resource languages. We provide human performance baselines, insight into task difficulty patterns, and an extensible evaluation framework that enables a more meaningful interpretation of the model and informs the development of both models and benchmarks. Our code, dataset, and leaderboard are publicly available at https://github.com/embeddings-benchmark/mteb.
Authors: Yibo Yang
Abstract: Deep learning has advanced NLP, but interpretability remains limited, especially in healthcare and finance. Concept bottleneck models tie predictions to human concepts in vision, but NLP versions either use binary activations that harm text representations or latent concepts that weaken semantics, and they rarely model dynamic concept interactions such as negation and context. We introduce the Concept Language Model Network (CLMN), a neural-symbolic framework that keeps both performance and interpretability. CLMN represents concepts as continuous, human-readable embeddings and applies fuzzy-logic reasoning to learn adaptive interaction rules that state how concepts affect each other and the final decision. The model augments original text features with concept-aware representations and automatically induces interpretable logic rules. Across multiple datasets and pre-trained language models, CLMN achieves higher accuracy than existing concept-based methods while improving explanation quality. These results show that integrating neural representations with symbolic reasoning in a unified concept space can yield practical, transparent NLP systems.
Authors: Hua Cai, Shuang Zhao, Liang Zhang, Xuli Shen, Qing Xu, Weilin Shen, Zihao Wen, Tianke Ban
Abstract: Reasoning-focused large language models (LLMs) are rapidly evolving across various domains, yet their capabilities in handling complex legal problems remains underexplored. In this paper, we introduce Unilaw-R1, a large language model tailored for legal reasoning. With a lightweight 7-billion parameter scale, Unilaw-R1 significantly reduces deployment cost while effectively tackling three core challenges in the legal domain: insufficient legal knowledge, unreliable reasoning logic, and weak business generalization. To address these issues, we first construct Unilaw-R1-Data, a high-quality dataset containing 17K distilled and screened chain-of-thought (CoT) samples. Based on this, we adopt a two-stage training strategy combining Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), which significantly boosts the performance on complex legal reasoning tasks and supports interpretable decision-making in legal AI applications. To assess legal reasoning ability, we also introduce Unilaw-R1-Eval, a dedicated benchmark designed to evaluate models across single- and multi-choice legal tasks. Unilaw-R1 demonstrates strong results on authoritative benchmarks, outperforming all models of similar scale and achieving performance on par with the much larger DeepSeek-R1-Distill-Qwen-32B (54.9%). Following domain-specific training, it also showed significant gains on LawBench and LexEval, exceeding Qwen-2.5-7B-Instruct (46.6%) by an average margin of 6.6%.
Authors: Wenqing Wang (Huazhong Agricultural University, China), Muhammad Asif Ali (King Abdullah University of Science and Technology, KSA), Ali Shoker (King Abdullah University of Science and Technology, KSA), Ruohan Yang (Huazhong Agricultural University, China), Junyang Chen (Shenzhen University, China), Ying Sha (Huazhong Agricultural University, China), Huan Wang (Huazhong Agricultural University, China)
Abstract: Human preferences are diverse and dynamic, shaped by regional, cultural, and social factors. Existing alignment methods like Direct Preference Optimization (DPO) and its variants often default to majority views, overlooking minority opinions and failing to capture latent user intentions in prompts. To address these limitations, we introduce \underline{\textbf{A}}daptive \textbf{\underline{I}}ntent-driven \textbf{\underline{P}}reference \textbf{\underline{O}}ptimization (\textbf{A-IPO}). Specifically,A-IPO introduces an intention module that infers the latent intent behind each user prompt and explicitly incorporates this inferred intent into the reward function, encouraging stronger alignment between the preferred model's responses and the user's underlying intentions. We demonstrate, both theoretically and empirically, that incorporating an intention--response similarity term increases the preference margin (by a positive shift of $\lambda\,\Delta\mathrm{sim}$ in the log-odds), resulting in clearer separation between preferred and dispreferred responses compared to DPO. For evaluation, we introduce two new benchmarks, Real-pref, Attack-pref along with an extended version of an existing dataset, GlobalOpinionQA-Ext, to assess real-world and adversarial preference alignment. Through explicit modeling of diverse user intents,A-IPO facilitates pluralistic preference optimization while simultaneously enhancing adversarial robustness in preference alignment. Comprehensive empirical evaluation demonstrates that A-IPO consistently surpasses existing baselines, yielding substantial improvements across key metrics: up to +24.8 win-rate and +45.6 Response-Intention Consistency on Real-pref; up to +38.6 Response Similarity and +52.2 Defense Success Rate on Attack-pref; and up to +54.6 Intention Consistency Score on GlobalOpinionQA-Ext.
Authors: Parthiv Chatterjee, Shivam Sonawane, Amey Hengle, Aditya Tanna, Sourish Dasgupta, Tanmoy Chakraborty
Abstract: Document summarization enables efficient extraction of user-relevant content but is inherently shaped by individual subjectivity, making it challenging to identify subjective salient information in multifaceted documents. This complexity underscores the necessity for personalized summarization. However, training models for personalized summarization has so far been challenging, particularly because diverse training data containing both user preference history (i.e., click-skip trajectory) and expected (gold-reference) summaries are scarce. The MS/CAS PENS dataset is a valuable resource but includes only preference history without target summaries, preventing end-to-end supervised learning, and its limited topic-transition diversity further restricts generalization. To address this, we propose $\mathrm{PerAugy}$, a novel cross-trajectory shuffling and summary-content perturbation based data augmentation technique that significantly boosts the accuracy of four state-of-the-art baseline (SOTA) user-encoders commonly used in personalized summarization frameworks (best result: $\text{0.132}$$\uparrow$ w.r.t AUC). We select two such SOTA summarizer frameworks as baselines and observe that when augmented with their corresponding improved user-encoders, they consistently show an increase in personalization (avg. boost: $\text{61.2\%}\uparrow$ w.r.t. PSE-SU4 metric). As a post-hoc analysis of the role of induced diversity in the augmented dataset by \peraugy, we introduce three dataset diversity metrics -- $\mathrm{TP}$, $\mathrm{RTC}$, and \degreed\ to quantify the induced diversity. We find that $\mathrm{TP}$ and $\mathrm{DegreeD}$ strongly correlate with user-encoder performance on the PerAugy-generated dataset across all accuracy metrics, indicating that increased dataset diversity is a key factor driving performance gains.
Authors: Renliang Sun, Wei Cheng, Dawei Li, Haifeng Chen, Wei Wang
Abstract: Chain-of-Thought (CoT) reasoning has driven recent gains of large language models (LLMs) on reasoning-intensive tasks by externalizing intermediate steps. However, excessive or redundant reasoning -- so-called overthinking -- can increase inference costs and lead LLMs toward incorrect conclusions. In this paper, we present REFRAIN ($\underline{REF}$lective-$\underline{R}$edundancy for $\underline{A}$daptive $\underline{IN}$ference), a training-free framework that adaptively determines when to stop reasoning to mitigate overthinking. REFRAIN integrates a two-stage stop discriminator to identify reflective yet redundant reasoning and a sliding-window Upper Confidence Bound (SW-UCB) multi-armed bandit controller to dynamically adjust stopping thresholds according to problem difficulty without supervision or fine-tuning. Across four representative benchmarks and two model families, REFRAIN reduces token usage by 20-55% while maintaining or improving accuracy compared to standard CoT prompting. Extensive ablation and robustness analyses demonstrate its stability across models, scorers, and prompt variations. In summary, our findings highlight when-to-stop as a new and practical axis of test-time scaling -- enabling models to reason not just more, but just enough.
Authors: Luyao Zhuang, Shengyuan Chen, Yilin Xiao, Huachi Zhou, Yujing Zhang, Hao Chen, Qinggang Zhang, Xiao Huang
Abstract: Retrieval-Augmented Generation (RAG) is widely used to mitigate hallucinations of Large Language Models (LLMs) by leveraging external knowledge. While effective for simple queries, traditional RAG systems struggle with large-scale, unstructured corpora where information is fragmented. Recent advances incorporate knowledge graphs to capture relational structures, enabling more comprehensive retrieval for complex, multi-hop reasoning tasks. However, existing graph-based RAG (GraphRAG) methods rely on unstable and costly relation extraction for graph construction, often producing noisy graphs with incorrect or inconsistent relations that degrade retrieval quality. In this paper, we revisit the pipeline of existing GraphRAG systems and propose LinearRAG (Linear Graph-based Retrieval-Augmented Generation), an efficient framework that enables reliable graph construction and precise passage retrieval. Specifically, LinearRAG constructs a relation-free hierarchical graph, termed Tri-Graph, using only lightweight entity extraction and semantic linking, avoiding unstable relation modeling. This new paradigm of graph construction scales linearly with corpus size and incurs no extra token consumption, providing an economical and reliable indexing of the original passages. For retrieval, LinearRAG adopts a two-stage strategy: (i) relevant entity activation via local semantic bridging, followed by (ii) passage retrieval through global importance aggregation. Extensive experiments on four datasets demonstrate that LinearRAG significantly outperforms baseline models.
Authors: Zilong Wang, Xiaoyu Shen
Abstract: Information extraction from copy-heavy documents, characterized by massive volumes of structurally similar content, represents a critical yet understudied challenge in enterprise document processing. We present a systematic framework that strategically combines OCR engines with Large Language Models (LLMs) to optimize the accuracy-efficiency trade-off inherent in repetitive document extraction tasks. Unlike existing approaches that pursue universal solutions, our method exploits document-specific characteristics through intelligent strategy selection. We implement and evaluate 25 configurations across three extraction paradigms (direct, replacement, and table-based) on identity documents spanning four formats (PNG, DOCX, XLSX, PDF). Through table-based extraction methods, our adaptive framework delivers outstanding results: F1=1.0 accuracy with 0.97s latency for structured documents, and F1=0.997 accuracy with 0.6 s for challenging image inputs when integrated with PaddleOCR, all while maintaining sub-second processing speeds. The 54 times performance improvement compared with multimodal methods over naive approaches, coupled with format-aware routing, enables processing of heterogeneous document streams at production scale. Beyond the specific application to identity extraction, this work establishes a general principle: the repetitive nature of copy-heavy tasks can be transformed from a computational burden into an optimization opportunity through structure-aware method selection.
Authors: Tingxu Han, Wei Song, Ziqi Ding, Ziming Li, Chunrong Fang, Yuekang Li, Dongfang Liu, Zhenyu Chen, Zhenting Wang
Abstract: Large language models (LLMs) increasingly mediate decisions in domains where unfair treatment of demographic groups is unacceptable. Existing work probes when biased outputs appear, but gives little insight into the mechanisms that generate them, leaving existing mitigations largely fragile. In this paper, we conduct a systematic investigation LLM unfairness and propose DiffHeads, a lightweight debiasing framework for LLMs. We first compare Direct-Answer (DA) prompting to Chain-of-Thought (CoT) prompting across eight representative open- and closed-source LLMs. DA will trigger the nature bias part of LLM and improve measured unfairness by 534.5%-391.9% in both one-turn and two-turn dialogues. Next, we define a token-to-head contribution score that traces each token's influence back to individual attention heads. This reveals a small cluster of bias heads that activate under DA but stay largely dormant with CoT, providing the first causal link between prompting strategy and bias emergence. Finally, building on this insight, we propose DiffHeads that identifies bias heads through differential activation analysis between DA and CoT, and selectively masks only those heads. DiffHeads reduces unfairness by 49.4%, and 40.3% under DA and CoT, respectively, without harming model utility.
Authors: Tsung-Min Pai, Jui-I Wang, Li-Chun Lu, Shao-Hua Sun, Hung-Yi Lee, Kai-Wei Chang
Abstract: Multi-LLM systems enhance the creativity of large language models by simulating human collective intelligence but suffer from significant drawbacks, such as high computational costs and inference latency. To address these limitations, we propose BILLY (BlendIng persona vectors for Large Language model creativitY), a training-free framework that captures the benefits of multi-LLM collaboration, i.e. inducing diverse perspectives and specialized expertise, within a single model. BILLY operates by extracting and blending multiple distinct persona vectors directly in the model's activation space. We steer the model's generation process with this merged vector while inference, enabling multi-perspective output without explicit multi-LLM communication. Our experiments across creativity-oriented benchmarks demonstrate that BILLY surpasses single model prompting and traditional multi-LLM approaches, while substantially reducing inference time and computational costs. Our analyses further reveal that distinct persona vectors can be blended to achieve both effective control over complementary aspects of generation and greater interpretability.
Authors: Jaap Jumelet, Abdellah Fourtassi, Akari Haga, Bastian Bunzeck, Bhargav Shandilya, Diana Galvan-Sosa, Faiz Ghifari Haznitrama, Francesca Padovani, Francois Meyer, Hai Hu, Julen Etxaniz, Laurent Pr\'evot, Linyang He, Mar\'ia Grandury, Mila Marcheva, Negar Foroutan, Nikitas Theodoropoulos, Pouya Sadeghi, Siyuan Song, Suchir Salhan, Susana Zhou, Yurii Paniv, Ziyin Zhang, Arianna Bisazza, Alex Warstadt, Leshem Choshen
Abstract: We present BabyBabelLM, a multilingual collection of datasets modeling the language a person observes from birth until they acquire a native language. We curate developmentally plausible pretraining data aiming to cover the equivalent of 100M English words of content in each of 45 languages. We compile evaluation suites and train baseline models in each language. BabyBabelLM aims to facilitate multilingual pretraining and cognitive modeling.
Authors: Liang Pang, Kangxi Wu, Sunhao Dai, Zihao Wei, Zenghao Duan, Jia Gu, Xiang Li, Zhiyi Yin, Jun Xu, Huawei Shen, Xueqi Cheng
Abstract: The rapid advancement of large language models (LLMs) has revolutionized artificial intelligence, shifting from supporting objective tasks (e.g., recognition) to empowering subjective decision-making (e.g., planning, decision). This marks the dawn of general and powerful AI, with applications spanning a wide range of fields, including programming, education, healthcare, finance, and law. However, their deployment introduces multifaceted risks. Due to the black-box nature of LLMs and the human-like quality of their generated content, issues such as hallucinations, bias, unfairness, and copyright infringement become particularly significant. In this context, sourcing information from multiple perspectives is essential. This survey presents a systematic investigation into provenance tracking for content generated by LLMs, organized around four interrelated dimensions that together capture both model- and data-centric perspectives. From the model perspective, Model Sourcing treats the model as a whole, aiming to distinguish content generated by specific LLMs from content authored by humans. Model Structure Sourcing delves into the internal generative mechanisms, analyzing architectural components that shape the outputs of model. From the data perspective, Training Data Sourcing focuses on internal attribution, tracing the origins of generated content back to the training data of model. In contrast, External Data Sourcing emphasizes external validation, identifying external information used to support or influence the responses of model. Moreover, we also propose a dual-paradigm taxonomy that classifies existing sourcing methods into prior-based (proactive traceability embedding) and posterior-based (retrospective inference) approaches. Traceability across these dimensions enhances the transparency, accountability, and trustworthiness of LLMs deployment in real-world applications.
Authors: Kedi Chen, Dezhao Ruan, Yuhao Dan, Yaoting Wang, Siyu Yan, Xuecheng Wu, Yinqi Zhang, Qin Chen, Jie Zhou, Liang He, Biqing Qi, Linyang Li, Qipeng Guo, Xiaoming Shi, Wei Zhang
Abstract: Reasoning is an important task for large language models (LLMs). Among all the reasoning paradigms, inductive reasoning is one of the fundamental types, which is characterized by its particular-to-general thinking process and the non-uniqueness of its answers. The inductive mode is crucial for knowledge generalization and aligns better with human cognition, so it is a fundamental mode of learning, hence attracting increasing interest. Despite the importance of inductive reasoning, there is no systematic summary of it. Therefore, this paper presents the first comprehensive survey of inductive reasoning for LLMs. First, methods for improving inductive reasoning are categorized into three main areas: post-training, test-time scaling, and data augmentation. Then, current benchmarks of inductive reasoning are summarized, and a unified sandbox-based evaluation approach with the observation coverage metric is derived. Finally, we offer some analyses regarding the source of inductive ability and how simple model architectures and data help with inductive tasks, providing a solid foundation for future research.
Authors: Lei Gu, Yinghao Zhu, Haoran Sang, Zixiang Wang, Dehao Sui, Wen Tang, Ewen Harrison, Junyi Gao, Lequan Yu, Liantao Ma
Abstract: While large language model (LLM)-based multi-agent systems show promise in simulating medical consultations, their evaluation is often confined to final-answer accuracy. This practice treats their internal collaborative processes as opaque "black boxes" and overlooks a critical question: is a diagnostic conclusion reached through a sound and verifiable reasoning pathway? The inscrutable nature of these systems poses a significant risk in high-stakes medical applications, potentially leading to flawed or untrustworthy conclusions. To address this, we conduct a large-scale empirical study of 3,600 cases from six medical datasets and six representative multi-agent frameworks. Through a rigorous, mixed-methods approach combining qualitative analysis with quantitative auditing, we develop a comprehensive taxonomy of collaborative failure modes. Our quantitative audit reveals four dominant failure patterns: flawed consensus driven by shared model deficiencies, suppression of correct minority opinions, ineffective discussion dynamics, and critical information loss during synthesis. This study demonstrates that high accuracy alone is an insufficient measure of clinical or public trust. It highlights the urgent need for transparent and auditable reasoning processes, a cornerstone for the responsible development and deployment of medical AI.
Authors: Bo Yuan, Yulin Chen, Yin Zhang
Abstract: Parameter-efficient fine-tuning (PEFT) large language models (LLMs) have shown impressive performance in various downstream tasks. However, in many real-world scenarios, the collected training data inevitably contains noisy labels. To learn from noisy labels, most solutions select samples with small losses for model training. However, the selected samples, in turn, impact the loss computation in the next iteration. An inaccurate initial selection can create a vicious cycle, leading to suboptimal performance. To break this cycle, we propose Delora, a novel framework that decouples the sample selection from model training. For sample selection, Delora establishes a noisy label detector by introducing clean and noisy LoRA. Benefiting from the memory effect, the clean LoRA is encouraged to memorize clean data, while the noisy LoRA is constrained to memorize mislabeled data, which serves as a learnable threshold for selecting clean and noisy samples. For model training, Delora can use carefully selected samples to fine-tune language models seamlessly. Experimental results on synthetic and real-world noisy datasets demonstrate the effectiveness of Delora in noisy label detection and text classification.
Authors: Yijie Xu, Huizai Yao, Zhiyu Guo, Weiyu Guo, Pengteng Li, Aiwei Liu, Xuming Hu, Hui Xiong
Abstract: Large language models (LLMs) are increasingly deployed in specialized domains such as finance, medicine, and agriculture, where they face significant distribution shifts from their training data. Domain-specific fine-tuning can mitigate this challenge but relies on high-quality labeled data that is expensive and slow to collect in expertise-limited settings. We study label-free test-time adaptation for language models and present SyTTA, an inference-time framework that adapts models on-the-fly without additional supervision. SyTTA couples two complementary uncertainty signals that arise under distribution shift: input-side perplexity, indicating mismatch with domain-specific terminology and patterns, and output-side predictive entropy, indicating diffuse and unstable token probabilities during generation. Across diverse model architectures and domain-specific benchmarks, SyTTA delivers consistent gains. Notably, on agricultural question answering, SyTTA improves Rouge-LSum by over 120% on Qwen-2.5-7B with only 4 extra tokens per query. These results show that effective test-time adaptation for language models is achievable without labeled examples, supporting deployment in label-scarce domains. The code will be made available upon acceptance.
Authors: Ruize An, Richong Zhang, Zhijie Nie, Zhanyu Wu, Yanzhao Zhang, Dingkun Long
Abstract: Unsupervised text representation learning (TRL) is a fundamental task in natural language processing, which is beneficial for improving search and recommendations with the web's unlabeled texts. A recent empirical study finds that the high-quality representation aligns with the key token of the input text, uncovering the potential connection between representation space and vocabulary space. Inspired by the findings, we revisit the generative tasks and develop an unsupervised generative framework for TRL, Text2Token. The framework is based on the token target prediction task, utilizing carefully constructed target token distribution as supervisory signals. To construct the high-quality target token distribution, we analyze the token-alignment properties with advanced embedders and identify two essential categories of key tokens: (1) the meaningful tokens in the text and (2) semantically derived tokens beyond the text. Based on these insights, we propose two methods -- data-driven and model-derived -- to construct synthetic token targets from data or the LLM backbone. Experiments on the MTEB v2 benchmark demonstrate that Text2Token achieves performance competitive with the state-of-the-art embedder with unsupervised contrastive learning, LLM2Vec. Our analysis further shows that vocabulary and representation spaces optimize together and toward the optimum solution during training, providing new ideas and insights for future work.
Authors: Kangyang Luo, Yuzhuo Bai, Shuzheng Si, Cheng Gao, Zhitong Wang, Yingli Shen, Wenhao Li, Zhu Liu, Yufeng Han, Jiayi Wu, Cunliang Kong, Maosong Sun
Abstract: Coreference Resolution (CR) is a critical task in Natural Language Processing (NLP). Current research faces a key dilemma: whether to further explore the potential of supervised neural methods based on small language models, whose detect-then-cluster pipeline still delivers top performance, or embrace the powerful capabilities of Large Language Models (LLMs). However, effectively combining their strengths remains underexplored. To this end, we propose \textbf{ImCoref-CeS}, a novel framework that integrates an enhanced supervised model with LLM-based reasoning. First, we present an improved CR method (\textbf{ImCoref}) to push the performance boundaries of the supervised neural method by introducing a lightweight bridging module to enhance long-text encoding capability, devising a biaffine scorer to comprehensively capture positional information, and invoking a hybrid mention regularization to improve training efficiency. Importantly, we employ an LLM acting as a multi-role Checker-Splitter agent to validate candidate mentions (filtering out invalid ones) and coreference results (splitting erroneous clusters) predicted by ImCoref. Extensive experiments demonstrate the effectiveness of ImCoref-CeS, which achieves superior performance compared to existing state-of-the-art (SOTA) methods.
Authors: Samir Abdaljalil, Erchin Serpedin, Khalid Qaraqe, Hasan Kurban
Abstract: Large language models (LLMs) often generate reasoning traces that appear coherent but rest on unsupported assumptions, leading to hallucinated conclusions. Prior work mainly addresses factual hallucinations or relies on post-hoc verification, leaving reasoning-induced hallucinations largely unaddressed. We propose Audit-of-Understanding (AoU), a framework that constrains inference to validated premises through three phases: (1) decomposing a query into candidate assumptions, (2) auditing their support, and (3) conditioning inference only on the validated subset. Formally, AoU is \emph{posterior-constrained inference}, connecting to selective prediction and rejection learning. Our contributions are threefold: (i) theoretical guarantees under perfect validation, (ii) excess-risk bounds under imperfect audits, and (iii) tractability analysis. Empirically, AoU improves both accuracy and faithfulness on GSM8K, MultiArith, and SVAMP, achieving up to +30% gains on GSM8K, +45% on MultiArith, and consistent +20--28% improvements on SVAMP over Chain-of-Thought, Self-Consistency, and CoT-Decoding. Code is available at https://anonymous.4open.science/r/audit-of-understanding-E28B.
URLs: https://anonymous.4open.science/r/audit-of-understanding-E28B.
Authors: Liang Lin, Miao Yu, Moayad Aloqaily, Zhenhong Zhou, Kun Wang, Linsey Pang, Prakhar Mehrotra, Qingsong Wen
Abstract: Backdoor attacks are a significant threat to large language models (LLMs), often embedded via public checkpoints, yet existing defenses rely on impractical assumptions about trigger settings. To address this challenge, we propose \ourmethod, a defense framework that requires no prior knowledge of trigger settings. \ourmethod is based on the key observation that when deliberately injecting known backdoors into an already-compromised model, both existing unknown and newly injected backdoors aggregate in the representation space. \ourmethod leverages this through a two-stage process: \textbf{first}, aggregating backdoor representations by injecting known triggers, and \textbf{then}, performing recovery fine-tuning to restore benign outputs. Extensive experiments across multiple LLM architectures demonstrate that: (I) \ourmethod reduces the average Attack Success Rate to 4.41\% across multiple benchmarks, outperforming existing baselines by 28.1\%$\sim$69.3\%$\uparrow$. (II) Clean accuracy and utility are preserved within 0.5\% of the original model, ensuring negligible impact on legitimate tasks. (III) The defense generalizes across different types of backdoors, confirming its robustness in practical deployment scenarios.
Authors: Yihong Liu, Mingyang Wang, Fran\c{c}ois Yvon, Hinrich Sch\"utze
Abstract: Multilingual large language models (LLMs) are expected to recall factual knowledge consistently across languages. However, the factors that give rise to such crosslingual consistency -- and its frequent failure -- remain poorly understood. In this work, we hypothesize that these inconsistencies may arise from failures in entity alignment, the process of mapping subject and object entities into a shared conceptual space across languages. To test this, we assess alignment through entity-level (subject and object) translation tasks, and find that consistency is strongly correlated with alignment across all studied models, with misalignment of subjects or objects frequently resulting in inconsistencies. Building on this insight, we propose SubSub and SubInj, two effective methods that integrate English translations of subjects into prompts across languages, leading to substantial gains in both factual recall accuracy and consistency. Finally, our mechanistic analysis reveals that these interventions reinforce the entity representation alignment in the conceptual space through model's internal pivot-language processing, offering effective and practical strategies for improving multilingual factual prediction.
Authors: Hongwei Chen, Yishu Lei, Dan Zhang, Bo Ke, Danxiang Zhu, Xuyi Chen, Yuxiang Lu, Zhengjie Huang, Shikun Feng, Jingzhou He, Yu Sun, Hua Wu, Haifeng Wang
Abstract: Test-time scaling has emerged as a promising paradigm in language modeling, wherein additional computational resources are allocated during inference to enhance model performance. Recent approaches, such as DeepConf, have demonstrated the efficacy of this strategy, however, they often incur substantial computational overhead to achieve competitive results. In this work, we propose MatryoshkaThinking, a novel method that significantly reduces computational cost while maintaining state-of-the-art performance. Specifically, MatryoshkaThinking attains a score of 99.79 on AIME2025 using only 4% of the computation required by DeepConf. The core of our approach lies in the recursive exploitation of the model's intrinsic capabilities in reasoning, verification, and summarization, which collectively enhance the retention of correct solutions and reduce the disparity between Pass@k and Pass@1. Comprehensive evaluations across multiple open-source models and challenging multi-modal reasoning benchmarks validate the effectiveness and generality of our method. These findings offer new insights into the design of efficient and scalable test-time inference strategies for advanced language models.
Authors: Ananya Malik, Nazanin Sabri, Melissa Karnaze, Mai Elsherief
Abstract: Large Language Models' (LLMs) ability to converse naturally is empowered by their ability to empathetically understand and respond to their users. However, emotional experiences are shaped by demographic and cultural contexts. This raises an important question: Can LLMs demonstrate equitable empathy across diverse user groups? We propose a framework to investigate how LLMs' cognitive and affective empathy vary across user personas defined by intersecting demographic attributes. Our study introduces a novel intersectional analysis spanning 315 unique personas, constructed from combinations of age, culture, and gender, across four LLMs. Results show that attributes profoundly shape a model's empathetic responses. Interestingly, we see that adding multiple attributes at once can attenuate and reverse expected empathy patterns. We show that they broadly reflect real-world empathetic trends, with notable misalignments for certain groups, such as those from Confucian culture. We complement our quantitative findings with qualitative insights to uncover model behaviour patterns across different demographic groups. Our findings highlight the importance of designing empathy-aware LLMs that account for demographic diversity to promote more inclusive and equitable model behaviour.
Authors: Nam Luu, Ond\v{r}ej Bojar
Abstract: Speech Translation (ST) is a machine translation task that involves converting speech signals from one language to the corresponding text in another language; this task has two different approaches, namely the traditional cascade and the more recent end-to-end. This paper explores a combined end-to-end architecture of pre-trained speech encoders and Large Language Models (LLMs) for performing both Automatic Speech Recognition (ASR) and ST simultaneously. Experiments with the English-to-German language pair show that our best model not only can achieve better translation results than SeamlessM4T, a large foundational end-to-end, multi-modal translation model, but can also match the performance of a cascaded system with Whisper and NLLB, with up to a score gain of 8% in $\text{COMET}^{\text{DA}}_{22}$ metric.
Authors: Hakyung Sung, Kristopher Kyle
Abstract: Argument structure constructions (ASCs) offer a theoretically grounded lens for analyzing second language (L2) proficiency, yet scalable and systematic tools for measuring their usage remain limited. This paper introduces the ASC analyzer, a publicly available Python package designed to address this gap. The analyzer automatically tags ASCs and computes 50 indices that capture diversity, proportion, frequency, and ASC-verb lemma association strength. To demonstrate its utility, we conduct both bivariate and multivariate analyses that examine the relationship between ASC-based indices and L2 writing scores.
Authors: Aashiq Muhamed, Leonardo F. R. Ribeiro, Markus Dreyer, Virginia Smith, Mona T. Diab
Abstract: The ability of language models in RAG systems to selectively refuse to answer based on flawed context is critical for safety, yet remains a significant failure point. Our large-scale study reveals that even frontier models struggle in this setting, with refusal accuracy dropping below 50% on multi-document tasks, while exhibiting either dangerous overconfidence or overcaution. Static benchmarks fail to reliably evaluate this capability, as models exploit dataset-specific artifacts and memorize test instances. We introduce RefusalBench, a generative methodology that programmatically creates diagnostic test cases through controlled linguistic perturbation. Our framework employs 176 distinct perturbation strategies across six categories of informational uncertainty and three intensity levels. Evaluation of over 30 models uncovers systematic failure patterns: refusal comprises separable detection and categorization skills, and neither scale nor extended reasoning improves performance. We find that selective refusal is a trainable, alignment-sensitive capability, offering a clear path for improvement. We release two benchmarks -- RefusalBench-NQ (single document) and RefusalBench-GaRAGe (multi-document) -- and our complete generation framework to enable continued, dynamic evaluation of this critical capability.
Authors: Kai Zhang, Xinyuan Zhang, Ejaz Ahmed, Hongda Jiang, Caleb Kumar, Kai Sun, Zhaojiang Lin, Sanat Sharma, Shereen Oraby, Aaron Colak, Ahmed Aly, Anuj Kumar, Xiaozhong Liu, Xin Luna Dong
Abstract: Accurate recall from large scale memories remains a core challenge for memory augmented AI assistants performing question answering (QA), especially in similarity dense scenarios where existing methods mainly rely on semantic distance to the query for retrieval. Inspired by how humans link information associatively, we propose AssoMem, a novel framework constructing an associative memory graph that anchors dialogue utterances to automatically extracted clues. This structure provides a rich organizational view of the conversational context and facilitates importance aware ranking. Further, AssoMem integrates multi-dimensional retrieval signals-relevance, importance, and temporal alignment using an adaptive mutual information (MI) driven fusion strategy. Extensive experiments across three benchmarks and a newly introduced dataset, MeetingQA, demonstrate that AssoMem consistently outperforms SOTA baselines, verifying its superiority in context-aware memory recall.
Authors: Geunyeong Jeong, Juoh Sun, Seonghee Lee, Harksoo Kim
Abstract: Large Language Models store extensive factual knowledge acquired during large-scale pre-training. However, this knowledge is inherently static, reflecting only the state of the world at the time of training. Knowledge editing has emerged as a promising solution for updating outdated or incorrect facts without full retraining. However, most existing locate-and-edit methods primarily focus on token-level likelihood optimization without addressing semantic coherence. Our analysis reveals that such edited knowledge is often encoded as isolated residual streams in the model's latent space, distinct from pre-existing knowledge and bypassing natural reasoning process. To address this, we propose \textsc{Steam}, a semantic-level knowledge editing framework that enhances integration of updated knowledge into the model's knowledge structure. \textsc{Steam} first identifies target representations as semantic anchors for the updated factual association, then guides the internal representation of the edited fact towards these anchors through an alignment loss during optimization. Experimental results demonstrate that \textsc{Steam} improves model's ability to reason with edited knowledge and enhances semantic coherence, underscoring the importance of latent-space alignment for reliable and coherent knowledge editing. The code is available at https://github.com/GY-Jeong/STEAM.
Authors: Federica Bologna, Tiffany Pan, Matthew Wilkens, Yue Guo, Lucy Lu Wang
Abstract: Evaluating long-form clinical question answering (QA) systems is resource-intensive and challenging: accurate judgments require medical expertise and achieving consistent human judgments over long-form text is difficult. We introduce LongQAEval, an evaluation framework and set of evaluation recommendations for limited-resource and high-expertise settings. Based on physician annotations of 300 real patient questions answered by physicians and LLMs, we compare coarse answer-level versus fine-grained sentence-level evaluation over the dimensions of correctness, relevance, and safety. We find that inter-annotator agreement (IAA) varies by dimension: fine-grained annotation improves agreement on correctness, coarse improves agreement on relevance, and judgments on safety remain inconsistent. Additionally, annotating only a small subset of sentences can provide reliability comparable to coarse annotations, reducing cost and effort.
Authors: Jingyi Chen, Zhimeng Guo, Jiyun Chun, Pichao Wang, Andrew Perrault, Micha Elsner
Abstract: Understanding emotion from speech requires sensitivity to both lexical and acoustic cues. However, it remains unclear whether large audio language models (LALMs) genuinely process acoustic information or rely primarily on lexical content. We present LISTEN (Lexical vs. Acoustic Speech Test for Emotion in Narratives), a controlled benchmark designed to disentangle lexical reliance from acoustic sensitivity in emotion understanding. Across evaluations of six state-of-the-art LALMs, we observe a consistent lexical dominance. Models predict "neutral" when lexical cues are neutral or absent, show limited gains under cue alignment, and fail to classify distinct emotions under cue conflict. In paralinguistic settings, performance approaches chance. These results indicate that current LALMs largely "transcribe" rather than "listen," relying heavily on lexical semantics while underutilizing acoustic cues. LISTEN offers a principled framework for assessing emotion understanding in multimodal models.
Authors: Zhichao Xu, Minheng Wang, Yawei Wang, Wenqian Ye, Yuntao Du, Yunpu Ma, Yijun Tian
Abstract: Retrieval-augmented generation (RAG) systems trained using reinforcement learning (RL) with reasoning are hampered by inefficient context management, where long, noisy retrieved documents increase costs and degrade performance. We introduce RECON (REasoning with CONdensation), a framework that integrates an explicit summarization module to compress evidence within the reasoning loop. Our summarizer is trained via a two-stage process: relevance pretraining on QA datasets, followed by multi-aspect distillation from proprietary LLMs to ensure factuality and clarity. Integrated into the Search-R1 pipeline, RECON reduces total context length by 35\%, leading to improved training speed and inference latency, while simultaneously improving RAG performance on downstream QA benchmarks. Notably, it boosts the average EM score of the 3B model by 14.5\% and the 7B model by 3.0\%, showing particular strength in multi-hop QA. RECON demonstrates that learned context compression is essential for building practical, scalable, and performant RAG systems. Our code implementation is made available at https://github.com/allfornancy/RECON.
Authors: Utsav Maskey, Mark Dras, Usman Naseem
Abstract: Safety alignment in large language models (LLMs) induces over-refusals -- where LLMs decline benign requests due to aggressive safety filters. We analyze this phenomenon in retrieval-augmented generation (RAG), where both the query intent and retrieved context properties influence refusal behavior. We construct RagRefuse, a domain-stratified benchmark spanning medical, chemical, and open domains, pairing benign and harmful queries with controlled context contamination patterns and sizes. Our analysis shows that context arrangement / contamination, domain of query and context, and harmful-text density trigger refusals even on benign queries, with effects depending on model-specific alignment choices. To mitigate over-refusals, we introduce \textsc{SafeRAG-Steering}, a model-centric embedding intervention that steers the embedding regions towards the confirmed safe, non-refusing output regions at inference time. This reduces over-refusals in contaminated RAG pipelines while preserving legitimate refusals.
Authors: Peng Fan, Wenping Wang, Fei Deng
Abstract: The mismatch of speech length and text length poses a challenge in automatic speech recognition (ASR). In previous research, various approaches have been employed to align text with speech, including the utilization of Connectionist Temporal Classification (CTC). In earlier work, a key frame mechanism (KFDS) was introduced, utilizing intermediate CTC outputs to guide downsampling and preserve keyframes, but traditional methods (CTC) failed to align speech and text appropriately when downsampling speech to a text-similar length. In this paper, we focus on speech recognition in those cases where the length of speech aligns closely with that of the corresponding text. To address this issue, we introduce two methods for alignment: a) Time Independence Loss (TIL) and b) Aligned Cross Entropy (AXE) Loss, which is based on edit distance. To enhance the information on keyframes, we incorporate frame fusion by applying weights and summing the keyframe with its context 2 frames. Experimental results on AISHELL-1 and AISHELL-2 dataset subsets show that the proposed methods outperform the previous work and achieve a reduction of at least 86\% in the number of frames.
Authors: Shaobo Wang, Cong Wang, Wenjie Fu, Yue Min, Mingquan Feng, Isabel Guan, Xuming Hu, Conghui He, Cunxiang Wang, Kexin Yang, Xingzhang Ren, Fei Huang, Dayiheng Liu, Linfeng Zhang
Abstract: As the demand for comprehensive evaluations of diverse model capabilities steadily increases, benchmark suites have correspondingly grown significantly in scale. Despite notable advances in redundancy reduction and subset-level performance prediction, a systematic framework that effectively integrates these methods to ensure both prediction accuracy and ranking consistency is still largely elusive. In this paper, we first perform a sample-level analysis of benchmark redundancy and identify several highly similar samples that can be eliminated. Besides, we frame benchmark compression as an optimization problem with the aim of score reconstruction. Building on these, we then propose EssenceBench, a coarse-to-fine framework utilizing an iterative Genetic Algorithm (GA), which takes the advantages of fitness-based subset search and attribution-based sample search. Compared to previous methods, our approach yields superior compression results with lower reconstruction error and markedly higher efficiency. In particular, on the HellaSwag benchmark (10K samples), our method preserves the ranking of all models shifting within 5% using 25x fewer samples, and achieves 95% ranking preservation shifting within 5% using only 200x fewer samples.
Authors: Prawaal Sharma, Poonam Goyal, Navneet Goyal, Vidisha Sharma
Abstract: Digital communication has become the cornerstone of modern interaction, enabling rapid, accessible, and interactive exchanges. However, individuals with lower academic literacy often face significant barriers, exacerbating the "digital divide". In this work, we introduce a novel, universal ideographic metalanguage designed as an innovative communication framework that transcends academic, linguistic, and cultural boundaries. Our approach leverages principles of Neuro-symbolic AI, combining neural-based large language models (LLMs) enriched with world knowledge and symbolic knowledge heuristics grounded in the linguistic theory of Natural Semantic Metalanguage (NSM). This enables the semantic decomposition of complex ideas into simpler, atomic concepts. Adopting a human-centric, collaborative methodology, we engaged over 200 semi-literate participants in defining the problem, selecting ideographs, and validating the system. With over 80\% semantic comprehensibility, an accessible learning curve, and universal adaptability, our system effectively serves underprivileged populations with limited formal education.
Authors: Qiran Zou, Hou Hei Lam, Wenhao Zhao, Yiming Tang, Tingting Chen, Samson Yu, Tianyi Zhang, Chang Liu, Xiangyang Ji, Dianbo Liu
Abstract: Large language models (LLMs) have sparked growing interest in automatic machine learning research agents. Among them, agents capable of autonomously proposing ideas and conducting machine learning experiments are particularly promising, as they maximize research automation and accelerate scientific progress by iteratively refining ideas based on experimental results. However, comprehensively evaluating such agents remains challenging. Existing benchmarks tend to overemphasize engineering aspects while neglecting academic rigor, creating barriers that obscure a clear assessment of an agent's scientific capabilities in machine learning research. They also suffer from limited task diversity, an overemphasis on application-oriented tasks over fundamental research problems, and limited scalability to realistic research settings. To address these limitations, we introduce FML-bench, a benchmark designed to evaluate automatic machine learning research agents on 8 diverse and fundamental machine learning research problems. It reduces coding burden, emphasizes fundamental problems rather than specific use cases, offers high task diversity, and is extensible to real-world machine learning GitHub repositories. Furthermore, we present a unified evaluation framework with five complementary metrics, designed to comprehensively assess agent performance on our benchmark. We evaluate state-of-the-art automatic research agents on FML-bench, and find that agents employing broad research exploration strategies outperform those focusing on narrow but deep exploration. These findings suggest that emphasizing the breadth of exploration may lead to more effective research outcomes than focusing solely on incremental refinement. Our benchmark is available at https://github.com/qrzou/FML-bench.
Authors: Jingyi Wu, Junying Liang
Abstract: Understanding what drives popularity is critical in today's digital service economy, where content creators compete for consumer attention. Prior studies have primarily emphasized the role of content features, yet creators often misjudge what audiences actually value. This study applies Latent Dirichlet Allocation (LDA) modeling to a large corpus of TED Talks, treating the platform as a case of digital service provision in which creators (speakers) and consumers (audiences) interact. By comparing the thematic supply of creators with the demand expressed in audience engagement, we identify persistent mismatches between producer offerings and consumer preferences. Our longitudinal analysis further reveals that temporal dynamics exert a stronger influence on consumer engagement than thematic content, suggesting that when content is delivered may matter more than what is delivered. These findings challenge the dominant assumption that content features are the primary drivers of popularity and highlight the importance of timing and contextual factors in shaping consumer responses. The results provide new insights into consumer attention dynamics on digital platforms and carry practical implications for marketers, platform managers, and content creators seeking to optimize audience engagement strategies.
Authors: A H M Rezaul Karim, Ozlem Uzuner
Abstract: Medical order extraction is essential for structuring actionable clinical information, supporting decision-making, and enabling downstream applications such as documentation and workflow automation. Orders may be embedded in diverse sources, including electronic health records, discharge summaries, and multi-turn doctor-patient dialogues, and can span categories such as medications, laboratory tests, imaging studies, and follow-up actions. The MEDIQA-OE 2025 shared task focuses on extracting structured medical orders from extended conversational transcripts, requiring the identification of order type, description, reason, and provenance. We present the MasonNLP submission, which ranked 5th among 17 participating teams with 105 total submissions. Our approach uses a general-purpose, instruction-tuned LLaMA-4 17B model without domain-specific fine-tuning, guided by a single in-context example. This few-shot configuration achieved an average F1 score of 37.76, with notable improvements in reason and provenance accuracy. These results demonstrate that large, non-domain-specific LLMs, when paired with effective prompt engineering, can serve as strong, scalable baselines for specialized clinical NLP tasks.
Authors: Guangxin He, Shen Nie, Fengqi Zhu, Yuankang Zhao, Tianyi Bai, Ran Yan, Jie Fu, Chongxuan Li, Binhang Yuan
Abstract: Diffusion LLMs have attracted growing interest, with plenty of recent work emphasizing their great potential in various downstream tasks; yet the long-context behavior of diffusion LLMs remains largely uncharted. We present a case study of post-training techniques for extending the context window of diffusion LLMs (i.e., LLaDA) without retraining from scratch. We show that a simple modification to the standard Rotary Positional Embeddings (RoPE) extension effectively accommodates the probabilistic modeling inherent in the diffusion process, enabling stable scaling to longer context ranges. We further compare masking strategies used during post-training and analyze their impact on optimization stability and long-range recall. Instantiating these insights, we introduce UltraLLaDA, a diffusion LLM with a 128K-token context window that, in our empirical evaluation on long-context tasks, significantly outperforms training-free baselines. Our experimental results highlight the special positional extension as a key lever for scaling diffusion LLMs to extended contexts and offer practical guidance for practitioners seeking 128K-scale context via efficient post-training.
Authors: Prawaal Sharma, Poonam Goyal, Vidisha Sharma, Navneet Goyal
Abstract: UNESCO has classified 2500 out of 7000 languages spoken worldwide as endangered. Attrition of a language leads to loss of traditional wisdom, folk literature, and the essence of the community that uses it. It is therefore imperative to bring digital inclusion to these languages and avoid its extinction. Low resource languages are at a greater risk of extinction. Lack of unsupervised Optical Character Recognition(OCR) methodologies for low resource languages is one of the reasons impeding their digital inclusion. We propose VOLTAGE - a contrastive learning based OCR methodology, leveraging auto-glyph feature recommendation for cluster-based labelling. We augment the labelled data for diversity and volume using image transformations and Generative Adversarial Networks. Voltage has been designed using Takri - a family of scripts used in 16th to 20th century in the Himalayan regions of India. We present results for Takri along with other Indic scripts (both low and high resource) to substantiate the universal behavior of the methodology. An accuracy of 95% for machine printed and 87% for handwritten samples on Takri script has been achieved. We conduct baseline and ablation studies along with building downstream use cases for Takri, demonstrating the usefulness of our work.
Authors: Heming Xia, Cunxiao Du, Rui Li, Chak Tou Leong, Yongqi Li, Wenjie Li
Abstract: Large reasoning models (LRMs) have demonstrated remarkable proficiency in tackling complex reasoning tasks through step-by-step thinking. However, such a lengthy reasoning process incurs substantial computational and latency overheads, hindering the practical deployment of these models. In this work, we present a new perspective on mitigating overthinking in LRMs via black-box adversarial prompting. By treating both open-source LRMs and closed-source APIs as black-box communicators, we investigate how to elicit concise responses without sacrificing accuracy. We introduce AdvPrompt, an iterative refinement framework that generates high-quality adversarial prompts from diverse perspectives. Experiments across multiple benchmarks demonstrate that AdvPrompt consistently reduces token usage while preserving performance. Notably, AdvPrompt achieves a 3x reduction in average response length on simple GSM8K questions for the Qwen3 model series, and delivers an average ~40% token reduction across four benchmarks. For closed-source APIs, AdvPrompt reduces token usage on MATH-500 by 35% for Claude-3.7 and 47% for Gemini-2.5. Further analysis reveals the generalizability of AdvPrompt across various model scales and families, underscoring the potential of black-box prompting as a practical and effective strategy for enhancing LRM efficiency.
Authors: Yujie Ren, Niklas Gruhlke, Anne Lauscher
Abstract: As large language models (LLMs) are increasingly applied in sensitive domains such as medicine and law, hallucination detection has become a critical task. Although numerous benchmarks have been proposed to advance research in this area, most of them are artificially constructed--either through deliberate hallucination induction or simulated interactions--rather than derived from genuine LLM-human dialogues. Consequently, these benchmarks fail to fully capture the characteristics of hallucinations that occur in real-world usage. To address this limitation, we introduce AuthenHallu, the first hallucination detection benchmark built entirely from authentic LLM-human interactions. For AuthenHallu, we select and annotate samples from genuine LLM-human dialogues, thereby providing a faithful reflection of how LLMs hallucinate in everyday user interactions. Statistical analysis shows that hallucinations occur in 31.4% of the query-response pairs in our benchmark, and this proportion increases dramatically to 60.0% in challenging domains such as Math & Number Problems. Furthermore, we explore the potential of using vanilla LLMs themselves as hallucination detectors and find that, despite some promise, their current performance remains insufficient in real-world scenarios.
Authors: Euhid Aman, Esteban Carlin, Hsing-Kuo Pao, Giovanni Beltrame, Ghaluh Indah Permata Sari, Yie-Tarng Chen
Abstract: Cross-attention transformers and other multimodal vision-language models excel at grounding and generation; however, their extensive, full-precision backbones make it challenging to deploy them on edge devices. Memory-augmented architectures enhance the utilization of past context; however, most works rarely pair them with aggressive edge-oriented quantization. We introduce BitMar, a quantized multimodal transformer that proposes an external human-like episodic memory for effective image-text generation on hardware with limited resources. BitMar utilizes 1.58-bit encoders, one for text (BitNet-style) and one for vision (DiNOv2-based), to create compact embeddings that are combined and used to query a fixed-size key-value episodic memory. During vector retrieval, the BitNet decoder applies per-layer conditioning, which increases the contextual relevance of generated content. The decoder also employs attention sinks with a sliding-window mechanism to process long or streaming inputs under tight memory budgets. The combination of per-layer conditioning and sliding-window attention achieves a strong quality-speed trade-off, delivering competitive captioning and multimodal understanding at low latency with a small model footprint. These characteristics make BitMar well-suited for edge deployment.
Authors: Di Wu abd Shuaidong Pan
Abstract: This paper proposes a modeling framework for dynamic topic evolution based on temporal large language models. The method first uses a large language model to obtain contextual embeddings of text and then introduces a temporal decay function and an attention mechanism. These components allow the model to adjust the importance of semantic units according to time intervals and capture topic variations across different periods. The temporal representations are then mapped into a latent topic space, where a state transition matrix is applied to describe the dynamic evolution of topics. A joint optimization objective constrains both semantic modeling and temporal consistency, ensuring diversity and smoothness in topic generation. The design emphasizes the unified modeling of semantic representation and temporal evolution, which improves topic coherence and diversity while enhancing stability and interpretability over time. Experiments on real-world corpora show that the framework effectively captures the generation, expansion, and decline of topics and outperforms existing models across multiple metrics. Overall, the proposed method provides a systematic solution for understanding dynamic semantic patterns in large-scale text, enriches the research paradigm of topic modeling, and supports complex text analysis tasks in multiple domains.
Authors: Bowei He, Lihao Yin, Huiling Zhen, Shuqi Liu, Han Wu, Xiaokun Zhang, Mingxuan Yuan, Chen Ma
Abstract: Post-training compression has been a widely employed approach to scale down large language model (LLM) and facilitate efficient inference. In various proposed compression methods, including pruning and quantization, calibration data plays a vital role by informing the weight importance and activation dynamic ranges. However, how calibration data impacts the LLM capability after compression is less explored. Few of the existing works, though recognizing the significance of this study, only investigate the language modeling or commonsense reasoning performance degradation from limited angles, like the data sources or sample amounts. More systematic research is still needed to examine the impacts on different LLM capabilities in terms of compositional properties and domain correspondence of calibration data. In this work, we aim at bridging this gap and further analyze underlying influencing mechanisms from the activation pattern perspective. Especially, we explore the calibration data's impacts on high-level complex reasoning capabilities, like math problem solving and code generation. Delving into the underlying mechanism, we find that the representativeness and diversity in activation space more fundamentally determine the quality of calibration data. Finally, we propose a calibration data curation framework based on such observations and analysis, enhancing the performance of existing post-training compression methods on preserving critical LLM capabilities. Our code is provided in \href{https://github.com/BokwaiHo/COLA.git}{Link}.
Authors: Guy Mor-Lan, Tamir Sheafer, Shaul R. Shenhav
Abstract: How is a factual claim made credible? We propose the novel task of Epistemic Appeal Identification, which identifies whether and how factual statements have been anchored by external sources or evidence. To advance research on this task, we present FactAppeal, a manually annotated dataset of 3,226 English-language news sentences. Unlike prior resources that focus solely on claim detection and verification, FactAppeal identifies the nuanced epistemic structures and evidentiary basis underlying these claims and used to support them. FactAppeal contains span-level annotations which identify factual statements and mentions of sources on which they rely. Moreover, the annotations include fine-grained characteristics of factual appeals such as the type of source (e.g. Active Participant, Witness, Expert, Direct Evidence), whether it is mentioned by name, mentions of the source's role and epistemic credentials, attribution to the source via direct or indirect quotation, and other features. We model the task with a range of encoder models and generative decoder models in the 2B-9B parameter range. Our best performing model, based on Gemma 2 9B, achieves a macro-F1 score of 0.73.
Authors: Guy Mor-Lan, Tamir Sheafer, Shaul R. Shenhav
Abstract: While media bias is widely studied, the epistemic strategies behind factual reporting remain computationally underexplored. This paper analyzes these strategies through a large-scale comparison of CNN and Fox News. To isolate reporting style from topic selection, we employ an article matching strategy to compare reports on the same events and apply the FactAppeal framework to a corpus of over 470K articles covering two highly politicized periods: the COVID-19 pandemic and the Israel-Hamas war. We find that CNN's reporting contains more factual statements and is more likely to ground them in external sources. The outlets also exhibit sharply divergent sourcing patterns: CNN builds credibility by citing Experts} and Expert Documents, constructing an appeal to formal authority, whereas Fox News favors News Reports and direct quotations. This work quantifies how partisan outlets use systematically different epistemic strategies to construct reality, adding a new dimension to the study of media bias.
Authors: Omid Reza Heidari, Siobhan Reid, Yassine Yaakoubi
Abstract: LLMs have advanced text-to-SQL generation, yet monolithic architectures struggle with complex reasoning and schema diversity. We propose AGENTIQL, an agent-inspired multi-expert framework that combines a reasoning agent for question decomposition, a coding agent for sub-query generation, and a refinement step for column selection. An adaptive router further balances efficiency and accuracy by selecting between our modular pipeline and a baseline parser. Several steps in the pipeline can be executed in parallel, making the framework scalable to larger workloads. Evaluated on the Spider benchmark, AGENTIQL improves execution accuracy and interpretability and achieves up to 86.07\% EX with 14B models using the Planner&Executor merging strategy. The attained performance is contingent upon the efficacy of the routing mechanism, thereby narrowing the gap to GPT-4-based SOTA (89.65% EX) while using much smaller open-source LLMs. Beyond accuracy, AGENTIQL enhances transparency by exposing intermediate reasoning steps, offering a robust, scalable, and interpretable approach to semantic parsing.
Authors: Zhengbo Zhang, Zhiheng Lyu, Junhao Gong, Hongzhu Yi, Xinming Wang, Yuxuan Zhou, Jiabing Yang, Ping Nie, Yan Huang, Wenhu Chen
Abstract: Efficiently solving real-world problems with LLMs increasingly hinges on their ability to interact with dynamic web environments and autonomously acquire external information. While recent research like Search-R1 and WebDancer demonstrates strong performance in solving web tasks, they heavily rely on additional tools to convert the interactive web environment into static text content. This is in contrast to human browsing behaviors, which involve diverse interactions with the browser, such as scrolling, clicking, and typing. In this paper, we propose BrowserAgent, a more interactive agent that solves complex tasks through human-inspired browser actions. BrowserAgent operates directly on raw web pages via Playwright through a set of predefined browser actions. We adopt a two-stage training (Supervised Fine-Tuning (SFT) and Rejection Fine-Tuning (RFT)) to improve the model's generalization abilities. Despite using significantly less training data than Search-R1, BrowserAgent achieves more competitive results across different Open-QA tasks. Additionally, we introduce an explicit memory mechanism to store key conclusions across steps, further enhancing the model's reasoning capabilities for long-horizon tasks. Notably, BrowserAgent-7B can achieve around 20\% improvement over Search-R1 on multi-hop QA tasks like HotpotQA, 2Wiki, and Bamboogle. These results indicate that BrowserAgent can serve as a more advanced framework for more interactive and scalable web agents.
Authors: Zhuowei Chen, Bowei Zhang, Nankai Lin, Tian Hou, Lianxi Wang
Abstract: Recent advances in LLMs have enhanced AI capabilities, but also increased the risk posed by malicious requests, highlighting the need for effective LLM safeguards to detect such queries. Existing approaches largely rely on classifier-based methods that lack interpretability and perform poorly on low-resource languages. To address these limitations, we propose ConsistentGuard, a novel reasoning-based multilingual safeguard, which enhances explainability via reasoning and boosts knowledge transfer between languages through alignment. With only 1,000 training samples, our method demonstrates superior performance on three datasets across six languages, outperforming larger models trained with significantly more data, and exhibits strong interpretability and generalization ability. We also contribute a multilingual benchmark extension and release our codes to support future research.
Authors: Zichun Yu, Chenyan Xiong
Abstract: High-quality pretraining data is the fossil fuel of large language models (LLMs), yet its reserves are running low for frontier models. In this paper, we introduce RePro, a novel web recycling method that trains a relatively small LM with reinforcement learning to generate effective and faithful rephrasings of pretraining data. Specifically, we design one quality reward and three faithfulness rewards, optimizing the LM rephraser to convert organic data into high-quality rephrasings while maintaining its core semantics and structure. In our experiment, we train a 4B rephraser to recycle 72B tokens sampled from DCLM-RefinedWeb. Pretraining results on 400M and 1.4B models demonstrate that RePro delivers 4.7%-14.0% relative accuracy gains over organic-only baseline on 22 downstream tasks. RePro also outperforms ReWire, the state-of-the-art web recycling method that prompts a 70B rephraser, as well as the organic baseline with a 4x larger data pool. Experiments with different amounts of recycled data highlight that RePro improves organic data efficiency by 2-3x. Individual and distributional analyses validate that RePro preserves more critical information and faithfully reflects the characteristics of organic data compared to prompting-based methods. Together, these results show that RePro provides an efficient and controllable path to effectively harness the fossil fuel of LLM pretraining. We open-source our code, rephraser, and recycled data at https://github.com/cxcscmu/RePro.
Authors: Manas Zambre (Supervisor), Sarika Bobade (Supervisor)
Abstract: Sarcasm is a nuanced and often misinterpreted form of communication, especially in text, where tone and body language are absent. This paper proposes a modular deep learning framework for sarcasm detection, leveraging Deep Convolutional Neural Networks (DCNNs) and contextual models such as BERT to analyze linguistic, emotional, and contextual cues. The system integrates sentiment analysis, contextual embeddings, linguistic feature extraction, and emotion detection through a multi-layer architecture. While the model is in the conceptual stage, it demonstrates feasibility for real-world applications such as chatbots and social media analysis.
Authors: Wenqing Zhang, Trang Nguyen, Elizabeth A. Stuart, Yiqun T. Chen
Abstract: Systematic reviews are crucial for synthesizing scientific evidence but remain labor-intensive, especially when extracting detailed methodological information. Large language models (LLMs) offer potential for automating methodological assessments, promising to transform evidence synthesis. Here, using causal mediation analysis as a representative methodological domain, we benchmarked state-of-the-art LLMs against expert human reviewers across 180 full-text scientific articles. Model performance closely correlated with human judgments (accuracy correlation 0.71; F1 correlation 0.97), achieving near-human accuracy on straightforward, explicitly stated methodological criteria. However, accuracy sharply declined on complex, inference-intensive assessments, lagging expert reviewers by up to 15%. Errors commonly resulted from superficial linguistic cues -- for instance, models frequently misinterpreted keywords like "longitudinal" or "sensitivity" as automatic evidence of rigorous methodological approache, leading to systematic misclassifications. Longer documents yielded lower model accuracy, whereas publication year showed no significant effect. Our findings highlight an important pattern for practitioners using LLMs for methods review and synthesis from full texts: current LLMs excel at identifying explicit methodological features but require human oversight for nuanced interpretations. Integrating automated information extraction with targeted expert review thus provides a promising approach to enhance efficiency and methodological rigor in evidence synthesis across diverse scientific fields.
Authors: James Ald Teves, Ray Daniel Cal, Josh Magdiel Villaluz, Jean Malolos, Mico Magtira, Ramon Rodriguez, Mideth Abisado, Joseph Marvin Imperial
Abstract: The language of Hiligaynon, spoken predominantly by the people of Panay Island, Negros Occidental, and Soccsksargen in the Philippines, remains underrepresented in language processing research due to the absence of annotated corpora and baseline models. This study introduces HiligayNER, the first publicly available baseline model for the task of Named Entity Recognition (NER) in Hiligaynon. The dataset used to build HiligayNER contains over 8,000 annotated sentences collected from publicly available news articles, social media posts, and literary texts. Two Transformer-based models, mBERT and XLM-RoBERTa, were fine-tuned on this collected corpus to build versions of HiligayNER. Evaluation results show strong performance, with both models achieving over 80% in precision, recall, and F1-score across entity types. Furthermore, cross-lingual evaluation with Cebuano and Tagalog demonstrates promising transferability, suggesting the broader applicability of HiligayNER for multilingual NLP in low-resource settings. This work aims to contribute to language technology development for underrepresented Philippine languages, specifically for Hiligaynon, and support future research in regional language processing.
Authors: Zhichao Wang, Cheng Wan, Dong Nie
Abstract: The performance gains of LLMs have historically been driven by scaling up model size and training data. However, the rapidly diminishing availability of high-quality training data is introducing a fundamental bottleneck, shifting the focus of research toward inference-time scaling. This paradigm uses additional computation at the time of deployment to substantially improve LLM performance on downstream tasks without costly model re-training. This review systematically surveys the diverse techniques contributing to this new era of inference-time scaling, organizing the rapidly evolving field into two comprehensive perspectives: Output-focused and Input-focused methods. Output-focused techniques encompass complex, multi-step generation strategies, including reasoning (e.g., CoT, ToT, ReAct), various search and decoding methods (e.g., MCTS, beam search), training for long CoT (e.g., RLVR, GRPO), and model ensemble methods. Input-focused techniques are primarily categorized by few-shot and RAG, with RAG as the central focus. The RAG section is further detailed through a structured examination of query expansion, data, retrieval and reranker, LLM generation methods, and multi-modal RAG.
Authors: Bahar \.Ilgen, Georges Hattab
Abstract: Text simplification is essential for making public health information accessible to diverse populations, including those with limited health literacy. However, commonly used evaluation metrics in Natural Language Processing (NLP), such as BLEU, FKGL, and SARI, mainly capture surface-level features and fail to account for human-centered qualities like clarity, trustworthiness, tone, cultural relevance, and actionability. This limitation is particularly critical in high-stakes health contexts, where communication must be not only simple but also usable, respectful, and trustworthy. To address this gap, we propose the Human-Centered Readability Score (HCRS), a five-dimensional evaluation framework grounded in Human-Computer Interaction (HCI) and health communication research. HCRS integrates automatic measures with structured human feedback to capture the relational and contextual aspects of readability. We outline the framework, discuss its integration into participatory evaluation workflows, and present a protocol for empirical validation. This work aims to advance the evaluation of health text simplification beyond surface metrics, enabling NLP systems that align more closely with diverse users' needs, expectations, and lived experiences.
Authors: Mihir Gupte, Paolo Giusto, Ramesh S
Abstract: Large Language Models (LLMs) are adept at generating responses based on information within their context. While this ability is useful for interacting with structured data like code files, another popular method, Retrieval-Augmented Generation (RAG), retrieves relevant documents to augment the model's in-context learning. However, it is not well-explored how to best represent this retrieved knowledge for generating responses on structured data, particularly hierarchical structures like trees. In this work, we propose a novel bottom-up method to linearize knowledge from tree-like structures (like a GitHub repository) by generating implicit, aggregated summaries at each hierarchical level. This approach enables the knowledge to be stored in a knowledge base and used directly with RAG. We then compare our method to using RAG on raw, unstructured code, evaluating the accuracy and quality of the generated responses. Our results show that while response quality is comparable across both methods, our approach generates over 68% fewer documents in the retriever, a significant gain in efficiency. This finding suggests that leveraging implicit, linearized knowledge may be a highly effective and scalable strategy for handling complex, hierarchical data structures.
Authors: Haeji Jung, Jinju Kim, Kyungjin Kim, Youjeong Roh, David R. Mortensen
Abstract: Transliteration has emerged as a promising means to bridge the gap between various languages in multilingual NLP, showing promising results especially for languages using non-Latin scripts. We investigate the degree to which shared script, overlapping token vocabularies, and shared phonology contribute to performance of multilingual models. To this end, we conduct controlled experiments using three kinds of transliteration (romanization, phonemic transcription, and substitution ciphers) as well as orthography. We evaluate each model on two downstream tasks -- named entity recognition (NER) and natural language inference (NLI) -- and find that romanization significantly outperforms other input types in 7 out of 8 evaluation settings, largely consistent with our hypothesis that it is the most effective approach. We further analyze how each factor contributed to the success, and suggest that having longer (subword) tokens shared with pre-trained languages leads to better utilization of the model.
Authors: Kaixuan Ren, Preslav Nakov, Usman Naseem
Abstract: As vision-language models become increasingly capable, maintaining a balance between safety and usefulness remains a central challenge. Safety mechanisms, while essential, can backfire, causing over-refusal, where models decline benign requests out of excessive caution. Yet, no existing benchmark has systematically addressed over-refusal in the visual modality. This setting introduces unique challenges, such as dual-use cases where an instruction is harmless, but the accompanying image contains harmful content. Models frequently fail in such scenarios, either refusing too conservatively or completing tasks unsafely, which highlights the need for more fine-grained alignment. The ideal behavior is safe completion, i.e., fulfilling the benign parts of a request while explicitly warning about any potentially harmful elements. To address this, we present DUAL-Bench, the first multimodal benchmark focused on over-refusal and safe completion in VLMs. We evaluated 18 VLMs across 12 hazard categories, with focus on their robustness under semantics-preserving visual perturbations. The results reveal substantial room for improvement: GPT-5-Nano achieves 12.9% safe completion, GPT-5 models average 7.9%, and Qwen models only 3.9%. We hope that DUAL-Bench will foster the development of more nuanced alignment strategies that ensure models remain both safe and useful in complex multimodal settings.
Authors: Jiajing Guo, Kenil Patel, Jorge Piazentin Ono, Wenbin He, Liu Ren
Abstract: Large language models (LLMs) are increasingly powering Text-to-SQL (Text2SQL) systems, enabling non-expert users to query industrial databases using natural language. While test-time scaling strategies have shown promise in LLM-based solutions, their effectiveness in real-world applications, especially with the latest reasoning models, remains uncertain. In this work, we benchmark six lightweight, industry-oriented test-time scaling strategies and four LLMs, including two reasoning models, evaluating their performance on the BIRD Mini-Dev benchmark. Beyond standard accuracy metrics, we also report inference latency and token consumption, providing insights relevant for practical system deployment. Our findings reveal that Divide-and-Conquer prompting and few-shot demonstrations consistently enhance performance for both general-purpose and reasoning-focused LLMs. However, introducing additional workflow steps yields mixed results, and base model selection plays a critical role. This work sheds light on the practical trade-offs between accuracy, efficiency, and complexity when deploying Text2SQL systems.
Authors: Yu Chao, Siyu Lin, xiaorong wang, Zhu Zhang, Zihan Zhou, Haoyu Wang, Shuo Wang, Jie Zhou, Zhiyuan Liu, Maosong Sun
Abstract: We introduce LLM x MapReduce-V3, a hierarchically modular agent system designed for long-form survey generation. Building on the prior work, LLM x MapReduce-V2, this version incorporates a multi-agent architecture where individual functional components, such as skeleton initialization, digest construction, and skeleton refinement, are implemented as independent model-context-protocol (MCP) servers. These atomic servers can be aggregated into higher-level servers, creating a hierarchically structured system. A high-level planner agent dynamically orchestrates the workflow by selecting appropriate modules based on their MCP tool descriptions and the execution history. This modular decomposition facilitates human-in-the-loop intervention, affording users greater control and customization over the research process. Through a multi-turn interaction, the system precisely captures the intended research perspectives to generate a comprehensive skeleton, which is then developed into an in-depth survey. Human evaluations demonstrate that our system surpasses representative baselines in both content depth and length, highlighting the strength of MCP-based modular planning.
Authors: Ki Jung Seo, Sehun Lim, Taeuk Kim
Abstract: Recent progress in large language models (LLMs) has enabled them to express their confidence in natural language, enhancing transparency and reliability. However, their confidence often exhibits overconfidence, the cause of which remains poorly understood. In this work, we conduct a detailed analysis of the dynamics underlying verbalized confidence and identify answer-independence as a key factor, defined as the model's failure to condition confidence on its own answer. To address this, we propose ADVICE (Answer-Dependent Verbalized Confidence Estimation), a fine-tuning framework that facilitates answer-grounded confidence estimation. Extensive experiments show that ADVICE substantially improves confidence calibration while preserving task performance. Further analyses confirm that ADVICE strengthens answer-groundedness, leading to more balanced and well-calibrated confidence distributions. Our findings shed light on the origin of overconfidence and establish a framework for more trustworthy confidence verbalization.
Authors: Yawen Yang, Fukun Ma, Shiao Meng, Aiwei Liu, Lijie Wen
Abstract: In biomedical fields, one named entity may consist of a series of non-adjacent tokens and overlap with other entities. Previous methods recognize discontinuous entities by connecting entity fragments or internal tokens, which face challenges of error propagation and decoding ambiguity due to the wide variety of span or word combinations. To address these issues, we deeply explore discontinuous entity structures and propose an effective Gap-aware grid tagging model for Discontinuous Named Entity Recognition, named GapDNER. Our GapDNER innovatively applies representation learning on the context gaps between entity fragments to resolve decoding ambiguity and enhance discontinuous NER performance. Specifically, we treat the context gap as an additional type of span and convert span classification into a token-pair grid tagging task. Subsequently, we design two interactive components to comprehensively model token-pair grid features from both intra- and inter-span perspectives. The intra-span regularity extraction module employs the biaffine mechanism along with linear attention to capture the internal regularity of each span, while the inter-span relation enhancement module utilizes criss-cross attention to obtain semantic relations among different spans. At the inference stage of entity decoding, we assign a directed edge to each entity fragment and context gap, then use the BFS algorithm to search for all valid paths from the head to tail of grids with entity tags. Experimental results on three datasets demonstrate that our GapDNER achieves new state-of-the-art performance on discontinuous NER and exhibits remarkable advantages in recognizing complex entity structures.
Authors: Katherine M. Collins, Cedegao E. Zhang, Graham Todd, Lance Ying, Mauricio Barba da Costa, Ryan Liu, Prafull Sharma, Adrian Weller, Ionatan Kuperwajs, Lionel Wong, Joshua B. Tenenbaum, Thomas L. Griffiths
Abstract: Reasoning is not just about solving problems -- it is also about evaluating which problems are worth solving at all. Evaluations of artificial intelligence (AI) systems primarily focused on problem solving, historically by studying how models play games such as chess and Go. In this paper, we advocate for a new paradigm that assesses AI systems' evaluation of games. First, we introduce a formalism for evaluating such evaluations. We then leverage a large-scale dataset of over $100$ novel board games and over 450 human judgments to compare evaluations produced by modern language and reasoning models against those of people and symbolic computational agents. We consider two kinds of evaluative queries: assessing the payoff (or fairness) and the funness of games. These queries span two dimensions relevant to the design of evaluations of AI evaluations: how complex a query is to compute and how difficult a query is to quantify. Our results show that reasoning models are generally more aligned to people in their evaluations of games than non-reasoning language models. However, we observe a non-monotonic relationship: as models get closer to game-theoretic optimal, their fit to human data weakens. We also observe more "jaggedness" across models for assessing funness, in line with the greater difficulty of quantifying this query. Across queries and games, reasoning models show highly variable and unpredictable resource usage when assessing queries, pointing to the importance of imbuing more resource-rational meta-reasoning in language and reasoning models.
Authors: Anirudh Ganesh, Jayavardhan Reddy
Abstract: We present a reproducibility study of the state-of-the-art neural architecture for sequence labeling proposed by Ma and Hovy (2016)\cite{ma2016end}. The original BiLSTM-CNN-CRF model combines character-level representations via Convolutional Neural Networks (CNNs), word-level context modeling through Bi-directional Long Short-Term Memory networks (BiLSTMs), and structured prediction using Conditional Random Fields (CRFs). This end-to-end approach eliminates the need for hand-crafted features while achieving excellent performance on named entity recognition (NER) and part-of-speech (POS) tagging tasks. Our implementation successfully reproduces the key results, achieving 91.18\% F1-score on CoNLL-2003 NER and demonstrating the model's effectiveness across sequence labeling tasks. We provide a detailed analysis of the architecture components and release an open-source PyTorch implementation to facilitate further research.
Authors: Eitan Klinger, Vivaan Wadhwa, Jungyeul Park
Abstract: This article presents a curated resource and evaluation suite for punctuation-aware treebank binarization. Standard binarization pipelines drop punctuation before head selection, which alters constituent shape and harms head-child identification. We release (1) a reproducible pipeline that preserves punctuation as sibling nodes prior to binarization, (2) derived artifacts and metadata (intermediate @X markers, reversibility signatures, alignment indices), and (3) an accompanying evaluation suite covering head-child prediction, round-trip reversibility, and structural compatibility with derivational resources (CCGbank). On the Penn Treebank, punctuation-aware preprocessing improves head prediction accuracy from 73.66\% (Collins rules) and 86.66\% (MLP) to 91.85\% with the same classifier, and achieves competitive alignment against CCGbank derivations. All code, configuration files, and documentation are released to enable replication and extension to other corpora.
Authors: Yejin Lee, Su-Hyeon Kim, Hyundong Jin, Dayoung Kim, Yeonsoo Kim, Yo-Sub Han
Abstract: Toxic content has become an increasingly critical social issue with the rapid expansion of online communication. While numerous studies explored methods for detecting and detoxifying such content, most have focused primarily on English, leaving low-resource language underrepresented. Consequently, Large Language Models~(LLMs) often struggle to identify and neutralize toxic expressions in these languages. This challenge becomes even more pronounced when user employ obfuscation techniques to evade detection systems. Therefore, we propose a \textbf{KOTOX: Korean Toxic Dataset} for deobfuscation and detoxicification to address this issue. We categorize various obfuscation approaches based on linguistic characteristics of Korean and define a set of transformation rules grounded in real-word examples. Using these rules, we construct three dataset versions (easy, normal, and hard) representing different levels of obfuscation difficulty. This is the first dataset that simultaneously supports deobfuscation and detoxification for the Korean language. We expect it to facilitate better understanding and mitigating of obfuscated toxic content in LLM for low-resource languages. Our code and data are available at https://github.com/leeyejin1231/KOTOX.
Authors: Jidong Li, Lingyong Fang, Haodong Zhao, Sufeng Duan, Gongshen Liu
Abstract: Multimodal large language models (MLLMs) have witnessed astonishing advancements in recent years. Despite these successes, MLLMs remain vulnerable to flase premise problems. However, existing benchmarks targeting this issue are limited in scope: they often lack fine-grained categorization, exhibit insufficient coverage, and thus fail to provide a rigorous evaluation of the ability of models to recognize false premises. To bridge this gap, we introduce a fully automated pipeline for constructing a comprehensive benchmark of false premise questions. Our method systematically categorizes the premises into three main types and thirteen subtypes according to the abilities required to identify the premises, resulting in the JBA dataset.Results show current MLLMs still struggle with false premise recognition. Building upon this benchmark, we further propose a recognition enhancement framework tailored to strengthen the robustness of MLLMs to detect false premises. Extensive experiments demonstrate that models trained with our framework achieve significant improvements in false premise recognition.
Authors: Yejin Lee, Hyeseon Ahn, Yo-Sub Han
Abstract: Hate speech remains prevalent in human society and continues to evolve in its forms and expressions. Modern advancements in internet and online anonymity accelerate its rapid spread and complicate its detection. However, hate speech datasets exhibit diverse characteristics primarily because they are constructed from different sources and platforms, each reflecting different linguistic styles and social contexts. Despite this diversity, prior studies on hate speech detection often rely on fixed methodologies without adapting to data-specific features. We introduce RV-HATE, a detection framework designed to account for the dataset-specific characteristics of each hate speech dataset. RV-HATE consists of multiple specialized modules, where each module focuses on distinct linguistic or contextual features of hate speech. The framework employs reinforcement learning to optimize weights that determine the contribution of each module for a given dataset. A voting mechanism then aggregates the module outputs to produce the final decision. RV-HATE offers two primary advantages: (1)~it improves detection accuracy by tailoring the detection process to dataset-specific attributes, and (2)~it also provides interpretable insights into the distinctive features of each dataset. Consequently, our approach effectively addresses implicit hate speech and achieves superior performance compared to conventional static methods. Our code is available at https://github.com/leeyejin1231/RV-HATE.
Authors: Zhiwen Ruan, Yixia Li, He Zhu, Yun Chen, Peng Li, Yang Liu, Guanhua Chen
Abstract: Large language models (LLMs) primarily rely on supervised fine-tuning (SFT) as a key method to adapt pre-trained models to domain-specific tasks such as mathematical reasoning. However, standard SFT uniformly penalizes all tokens, neglecting that only a small subset of critical tokens determines reasoning correctness. This uniform supervision often causes reduced output diversity and limited generalization. We propose Critical Token Fine-tuning (CFT), a simple yet effective approach that updates only tokens identified as functionally indispensable via counterfactual perturbations. By focusing gradient signals on these decisive reasoning steps while preserving the diversity of non-critical tokens, CFT can enhance both generation and diversity. Extensive experiments on five models across three families (Qwen, OLMo, LLaMA) and eleven mathematical reasoning benchmarks show that CFT, despite fine-tuning on less than 12% of tokens, consistently outperforms standard SFT. Moreover, CFT enables test-time scaling through improved sampling diversity and provides a stronger initialization for reinforcement learning, sustaining performance gains in later training stages while maintaining higher entropy for better exploration. These results highlight CFT as a practical and general framework for efficient and robust LLM fine-tuning.
Authors: Wei-Chieh Huang, Henry Peng Zou, Yaozu Wu, Dongyuan Li, Yankai Chen, Weizhi Zhang, Yangning Li, Angelo Zangari, Jizhou Guo, Chunyu Miao, Liancheng Fang, Langzhou He, Renhe Jiang, Philip S. Yu
Abstract: Deep research frameworks have shown promising capabilities in synthesizing comprehensive reports from web sources. While deep research possesses significant potential to address complex issues through planning and research cycles, existing frameworks are deficient in sufficient evaluation procedures and stage-specific protections. They typically treat evaluation as exact match accuracy of question-answering, but overlook crucial aspects of report quality such as credibility, coherence, breadth, depth, and safety. This oversight may result in hazardous or malicious sources being integrated into the final report. To address these issues, we introduce DEEPRESEARCHGUARD, a comprehensive framework featuring four-stage safeguards with open-domain evaluation of references and reports. We assess performance across multiple metrics, e.g., defense success rate and over-refusal rate, and five key report dimensions. In the absence of a suitable safety benchmark, we introduce DRSAFEBENCH, a stage-wise benchmark for deep research safety. Our evaluation spans diverse state-of-the-art LLMs, including GPT-4o, Gemini-2.5-flash, DeepSeek-v3, and o4-mini. DEEPRESEARCHGUARD achieves an average defense success rate improvement of 18.16% while reducing over-refusal rate by 6%. The input guard provides the most substantial early-stage protection by filtering out obvious risks, while the plan and research guards enhance citation discipline and source credibility. Through extensive experiments, we show that DEEPRESEARCHGUARD enables comprehensive open-domain evaluation and stage-aware defenses that effectively block harmful content propagation, while systematically improving report quality without excessive over-refusal rates. The code can be found via https://github.com/Jasonya/DeepResearchGuard.
Authors: Mahika Phutane, Hayoung Jung, Matthew Kim, Tanushree Mitra, Aditya Vashistha
Abstract: Large language models (LLMs) are increasingly under scrutiny for perpetuating identity-based discrimination in high-stakes domains such as hiring, particularly against people with disabilities (PwD). However, existing research remains largely Western-centric, overlooking how intersecting forms of marginalization--such as gender and caste--shape experiences of PwD in the Global South. We conduct a comprehensive audit of six LLMs across 2,820 hiring scenarios spanning diverse disability, gender, nationality, and caste profiles. To capture subtle intersectional harms and biases, we introduce ABLEIST (Ableism, Inspiration, Superhumanization, and Tokenism), a set of five ableism-specific and three intersectional harm metrics grounded in disability studies literature. Our results reveal significant increases in ABLEIST harms towards disabled candidates--harms that many state-of-the-art models failed to detect. These harms were further amplified by sharp increases in intersectional harms (e.g., Tokenism) for gender and caste-marginalized disabled candidates, highlighting critical blind spots in current safety tools and the need for intersectional safety evaluations of frontier models in high-stakes domains like hiring.
Authors: Tieyuan Chen, Xiaodong Chen, Haoxing Chen, Zhenzhong Lan, Weiyao Lin, Jianguo Li
Abstract: We introduce Dynamic Nested Depth (DND), a novel method that improves performance for off-the-shelf LLMs by selecting critical tokens to reprocess in a nested depth manner. Specifically, at the end of the given transformer layer, DND identifies more critical tokens with a router and feeds them back for an extra round of processing, effectively ``reviewing" difficult tokens while avoiding redundant computation for easier ones. The dynamic selection mechanism is tailored for precise control via two novel strategies: a router controlling loss to enhance token selection distinguishability, and a threshold control scheme to ensure selection stability. We demonstrate the effectiveness of DND by directly integrating it into pre-trained dense and MoE models during a post-training phase. On diverse benchmarks, this approach boosts the performances of the dense Qwen3-1.7B by 1.88% and the MoE Qwen3-30B-A3B by 0.87%, all with a minimal parameter and computing increase.
Authors: Yiwei Liu, Yucheng Li, Xiao Li, Gong Cheng
Abstract: Joint logical-numerical reasoning remains a major challenge for language models, yet existing datasets rely on fixed rule sets and offer limited control over task complexity, constraining their generalizability for evaluation and training. We present LogiNumSynth, a flexible natural language problem synthesizer that synthesizes tasks requiring proficiency in joint logical reasoning (e.g., rule-based reasoning) and numerical reasoning (e.g., arithmetic computation). LogiNumSynth supports fine-grained control over reasoning world richness, logical reasoning depth, and the complexity of numerical computations, enabling flexible data synthesis across difficulty levels. We demonstrate three key contributions: (1) Synthesizer -- synthesizing fully controllable joint reasoning tasks over natural language; (2) Evaluation & Process Analysis -- evaluating both process accuracy and answer accuracy; (3) Targeted Training -- using synthesized data to enhance LLMs' reasoning performance. Experiments with multiple LLMs highlight persistent weaknesses in logical-numerical reasoning, showing that LogiNumSynth can serve as both a diagnostic tool and a source of targeted supervision for advancing integrated reasoning skills.
Authors: Wenya Xie, Qingying Xiao, Yu Zheng, Xidong Wang, Junying Chen, Ke Ji, Anningzhe Gao, Prayag Tiwari, Xiang Wan, Feng Jiang, Benyou Wang
Abstract: The rise of large language models (LLMs) has transformed healthcare by offering clinical guidance, yet their direct deployment to patients poses safety risks due to limited domain expertise. To mitigate this, we propose repositioning LLMs as clinical assistants that collaborate with experienced physicians rather than interacting with patients directly. We conduct a two-stage inspiration-feedback survey to identify real-world needs in clinical workflows. Guided by this, we construct DoctorFLAN, a large-scale Chinese medical dataset comprising 92,000 Q&A instances across 22 clinical tasks and 27 specialties. To evaluate model performance in doctor-facing applications, we introduce DoctorFLAN-test (550 single-turn Q&A items) and DotaBench (74 multi-turn conversations). Experimental results with over ten popular LLMs demonstrate that DoctorFLAN notably improves the performance of open-source LLMs in medical contexts, facilitating their alignment with physician workflows and complementing existing patient-oriented models. This work contributes a valuable resource and framework for advancing doctor-centered medical LLM development
Authors: Qinglin Zhu, Yizhen Yao, Runcong Zhao, Yanzheng Xiang, Amrutha Saseendran, Chen Jin, Philip Alexander Teare, Bin Liang, Yulan He, Lin Gui
Abstract: Autoregressive (AR) models remain the standard for natural language generation but still suffer from high latency due to strictly sequential decoding. Recent diffusion-inspired approaches, such as LlaDA and Dream, mitigate this by generating in parallel, yet they suffer from two core limitations: information loss, as predictive distributions for non-finalized tokens are discarded at each step, and premature commitment, where local decisions are made without sufficient global coordination. We introduce Latent Refinement Decoding (LRD), a two-stage framework with Latent Refinement and a Predictive Feedback Loop. The first stage maintains masked positions as distributional mixtures of predicted tokens and the mask embedding, allowing the model to establish more globally consistent beliefs. The second stage progressively finalizes confident tokens while retaining uncertain ones for iterative feedback. KL-divergence dynamics provide a principled and reliable criterion for convergence and early stopping. Experiments across coding (HumanEval +6.3, MBPP +2.6) and reasoning (GSM8K +2.9, MATH500 +3.8) show that LRD improves accuracy while delivering speedups of up to 10.6x, making it a strong and versatile alternative for parallel sequence generation.
Authors: Junjie Lu, Yuliang Liu, Chaofeng Qu, Wei Shen, Zhouhan Lin, Min Xu
Abstract: Current approaches for strengthening LLM reasoning tend to introduce a training bias toward human-like reasoning trajectories. In step-wise preference optimization, in particular, dependence on human or higher-capacity model annotations for intermediate steps limits exploration of alternative, non-human-like reasoning paths and thus constrains achievable performance. Furthermore, through a small-scale pilot study, we observed that in approximately 75% of cases, the model's first erroneous step occurs after the lowest-confidence point. This suggests that guiding the model at its lowest-confidence point before an error provides more accurate supervision than locating the first explicit error. In this paper, we propose Confidence-Guided Reasoning Path Preference Optimization (CGPO), a method that leverages a confidence signal to identify points of maximal uncertainty in the model's reasoning process and applies self-generated, non-human-like reasoning-path guidance to mitigate trajectory drift. Our experiments span diverse models applied to both code and mathematical reasoning tasks. The results show that, with the same amount of training data, our method using data generated by a small model can achieve better performance in most cases compared with approaches using data generated by a strong model or human-annotated.
Authors: Alexander Sternfeld, Andrei Kucharavy, Ljiljana Dolamic
Abstract: Large language Models (LLMs) have shown remarkable proficiency in code generation tasks across various programming languages. However, their outputs often contain subtle but critical vulnerabilities, posing significant risks when deployed in security-sensitive or mission-critical systems. This paper introduces TypePilot, an agentic AI framework designed to enhance the security and robustness of LLM-generated code by leveraging strongly typed and verifiable languages, using Scala as a representative example. We evaluate the effectiveness of our approach in two settings: formal verification with the Stainless framework and general-purpose secure code generation. Our experiments with leading open-source LLMs reveal that while direct code generation often fails to enforce safety constraints, just as naive prompting for more secure code, our type-focused agentic pipeline substantially mitigates input validation and injection vulnerabilities. The results demonstrate the potential of structured, type-guided LLM workflows to improve the SotA of the trustworthiness of automated code generation in high-assurance domains.
Authors: Jens Van Nooten, Andriy Kosar, Guy De Pauw, Walter Daelemans
Abstract: Distance-based unsupervised text classification is a method within text classification that leverages the semantic similarity between a label and a text to determine label relevance. This method provides numerous benefits, including fast inference and adaptability to expanding label sets, as opposed to zero-shot, few-shot, and fine-tuned neural networks that require re-training in such cases. In multi-label distance-based classification and information retrieval algorithms, thresholds are required to determine whether a text instance is "similar" to a label or query. Similarity between a text and label is determined in a dense embedding space, usually generated by state-of-the-art sentence encoders. Multi-label classification complicates matters, as a text instance can have multiple true labels, unlike in multi-class or binary classification, where each instance is assigned only one label. We expand upon previous literature on this underexplored topic by thoroughly examining and evaluating the ability of sentence encoders to perform distance-based classification. First, we perform an exploratory study to verify whether the semantic relationships between texts and labels vary across models, datasets, and label sets by conducting experiments on a diverse collection of realistic multi-label text classification (MLTC) datasets. We find that similarity distributions show statistically significant differences across models, datasets and even label sets. We propose a novel method for optimizing label-specific thresholds using a validation set. Our label-specific thresholding method achieves an average improvement of 46% over normalized 0.5 thresholding and outperforms uniform thresholding approaches from previous work by an average of 14%. Additionally, the method demonstrates strong performance even with limited labeled examples.
Authors: Paloma Piot, Jos\'e Ramom Pichel Campos, Javier Parapar
Abstract: Hate speech poses a serious threat to social cohesion and individual well-being, particularly on social media, where it spreads rapidly. While research on hate speech detection has progressed, it remains largely focused on English, resulting in limited resources and benchmarks for low-resource languages. Moreover, many of these languages have multiple linguistic varieties, a factor often overlooked in current approaches. At the same time, large language models require substantial amounts of data to perform reliably, a requirement that low-resource languages often cannot meet. In this work, we address these gaps by compiling a meta-collection of hate speech datasets for European Spanish, standardised with unified labels and metadata. This collection is based on a systematic analysis and integration of existing resources, aiming to bridge the data gap and support more consistent and scalable hate speech detection. We extended this collection by translating it into European Portuguese and into a Galician standard that is more convergent with Spanish and another Galician variant that is more convergent with Portuguese, creating aligned multilingual corpora. Using these resources, we establish new benchmarks for hate speech detection in Iberian languages. We evaluate state-of-the-art large language models in zero-shot, few-shot, and fine-tuning settings, providing baseline results for future research. Moreover, we perform a cross-lingual analysis with our target languages. Our findings underscore the importance of multilingual and variety-aware approaches in hate speech detection and offer a foundation for improved benchmarking in underrepresented European languages.
Authors: Johannes Moll, Markus Graf, Tristan Lemke, Nicolas Lenhart, Daniel Truhn, Jean-Benoit Delbrouck, Jiazhen Pan, Daniel Rueckert, Lisa C. Adams, Keno K. Bressem
Abstract: Vision-language models (VLMs) often produce chain-of-thought (CoT) explanations that sound plausible yet fail to reflect the underlying decision process, undermining trust in high-stakes clinical use. Existing evaluations rarely catch this misalignment, prioritizing answer accuracy or adherence to formats. We present a clinically grounded framework for chest X-ray visual question answering (VQA) that probes CoT faithfulness via controlled text and image modifications across three axes: clinical fidelity, causal attribution, and confidence calibration. In a reader study (n=4), evaluator-radiologist correlations fall within the observed inter-radiologist range for all axes, with strong alignment for attribution (Kendall's $\tau_b=0.670$), moderate alignment for fidelity ($\tau_b=0.387$), and weak alignment for confidence tone ($\tau_b=0.091$), which we report with caution. Benchmarking six VLMs shows that answer accuracy and explanation quality are decoupled, acknowledging injected cues does not ensure grounding, and text cues shift explanations more than visual cues. While some open-source models match final answer accuracy, proprietary models score higher on attribution (25.0% vs. 1.4%) and often on fidelity (36.1% vs. 31.7%), highlighting deployment risks and the need to evaluate beyond final answer accuracy.
Authors: Yisong Miao, Min-Yen Kan
Abstract: Which components in transformer language models are responsible for discourse understanding? We hypothesize that sparse computational graphs, termed as discursive circuits, control how models process discourse relations. Unlike simpler tasks, discourse relations involve longer spans and complex reasoning. To make circuit discovery feasible, we introduce a task called Completion under Discourse Relation (CuDR), where a model completes a discourse given a specified relation. To support this task, we construct a corpus of minimal contrastive pairs tailored for activation patching in circuit discovery. Experiments show that sparse circuits ($\approx 0.2\%$ of a full GPT-2 model) recover discourse understanding in the English PDTB-based CuDR task. These circuits generalize well to unseen discourse frameworks such as RST and SDRT. Further analysis shows lower layers capture linguistic features such as lexical semantics and coreference, while upper layers encode discourse-level abstractions. Feature utility is consistent across frameworks (e.g., coreference supports Expansion-like relations).
Authors: Chris Xing Tian, Weihao Xie, Zhen Chen, Zhengyuan Yi, Hui Liu, Haoliang Li, Shiqi Wang, Siwei Ma
Abstract: Retrieval-Augmented Generation (RAG) combines the language understanding and reasoning power of large language models (LLMs) with external retrieval to enable domain-grounded responses. Effectively adapting RAG systems to domain-specific settings requires specialized, context-rich training data beyond general-purpose question-answering. Here, we propose RAGen, a scalable and modular framework for generating domain-grounded question-answer-context (QAC) triples tailored to diverse RAG adaptation approaches. RAGen produces these QAC triples by identifying key concepts in documents, generating diverse questions guided by Bloom's Taxonomy-inspired principles, and pairing them with precise answers extracted from relevant contexts. RAGen supports multiple RAG adaptation strategies, including the optimization of key components such as the LLM, retriever, and embedding model, etc. Its modular pipeline features semantic chunking, hierarchical concept extraction, and multi-chunk retrieval, along with the introduction of curated distractor contexts to promote robust reasoning. Designed for scalability, RAGen efficiently handles large and evolving document corpora without redundant processing, making it especially suitable for dynamic evolving domains such as scientific research and enterprise knowledge bases.
Authors: Saad Obaid ul Islam, Anne Lauscher, Goran Glava\v{s}
Abstract: Large language models (LLMs) can correctly answer "When was Einstein born?" yet fail to provide the same date when writing about Einstein's life revealing a fundamental inconsistency in how models access factual knowledge across task complexities. While models display impressive accuracy on factual question-answering benchmarks, the reliability gap between simple and complex queries remains poorly understood, eroding their trustworthiness. In this work, we introduce Short-Long Form Alignment for Factual Question Answering (SLAQ), a controlled evaluation framework that compares LLMs' answers to the same factual questions asked (a) in isolation (short) vs. (b) integrated into complex queries (long). Looking at 16 LLMs across 600 queries, we find a systematic misalignment of answers to the corresponding short and long queries. We further uncover position-dependent accuracy loss and momentum effects where consecutive correct or incorrect answers create self-reinforcing patterns. Through mechanistic analysis, we find that aligned facts activate overlapping model internals, and that metrics based on mechanistic similarity can predict short-long answer alignment with up to 78% accuracy. Our work establishes factual consistency over query complexity as an important aspect of LLMs' trustworthiness and challenges current evaluation practices, which implicitly assume that good performance for simple factual queries implies reliability in more complex knowledge-seeking tasks too.
Authors: Tao Li, Jinlong Hu, Yang Wang, Junfeng Liu, Xuejun Liu
Abstract: LLM-brained web agents offer powerful capabilities for web automation but face a critical cost-performance trade-off. The challenge is amplified by web agents' inherently complex prompts that include goals, action histories, and environmental states, leading to degraded LLM ensemble performance. To address this, we introduce WebRouter, a novel query-specific router trained from an information-theoretic perspective. Our core contribution is a cost-aware Variational Information Bottleneck (ca-VIB) objective, which learns a compressed representation of the input prompt while explicitly penalizing the expected operational cost. Experiments on five real-world websites from the WebVoyager benchmark show that WebRouter reduces operational costs by a striking 87.8\% compared to a GPT-4o baseline, while incurring only a 3.8\% accuracy drop.
Authors: Battemuulen Naranbat, Seyed Sahand Mohammadi Ziabari, Yousuf Nasser Al Husaini, Ali Mohammed Mansoor Alsahag
Abstract: Ensuring fairness in natural language processing for moral sentiment classification is challenging, particularly under cross-domain shifts where transformer models are increasingly deployed. Using the Moral Foundations Twitter Corpus (MFTC) and Moral Foundations Reddit Corpus (MFRC), this work evaluates BERT and DistilBERT in a multi-label setting with in-domain and cross-domain protocols. Aggregate performance can mask disparities: we observe pronounced asymmetry in transfer, with Twitter->Reddit degrading micro-F1 by 14.9% versus only 1.5% for Reddit->Twitter. Per-label analysis reveals fairness violations hidden by overall scores; notably, the authority label exhibits Demographic Parity Differences of 0.22-0.23 and Equalized Odds Differences of 0.40-0.41. To address this gap, we introduce the Moral Fairness Consistency (MFC) metric, which quantifies the cross-domain stability of moral foundation detection. MFC shows strong empirical validity, achieving a perfect negative correlation with Demographic Parity Difference (rho = -1.000, p < 0.001) while remaining independent of standard performance metrics. Across labels, loyalty demonstrates the highest consistency (MFC = 0.96) and authority the lowest (MFC = 0.78). These findings establish MFC as a complementary, diagnosis-oriented metric for fairness-aware evaluation of moral reasoning models, enabling more reliable deployment across heterogeneous linguistic contexts. .
Authors: Hayate Funakura, Hyunsoo Kim, Koji Mineshima
Abstract: Graph-matching metrics such as Smatch are the de facto standard for evaluating neural semantic parsers, yet they capture surface overlap rather than logical equivalence. We reassess evaluation by pairing graph-matching with automated theorem proving. We compare two approaches to building parsers: supervised fine-tuning (T5-Small/Base) and few-shot in-context learning (GPT-4o/4.1/5), under normalized and unnormalized targets. We evaluate outputs using graph-matching, bidirectional entailment between source and target formulas with a first-order logic theorem prover, and well-formedness. Across settings, we find that models performing well on graph-matching often fail to produce logically equivalent formulas. Normalization reduces incidental target variability, improves well-formedness, and strengthens logical adequacy. Error analysis shows performance degrades with increasing formula complexity and with coordination, prepositional phrases, and passive voice; the dominant failures involve variable binding and indexing, and predicate naming. These findings highlight limits of graph-based metrics for reasoning-oriented applications and motivate logic-sensitive evaluation and training objectives together with simplified, normalized target representations. All code and data for our experiments are publicly available.
Authors: Jinyuan Xu, Tian Lan, Xintao Yu, Xue He, Hezhi Zhang, Ying Wang, Pierre Magistry, Mathieu Valette, Lei Li
Abstract: Depression is a pressing global public health issue, yet publicly available Chinese-language resources for risk detection remain scarce and are mostly limited to binary classification. To address this limitation, we release CNSocialDepress, a benchmark dataset for depression risk detection from Chinese social media posts. The dataset contains 44,178 texts from 233 users, within which psychological experts annotated 10,306 depression-related segments. CNSocialDepress provides binary risk labels together with structured multi-dimensional psychological attributes, enabling interpretable and fine-grained analysis of depressive signals. Experimental results demonstrate its utility across a wide range of NLP tasks, including structured psychological profiling and fine-tuning of large language models for depression detection. Comprehensive evaluations highlight the dataset's effectiveness and practical value for depression risk identification and psychological analysis, thereby providing insights to mental health applications tailored for Chinese-speaking populations.
Authors: Haoqi Yang, Yao Yao, Zuchao Li, Baoyuan Qi, Guoming Liu, Hai Zhao
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse natural language processing tasks. However, their extensive memory requirements, particularly due to KV cache growth during long-text understanding and generation, present significant challenges for deployment in resource-constrained environments. Quantization has emerged as a promising solution to reduce memory consumption while preserving historical information. We propose XQuant, a training-free and plug-and-play framework that achieves ultra-low equivalent bit-width KV cache quantization. XQuant introduces two key innovations: a computationally negligible data-free calibration method and cross-layer KV cache compression, enabling quantization to sub-1.4 bits. Extensive experiments on TruthfulQA and LongBench demonstrate that XQuant outperforms state-of-the-art methods (e.g., KIVI-2bit and AsymKV-1.5bit) by achieving lower bit-width while maintaining superior performance, establishing a better trade-off between memory efficiency and model accuracy.
Authors: Michael Schlichtkrull
Abstract: When AI agents retrieve and reason over external documents, adversaries can manipulate the data they receive to subvert their behaviour. Previous research has studied indirect prompt injection, where the attacker injects malicious instructions. We argue that injection of instructions is not necessary to manipulate agents - attackers could instead supply biased, misleading, or false information. We term this an attack by content. Existing defenses, which focus on detecting hidden commands, are ineffective against attacks by content. To defend themselves and their users, agents must critically evaluate retrieved information, corroborating claims with external evidence and evaluating source trustworthiness. We argue that this is analogous to an existing NLP task, automated fact-checking, which we propose to repurpose as a cognitive self-defense tool for agents.
Authors: Jana Jung, Marlene Lutz, Indira Sen, Markus Strohmaier
Abstract: Psychometric tests are increasingly used to assess psychological constructs in large language models (LLMs). However, it remains unclear whether these tests -- originally developed for humans -- yield meaningful results when applied to LLMs. In this study, we systematically evaluate the reliability and validity of human psychometric tests for three constructs: sexism, racism, and morality. We find moderate reliability across multiple item and prompt variations. Validity is evaluated through both convergent (i.e., testing theory-based inter-test correlations) and ecological approaches (i.e., testing the alignment between tests scores and behavior in real-world downstream tasks). Crucially, we find that psychometric test scores do not align, and in some cases even negatively correlate with, model behavior in downstream tasks, indicating low ecological validity. Our results highlight that systematic evaluations of psychometric tests is essential before interpreting their scores. They also suggest that psychometric tests designed for humans cannot be applied directly to LLMs without adaptation.
Authors: Guangyu Wei, Ke Han, Yueming Lyu, Yu Luo, Yue Jiang, Caifeng Shan, Nicu Sebe
Abstract: Fake news detection becomes particularly challenging in real-time scenarios, where emerging events often lack sufficient supporting evidence. Existing approaches often rely heavily on external evidence and therefore struggle to generalize under evidence scarcity. To address this issue, we propose Evaluation-Aware Selection of Experts (EASE), a novel framework for real-time fake news detection that dynamically adapts its decision-making process according to the assessed sufficiency of available evidence. EASE introduces a sequential evaluation mechanism comprising three independent perspectives: (1) Evidence-based evaluation, which assesses evidence and incorporates it into decision-making only when the evidence is sufficiently supportive; (2) Reasoning-based evaluation, which leverages the world knowledge of large language models (LLMs) and applies them only when their reliability is adequately established; and (3) Sentiment-based fallback, which integrates sentiment cues when neither evidence nor reasoning is reliable. To enhance the accuracy of evaluation processes, EASE employs instruction tuning with pseudo labels to guide each evaluator in justifying its perspective-specific knowledge through interpretable reasoning. Furthermore, the expert modules integrate the evaluators' justified assessments with the news content to enable evaluation-aware decision-making, thereby enhancing overall detection accuracy. Moreover, we introduce RealTimeNews-25, a new benchmark comprising recent news for evaluating model generalization on emerging news with limited evidence. Extensive experiments demonstrate that EASE not only achieves state-of-the-art performance across multiple benchmarks, but also significantly improves generalization to real-time news. The code and dataset are available: https://github.com/wgyhhhh/EASE.
Authors: Nikita Afonin, Nikita Andriyanov, Nikhil Bageshpura, Kyle Liu, Kevin Zhu, Sunishchal Dev, Ashwinee Panda, Alexander Panchenko, Oleg Rogov, Elena Tutubalina, Mikhail Seleznyov
Abstract: Recent work has shown that narrow finetuning can produce broadly misaligned LLMs, a phenomenon termed emergent misalignment (EM). While concerning, these findings were limited to finetuning and activation steering, leaving out in-context learning (ICL). We therefore ask: does EM emerge in ICL? We find that it does: across three datasets, three frontier models produce broadly misaligned responses at rates between 2% and 17% given 64 narrow in-context examples, and up to 58% with 256 examples. We also examine mechanisms of EM by eliciting step-by-step reasoning (while leaving in-context examples unchanged). Manual analysis of the resulting chain-of-thought shows that 67.5% of misaligned traces explicitly rationalize harmful outputs by adopting a reckless or dangerous ''persona'', echoing prior results on finetuning-induced EM.
Authors: Ruirui Chen, Weifeng Jiang, Chengwei Qin, Bo Xiong, Fiona Liausvia, Dongkyu Choi, Boon Kiat Quek
Abstract: Knowledge graphs (KGs) are vital for knowledge-intensive tasks and have shown promise in reducing hallucinations in large language models (LLMs). However, constructing high-quality KGs remains difficult, requiring accurate information extraction and structured representations that support interpretability and downstream utility. Existing LLM-based approaches often focus narrowly on entity and relation extraction, limiting coverage to sentence-level contexts or relying on predefined schemas. We propose a hierarchical extraction framework that organizes information at multiple levels, enabling the creation of semantically rich and well-structured KGs. Using state-of-the-art LLMs, we extract and construct knowledge graphs and evaluate them comprehensively from both structural and semantic perspectives. Our results highlight the strengths and shortcomings of current LLMs in KG construction and identify key challenges for future work. To advance research in this area, we also release a curated dataset of LLM-generated KGs derived from research papers on children's mental well-being. This resource aims to foster more transparent, reliable, and impactful applications in high-stakes domains such as healthcare.
Authors: Sabrina McCallum, Amit Parekh, Alessandro Suglia
Abstract: Current approaches to embodied AI tend to learn policies from expert demonstrations. However, without a mechanism to evaluate the quality of demonstrated actions, they are limited to learning from optimal behaviour, or they risk replicating errors and inefficiencies. While reinforcement learning offers one alternative, the associated exploration typically results in sacrificing data efficiency. This work explores how agents trained with imitation learning can learn robust representations from both optimal and suboptimal demonstrations when given access to constructive language feedback as a means to contextualise different modes of behaviour. We directly provide language feedback embeddings as part of the input sequence into a Transformer-based policy, and optionally complement the traditional next action prediction objective with auxiliary self-supervised learning objectives for feedback prediction. We test our approach on a range of embodied Vision-and-Language tasks in our custom BabyAI-XGen environment and show significant improvements in agents' compositional generalisation abilities and robustness, suggesting that our data-efficient method allows models to successfully convert suboptimal behaviour into learning opportunities. Overall, our results suggest that language feedback is a competitive and intuitive alternative to intermediate scalar rewards for language-specified embodied tasks.
Authors: Belkiss Souayed, Sarah Ebling, Yingqiang Gao
Abstract: Individuals with intellectual disabilities often have difficulties in comprehending complex texts. While many text-to-image models prioritize aesthetics over accessibility, it is not clear how visual illustrations relate to text simplifications (TS) generated from them. This paper presents a structured vision-language model (VLM) prompting framework for generating accessible images from simplified texts. We designed five prompt templates, i.e., Basic Object Focus, Contextual Scene, Educational Layout, Multi-Level Detail, and Grid Layout, each following distinct spatial arrangements while adhering to accessibility constraints such as object count limits, spatial separation, and content restrictions. Using 400 sentence-level simplifications from four established TS datasets (OneStopEnglish, SimPA, Wikipedia, and ASSET), we conducted a two-phase evaluation: Phase 1 assessed prompt template effectiveness with CLIPScores, and Phase 2 involved human annotation of generated images across ten visual styles by four accessibility experts. Results show that the Basic Object Focus prompt template achieved the highest semantic alignment, indicating that visual minimalism enhances language accessibility. Expert evaluation further identified Retro style as the most accessible and Wikipedia as the most effective data source. Inter-annotator agreement varied across dimensions, with Text Simplicity showing strong reliability and Image Quality proving more subjective. Overall, our framework offers practical guidelines for accessible content generation and underscores the importance of structured prompting in AI-generated visual accessibility tools.
Authors: Chenxi Wang, Yixuan Zhang, Ruiji Yu, Yufei Zheng, Lang Gao, Zirui Song, Zixiang Xu, Gus Xia, Huishuai Zhang, Dongyan Zhao, Xiuying Chen
Abstract: As the demand for emotional intelligence in large language models (LLMs) grows, a key challenge lies in understanding the internal mechanisms that give rise to emotional expression and in controlling emotions in generated text. This study addresses three core questions: (1) Do LLMs contain context-agnostic mechanisms shaping emotional expression? (2) What form do these mechanisms take? (3) Can they be harnessed for universal emotion control? We first construct a controlled dataset, SEV (Scenario-Event with Valence), to elicit comparable internal states across emotions. Subsequently, we extract context-agnostic emotion directions that reveal consistent, cross-context encoding of emotion (Q1). We identify neurons and attention heads that locally implement emotional computation through analytical decomposition and causal analysis, and validate their causal roles via ablation and enhancement interventions. Next, we quantify each sublayer's causal influence on the model's final emotion representation and integrate the identified local components into coherent global emotion circuits that drive emotional expression (Q2). Directly modulating these circuits achieves 99.65% emotion-expression accuracy on the test set, surpassing prompting- and steering-based methods (Q3). To our knowledge, this is the first systematic study to uncover and validate emotion circuits in LLMs, offering new insights into interpretability and controllable emotional intelligence.
Authors: Hengran Zhang, Keping Bi, Jiafeng Guo, Jiaming Zhang, Shuaiqiang Wang, Dawei Yin, Xueqi Cheng
Abstract: Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating external knowledge. While traditional retrieval focuses on relevance, RAG's effectiveness depends on the utility of retrieved passages, i.e., the usefulness in facilitating the generation of an accurate and comprehensive answer. Existing studies often treat utility as a generic attribute, ignoring the fact that different LLMs may benefit differently from the same passage due to variations in internal knowledge and comprehension ability. In this work, we introduce and systematically investigate the notion of LLM-specific utility. Through large-scale experiments across multiple datasets and LLMs, we demonstrate that human-annotated passages are not optimal for LLMs and that ground-truth utilitarian passages are not transferable across different LLMs. These findings highlight the necessity of adopting the LLM-specific utility in RAG research. Our findings indicate that some human-annotated passages are not ground-truth utilitarian passages for specific LLMs, partially due to the varying readability of queries and passages for LLMs, a tendency for which perplexity is a key metric. Based on these findings, we propose a benchmarking procedure for LLM-specific utility judgments. We evaluate existing utility judgment methods on six datasets and find that while verbalized methods using pseudo-answers perform robustly, LLMs struggle to assess utility effectively-failing to reject all passages for known queries and to select truly useful ones for unknown queries.
Authors: Wenhan Ma, Hailin Zhang, Liang Zhao, Yifan Song, Yudong Wang, Zhifang Sui, Fuli Luo
Abstract: Reinforcement learning (RL) has emerged as a crucial approach for enhancing the capabilities of large language models. However, in Mixture-of-Experts (MoE) models, the routing mechanism often introduces instability, even leading to catastrophic RL training collapse. We analyze the training-inference consistency of MoE models and identify a notable discrepancy in routing behaviors between the two phases. Moreover, even under identical conditions, the routing framework can yield divergent expert selections across repeated forward passes. To address this foundational inconsistency, we propose Rollout Routing Replay (R3), a method that records routing distributions from the inference engine and replays them during training. R3 significantly reduces training-inference policy KL divergence and mitigates extreme discrepancies without compromising training speed. Extensive experiments on various settings confirm that R3 succeeds in stabilizing RL training, preventing collapse and outperforming methods such as GSPO and TIS. We believe this work can offer a new solution for stabilizing RL in MoE models.
Authors: Dean L. Slack, Noura Al Moubayed
Abstract: Although large language models excel across many tasks, they can memorise training data and thereby expose private or copyrighted text. Most defences target the pre-training stage, leaving memorisation during fine-tuning, especially for domain adaptation and instruction tuning, poorly understood. We fine-tune Pythia, Llama3, and Mistral models spanning 1.4B-70B parameters on common evaluation datasets and track verbatim memorisation throughout training. We find that memorisation increases dramatically in the first few epochs, often significantly before either validation perplexity or evaluation performance is optimised. We use a simple but effective n-gram memorisation score which reliably precedes verbatim memorisation; using it as an early-stopping criterion mitigates memorisation with minimal performance loss. Further, we introduce an n-gram-aware loss regulariser and show that it reduces memorisation across all model families tested by up to 40% while minimising evaluation performance trade-offs when compared to an existing memorisation mitigation strategy. These results yield practical, scalable insights into memorisation dynamics during language model fine-tuning.
Authors: Zirui Song, Yuan Huang, Junchang Liu, Haozhe Luo, Chenxi Wang, Lang Gao, Zixiang Xu, Mingfei Han, Xiaojun Chang, Xiuying Chen
Abstract: Social deduction games like Werewolf combine language, reasoning, and strategy, providing a testbed for studying natural language and social intelligence. However, most studies reduce the game to LLM-based self-play, yielding templated utterances and anecdotal cases that overlook the richness of social gameplay. Evaluation further relies on coarse metrics such as survival time or subjective scoring due to the lack of quality reference data. To address these gaps, we curate a high-quality, human-verified multimodal Werewolf dataset containing over 100 hours of video, 32.4M utterance tokens, and 15 rule variants. Based on this dataset, we propose a novel strategy-alignment evaluation that leverages the winning faction's strategies as ground truth in two stages: 1) Speech evaluation, formulated as multiple-choice-style tasks that assess whether the model can adopt appropriate stances across five dimensions of social ability; and 2) Decision evaluation, which assesses the model's voting choices and opponent-role inferences. This framework enables a fine-grained evaluation of models' linguistic and reasoning capabilities, while capturing their ability to generate strategically coherent gameplay. Our experiments show that state-of-the-art LLMs show diverse performance, with roughly half remain below 0.50, revealing clear gaps in deception and counterfactual reasoning. We hope our dataset further inspires research on language, reasoning, and strategy in multi-agent interaction.
Authors: Sahil Kale, Devendra Singh Dhami
Abstract: Truly reliable AI requires more than simply scaling up knowledge; it demands the ability to know what it knows and when it does not. Yet recent research shows that even the best LLMs misjudge their own competence in more than one in five cases, making any response born of such internal uncertainty impossible to fully trust. Inspired by self-improvement reinforcement learning techniques that require minimal data, we present a simple but powerful framework KnowRL that strengthens a model's internal understanding of its own feasibility boundaries, enabling safer and more responsible behaviour. Our framework combines two components: (i) introspection, where the model generates and classifies tasks it judges feasible or infeasible, and (ii) consensus-based rewarding, where stability of self-knowledge assessment is reinforced through internal agreement. By using internally generated data, this design strengthens consistency in self-knowledge and entirely avoids costly external supervision. In experiments on LLaMA-3.1-8B and Qwen-2.5-7B, KnowRL steadily improved self-knowledge, validated by both intrinsic self-consistency and extrinsic benchmarking. With nothing more than a small seed set and no external supervision, our method drove gains as high as 28% in accuracy and 12% in F1, outperforming baselines in just a few iterations. Our framework essentially unlocks the untapped capacity of LLMs to self-improve their knowledge awareness, opening the door to reliable, more accountable AI and safer deployment in critical applications. Owing to its simplicity and independence from external effort, we encourage applying this reliability-enhancing process to all future models.
Authors: Stefan Krsteski, Giuseppe Russo, Serina Chang, Robert West, Kristina Gligori\'c
Abstract: Surveys provide valuable insights into public opinion and behavior, but their execution is costly and slow. Large language models (LLMs) have been proposed as a scalable, low-cost substitute for human respondents, but their outputs are often biased and yield invalid estimates. We study the interplay between synthesis methods that use LLMs to generate survey responses and rectification methods that debias population estimates, and explore how human responses are best allocated between them. Using two panel surveys with questions on nutrition, politics, and economics, we find that synthesis alone introduces substantial bias (24-86%), whereas combining it with rectification reduces bias below 5% and increases effective sample size by up to 14%. Overall, we challenge the common practice of using all human responses for fine-tuning, showing that under a fixed budget, allocating most to rectification results in far more effective estimation.
Authors: Dana Sotto Porat, Ella Rabinovich
Abstract: Generative large language models (LLMs) have become central to everyday life, producing human-like text across diverse domains. A growing body of research investigates whether these models also exhibit personality- and demographic-like characteristics in their language. In this work, we introduce a novel, data-driven methodology for assessing LLM personality without relying on self-report questionnaires, applying instead automatic personality and gender classifiers to model replies on open-ended questions collected from Reddit. Comparing six widely used models to human-authored responses, we find that LLMs systematically express higher Agreeableness and lower Neuroticism, reflecting cooperative and stable conversational tendencies. Gendered language patterns in model text broadly resemble those of human writers, though with reduced variation, echoing prior findings on automated agents. We contribute a new dataset of human and model responses, along with large-scale comparative analyses, shedding new light on the topic of personality and demographic patterns of generative AI.
Authors: Yawen Yang, Fukun Ma, Shiao Meng, Aiwei Liu, Lijie Wen
Abstract: Traditional named entity recognition (NER) aims to identify text mentions into pre-defined entity types. Continual Named Entity Recognition (CNER) is introduced since entity categories are continuously increasing in various real-world scenarios. However, existing continual learning (CL) methods for NER face challenges of catastrophic forgetting and semantic shift of non-entity type. In this paper, we propose GenCNER, a simple but effective Generative framework for CNER to mitigate the above drawbacks. Specifically, we skillfully convert the CNER task into sustained entity triplet sequence generation problem and utilize a powerful pre-trained seq2seq model to solve it. Additionally, we design a type-specific confidence-based pseudo labeling strategy along with knowledge distillation (KD) to preserve learned knowledge and alleviate the impact of label noise at the triplet level. Experimental results on two benchmark datasets show that our framework outperforms previous state-of-the-art methods in multiple CNER settings, and achieves the smallest gap compared with non-CL results.
Authors: Marco Braga, Gian Carlo Milanese, Gabriella Pasi
Abstract: Text preprocessing is a fundamental component of Natural Language Processing, involving techniques such as stopword removal, stemming, and lemmatization to prepare text as input for further processing and analysis. Despite the context-dependent nature of the above techniques, traditional methods usually ignore contextual information. In this paper, we investigate the idea of using Large Language Models (LLMs) to perform various preprocessing tasks, due to their ability to take context into account without requiring extensive language-specific annotated resources. Through a comprehensive evaluation on web-sourced data, we compare LLM-based preprocessing (specifically stopword removal, lemmatization and stemming) to traditional algorithms across multiple text classification tasks in six European languages. Our analysis indicates that LLMs are capable of replicating traditional stopword removal, lemmatization, and stemming methods with accuracies reaching 97%, 82%, and 74%, respectively. Additionally, we show that ML algorithms trained on texts preprocessed by LLMs achieve an improvement of up to 6% with respect to the $F_1$ measure compared to traditional techniques. Our code, prompts, and results are publicly available at https://github.com/GianCarloMilanese/llm_pipeline_wi-iat.
URLs: https://github.com/GianCarloMilanese/llm_pipeline_wi-iat.
Authors: Yusheng Song, Lirong Qiu, Xi Zhang, Zhihao Tang
Abstract: The detection of sophisticated hallucinations in Large Language Models (LLMs) is hampered by a ``Detection Dilemma'': methods probing internal states (Internal State Probing) excel at identifying factual inconsistencies but fail on logical fallacies, while those verifying externalized reasoning (Chain-of-Thought Verification) show the opposite behavior. This schism creates a task-dependent blind spot: Chain-of-Thought Verification fails on fact-intensive tasks like open-domain QA where reasoning is ungrounded, while Internal State Probing is ineffective on logic-intensive tasks like mathematical reasoning where models are confidently wrong. We resolve this with a unified framework that bridges this critical gap. However, unification is hindered by two fundamental challenges: the Signal Scarcity Barrier, as coarse symbolic reasoning chains lack signals directly comparable to fine-grained internal states, and the Representational Alignment Barrier, a deep-seated mismatch between their underlying semantic spaces. To overcome these, we introduce a multi-path reasoning mechanism to obtain more comparable, fine-grained signals, and a segment-aware temporalized cross-attention module to adaptively fuse these now-aligned representations, pinpointing subtle dissonances. Extensive experiments on three diverse benchmarks and two leading LLMs demonstrate that our framework consistently and significantly outperforms strong baselines. Our code is available: https://github.com/peach918/HalluDet.
Authors: Ba-Quang Nguyen
Abstract: We propose a novel neural architecture named TextGraphFuseGAT, which integrates a pretrained transformer encoder (PhoBERT) with Graph Attention Networks for token-level classification tasks. The proposed model constructs a fully connected graph over the token embeddings produced by PhoBERT, enabling the GAT layer to capture rich inter-token dependencies beyond those modeled by sequential context alone. To further enhance contextualization, a Transformer-style self-attention layer is applied on top of the graph-enhanced embeddings. The final token representations are passed through a classification head to perform sequence labeling. We evaluate our approach on three Vietnamese benchmark datasets: PhoNER-COVID19 for named entity recognition in the COVID-19 domain, PhoDisfluency for speech disfluency detection, and VietMed-NER for medical-domain NER. VietMed-NER is the first Vietnamese medical spoken NER dataset, featuring 18 entity types collected from real-world medical speech transcripts and annotated with the BIO tagging scheme. Its specialized vocabulary and domain-specific expressions make it a challenging benchmark for token-level classification models. Experimental results show that our method consistently outperforms strong baselines, including transformer-only and hybrid neural models such as BiLSTM + CNN + CRF, confirming the effectiveness of combining pretrained semantic features with graph-based relational modeling for improved token classification across multiple domains.
Authors: Jiayu Ding, Lei Cui, Li Dong, Nanning Zheng, Furu Wei
Abstract: Recent advances in Large Language Models (LLMs) show that extending the length of reasoning chains significantly improves performance on complex tasks. While revealing these reasoning traces helps users better follow, verify, and learn from the model's problem-solving process, it also makes them highly vulnerable to unauthorized distillation. To mitigate this risk, proprietary model providers often adopt aggressive protection strategies, such as replacing detailed reasoning with brief summaries, which deprive users of valuable intermediate information. To address this trade-off, we propose PART, an information-preserving antidistillation reformulation of reasoning traces. Motivated by the difference between how humans understand reasoning traces and how LLMs exploit them for supervised fine-tuning, we design a simple but effective two-step reformulation: removing self-talk behaviors and reordering sub-conclusions. A small auxiliary model is trained to perform this reformulation, incurring minimal computational overhead. Extensive experiments demonstrate that PART consistently disrupts distillation across student models of different sizes and types on various reasoning benchmarks. For instance, when training on reformulated traces, even the performance of a large 32B student model decreases from 54.17 to 46.88 on AIME 2024, corresponding to a 13.5% degradation.
Authors: Saurabh Khanna, Xinxu Li
Abstract: Large Language Models are trained on massive multilingual corpora, yet this abundance masks a profound crisis: of the world's 7,613 living languages, approximately 2,000 languages with millions of speakers remain effectively invisible in digital ecosystems. We propose a critical framework connecting empirical measurements of language vitality (real world demographic strength) and digitality (online presence) with postcolonial theory and epistemic injustice to explain why linguistic inequality in AI systems is not incidental but structural. Analyzing data across all documented human languages, we identify four categories: Strongholds (33%, high vitality and digitality), Digital Echoes (6%, high digitality despite declining vitality), Fading Voices (36%, low on both dimensions), and critically, Invisible Giants (27%, high vitality but near-zero digitality) - languages spoken by millions yet absent from the LLM universe. We demonstrate that these patterns reflect continuities from colonial-era linguistic hierarchies to contemporary AI development, constituting what we term digital epistemic injustice. Our analysis reveals that English dominance in AI is not a technical necessity but an artifact of power structures that systematically exclude marginalized linguistic knowledge. We conclude with implications for decolonizing language technology and democratizing access to AI benefits.
Authors: Shreya Havaldar, Sunny Rai, Young-Min Cho, Lyle Ungar
Abstract: Existing benchmarks that measure cultural adaptation in LLMs are misaligned with the actual challenges these models face when interacting with users from diverse cultural backgrounds. In this work, we introduce the first framework and benchmark designed to evaluate LLMs in realistic, multicultural conversational settings. Grounded in sociocultural theory, our framework formalizes how linguistic style - a key element of cultural communication - is shaped by situational, relational, and cultural context. We construct a benchmark dataset based on this framework, annotated by culturally diverse raters, and propose a new set of desiderata for cross-cultural evaluation in NLP: conversational framing, stylistic sensitivity, and subjective correctness. We evaluate today's top LLMs on our benchmark and show that these models struggle with cultural adaptation in a conversational setting.
Authors: Ting Li, Yang Yang, Yipeng Yu, Liang Yao, Guoqing Chao, Ruifeng Xu
Abstract: Adversarial attacks on knowledge graph embeddings (KGE) aim to disrupt the model's ability of link prediction by removing or inserting triples. A recent black-box method has attempted to incorporate textual and structural information to enhance attack performance. However, it is unable to generate human-readable explanations, and exhibits poor generalizability. In the past few years, large language models (LLMs) have demonstrated powerful capabilities in text comprehension, generation, and reasoning. In this paper, we propose LLMAtKGE, a novel LLM-based framework that selects attack targets and generates human-readable explanations. To provide the LLM with sufficient factual context under limited input constraints, we design a structured prompting scheme that explicitly formulates the attack as multiple-choice questions while incorporating KG factual evidence. To address the context-window limitation and hesitation issues, we introduce semantics-based and centrality-based filters, which compress the candidate set while preserving high recall of attack-relevant information. Furthermore, to efficiently integrate both semantic and structural information into the filter, we precompute high-order adjacency and fine-tune the LLM with a triple classification task to enhance filtering performance. Experiments on two widely used knowledge graph datasets demonstrate that our attack outperforms the strongest black-box baselines and provides explanations via reasoning, and showing competitive performance compared with white-box methods. Comprehensive ablation and case studies further validate its capability to generate explanations.
Authors: Georg Ahnert, Anna-Carolina Haensch, Barbara Plank, Markus Strohmaier
Abstract: Many in-silico simulations of human survey responses with large language models (LLMs) focus on generating closed-ended survey responses, whereas LLMs are typically trained to generate open-ended text instead. Previous research has used a diverse range of methods for generating closed-ended survey responses with LLMs, and a standard practice remains to be identified. In this paper, we systematically investigate the impact that various Survey Response Generation Methods have on predicted survey responses. We present the results of 32 mio. simulated survey responses across 8 Survey Response Generation Methods, 4 political attitude surveys, and 10 open-weight language models. We find significant differences between the Survey Response Generation Methods in both individual-level and subpopulation-level alignment. Our results show that Restricted Generation Methods perform best overall, and that reasoning output does not consistently improve alignment. Our work underlines the significant impact that Survey Response Generation Methods have on simulated survey responses, and we develop practical recommendations on the application of Survey Response Generation Methods.
Authors: Bo Cheng, Xu Wang, Jinda Liu, Yi Chang, Yuan Wu
Abstract: Low-Rank Adaptation (LoRA) has emerged as one of the most widely used parameter-efficient fine-tuning (PEFT) methods for adapting large language models (LLMs) to downstream tasks. While highly effective in single-task settings, it struggles to efficiently leverage inter-task knowledge in complex multi-task learning scenarios, often requiring substantial task-specific data to achieve optimal performance. To address this limitation, we introduce MeTA-LoRA, a two-stage optimization framework that significantly improves data efficiency in multi-task adaptation. In the first stage, task-specific LoRA adapters are learned using only a few samples from each involved dataset, enabling rapid adaptation without large-scale supervision. In the second stage, the shared LoRA adapter is updated by aggregating gradients from multiple tasks to promote knowledge transfer across tasks, further reducing data usage by leveraging common patterns. In both multi-task learning and multilingual learning scenarios, our method matches or surpasses the performance of traditional full-data LoRA fine-tuning approaches, while using significantly less task-specific data.
Authors: Marc Brinner, Sina Zarrie{\ss}
Abstract: We propose SemCSE-Multi, a novel unsupervised framework for generating multifaceted embeddings of scientific abstracts, evaluated in the domains of invasion biology and medicine. These embeddings capture distinct, individually specifiable aspects in isolation, thus enabling fine-grained and controllable similarity assessments as well as adaptive, user-driven visualizations of scientific domains. Our approach relies on an unsupervised procedure that produces aspect-specific summarizing sentences and trains embedding models to map semantically related summaries to nearby positions in the embedding space. We then distill these aspect-specific embedding capabilities into a unified embedding model that directly predicts multiple aspect embeddings from a scientific abstract in a single, efficient forward pass. In addition, we introduce an embedding decoding pipeline that decodes embeddings back into natural language descriptions of their associated aspects. Notably, we show that this decoding remains effective even for unoccupied regions in low-dimensional visualizations, thus offering vastly improved interpretability in user-centric settings.
Authors: Huiyin Xue, Nafise Sadat Moosavi, Nikolaos Aletras
Abstract: The success of Transformer language models is widely credited to their dot-product attention mechanism, which interweaves a set of key design principles: mixing information across positions (enabling multi-token interactions), sequence-dependent activations (where attention weights adapt to each input), a specific mathematical form (dot-product similarities plus softmax weighting), and coupling of queries and keys to evolving hidden states (grounding attention in the current layer). However, the necessity of each of these principles remains largely untested. In this work, we systematically deconstruct attention by designing controlled variants that selectively relax these principles, applied both uniformly across all layers and in hybrid architectures where only some layers retain standard attention. Our empirical analysis reveals that mechanisms for mixing tokens are indispensable, as their absence collapses models to near-random behavior, while the exact mathematical form and sequence dependency can be substantially relaxed, especially when preserved in just a subset of layers. Surprisingly, even variants that fail in isolation can achieve robust performance when interleaved with standard attention, highlighting a cooperative effect. These findings deepen our understanding of what truly underpins attention's effectiveness and open new avenues for simplifying language models without sacrificing performance.
Authors: Xurong Xie, Zhucun Xue, Jiafu Wu, Jian Li, Yabiao Wang, Xiaobin Hu, Yong Liu, Jiangning Zhang
Abstract: Knowledge distillation (KD) is a key technique for compressing large-scale language models (LLMs), yet prevailing logit-based methods typically employ static strategies that are misaligned with the dynamic learning process of student models. These methods typically treat all tokens indiscriminately and apply a single, fixed temperature, resulting in suboptimal knowledge transfer. To address these limitations, we propose LLM-Oriented Token-Adaptive Knowledge Distillation (AdaKD), a novel framework that adapts the distillation process to the real-time learning state of each token. AdaKD consists of two synergistic modules driven by a unified token difficulty metric. First, our Loss-Driven Adaptive Token Focusing (LATF) module dynamically adjusts the distillation focus by monitoring the student's learning stability, concentrating computational resources on the most valuable tokens at each training phase. Second, we introduce Inverse Difficulty Temperature Scaling (IDTS), a counterintuitive yet effective token-level temperature strategy. It employs low temperatures for difficult tokens for targeted error correction, and high temperatures for easy tokens to encourage students to learn from the teacher's complete and smooth output distribution, thereby enhancing generalization. As a plug-and-play framework, AdaKD can consistently improve the performance of various distillation methods on multiple model architectures and benchmarks.
Authors: Zehao Chen, Rong Pan, Haoran Li
Abstract: Human writers often begin their stories with an overarching mental scene, where they envision the interactions between characters and their environment. Inspired by this creative process, we propose a novel approach to long-form story generation, termed hybrid bottom-up long-form story generation, using multi-agent simulations. In our method, agents interact within a dynamic sandbox environment, where their behaviors and interactions with one another and the environment generate emergent events. These events form the foundation for the story, enabling organic character development and plot progression. Unlike traditional top-down approaches that impose rigid structures, our hybrid bottom-up approach allows for the natural unfolding of events, fostering more spontaneous and engaging storytelling. The system is capable of generating stories exceeding 10,000 words while maintaining coherence and consistency, addressing some of the key challenges faced by current story generation models. We achieve state-of-the-art performance across several metrics. This approach offers a scalable and innovative solution for creating dynamic, immersive long-form stories that evolve organically from agent-driven interactions.
Authors: Siheng Xiong, Ali Payani, Faramarz Fekri
Abstract: Inference-time scaling enhances the reasoning ability of a language model (LM) by extending its chain-of-thought (CoT). However, existing approaches typically generate the entire reasoning chain in a single forward pass, which often leads to CoT derailment, i.e., the reasoning trajectory drifting off course due to compounding errors. This problem is particularly severe for smaller LMs with long CoTs due to their limited capacity. To address this, we analyze raw long CoTs and uncover a reasoning hierarchy consisting of planning and execution steps. Our analysis reveals that most reasoning errors stem from incorrect planning. Motivated by this observation, we propose Multi-Path Plan Aggregation (MPPA), a framework that augments single-pass reasoning with plan exploration and aggregation. Following a variable interval schedule based on the token position, MPPA generates multiple candidate plans and aggregates them into a refined planning step. To maintain efficiency, we adopt a minimal design in which the base LM serves as the primary policy, while a lightweight LoRA module implements the plan aggregation policy. We further observe that outcome-reward RL is inefficient for long trajectories (e.g., exceeding 4K tokens). To overcome this, we introduce online Step-DPO, a process-level preference optimization scheme that leverages Twisted Sequential Monte Carlo (TSMC) to provide scalable stepwise supervision using small LMs. This yields more efficient training, improved stability, and higher accuracy. Extensive experiments on challenging math, science, and logical reasoning benchmarks demonstrate that, with only 10% SFT data and 5% of preference pairs, our method outperforms both the DeepSeek-R1 distillation baseline and the outcome-reward RL baseline across multiple base models and tasks.
Authors: Xin Gui, King Zhu, JinCheng Ren, Qianben Chen, Zekun Moore Wang, Yizhi LI, Xinpeng Liu, Xiaowan Li, Wenli Ren, Linyu Miao, Tianrui Qin, Ziqi Shu, He Zhu, Xiangru Tang, Dingfeng Shi, Jiaheng Liu, Yuchen Eleanor Jiang, Minghao Liu, Ge Zhang, Wangchunshu Zhou
Abstract: In recent years, the research focus of large language models (LLMs) and agents has shifted increasingly from demonstrating novel capabilities to complex reasoning and tackling challenging tasks. However, existing evaluations focus mainly on math/code contests or general tasks, while existing multi-domain academic benchmarks lack sufficient reasoning depth, leaving the field without a rigorous benchmark for high-level reasoning. To fill this gap, we introduce the Acadreason benchmark, designed to evaluate the ability of LLMs and agents to acquire and reason over academic knowledge. It consists of 50 expert-annotated academic problems across five high-reasoning domains, including computer science, economics, law, mathematics, and philosophy. All questions are sourced from top-tier publications in recent years and undergo rigorous annotation and quality control to ensure they are both challenging and answerable. We conduct systematic evaluations of over 10 mainstream LLMs and agents. The results show that most LLMs scored below 20 points, with even the cutting-edge GPT-5 achieving only 16 points. While agents achieved higher scores, none exceeded 40 points. This demonstrates the current capability gap between LLMs and agents in super-intelligent academic research tasks and highlights the challenges of Acadreason.
Authors: Chenghao Xiao, Hou Pong Chan, Hao Zhang, Weiwen Xu, Mahani Aljunied, Yu Rong
Abstract: Recent multimodal embedding approaches leveraging multimodal large language models (MLLMs) fine-tuned with contrastive learning (CL) have shown promising results, yet the underlying reasons behind their superiority remain underexplored. This work argues that a crucial advantage of MLLM-based approaches stems from implicit cross-modal alignment achieved during generative pretraining, where the language decoder learns to exploit multimodal signals within a shared representation space for generating unimodal outputs. Through analysis of anisotropy and kernel similarity structure, we empirically confirm that latent alignment emerges within MLLM representations, allowing CL to serve as a lightweight refinement stage. Leveraging this insight, we propose a Language-Centric Omnimodal Embedding framework, termed LCO-Emb. Extensive experiments across diverse backbones and benchmarks demonstrate its effectiveness, achieving state-of-the-art performance across modalities. Furthermore, we identify a Generation-Representation Scaling Law (GRSL), showing that the representational capabilities gained through contrastive refinement scales positively with the MLLM's generative capabilities. This suggests that improving generative abilities evolves as an effective paradigm for enhancing representation quality. We provide a theoretical explanation of GRSL, which formally links the MLLM's generative quality to the upper bound on its representation performance, and validate it on a challenging, low-resource visual-document retrieval task, showing that continual generative pretraining before CL can further enhance the potential of a model's embedding capabilities. Codes, models, and resources are available at https://github.com/LCO-Embedding/LCO-Embedding.
Authors: Lingfei Qian, Xueqing Peng, Yan Wang, Vincent Jim Zhang, Huan He, Hanley Smith, Yi Han, Yueru He, Haohang Li, Yupeng Cao, Yangyang Yu, Alejandro Lopez-Lira, Peng Lu, Jian-Yun Nie, Guojun Xiong, Jimin Huang, Sophia Ananiadou
Abstract: Although Large Language Model (LLM)-based agents are increasingly used in financial trading, it remains unclear whether they can reason and adapt in live markets, as most studies test models instead of agents, cover limited periods and assets, and rely on unverified data. To address these gaps, we introduce Agent Market Arena (AMA), the first lifelong, real-time benchmark for evaluating LLM-based trading agents across multiple markets. AMA integrates verified trading data, expert-checked news, and diverse agent architectures within a unified trading framework, enabling fair and continuous comparison under real conditions. It implements four agents, including InvestorAgent as a single-agent baseline, TradeAgent and HedgeFundAgent with different risk styles, and DeepFundAgent with memory-based reasoning, and evaluates them across GPT-4o, GPT-4.1, Claude-3.5-haiku, Claude-sonnet-4, and Gemini-2.0-flash. Live experiments on both cryptocurrency and stock markets demonstrate that agent frameworks display markedly distinct behavioral patterns, spanning from aggressive risk-taking to conservative decision-making, whereas model backbones contribute less to outcome variation. AMA thus establishes a foundation for rigorous, reproducible, and continuously evolving evaluation of financial reasoning and trading intelligence in LLM-based agents.
Authors: Zhaochen Yu, Ling Yang, Jiaru Zou, Shuicheng Yan, Mengdi Wang
Abstract: Recently, the emergence of agentic RL has showcased that RL could also effectively improve the agentic reasoning ability of LLMs, yet the key design principles and optimal practices remain unclear. In this work, we conduct a comprehensive and systematic investigation to demystify reinforcement learning in agentic reasoning from three key perspectives: data, algorithm, and reasoning mode. We highlight our key insights: (i) Replacing stitched synthetic trajectories with real end-to-end tool-use trajectories yields a far stronger SFT initialization; high-diversity, model-aware datasets sustain exploration and markedly improve RL performance. (ii) Exploration-friendly techniques are crucial for agentic RL, such as clip higher, overlong reward shaping, and maintaining adequate policy entropy could improve the training efficiency. (iii) A deliberative strategy with fewer tool calls outperforms frequent tool calls or verbose self-reasoning, improving tool efficiency and final accuracy. Together, these simple practices consistently enhance agentic reasoning and training efficiency, achieving strong results on challenging benchmarks with smaller models, and establishing a practical baseline for future agentic RL research. Beyond these empirical insights, we further contribute a high-quality, real end-to-end agentic SFT dataset along with a high-quality RL dataset, and demonstrate the effectiveness of our insights in boosting the agentic reasoning ability of LLMs across four challenging benchmarks, including AIME2024/AIME2025, GPQA-Diamond, and LiveCodeBench-v6. With our recipes, 4B-sized models could also achieve superior agentic reasoning performance compared to 32B-sized models. Code and models: https://github.com/Gen-Verse/Open-AgentRL
Authors: Tsung-Han Wu, Mihran Miroyan, David M. Chan, Trevor Darrell, Narges Norouzi, Joseph E. Gonzalez
Abstract: Large Reasoning Models (LRMs) excel at complex reasoning but are traditionally evaluated in static, "frozen world" settings: model responses are assumed to be instantaneous, and the context of a request is presumed to be immutable over the duration of the response. While generally true for short-term tasks, the "frozen world" assumption breaks down in modern reasoning tasks such as assistive programming, where models may take hours to think through problems and code may change dramatically from the time the model starts thinking to the model's final output. In this work, we challenge the frozen world assumption and evaluate LRM robustness under two realistic dynamic scenarios: interruptions, which test the quality of the model's partial outputs on a limited budget, and dynamic context, which tests model adaptation to in-flight changes. Across mathematics and programming benchmarks that require long-form reasoning, static evaluations consistently overestimate robustness: even state-of-the-art LRMs, which achieve high accuracy in static settings, can fail unpredictably when interrupted or exposed to changing context, with performance dropping by up to 60% when updates are introduced late in the reasoning process. Our analysis further reveals several novel failure modes, including reasoning leakage, where models fold the reasoning into their final answer when interrupted; panic, where under time pressure models abandon reasoning entirely and return incorrect answers; and self-doubt, where performance degrades while incorporating updated information.
Authors: Jianghao Lin, Rong Shan, Jiachen Zhu, Yunjia Xi, Yong Yu, Weinan Zhang
Abstract: Survey papers are foundational to the scholarly progress of research communities, offering structured overviews that guide both novices and experts across disciplines. However, the recent surge of AI-generated surveys, especially enabled by large language models (LLMs), has transformed this traditionally labor-intensive genre into a low-effort, high-volume output. While such automation lowers entry barriers, it also introduces a critical threat: the phenomenon we term the "survey paper DDoS attack" to the research community. This refers to the unchecked proliferation of superficially comprehensive but often redundant, low-quality, or even hallucinated survey manuscripts, which floods preprint platforms, overwhelms researchers, and erodes trust in the scientific record. In this position paper, we argue that we must stop uploading massive amounts of AI-generated survey papers (i.e., survey paper DDoS attack) to the research community, by instituting strong norms for AI-assisted review writing. We call for restoring expert oversight and transparency in AI usage and, moreover, developing new infrastructures such as Dynamic Live Surveys, community-maintained, version-controlled repositories that blend automated updates with human curation. Through quantitative trend analysis, quality audits, and cultural impact discussion, we show that safeguarding the integrity of surveys is no longer optional but imperative to the research community.
Authors: Zhao Tong, Chunlin Gong, Yimeng Gu, Haichao Shi, Qiang Liu, Shu Wu, Xiao-Yu Zhang
Abstract: The spread of fake news online distorts public judgment and erodes trust in social media platforms. Although recent fake news detection (FND) models perform well in standard settings, they remain vulnerable to adversarial comments-authored by real users or by large language models (LLMs)-that subtly shift model decisions. In view of this, we first present a comprehensive evaluation of comment attacks to existing fake news detectors and then introduce a group-adaptive adversarial training strategy to improve the robustness of FND models. To be specific, our approach comprises three steps: (1) dividing adversarial comments into three psychologically grounded categories: perceptual, cognitive, and societal; (2) generating diverse, category-specific attacks via LLMs to enhance adversarial training; and (3) applying a Dirichlet-based adaptive sampling mechanism (InfoDirichlet Adjusting Mechanism) that dynamically adjusts the learning focus across different comment categories during training. Experiments on benchmark datasets show that our method maintains strong detection accuracy while substantially increasing robustness to a wide range of adversarial comment perturbations.
Authors: Jiale Guo, Suizhi Huang, Mei Li, Dong Huang, Xingsheng Chen, Regina Zhang, Zhijiang Guo, Han Yu, Siu-Ming Yiu, Christian Jensen, Pietro Lio, Kwok-Yan Lam
Abstract: The integration of LLMs into software engineering has catalyzed a paradigm shift from traditional rule-based systems to sophisticated agentic systems capable of autonomous problem-solving. Despite this transformation, the field lacks a comprehensive understanding of how benchmarks and solutions interconnect, hindering systematic progress and evaluation. This survey presents the first holistic analysis of LLM-empowered software engineering, bridging the critical gap between evaluation and solution approaches. We analyze 150+ recent papers and organize them into a comprehensive taxonomy spanning two major dimensions: (1) Solutions, categorized into prompt-based, fine-tuning-based, and agent-based paradigms, and (2) Benchmarks, covering code generation, translation, repair, and other tasks. Our analysis reveals how the field has evolved from simple prompt engineering to complex agentic systems incorporating planning and decomposition, reasoning and self-refinement, memory mechanisms, and tool augmentation. We present a unified pipeline that illustrates the complete workflow from task specification to final deliverables, demonstrating how different solution paradigms address varying complexity levels across software engineering tasks. Unlike existing surveys that focus on isolated aspects, we provide full-spectrum coverage connecting 50+ benchmarks with their corresponding solution strategies, enabling researchers to identify optimal approaches for specific evaluation criteria. Furthermore, we identify critical research gaps and propose actionable future directions, including multi-agent collaboration frameworks, self-evolving code generation systems, and integration of formal verification with LLM-based methods. This survey serves as a foundational resource for researchers and practitioners seeking to understand, evaluate, and advance LLM-empowered software engineering systems.
Authors: Gregory D. Baker
Abstract: In this paper, we introduce Narrative Learning, a methodology where models are defined entirely in natural language and iteratively refine their classification criteria using explanatory prompts rather than traditional numerical optimisation. We report on experiments to evaluate the accuracy and potential of this approach using 3 synthetic and 3 natural datasets and compare them against 7 baseline explainable machine learning models. We demonstrate that on 5 out of 6 of these datasets, Narrative Learning became more accurate than the baseline explainable models in 2025 or earlier because of improvements in language models. We also report on trends in the lexicostatistics of these models' outputs as a proxy for the comprehensibility of the explanations.
Authors: Ayoub Bouguettaya, Elizabeth M. Stuart
Abstract: The lexical hypothesis posits that personality traits are encoded in language and is foundational to models like the Big Five. We created a bottom-up personality model from a classic adjective list using machine learning and compared its descriptive utility against the Big Five by analyzing one million Reddit comments. The Big Five, particularly Agreeableness, Conscientiousness, and Neuroticism, provided a far more powerful and interpretable description of these online communities. In contrast, our machine-learning clusters provided no meaningful distinctions, failed to recover the Extraversion trait, and lacked the psychometric coherence of the Big Five. These results affirm the robustness of the Big Five and suggest personality's semantic structure is context-dependent. Our findings show that while machine learning can help check the ecological validity of established psychological theories, it may not be able to replace them.
Authors: Yue Huang, Hang Hua, Yujun Zhou, Pengcheng Jing, Manish Nagireddy, Inkit Padhi, Greta Dolcetti, Zhangchen Xu, Subhajit Chaudhury, Ambrish Rawat, Liubov Nedoshivina, Pin-Yu Chen, Prasanna Sattigeri, Xiangliang Zhang
Abstract: While LLM agents can plan multi-step tasks, intervening at the planning stage-before any action is executed-is often the safest way to prevent harm, since certain risks can lead to severe consequences once carried out. However, existing guardrails mostly operate post-execution, which is difficult to scale and leaves little room for controllable supervision at the plan level. To address this challenge, we highlight three critical gaps in current research: data gap, model gap, and evaluation gap. To close the data gap, we introduce AuraGen, a controllable engine that (i) synthesizes benign trajectories, (ii) injects category-labeled risks with calibrated difficulty, and (iii) filters outputs via an automated reward model, producing large and reliable corpora for pre-execution safety. To close the guardian model gap, we propose a foundational guardrail Safiron, combining a cross-planner adapter with a compact guardian model. The adapter unifies different input formats, while Safiron flags risky cases, assigns risk types, and generates rationales; trained in two stages with a broadly explored data recipe, Safiron achieves robust transfer across settings. To close the evaluation gap, we release Pre-Exec Bench, a realistic benchmark covering diverse tools and branching trajectories, which measures detection, fine-grained categorization, explanation, and cross-planner generalization in human-verified scenarios. Extensive experiments demonstrate consistent gains of the proposed guardrail over strong baselines on Pre-Exec Bench, and ablations further distill actionable practices, providing a practical template for safer agentic systems.
Authors: Yufa Zhou, Yixiao Wang, Xunjian Yin, Shuyan Zhou, Anru R. Zhang
Abstract: We study how large language models (LLMs) ``think'' through their representation space. We propose a novel geometric framework that models an LLM's reasoning as flows -- embedding trajectories evolving where logic goes. We disentangle logical structure from semantics by employing the same natural deduction propositions with varied semantic carriers, allowing us to test whether LLMs internalize logic beyond surface form. This perspective connects reasoning with geometric quantities such as position, velocity, and curvature, enabling formal analysis in representation and concept spaces. Our theory establishes: (1) LLM reasoning corresponds to smooth flows in representation space, and (2) logical statements act as local controllers of these flows' velocities. Using learned representation proxies, we design controlled experiments to visualize and quantify reasoning flows, providing empirical validation of our theoretical framework. Our work serves as both a conceptual foundation and practical tools for studying reasoning phenomenon, offering a new lens for interpretability and formal analysis of LLMs' behavior.
Authors: Weiqing Luo, Zhen Tan, Yifan Li, Xinyu Zhao, Kwonjoon Lee, Behzad Dariush, Tianlong Chen
Abstract: Real-world vision-language applications demand varying levels of perceptual granularity. However, most existing visual large language models (VLLMs), such as LLaVA, pre-assume a fixed resolution for downstream tasks, which leads to subpar performance. To address this problem, we first conduct a comprehensive and pioneering investigation into the resolution preferences of different vision-language tasks, revealing a correlation between resolution preferences with image complexity, and uncertainty variance of the VLLM at different image input resolutions. Building on this insight, we propose an empirical formula to determine the optimal resolution for a given vision-language task, combining these two factors. Second, based on rigorous experiments, we propose a novel parameter-efficient fine-tuning technique to extend the visual input resolution of pre-trained VLLMs to the identified optimal resolution. Extensive experiments on various vision-language tasks validate the effectiveness of our method.
Authors: Xi Fang, Weijie Xu, Yuchong Zhang, Stephanie Eckman, Scott Nickleach, Chandan K. Reddy
Abstract: When an AI assistant remembers that Sarah is a single mother working two jobs, does it interpret her stress differently than if she were a wealthy executive? As personalized AI systems increasingly incorporate long-term user memory, understanding how this memory shapes emotional reasoning is critical. We investigate how user memory affects emotional intelligence in large language models (LLMs) by evaluating 15 models on human validated emotional intelligence tests. We find that identical scenarios paired with different user profiles produce systematically divergent emotional interpretations. Across validated user independent emotional scenarios and diverse user profiles, systematic biases emerged in several high-performing LLMs where advantaged profiles received more accurate emotional interpretations. Moreover, LLMs demonstrate significant disparities across demographic factors in emotion understanding and supportive recommendations tasks, indicating that personalization mechanisms can embed social hierarchies into models emotional reasoning. These results highlight a key challenge for memory enhanced AI: systems designed for personalization may inadvertently reinforce social inequalities.
Authors: Stefan Pasch
Abstract: Organizational efforts to utilize and operationalize artificial intelligence (AI) are often accompanied by substantial challenges, including scalability, maintenance, and coordination across teams. In response, the concept of Machine Learning Operations (MLOps) has emerged as a set of best practices that integrate software engineering principles with the unique demands of managing the ML lifecycle. Yet, empirical evidence on whether and how these practices support users in developing and operationalizing AI applications remains limited. To address this gap, this study analyzes over 8,000 user reviews of AI development platforms from G2.com. Using zero-shot classification, we measure review sentiment toward nine established MLOps practices, including continuous integration and delivery (CI/CD), workflow orchestration, reproducibility, versioning, collaboration, and monitoring. Seven of the nine practices show a significant positive relationship with user satisfaction, suggesting that effective MLOps implementation contributes tangible value to AI development. However, organizational context also matters: reviewers from small firms discuss certain MLOps practices less frequently, suggesting that organizational context influences the prevalence and salience of MLOps, though firm size does not moderate the MLOps-satisfaction link. This indicates that once applied, MLOps practices are perceived as universally beneficial across organizational settings.
Authors: Hehe Fan, Yi Yang, Mohan Kankanhalli, Fei Wu
Abstract: When modeling a given type of data, we consider it to involve two key aspects: 1) identifying relevant elements (e.g., image pixels or textual words) to a central element, as in a convolutional receptive field, or to a query element, as in self-attention, and 2) encoding these tokens effectively. Self-attention can adaptively identify these elements but relies on absolute positional embedding for structural representation learning. In contrast, convolution encodes elements in a relative manner, yet their fixed kernel size limits their ability to adaptively select the relevant elements. In this paper, we introduce Translution, an operation that unifies the adaptive identification capability of self-attention and the relative encoding advantage of convolution. However, this integration leads to a substantial increase in the number of parameters, exceeding most currently available computational resources. Therefore, we propose a lightweight variant of Translution, named {\alpha}-Translution. Experiments on computer vision and natural language processing tasks show that Translution (including {\alpha}-Translution) achieves superior accuracy compared to self-attention. The code is available at https://github.com/hehefan/Translution.
Authors: Peiyuan Gong, Feiran Zhu, Yaqi Yin, Chenglei Dai, Chao Zhang, Kai Zheng, Wentian Bao, Jiaxin Mao, Yi Zhang
Abstract: Short-video platforms have rapidly become a new generation of information retrieval systems, where users formulate queries to access desired videos. However, user queries, especially long-tail ones, often suffer from spelling errors, incomplete phrasing, and ambiguous intent, resulting in mismatches between user expectations and retrieved results. While large language models (LLMs) have shown success in long-tail query rewriting within e-commerce, they struggle on short-video platforms, where proprietary content such as short videos, live streams, micro dramas, and user social networks falls outside their training distribution. To address this challenge, we introduce \textbf{CardRewriter}, an LLM-based framework that incorporates domain-specific knowledge to enhance long-tail query rewriting. For each query, our method aggregates multi-source knowledge relevant to the query and summarizes it into an informative and query-relevant knowledge card. This card then guides the LLM to better capture user intent and produce more effective query rewrites. We optimize CardRewriter using a two-stage training pipeline: supervised fine-tuning followed by group relative policy optimization, with a tailored reward system balancing query relevance and retrieval effectiveness. Offline experiments show that CardRewriter substantially improves rewriting quality for queries targeting proprietary content. Online A/B testing further confirms significant gains in long-view rate (LVR) and click-through rate (CTR), along with a notable reduction in initiative query reformulation rate (IQRR). Since September 2025, CardRewriter has been deployed on Kuaishou, one of China's largest short-video platforms, serving hundreds of millions of users daily.
Authors: Jinghao Zhang, Naishan Zheng, Ruilin Li, Dongzhou Cheng, Zheming Liang, Feng Zhao, Jiaqi Wang
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has recently emerged as a promising framework for improving reasoning abilities in Large Language Models (LLMs). However, policy optimized with binary verification prone to overlook potential valuable exploration in reasoning trajectory. In view of heavy annotation cost of golden Process Reward Models (PRMs), recent works attempt using auxiliary signals for reward shaping of process tokens, involving entropy and likelihood collected from logit space. In this work, we offer a novel perspective on shaping RLVR with flow rewards derived from latent space, and propose RLFR, where the flow fields of model latents are constructed from either off-policy high-quality data and on-policy rejection sampling data, and the velocity deviations of policy latents within it are quantified to serve as a reward signal. RLFR first demonstrates that a well-established flow field can be a sound environment for reward signal collection, highlighting the expressive latent space is much underexplored. Moreover, RLFR is able to compress any off-policy expert data as reference for constituting reward signals, and we show that the efficient context dependence compressed within the hidden states are utilized, rather than individual token-level denotation for context comprehending. Experiments on both language and multimodal reasoning benchmarks demonstrate the reliability of flow rewards, and suggesting a promising paradigm for reward shaping with auxiliary signals.
Authors: Guan-Yan Yang, Tzu-Yu Cheng, Ya-Wen Teng, Farn Wanga, Kuo-Hui Yeh
Abstract: The integration of Large Language Models (LLMs) into computer applications has introduced transformative capabilities but also significant security challenges. Existing safety alignments, which primarily focus on semantic interpretation, leave LLMs vulnerable to attacks that use non-standard data representations. This paper introduces ArtPerception, a novel black-box jailbreak framework that strategically leverages ASCII art to bypass the security measures of state-of-the-art (SOTA) LLMs. Unlike prior methods that rely on iterative, brute-force attacks, ArtPerception introduces a systematic, two-phase methodology. Phase 1 conducts a one-time, model-specific pre-test to empirically determine the optimal parameters for ASCII art recognition. Phase 2 leverages these insights to launch a highly efficient, one-shot malicious jailbreak attack. We propose a Modified Levenshtein Distance (MLD) metric for a more nuanced evaluation of an LLM's recognition capability. Through comprehensive experiments on four SOTA open-source LLMs, we demonstrate superior jailbreak performance. We further validate our framework's real-world relevance by showing its successful transferability to leading commercial models, including GPT-4o, Claude Sonnet 3.7, and DeepSeek-V3, and by conducting a rigorous effectiveness analysis against potential defenses such as LLaMA Guard and Azure's content filters. Our findings underscore that true LLM security requires defending against a multi-modal space of interpretations, even within text-only inputs, and highlight the effectiveness of strategic, reconnaissance-based attacks. Content Warning: This paper includes potentially harmful and offensive model outputs.
Authors: Michael Y. Hu, Benjamin Van Durme, Jacob Andreas, Harsh Jhamtani
Abstract: Language model (LM) agents deployed in novel environments often exhibit poor sample efficiency when learning from sequential interactions. This significantly hinders the usefulness of such agents in environments where interaction is costly (for example, when they interact with humans or reset physical systems). While a number of existing LM agent architectures incorporate various mechanisms for experience storage and reflection, they make limited use of LMs' abilities to directly generate or reason about full counterfactual trajectories. We introduce ECHO (Experience Consolidation via Hindsight Optimization), a prompting framework that adapts hindsight experience replay from reinforcement learning for language model agents. ECHO generates optimized trajectories for alternative goals that could have been achieved during failed attempts, effectively creating synthetic positive examples from unsuccessful interactions. Our approach consists of two components: a hindsight rule that uses the language model itself to identify relevant subgoals and generate optimized trajectories, and an update rule that maintains compressed trajectory representations in memory. We evaluate ECHO on stateful versions of XMiniGrid, a text-based navigation and planning benchmark, and PeopleJoinQA, a collaborative information-gathering enterprise simulation. Across both domains, ECHO outperforms vanilla language agent baselines by up to 80%; in XMiniGrid, it also outperforms a number of sophisticated agent architectures including Reflexion and AWM, demonstrating faster adaptation to novel environments through more effective utilization of past experiences.
Authors: Hong Su
Abstract: Existing studies have introduced method-based reasoning and scope extension as approaches to enhance Large Language Model (LLM) performance beyond direct matrix mappings. Building on these foundations, this paper summarizes and integrates these ideas into a unified Intuition-Method Layered Model with Scope Extension, designed to address indirected (unseen) issues more systematically. In this framework, intuition-based thinking provides rapid first-reaction answers, while method-based thinking decouples questions and solutions into transferable reasoning units. Scope extension is then applied to broaden applicability, including vertical (cause analysis), horizontal (parallel and generalized issues), and for the first time, temporal and spatial extensions, which expand reasoning across time and contextual dimensions. These extensions are organized into systematic knowledge trees that interconnect into a knowledge network, thereby increasing adaptability. To quantitatively evaluate this process, we propose the entropy of method extension, which measures the independence and diversity of extensions as an indicator of the system's capacity to solve unseen questions. By logically connecting existing approaches with new extensions and introducing an entropy-based evaluation framework, this work advances toward a more robust and extensible reasoning paradigm for LLMs in real-world problem-solving.
Authors: Mukul Lokhande, Tanushree Dewangan, Mohd Sharik Mansoori, Tejas Chaudhari, Akarsh J., Damayanti Lokhande, Adam Teman, Santosh Kumar Vishvakarma
Abstract: This paper introduces Bhasha-Rupantarika, a light and efficient multilingual translation system tailored through algorithm-hardware codesign for resource-limited settings. The method investigates model deployment at sub-octet precision levels (FP8, INT8, INT4, and FP4), with experimental results indicating a 4.1x reduction in model size (FP4) and a 4.2x speedup in inference speed, which correlates with an increased throughput of 66 tokens/s (improvement by 4.8x). This underscores the importance of ultra-low precision quantization for real-time deployment in IoT devices using FPGA accelerators, achieving performance on par with expectations. Our evaluation covers bidirectional translation between Indian and international languages, showcasing its adaptability in low-resource linguistic contexts. The FPGA deployment demonstrated a 1.96x reduction in LUTs and a 1.65x decrease in FFs, resulting in a 2.2x enhancement in throughput compared to OPU and a 4.6x enhancement compared to HPTA. Overall, the evaluation provides a viable solution based on quantisation-aware translation along with hardware efficiency suitable for deployable multilingual AI systems. The entire codes [https://github.com/mukullokhande99/Bhasha-Rupantarika/] and dataset for reproducibility are publicly available, facilitating rapid integration and further development by researchers.
URLs: https://github.com/mukullokhande99/Bhasha-Rupantarika/]
Authors: Meiru Zhang, Philipp Borchert, Milan Gritta, Gerasimos Lampouras
Abstract: Automating the formalization of mathematical statements for theorem proving remains a major challenge for Large Language Models (LLMs). LLMs struggle to identify and utilize the prerequisite mathematical knowledge and its corresponding formal representation in languages like Lean. Current retrieval-augmented autoformalization methods query external libraries using the informal statement directly, but overlook a fundamental limitation: informal mathematical statements are often complex and offer limited context on the underlying math concepts. To address this, we introduce DRIFT, a novel framework that enables LLMs to decompose informal mathematical statements into smaller, more tractable ''sub-components''. This facilitates targeted retrieval of premises from mathematical libraries such as Mathlib. Additionally, DRIFT retrieves illustrative theorems to help models use premises more effectively in formalization tasks. We evaluate DRIFT across diverse benchmarks (ProofNet, ConNF, and MiniF2F-test) and find that it consistently improves premise retrieval, nearly doubling the F1 score compared to the DPR baseline on ProofNet. Notably, DRIFT demonstrates strong performance on the out-of-distribution ConNF benchmark, with BEq+@10 improvements of 37.14% and 42.25% using GPT-4.1 and DeepSeek-V3.1, respectively. Our analysis shows that retrieval effectiveness in mathematical autoformalization depends heavily on model-specific knowledge boundaries, highlighting the need for adaptive retrieval strategies aligned with each model's capabilities.
Authors: Hengyuan Zhang, Shiping Yang, Xiao Liang, Chenming Shang, Yuxuan Jiang, Chaofan Tao, Jing Xiong, Hayden Kwok-Hay So, Ruobing Xie, Angel X. Chang, Ngai Wong
Abstract: Training student models on synthetic data generated by strong teacher models is a promising way to distilling the capabilities of teachers. However, recent studies show that stronger models are not always optimal teachers, revealing a mismatch between teacher outputs and student learnability. To address this issue, we propose PerSyn (Personalized data Synthesis), a novel synthesis strategy that operates under a new ``Route then Generate'' paradigm to create data tailored to each student model, enabling it to learn more effectively. Specifically, PerSyn first assigns each prompt to its optimal teacher via a query-level router that jointly considers student learnability and teacher response quality. Each teacher then synthesizes data only for its assigned prompts, making the process more efficient than the conventional ``Generate then Select'' paradigm, where all teachers must generate parallel responses for the entire prompt set before constructing the final dataset. Extensive experiments across different model families and scales demonstrate that PerSyn consistently achieves superior or comparable performance to all baselines in instruct tuning and math reasoning settings. Further analysis verifies the effectiveness of PerSyn and offers extra insights to propel future research.
Authors: Thi-Nhung Nguyen, Linhao Luo, Thuy-Trang Vu, Dinh Phung
Abstract: Bias in large language models (LLMs) remains a persistent challenge, manifesting in stereotyping and unfair treatment across social groups. While prior research has primarily focused on individual models, the rise of multi-agent systems (MAS), where multiple LLMs collaborate and communicate, introduces new and largely unexplored dynamics in bias emergence and propagation. In this work, we present a comprehensive study of stereotypical bias in MAS, examining how internal specialization, underlying LLMs and inter-agent communication protocols influence bias robustness, propagation, and amplification. We simulate social contexts where agents represent different social groups and evaluate system behavior under various interaction and adversarial scenarios. Experiments on three bias benchmarks reveal that MAS are generally less robust than single-agent systems, with bias often emerging early through in-group favoritism. However, cooperative and debate-based communication can mitigate bias amplification, while more robust underlying LLMs improve overall system stability. Our findings highlight critical factors shaping fairness and resilience in multi-agent LLM systems.
Authors: Xiaoyun Zhang, Xiaojian Yuan, Di Huang, Wang You, Chen Hu, Jingqing Ruan, Kejiang Chen, Xing Hu
Abstract: Reasoning ability has become a defining capability of Large Language Models (LLMs), with Reinforcement Learning with Verifiable Rewards (RLVR) emerging as a key paradigm to enhance it. However, RLVR training often suffers from policy entropy collapse, where the policy becomes overly deterministic, hindering exploration and limiting reasoning performance. While entropy regularization is a common remedy, its effectiveness is highly sensitive to the fixed coefficient, making it unstable across tasks and models. In this work, we revisit entropy regularization in RLVR and argue that its potential has been largely underestimated. Our analysis shows that (i) tasks of varying difficulty demand distinct exploration intensities, and (ii) balanced exploration may require the policy entropy to be maintained within a moderate range below its initial level. Therefore, we propose Adaptive Entropy Regularization (AER)--a framework that dynamically balances exploration and exploitation via three components: difficulty-aware coefficient allocation, initial-anchored target entropy, and dynamic global coefficient adjustment. Experiments on multiple mathematical reasoning benchmarks show that AER consistently outperforms baselines, improving both reasoning accuracy and exploration capability.
Authors: Taiqiang Wu, Runming Yang, Tao Liu, Jiahao Wang, Ngai Wong
Abstract: Model merging, typically on Instruct and Thinking models, has shown remarkable performance for efficient reasoning. In this paper, we systematically revisit the simplest merging method that interpolates two weights directly. Particularly, we observe that model interpolation follows a three-stage evolutionary paradigm with distinct behaviors on the reasoning trajectory. These dynamics provide a principled guide for navigating the performance-cost trade-off. Empirical results demonstrate that a strategically interpolated model surprisingly surpasses sophisticated model merging baselines on both efficiency and effectiveness. We further validate our findings with extensive ablation studies on model layers, modules, and decoding strategies. Ultimately, this work demystifies model interpolation and offers a practical framework for crafting models with precisely targeted reasoning capabilities. Code is available at \href{https://github.com/wutaiqiang/MI}{Github}.
Authors: Tianze Wang, Zhaoyu Chen, Jian Du, Yingtai Xiao, Linjun Zhang, Qiang Yan
Abstract: Text data has become extremely valuable on large language models (LLMs) and even lead to general artificial intelligence (AGI). A lot of high-quality text in the real world is private and cannot be freely used due to privacy concerns. Therefore, differentially private (DP) synthetic text generation has been proposed, aiming to produce high-utility synthetic data while protecting sensitive information. However, existing DP synthetic text generation imposes uniform guarantees that often overprotect non-sensitive content, resulting in substantial utility loss and computational overhead. Therefore, we propose Secret-Protected Evolution (SecPE), a novel framework that extends private evolution with secret-aware protection. Theoretically, we show that SecPE satisfies $(\mathrm{p}, \mathrm{r})$-secret protection, constituting a relaxation of Gaussian DP that enables tighter utility-privacy trade-offs, while also substantially reducing computational complexity relative to baseline methods. Empirically, across the OpenReview, PubMed, and Yelp benchmarks, SecPE consistently achieves lower Fr\'echet Inception Distance (FID) and higher downstream task accuracy than GDP-based Aug-PE baselines, while requiring less noise to attain the same level of protection. Our results highlight that secret-aware guarantees can unlock more practical and effective privacy-preserving synthetic text generation.
Authors: Huanjin Yao, Ruifei Zhang, Jiaxing Huang, Jingyi Zhang, Yibo Wang, Bo Fang, Ruolin Zhu, Yongcheng Jing, Shunyu Liu, Guanbin Li, Dacheng Tao
Abstract: With the recent emergence of revolutionary autonomous agentic systems, research community is witnessing a significant shift from traditional static, passive, and domain-specific AI agents toward more dynamic, proactive, and generalizable agentic AI. Motivated by the growing interest in agentic AI and its potential trajectory toward AGI, we present a comprehensive survey on Agentic Multimodal Large Language Models (Agentic MLLMs). In this survey, we explore the emerging paradigm of agentic MLLMs, delineating their conceptual foundations and distinguishing characteristics from conventional MLLM-based agents. We establish a conceptual framework that organizes agentic MLLMs along three fundamental dimensions: (i) Agentic internal intelligence functions as the system's commander, enabling accurate long-horizon planning through reasoning, reflection, and memory; (ii) Agentic external tool invocation, whereby models proactively use various external tools to extend their problem-solving capabilities beyond their intrinsic knowledge; and (iii) Agentic environment interaction further situates models within virtual or physical environments, allowing them to take actions, adapt strategies, and sustain goal-directed behavior in dynamic real-world scenarios. To further accelerate research in this area for the community, we compile open-source training frameworks, training and evaluation datasets for developing agentic MLLMs. Finally, we review the downstream applications of agentic MLLMs and outline future research directions for this rapidly evolving field. To continuously track developments in this rapidly evolving field, we will also actively update a public repository at https://github.com/HJYao00/Awesome-Agentic-MLLMs.
Authors: Haoran Liang, Yufa Zhou, Mohammad Talebi Kalaleh, Qipei Mei
Abstract: We introduce $\textbf{MASSE}$, the first Multi-Agent System for Structural Engineering, effectively integrating large language model (LLM)-based agents with real-world engineering workflows. Structural engineering is a fundamental yet traditionally stagnant domain, with core workflows remaining largely unchanged for decades despite its substantial economic impact and global market size. Recent advancements in LLMs have significantly enhanced their ability to perform complex reasoning, long-horizon planning, and precise tool utilization -- capabilities well aligned with structural engineering tasks such as interpreting design codes, executing load calculations, and verifying structural capacities. We present a proof-of-concept showing that most real-world structural engineering workflows can be fully automated through a training-free LLM-based multi-agent system. MASSE enables immediate deployment in professional environments, and our comprehensive validation on real-world case studies demonstrates that it can reduce expert workload from approximately two hours to mere minutes, while enhancing both reliability and accuracy in practical engineering scenarios.
Authors: Jiliang Hu, Wenfu Wang, Zuchao Li, Chenxing Li, Yiyang Zhao, Hanzhao Li, Liqiang Zhang, Meng Yu, Dong Yu
Abstract: Recent advances in large audio language models (LALMs) have greatly enhanced multimodal conversational systems. However, existing benchmarks remain limited -- they are mainly English-centric, rely on synthetic speech, and lack comprehensive, discriminative evaluation across multiple dimensions. To address these gaps, we present Voice Chat Bot Bench (VCB Bench) -- a high-quality Chinese benchmark built entirely on real human speech. VCB Bench evaluates LALMs from three complementary perspectives: instruction following (including speech-level control beyond text commands), knowledge understanding (general knowledge, reasoning, and daily dialogue), and robustness (stability under perturbations in content, environment, and speaker traits). Experiments on representative LALMs reveal notable performance gaps and highlight future directions for improvement. VCB Bench provides a reproducible and fine-grained evaluation framework, offering standardized methodology and practical insights for advancing Chinese voice conversational models.
Authors: Gautier Dagan, Frank Keller, Alex Lascarides
Abstract: An agent facing a planning problem can use answers to how-to questions to reduce uncertainty and fill knowledge gaps, helping it solve both current and future tasks. However, their open ended nature, where valid answers to "How do I X?" range from executable actions to high-level descriptions of X's sub-goals, makes them challenging for AI agents to ask, and for AI experts to answer, in ways that support efficient planning. We introduce $How^{2}$, a memory agent framework that enables agents to ask how-to questions, store the answers, and reuse them for lifelong learning in interactive environments. We evaluate our approach in Plancraft, a Minecraft crafting environment, where agents must complete an assembly task by manipulating inventory items. Using teacher models that answer at varying levels of abstraction, from executable action sequences to high-level subgoal descriptions, we show that lifelong learning agents benefit most from answers that are abstracted and decoupled from the current state. $How^{2}$ offers a way for LLM-based agents to improve their planning capabilities over time by asking questions in interactive environments.
Authors: Jinbin Zhang, Nasib Ullah, Erik Schultheis, Rohit Babbar
Abstract: Large output spaces, also referred to as Extreme multilabel classification (XMC), is a setting that arises, e.g., in large-scale tagging and product-to-product recommendation, and is characterized by the number of labels ranging from hundreds of thousands to millions. This means that the linear classification head, usually only a tiny fraction of the overall model, turns into the main driver for compute and memory demand. Current state-of-the-art XMC methods predominantly rely on FP16-FP32 mixed-precision training, which we show can be unstable, and inefficient in terms of memory usage and computational overhead. Meanwhile, existing low-precision methods typically retain higher precision for the classification layer. In this work, we propose ELMO, a pure low-precision training framework for XMC models using BFloat16 and Float8 data types. By leveraging Kahan summation and stochastic rounding, we demonstrate that XMC models can be effectively trained entirely in Float8, without relying on single-precision master weights or tensor scaling. Low-precision training, combined with our proposed memory optimizations -- gradient fusion and chunking -- enables significant reductions in GPU memory usage. For example, we train a 3-million-label XMC model with only 6.6 GiB of GPU memory, compared to the 39.7 GiB required by the optimized SOTA method, Renee without compromising accuracy.
Authors: Daniel Scalena, Leonidas Zotos, Elisabetta Fersini, Malvina Nissim, Ahmet \"Ust\"un
Abstract: With the rise of reasoning language models and test-time scaling methods as a paradigm for improving model performance, substantial computation is often required to generate multiple candidate sequences from the same prompt. This enables exploration of different reasoning paths toward the correct solution, however, allocates the same compute budget for each prompt. Grounded on the assumption that different prompts carry different degrees of complexity, and thus different computation needs, we propose EAGer, a training-free generation method that leverages model uncertainty through token-wise entropy distribution to reduce redundant computation and concurrently improve overall performance. EAGer allows branching to multiple reasoning paths only in the presence of high-entropy tokens, and then reallocates the saved compute budget to the instances where exploration of alternative paths is most needed. We find that across multiple open-source models on complex reasoning benchmarks such as AIME 2025, EAGer can reallocate the budget without accessing target labels, achieving the best efficiency-performance trade-off in terms of reasoning length and Pass@k. When target labels are accessible, EAGer generates up to 65% fewer tokens (hence saving compute) and achieves up to 37% improvement in Pass@k compared to the Full Parallel Sampling.
Authors: Zhengyu Chen, Jinluan Yang, Teng Xiao, Ruochen Zhou, Luan Zhang, Xiangyu Xi, Xiaowei Shi, Wei Wang, Jinggang Wang
Abstract: Recent advances in large language models (LLMs) have demonstrated remarkable capabilities in reasoning and tool utilization. However, the generalization of tool-augmented reinforcement learning (RL) across diverse domains remains underexplored. In this work, we investigate the cross-domain generalization of an LLM agent equipped with a code interpreter tool, which is exclusively trained on mathematical problem-solving tasks. Despite the restricted training domain, we evaluate the agent's performance across several distinct reasoning domains. The results reveal that RL-based tool usage learned from mathematical tasks can be effectively transferred to complex tasks in other domains, enabling great task performance and high token efficiency. To facilitate this cross-domain transfer, we propose a Tool Generalization Reinforcement Learning (TGRL) framework designed to promote domain-agnostic learning and skill migration, encompassing: (i) a standardized tool interface that abstracts domain-specific nuances through consistent formatting and explicit termination, fostering transferable invocation patterns; (ii) a dual-component reward system that decomposes rewards to incentivize generalizable behaviors like tool efficiency and reasoning abstraction, ensuring alignment and robustness across domain shifts; and (iii) an XML-based prompt template that separates thinking, tool calls, and responses to encourage modular, domain-invariant planning and coherent multi-turn interactions. Extensive experiments across diverse benchmarks validate our approach, achieving state-of-the-art performance and highlighting the cross-domain potential of Tool RL for LLM reasoning.
Authors: Gareth Seneque, Lap-Hang Ho, Nafise Erfanian Saeedi, Jeffrey Molendijk, Ariel Kupermann, Tim Elson
Abstract: We present Entropic Mutual-Information Geometry Large-Language Model Alignment (ENIGMA), a novel approach to Large-Language Model (LLM) training that jointly improves reasoning, alignment and robustness by treating an organisation's policies/principles as directions to move on a model's information manifold. Our single-loop trainer combines Group-Relative Policy Optimisation (GRPO), an on-policy, critic-free RL method with Chain-of-Thought (CoT)-format only rewards; a Self-Supervised Alignment with Mutual Information (SAMI)-style symmetric InfoNCE auxiliary; and an entropic Sinkhorn optimal-transport regulariser on hidden-state distributions to bound geometry drift. We also introduce infoNCE metrics that specialise to a standard MI lower bound under matched negatives to measure how strongly a model's CoT encodes these policies. These metrics include a Sufficiency Index (SI) that enables the selection and creation of principles that maximise downstream performance prior to training. In our experiments using small (1B) LLMs, high-SI principles predict steadier training dynamics and improved benchmark performance over GRPO ablations. Our information-geometry analysis of trained models validates desirable structural change in the manifold. These results support our hypothesis that reasoning, alignment, and robustness are projections of a single informationgeometric objective, and that models trained using ENIGMA demonstrate principled reasoning without the use of a reward model, offering a path to trusted capability
Authors: KiHyun Nam, Jongmin Choi, Hyeongkeun Lee, Jungwoo Heo, Joon Son Chung
Abstract: Contrastive audio-language pretraining yields powerful joint representations, yet a persistent audio-text modality gap limits the benefits of coupling multimodal encoders with large language models (LLMs). We present Diffusion-Link, a diffusion-based modality-bridging module that generatively maps audio embeddings into the text-embedding distribution. The module is trained at the output embedding from the frozen multimodal encoder and implemented as a lightweight network with three residual MLP blocks. To assess the effect of Diffusion-Link on multimodal encoder-LLM coupling, we evaluate on Automatic Audio Captioning (AAC); to our knowledge, this is the first application of diffusion-based modality bridging to AAC. We report two results. (1) Modality-gap analysis: on similarity and geometric criteria, Diffusion-Link reduces the modality gap the most among prior diffusion-based methods and shows a collective migration of audio embeddings toward the text distribution. (2) Downstream AAC: attaching Diffusion-Link to the same multimodal LLM baseline achieves state-of-the-art on AudioCaps in both zero-shot and fully supervised captioning without external knowledge, with relative gains up to 52.5% and 7.5%, respectively. These findings show that closing the modality gap is pivotal for effective coupling between multimodal encoders and LLMs, and diffusion-based modality bridging offers a promising direction beyond knowledge-retrieval-centric designs. Code will be released upon acceptance https://github.com/DevKiHyun/Diffusion-Link
Authors: Junpeng Liu, Yuzhong Zhao, Bowen Cao, Jiayu Ding, Yilin Jia, Tengchao Lv, Yupan Huang, Shaohan Huang, Nan Yang, Li Dong, Lei Cui, Tao Ge, Xun Wang, Huitian Jiao, Sun Mao, FNU Kartik, Si-Qing Chen, Wai Lam, Furu Wei
Abstract: Recent advances in agentic workflows have enabled the automation of tasks such as professional document generation. However, they primarily focus on textual quality, neglecting visual structure and style, which are crucial for readability and engagement. This gap arises mainly from the absence of suitable reward models to guide agentic workflows toward producing documents with stronger structural and stylistic quality. To address this, we propose DocReward, a document reward model that evaluates documents based on their structure and style. We construct a multi-domain dataset DocPair of 117K paired documents, covering 32 domains and 267 document types, each including a high- and low-professionalism document with identical content but different structure and style. This enables the model to evaluate professionalism comprehensively, and in a textual-quality-agnostic way. DocReward is trained using the Bradley-Terry loss to score documents, penalizing predictions that contradict the annotated ranking. To assess the performance of reward models, we create a test dataset containing document bundles ranked by well-educated human evaluators. Notably, DocReward outperforms GPT-4o and GPT-5 in accuracy by 30.6 and 19.4 percentage points, respectively, demonstrating its superiority over baselines. In an extrinsic evaluation of document generation, DocReward achieves a significantly higher win rate of 60.8%, compared to GPT-5's 37.7% win rate, demonstrating its utility in guiding generation agents toward producing human-preferred documents.
Authors: Jiaying Wu, Zihang Fu, Haonan Wang, Fanxiao Li, Min-Yen Kan
Abstract: Community Notes, the crowd-sourced misinformation governance system on X (formerly Twitter), enables users to flag misleading posts, attach contextual notes, and vote on their helpfulness. However, our analysis of 30.8K health-related notes reveals significant latency, with a median delay of 17.6 hours before the first note receives a helpfulness status. To improve responsiveness during real-world misinformation surges, we propose CrowdNotes+, a unified framework that leverages large language models (LLMs) to augment Community Notes for faster and more reliable health misinformation governance. CrowdNotes+ integrates two complementary modes: (1) evidence-grounded note augmentation and (2) utility-guided note automation, along with a hierarchical three-step evaluation that progressively assesses relevance, correctness, and helpfulness. We instantiate the framework through HealthNotes, a benchmark of 1.2K helpfulness-annotated health notes paired with a fine-tuned helpfulness judge. Experiments on fifteen LLMs reveal an overlooked loophole in current helpfulness evaluation, where stylistic fluency is mistaken for factual accuracy, and demonstrate that our hierarchical evaluation and LLM-augmented generation jointly enhance factual precision and evidence utility. These results point toward a hybrid human-AI governance model that improves both the rigor and timeliness of crowd-sourced fact-checking.
Authors: Yuhang Li, Chenchen Zhang, Ruilin Lv, Ao Liu, Ken Deng, Yuanxing Zhang, Jiaheng Liu, Wiggin Zhou, Bo Zhou
Abstract: While Large Language Models (LLMs) excel at algorithmic code generation, they struggle with front-end development, where correctness is judged on rendered pixels and interaction. We present ReLook, an agentic, vision-grounded reinforcement learning framework that empowers an agent to close a robust generate--diagnose--refine loop by invoking a multimodal LLM (MLLM) as a tool. During training, the agent uses the MLLM-in-the-loop both as a visual critic--scoring code with screenshots--and as a source of actionable, vision-grounded feedback; a strict zero-reward rule for invalid renders anchors renderability and prevents reward hacking. To prevent behavioral collapse, we introduce Forced Optimization, a strict acceptance rule that admits only improving revisions, yielding monotonically better trajectories. At inference, we decouple the critic and run a lightweight, critic-free self-edit cycle, keeping latency comparable to base decoding while retaining most of the gains. Across three widely used benchmarks, ReLook consistently outperforms strong baselines in vision-grounded front-end code generation, highlighting the benefits of agentic perception, visual rewards, and training-inference decoupling.
Authors: Shuo Chen, Zhen Han, Haokun Chen, Bailan He, Shengyun Si, Jingpei Wu, Philip Torr, Volker Tresp, Jindong Gu
Abstract: Recent reasoning-based safety guardrails for Large Reasoning Models (LRMs), such as deliberative alignment, have shown strong defense against jailbreak attacks. By leveraging LRMs' reasoning ability, these guardrails help the models to assess the safety of user inputs before generating final responses. The powerful reasoning ability can analyze the intention of the input query and will refuse to assist once it detects the harmful intent hidden by the jailbreak methods. Such guardrails have shown a significant boost in defense, such as the near-perfect refusal rates on the open-source gpt-oss series. Unfortunately, we find that these powerful reasoning-based guardrails can be extremely vulnerable to subtle manipulation of the input prompts, and once hijacked, can lead to even more harmful results. Specifically, we first uncover a surprisingly fragile aspect of these guardrails: simply adding a few template tokens to the input prompt can successfully bypass the seemingly powerful guardrails and lead to explicit and harmful responses. To explore further, we introduce a bag of jailbreak methods that subvert the reasoning-based guardrails. Our attacks span white-, gray-, and black-box settings and range from effortless template manipulations to fully automated optimization. Along with the potential for scalable implementation, these methods also achieve alarmingly high attack success rates (e.g., exceeding 90% across 5 different benchmarks on gpt-oss series on both local host models and online API services). Evaluations across various leading open-source LRMs confirm that these vulnerabilities are systemic, underscoring the urgent need for stronger alignment techniques for open-sourced LRMs to prevent malicious misuse. Code is open-sourced at https://chenxshuo.github.io/bag-of-tricks.
Authors: Shubham Chatterjee, Jeff Dalton
Abstract: Neural IR has advanced through two distinct paths: entity-oriented approaches leveraging knowledge graphs and multi-vector models capturing fine-grained semantics. We introduce QDER, a neural re-ranking model that unifies these approaches by integrating knowledge graph semantics into a multi-vector model. QDER's key innovation lies in its modeling of query-document relationships: rather than computing similarity scores on aggregated embeddings, we maintain individual token and entity representations throughout the ranking process, performing aggregation only at the final scoring stage - an approach we call "late aggregation." We first transform these fine-grained representations through learned attention patterns, then apply carefully chosen mathematical operations for precise matches. Experiments across five standard benchmarks show that QDER achieves significant performance gains, with improvements of 36% in nDCG@20 over the strongest baseline on TREC Robust 2004 and similar improvements on other datasets. QDER particularly excels on difficult queries, achieving an nDCG@20 of 0.70 where traditional approaches fail completely (nDCG@20 = 0.0), setting a foundation for future work in entity-aware retrieval.
Authors: Shubham Chatterjee
Abstract: Current neural re-rankers often struggle with complex information needs and long, content-rich documents. The fundamental issue is not computational--it is intelligent content selection: identifying what matters in lengthy, multi-faceted texts. While humans naturally anchor their understanding around key entities and concepts, neural models process text within rigid token windows, treating all interactions as equally important and missing critical semantic signals. We introduce REGENT, a neural re-ranking model that mimics human-like understanding by using entities as a "semantic skeleton" to guide attention. REGENT integrates relevance guidance directly into the attention mechanism, combining fine-grained lexical matching with high-level semantic reasoning. This relevance-guided attention enables the model to focus on conceptually important content while maintaining sensitivity to precise term matches. REGENT achieves new state-of-the-art performance in three challenging datasets, providing up to 108% improvement over BM25 and consistently outperforming strong baselines including ColBERT and RankVicuna. To our knowledge, this is the first work to successfully integrate entity semantics directly into neural attention, establishing a new paradigm for entity-aware information retrieval.
Authors: Daniel Berhane Araya, Duoduo Liao
Abstract: Financial markets face growing threats from misinformation that can trigger billions in losses in minutes. Most existing approaches lack transparency in their decision-making and provide limited attribution to credible sources. We introduce FinVet, a novel multi-agent framework that integrates two Retrieval-Augmented Generation (RAG) pipelines with external fact-checking through a confidence-weighted voting mechanism. FinVet employs adaptive three-tier processing that dynamically adjusts verification strategies based on retrieval confidence, from direct metadata extraction to hybrid reasoning to full model-based analysis. Unlike existing methods, FinVet provides evidence-backed verdicts, source attribution, confidence scores, and explicit uncertainty flags when evidence is insufficient. Experimental evaluation on the FinFact dataset shows that FinVet achieves an F1 score of 0.85, which is a 10.4% improvement over the best individual pipeline (fact-check pipeline) and 37% improvement over standalone RAG approaches.
Authors: Nianyi Lin, Jiajie Zhang, Lei Hou, Juanzi Li
Abstract: A key challenge in applying reinforcement learning (RL) to diffusion large language models (dLLMs) lies in the intractability of their likelihood functions, which are essential for the RL objective, necessitating corresponding approximation in each training step. While existing methods approximate the log-likelihoods by their evidence lower bounds (ELBOs) via customized Monte Carlo (MC) sampling, the forward computational graphs of all MC samples need to be retained for the gradient computation of non-linear terms in the RL objective, resulting in significant memory overhead. This constraint restricts feasible sample sizes, leading to imprecise likelihood approximations and ultimately distorting the RL objective. To overcome this limitation, we propose \emph{Boundary-Guided Policy Optimization} (BGPO), a memory-efficient RL algorithm that maximizes a specially constructed lower bound of the ELBO-based objective. This lower bound is carefully designed to satisfy two key properties: (1) Linearity: it is formulated in a linear sum where each term depends only on a single MC sample, thereby enabling gradient accumulation across samples and ensuring constant memory usage; (2) Equivalence: Both the value and gradient of this lower bound are equal to those of the ELBO-based objective in on-policy training, making it also an effective approximation for the original RL objective. These properties allow BGPO to adopt a large MC sample size, resulting in more accurate likelihood approximations and improved RL objective estimation, which in turn leads to enhanced performance. Experiments show that BGPO significantly outperforms previous RL algorithms for dLLMs in math problem solving, code generation, and planning tasks.
Authors: Wei Huang, Yi Ge, Shuai Yang, Yicheng Xiao, Huizi Mao, Yujun Lin, Hanrong Ye, Sifei Liu, Ka Chun Cheung, Hongxu Yin, Yao Lu, Xiaojuan Qi, Song Han, Yukang Chen
Abstract: We propose QeRL, a Quantization-enhanced Reinforcement Learning framework for large language models (LLMs). While RL is essential for LLMs' reasoning capabilities, it is resource-intensive, requiring substantial GPU memory and long rollout durations. QeRL addresses these issues by combining NVFP4 quantization with Low-Rank Adaptation (LoRA), accelerating rollout phase of RL while reducing memory overhead. Beyond efficiency, our findings show that quantization noise increases policy entropy, enhancing exploration, and enabling the discovery of better strategies during RL. To further optimize exploration, QeRL introduces an Adaptive Quantization Noise (AQN) mechanism, which dynamically adjusts noise during training. Experiments demonstrate that QeRL delivers over 1.5 times speedup in the rollout phase. Moreover, this is the first framework to enable RL training of a 32B LLM on a single H100 80GB GPU, while delivering overall speedups for RL training. It also achieves faster reward growth and higher final accuracy than 16-bit LoRA and QLoRA, while matching the performance of full-parameter fine-tuning on mathematical benchmarks such as GSM8K (90.8%) and MATH 500 (77.4%) in the 7B model. These results establish QeRL as an efficient and effective framework for RL training in LLMs.
Authors: Shuzheng Si, Zefan Cai, Shuang Zeng, Guoqiang Feng, Jiaxing Lin, Baobao Chang
Abstract: Distantly-Supervised Named Entity Recognition effectively alleviates the burden of time-consuming and expensive annotation in the supervised setting. But the context-free matching process and the limited coverage of knowledge bases introduce inaccurate and incomplete annotation noise respectively. Previous studies either considered only incomplete annotation noise or indiscriminately handle two types of noise with the same strategy. In this paper, we argue that the different causes of two types of noise bring up the requirement of different strategies in model architecture. Therefore, we propose the SANTA to handle these two types of noise separately with (1) Memory-smoothed Focal Loss and Entity-aware KNN to relieve the entity ambiguity problem caused by inaccurate annotation, and (2) Boundary Mixup to alleviate decision boundary shifting problem caused by incomplete annotation and a noise-tolerant loss to improve the robustness. Benefiting from our separate tailored strategies, we confirm in the experiment that the two types of noise are well mitigated. SANTA also achieves a new state-of-the-art on five public datasets.
Authors: Ahmet Yavuz Uluslu, Gerold Schneider
Abstract: This paper presents the first application of Native Language Identification (NLI) for the Turkish language. NLI is the task of automatically identifying an individual's native language (L1) based on their writing or speech in a non-native language (L2). While most NLI research has focused on L2 English, our study extends this scope to L2 Turkish by analyzing a corpus of texts written by native speakers of Albanian, Arabic and Persian. We leverage a cleaned version of the Turkish Learner Corpus and demonstrate the effectiveness of syntactic features, comparing a structural Part-of-Speech n-gram model to a hybrid model that retains function words. Our models achieve promising results, and we analyze the most predictive features to reveal L1-specific transfer effects. We make our data and code publicly available for further study.
Authors: Yunshi Lan, Xinyuan Li, Hanyue Du, Xuesong Lu, Ming Gao, Weining Qian, Aoying Zhou
Abstract: Natural Language Processing (NLP) aims to analyze text or speech via techniques in the computer science field. It serves applications in the domains of healthcare, commerce, education, and so on. Particularly, NLP has been widely applied to the education domain and its applications have enormous potential to help teaching and learning. In this survey, we review recent advances in NLP with a focus on solving problems relevant to the education domain. In detail, we begin with introducing the related background and the real-world scenarios in education to which NLP techniques could contribute. Then, we present a taxonomy of NLP in the education domain and highlight typical NLP applications including question answering, question construction, automated assessment, and error correction. Next, we illustrate the task definition, challenges, and corresponding cutting-edge techniques based on the above taxonomy. In particular, LLM-involved methods are included for discussion due to the wide usage of LLMs in diverse NLP applications. After that, we showcase some off-the-shelf demonstrations in this domain, which are designed for educators or researchers. At last, we conclude with five promising directions for future research, including generalization over subjects and languages, deployed LLM-based systems for education, adaptive learning for teaching and learning, interpretability for education, and ethical consideration of NLP techniques. We organize all relevant datasets and papers in the open-available Github Link for better review https://github.com/LiXinyuan1015/NLP-for-Education.
Authors: Amin Dada, Marie Bauer, Amanda Butler Contreras, Osman Alperen Kora\c{s}, Constantin Marc Seibold, Kaleb E Smith, Jens Kleesiek
Abstract: Large Language Models (LLMs) are expected to significantly contribute to patient care, diagnostics, and administrative processes. Emerging biomedical LLMs aim to address healthcare-specific challenges, including privacy demands and computational constraints. Assessing the models' suitability for this sensitive application area is of the utmost importance. However, biomedical training has not been systematically evaluated on medical tasks. This study investigates the effect of biomedical training in the context of six practical medical tasks evaluating $25$ models. In contrast to previous evaluations, our results reveal a performance decline in nine out of twelve biomedical models after fine-tuning, particularly on tasks involving hallucinations, ICD10 coding, and instruction adherence. General-domain models like Meta-Llama-3.1-70B-Instruct outperformed their biomedical counterparts, indicating a trade-off between domain-specific fine-tuning and general medical task performance. We open-source all evaluation scripts and datasets at https://github.com/TIO-IKIM/CLUE to support further research in this critical area.
Authors: Yihang Wang, Xu Huang, Bowen Tian, Yueyang Su, Lei Yu, Huaming Liao, Yixing Fan, Jiafeng Guo, Xueqi Cheng
Abstract: Generative LLM have achieved remarkable success in various industrial applications, owing to their promising In-Context Learning capabilities. However, the issue of long context in complex tasks poses a significant barrier to their wider adoption, manifested in two main aspects: (i) The excessively long context leads to high costs and inference delays. (ii) A substantial amount of task-irrelevant information introduced by long contexts exacerbates the "lost in the middle" problem. Existing methods compress context by removing redundant tokens using metrics such as self-information or PPL, which is inconsistent with the objective of retaining the most important tokens when conditioning on a given query. In this study, we introduce information bottleneck theory (IB) to model the problem, offering a novel perspective that thoroughly addresses the essential properties required for context compression. Additionally, we propose a cross-attention-based approach to approximate mutual information in IB, which can be flexibly replaced with suitable alternatives in different scenarios. Extensive experiments on four datasets demonstrate that our method achieves a 25% increase in compression rate compared to the state-of-the-art, while maintaining question answering performance. In particular, the context compressed by our method even outperform the full context in some cases.
Authors: Chaithanya Bandi, Abir Harrasse
Abstract: The evaluation of Large Language Models (LLMs) remains challenging due to inconsistency, bias, and the absence of transparent decision criteria in automated judging. We present Debate, Deliberate, Decide (D3), a cost-aware, adversarial multi-agent framework that orchestrates structured debate among role-specialized agents (advocates, a judge, and an optional jury) to produce reliable and interpretable evaluations. D3 instantiates two complementary protocols: (1) Multi-Advocate One-Round Evaluation (MORE), which elicits k parallel defenses per answer to amplify signal via diverse advocacy, and (2) Single-Advocate Multi-Round Evaluation (SAMRE) with budgeted stopping, which iteratively refines arguments under an explicit token budget and convergence checks. We develop a probabilistic model of score gaps that (i) characterizes reliability and convergence under iterative debate and (ii) explains the separation gains from parallel advocacy. Under mild assumptions, the posterior distribution of the round-r gap concentrates around the true difference and the probability of mis-ranking vanishes; moreover, aggregating across k advocates provably increases expected score separation. We complement theory with a rigorous experimental suite across MT-Bench, AlignBench, and AUTO-J, showing state-of-the-art agreement with human judgments (accuracy and Cohen's kappa), reduced positional and verbosity biases via anonymization and role diversification, and a favorable cost-accuracy frontier enabled by budgeted stopping. Ablations and qualitative analyses isolate the contributions of debate, aggregation, and anonymity. Together, these results establish D3 as a principled, practical recipe for reliable, interpretable, and cost-aware LLM evaluation.
Authors: Ananya Malik, Kartik Sharma, Shaily Bhatt, Lynnette Hui Xian Ng
Abstract: Large Language Models (LLMs) offer a lucrative promise for scalable content moderation, including hate speech detection. However, they are also known to be brittle and biased against marginalised communities and dialects. This requires their applications to high-stakes tasks like hate speech detection to be critically scrutinized. In this work, we investigate the robustness of hate speech classification using LLMs particularly when explicit and implicit markers of the speaker's ethnicity are injected into the input. For explicit markers, we inject a phrase that mentions the speaker's linguistic identity. For the implicit markers, we inject dialectal features. By analysing how frequently model outputs flip in the presence of these markers, we reveal varying degrees of brittleness across 3 LLMs and 1 LM and 5 linguistic identities. We find that the presence of implicit dialect markers in inputs causes model outputs to flip more than the presence of explicit markers. Further, the percentage of flips varies across ethnicities. Finally, we find that larger models are more robust. Our findings indicate the need for exercising caution in deploying LLMs for high-stakes tasks like hate speech detection.
Authors: Ivan Srba, Olesya Razuvayevskaya, Jo\~ao A. Leite, Robert Moro, Ipek Baris Schlicht, Sara Tonelli, Francisco Moreno Garc\'ia, Santiago Barrio Lottmann, Denis Teyssou, Valentin Porcellini, Carolina Scarton, Kalina Bontcheva, Maria Bielikova
Abstract: In the age of social media and generative AI, the ability to automatically assess the credibility of online content has become increasingly critical, complementing traditional approaches to false information detection. Credibility assessment relies on aggregating diverse credibility signals - small units of information, such as content subjectivity, bias, or a presence of persuasion techniques - into a final credibility label/score. However, current research in automatic credibility assessment and credibility signals detection remains highly fragmented, with many signals studied in isolation and lacking integration. Notably, there is a scarcity of approaches that detect and aggregate multiple credibility signals simultaneously. These challenges are further exacerbated by the absence of a comprehensive and up-to-date overview of research works that connects these research efforts under a common framework and identifies shared trends, challenges, and open problems. In this survey, we address this gap by presenting a systematic and comprehensive literature review of 175 research papers, focusing on textual credibility signals within the field of Natural Language Processing (NLP), which undergoes a rapid transformation due to advancements in Large Language Models (LLMs). While positioning the NLP research into the the broader multidisciplinary landscape, we examine both automatic credibility assessment methods as well as the detection of nine categories of credibility signals. We provide an in-depth analysis of three key categories: 1) factuality, subjectivity and bias, 2) persuasion techniques and logical fallacies, and 3) check-worthy and fact-checked claims. In addition to summarising existing methods, datasets, and tools, we outline future research direction and emerging opportunities, with particular attention to evolving challenges posed by generative AI.
Authors: Matej Martinc, Hanh Thi Hong Tran, Senja Pollak, Boshko Koloski
Abstract: Keyword extraction involves identifying the most descriptive words in a document, allowing automatic categorisation and summarisation of large quantities of diverse textual data. Relying on the insight that real-world keyword detection often requires handling of diverse content, we propose a novel supervised keyword extraction approach based on the mixture of experts (MoE) technique. MoE uses a learnable routing sub-network to direct information to specialised experts, allowing them to specialise in distinct regions of the input space. SEKE, a mixture of Specialised Experts for supervised Keyword Extraction, uses DeBERTa as the backbone model and builds on the MoE framework, where experts attend to each token, by integrating it with a bidirectional Long short-term memory (BiLSTM) network, to allow successful extraction even on smaller corpora, where specialisation is harder due to lack of training data. The MoE framework also provides an insight into inner workings of individual experts, enhancing the explainability of the approach. We benchmark SEKE on multiple English datasets, achieving state-of-the-art performance compared to strong supervised and unsupervised baselines. Our analysis reveals that depending on data size and type, experts specialise in distinct syntactic and semantic components, such as punctuation, stopwords, parts-of-speech, or named entities. Code is available at https://github.com/matejMartinc/SEKE_keyword_extraction
URLs: https://github.com/matejMartinc/SEKE_keyword_extraction
Authors: Pietro Bernardelle, Leon Fr\"ohling, Stefano Civelli, Gianluca Demartini
Abstract: As increasingly capable large language models (LLMs) emerge, researchers have begun exploring their potential for subjective tasks. While recent work demonstrates that LLMs can be aligned with diverse human perspectives, evaluating this alignment on downstream tasks (e.g., hate speech detection) remains challenging due to the use of inconsistent datasets across studies. To address this issue, in this resource paper we propose a two-step framework: we (1) introduce SubData, an open-source Python library designed for standardizing heterogeneous datasets to evaluate LLMs perspective alignment; and (2) present a theory-driven approach leveraging this library to test how differently-aligned LLMs (e.g., aligned with different political viewpoints) classify content targeting specific demographics. SubData's flexible mapping and taxonomy enable customization for diverse research needs, distinguishing it from existing resources. We illustrate its usage with an example application and invite contributions to extend our initial release into a multi-construct benchmark suite for evaluating LLMs perspective alignment on natural language processing tasks.
Authors: Xiaopeng Li, Shanwen Wang, Shasha Li, Shezheng Song, Bin Ji, Jun Ma, Jie Yu
Abstract: Model editing enables targeted updates to the knowledge of large language models (LLMs) with minimal retraining. Among existing approaches, locate-then-edit methods constitute a prominent paradigm: they first identify critical layers, then compute residuals at the final critical layer based on the target edit, and finally apply least-squares-based multi-layer updates via $\textbf{residual distribution}$. While empirically effective, we identify a counterintuitive failure mode: residual distribution, a core mechanism in these methods, introduces weight shift errors that undermine editing precision. Through theoretical and empirical analysis, we show that such errors increase with the distribution distance, batch size, and edit sequence length, ultimately leading to inaccurate or suboptimal edits. To address this, we propose the $\textbf{B}$oundary $\textbf{L}$ayer $\textbf{U}$pdat$\textbf{E (BLUE)}$ strategy to enhance locate-then-edit methods. Sequential batch editing experiments on three LLMs and two datasets demonstrate that BLUE not only delivers an average performance improvement of 35.59\%, significantly advancing the state of the art in model editing, but also enhances the preservation of LLMs' general capabilities. Our code is available at https://github.com/xpq-tech/BLUE.
Authors: Changhao Jiang, Ming Zhang, Yifei Cao, Junjie Ye, Xiaoran Fan, Shihan Dou, Zhiheng Xi, Jiajun Sun, Yi Dong, Yujiong Shen, Jingqi Tong, Baoyu Fan, Qi Zhang, Tao Gui, Xuanjing Huang
Abstract: The GPT-4 technical report suggests that downstream performance can be predicted from pre-training signals, but offers little methodological detail on how to quantify this. This work address this gap by modeling knowledge retention, the capacity of a pre-trained language model to memorize factual information from its corpus, and introduce a principled method to estimate it prior to training. We propose Size-dependent Mutual Information (SMI), an information-theoretic predictor that integrates knowledge frequency, knowledge specificity, and model size to forecast closed-book question answering (QA) accuracy. SMI is validated through large-scale document retrieval over the disclosed pre-training corpora of 21 public and 3 custom models, combined with a robust multi-template QA evaluation. Experiments show that SMI significantly outperforms repetition-based baselines and achieves $R^2$ > 0.7 in predicting QA accuracy for models above 1B parameters, without additional training. The analysis further reveals diminishing returns from scaling data and model size and provides evidence for an intrinsic upper bound on knowledge retention achievable by pre-training alone, motivating retrieval and other augmentation strategies.
Authors: Selim Furkan Tekin, Fatih Ilhan, Gaowen Liu, Ramana Rao Kompella, Ling Liu
Abstract: The advancement of LLMs and their accessibility have triggered renewed interest in multi-agent reinforcement learning as robust and adaptive frameworks for dynamically changing environments. This paper introduces RL-Focal, a two-stage RL agent framework that routes and ensembles LLMs. First, we develop the Decider RL-agent, which learns to dynamically select an ensemble of small size ($m_i$) among $N$ LLMs ($m_i \ll N$) for incoming queries from a user-defined downstream task $i$, by maximizing both error-diversity and reasoning-performance of the selected ensemble through iterative updates of task-adaptive rewards and policy. Second, to enable effective fusion of dynamically selected LLMs, we develop the stage-2 Fusion RL-agent, which learns to resolve reasoning conflicts from different LLMs and dynamically adapts to different ensemble teams composed by the Decider Agent for different downstream tasks. Third, we introduce the focal diversity metric to better model the error correlations among multiple LLMs, further improving the generalization performance of the Decider Agent, which actively prunes the ensemble combinations. By focal diversity, we enhance performance across tasks by effectively promoting reward-aware and policy-adaptive ensemble selection and inference fusion. Extensive evaluations on five benchmarks show that RL-Focal achieves the performance improvement of 8.48\% with an ensemble of small size compared to the best individual LLM in a pool and offers stronger robustness. Code is available at https://github.com/sftekin/rl-focal
Authors: Shuhaib Mehri, Xiusi Chen, Heng Ji, Dilek Hakkani-T\"ur
Abstract: High-quality instruction-tuning data is crucial for developing Large Language Models (LLMs) that can effectively navigate real-world tasks and follow human instructions. While synthetic data generation offers a scalable approach for creating such datasets, it imposes a quality ceiling where models trained on the data cannot outperform the LLM generating it. To overcome this limitation, we introduce Reference-Level Feedback, a paradigm that extracts desirable characteristics from carefully curated reference samples to guide the synthesis of higher-quality instruction-response pairs. Using this approach, we synthesize REFED, a dataset of 10K instruction-response pairs. Fine-tuning Llama-3.1-8B-Instruct and Mistral-7B-Instruct on REFED demonstrate state-of-the-art performance among similarly sized models, notably reaching a 43.96\% length-controlled win-rate on AlpacaEval 2.0. Extensive experiments demonstrate that Reference-Level Feedback consistently outperforms traditional sample-level feedback methods, generalizes across model architectures, and produces high-quality and diverse data at low cost.
Authors: Yongcheng Zeng, Xinyu Cui, Xuanfa Jin, Qirui Mi, Guoqing Liu, Zexu Sun, Mengyue Yang, Dong Li, Weiyu Ma, Ning Yang, Jian Zhao, Jianye Hao, Haifeng Zhang, Jun Wang
Abstract: Self-Refinement refers to a model's ability to revise its own responses to produce improved outputs. This capability can also serve as a fundamental mechanism for Self-Improvement, for example, by reconstructing datasets with refined results to enhance intrinsic model performance. However, our comprehensive experiments reveal that large language models (LLMs) show no clear evidence of inherent Self-Refinement and may even experience response quality degradation after Self-Refinement. To address this issue, we propose EVOLVE, a simple and effective framework for eliciting and tracking the evolution of Self-Refinement through iterative training. We first explore optimization methods during training to activate the model's Self-Refinement capability. Then, at inference, we investigate various generation strategies to further enhance and utilize Self-Refinement while supplying the necessary data for training. Through synergistic optimization of training and inference stages, we continually evolve the model's Self-Refinement ability, enabling it to better refine its own responses. Moreover, we demonstrate the potential of leveraging Self-Refinement to achieve broader Self-Improvement of intrinsic model abilities. Experiments show that the evolved Self-Refinement ability enables the Llama-3.1-8B base model to surpass GPT-4o, achieving 62.3% length-controlled and 63.3% raw win rates on AlpacaEval 2, and 50.3% on Arena-Hard. It also generalizes effectively to out-of-domain reasoning tasks, improving performance on mathematical reasoning benchmarks such as GSM8K and MATH.
Authors: Jonathan Pofcher, Christopher M. Homan, Randall Sell, Ashiqur R. KhudaBukhsh
Abstract: This paper makes three contributions. First, via a substantial corpus of 1,419,047 comments posted on 3,161 YouTube news videos of major US cable news outlets, we analyze how users engage with LGBTQ+ news content. Our analyses focus both on positive and negative content. In particular, we construct a fine-grained hope speech classifier that detects positive (hope speech), negative, neutral, and irrelevant content. Second, in consultation with a public health expert specializing on LGBTQ+ health, we conduct an annotation study with a balanced and diverse political representation and release a dataset of 3,750 instances with fine-grained labels and detailed annotator demographic information. Finally, beyond providing a vital resource for the LGBTQ+ community, our annotation study and subsequent in-the-wild assessments reveal (1) strong association between rater political beliefs and how they rate content relevant to a marginalized community; (2) models trained on individual political beliefs exhibit considerable in-the-wild disagreement; and (3) zero-shot large language models (LLMs) align more with liberal raters.
Authors: Seojin Hwang, Yumin Kim, Byeongjeong Kim, Donghoon Shin, Hwanhee Lee
Abstract: Large Language Models (LLMs) are integral to applications such as conversational agents and content creation, where precise control over a model's personality is essential for maintaining tone, consistency, and user engagement. However, prevailing prompt-based or fine-tuning approaches either lack robustness or demand large-scale training data, making them costly and impractical. In this paper, we present PALETTE (Personality Adjustment by LLM SElf-TargeTed quEries), a novel method for personality editing in LLMs. Our approach introduces adjustment queries, where self-referential statements grounded in psychological constructs are treated analogously to factual knowledge, enabling direct editing of personality-related responses. Unlike fine-tuning, PALETTE requires only 12 editing samples to achieve substantial improvements in personality alignment across personality dimensions. Experimental results from both automatic and human evaluations demonstrate that our method enables more stable and well-balanced personality control in LLMs.
Authors: Reza Averly, Frazier N. Baker, Ian A. Watson, Xia Ning
Abstract: Drug discovery is a long, expensive, and complex process, relying heavily on human medicinal chemists, who can spend years searching the vast space of potential therapies. Recent advances in artificial intelligence for chemistry have sought to expedite individual drug discovery tasks; however, there remains a critical need for an intelligent agent that can navigate the drug discovery process. Towards this end, we introduce LIDDIA, an autonomous agent capable of intelligently navigating the drug discovery process in silico. By leveraging the reasoning capabilities of large language models, LIDDIA serves as a low-cost and highly-adaptable tool for autonomous drug discovery. We comprehensively examine LIDDIA , demonstrating that (1) it can generate molecules meeting key pharmaceutical criteria on over 70% of 30 clinically relevant targets, (2) it intelligently balances exploration and exploitation in the chemical space, and (3) it identifies one promising novel candidate on AR/NR3C4, a critical target for both prostate and breast cancers. Code and dataset are available at https://github.com/ninglab/LIDDiA
Authors: Chengyin Xu, Kaiyuan Chen, Xiao Li, Ke Shen, Chenggang Li
Abstract: The escalating scale and cost of Large Language Models (LLMs) training necessitate accurate pre-training prediction of downstream task performance for efficient resource allocation. This is challenged by: 1) the emergence phenomenon, where metrics become meaningful only after extensive training, hindering prediction by smaller models; and 2) uneven task difficulty and inconsistent performance scaling patterns, leading to high metric variability. Current prediction methods lack accuracy and reliability. We propose a Clustering-On-Difficulty (COD) framework for downstream performance prediction. The COD framework clusters tasks by their difficulty scaling features, thereby establishing a more stable and predictable support subset through the exclusion of tasks exhibiting non-emergent behavior or irregular scaling. We adopt a performance scaling law to predict cluster-wise performance with theoretical support. Predictable subset performance acts as an intermediate predictor for the full evaluation set. We further derive a mapping function to accurately extrapolate the performance of the subset to the full set. Applied to an LLM with 70B parameters, COD achieved a 1.36% average prediction error across eight key LLM benchmarks, offering actionable insights for resource allocation and training monitoring of LLMs pretraining.
Authors: Wenkai Yang, Shuming Ma, Yankai Lin, Furu Wei
Abstract: Recent studies have shown that making a model spend more time thinking through longer Chain of Thoughts (CoTs) enables it to gain significant improvements in complex reasoning tasks. While current researches continue to explore the benefits of increasing test-time compute by extending the CoT lengths of Large Language Models (LLMs), we are concerned about a potential issue hidden behind the current pursuit of test-time scaling: Would excessively scaling the CoT length actually bring adverse effects to a model's reasoning performance? Our explorations on mathematical reasoning tasks reveal an unexpected finding that scaling with longer CoTs can indeed impair the reasoning performance of LLMs in certain domains. Moreover, we discover that there exists an optimal scaled length distribution that differs across different domains. Based on these insights, we propose a Thinking-Optimal Scaling strategy. Our method first uses a small set of seed data with varying response length distributions to teach the model to adopt different reasoning efforts for deep thinking. Then, the model selects its shortest correct response under different reasoning efforts on additional problems for self-improvement. Our self-improved models built upon Qwen2.5-32B-Instruct outperform other distillation-based 32B o1-like models across various math benchmarks, and achieve performance on par with the teacher model QwQ-32B-Preview that produces the seed data.
Authors: Jakub Macina, Nico Daheim, Ido Hakimi, Manu Kapur, Iryna Gurevych, Mrinmaya Sachan
Abstract: Evaluating the pedagogical capabilities of AI-based tutoring models is critical for making guided progress in the field. Yet, we lack a reliable, easy-to-use, and simple-to-run evaluation that reflects the pedagogical abilities of models. To fill this gap, we present MathTutorBench, an open-source benchmark for holistic tutoring model evaluation. MathTutorBench contains a collection of datasets and metrics that broadly cover tutor abilities as defined by learning sciences research in dialog-based teaching. To score the pedagogical quality of open-ended teacher responses, we train a reward model and show it can discriminate expert from novice teacher responses with high accuracy. We evaluate a wide set of closed- and open-weight models on MathTutorBench and find that subject expertise, indicated by solving ability, does not immediately translate to good teaching. Rather, pedagogy and subject expertise appear to form a trade-off that is navigated by the degree of tutoring specialization of the model. Furthermore, tutoring appears to become more challenging in longer dialogs, where simpler questioning strategies begin to fail. We release the benchmark, code, and leaderboard openly to enable rapid benchmarking of future models.
Authors: Siyuan Zhang, Yichi Zhang, Yinpeng Dong, Hang Su
Abstract: Large Language Models (LLMs) often struggle to align their responses with objective facts, resulting in the issue of factual hallucinations, which can be difficult to detect and mislead users without relevant knowledge. Although post-training techniques have been employed to mitigate the issue, existing methods usually suffer from poor generalization and trade-offs in other different capabilities. In this paper, we propose to address these by directly augmenting LLM's fundamental ability to precisely leverage its knowledge and introduce PKUE (Precise Knowledge Utilization Enhancement), which fine-tunes the model on self-generated responses to precise and simple factual questions through preference optimization. Furthermore, we construct FactualBench, a comprehensive and precise factual QA dataset containing 181k Chinese data spanning 21 domains, to facilitate both evaluation and training. Extensive experiments demonstrate that PKUE significantly improves LLM overall performance, with consistent enhancement across factual tasks of various forms, general tasks beyond factuality, and tasks in different language.
Authors: Zixuan Gong, Shijia Li, Yong Liu, Jiaye Teng
Abstract: Transformers may exhibit two-stage training dynamics during the real-world training process. For instance, when training GPT-2 on the Counterfact dataset, the answers progress from syntactically incorrect to syntactically correct to semantically correct. However, existing theoretical analyses hardly account for this feature-level two-stage phenomenon, which originates from the disentangled two-type features like syntax and semantics. In this paper, we theoretically demonstrate how the two-stage training dynamics potentially occur in transformers. Specifically, we analyze the feature learning dynamics induced by the aforementioned disentangled two-type feature structure, grounding our analysis in a simplified yet illustrative setting that comprises a normalized ReLU self-attention layer and structured data. Such disentanglement of feature structure is general in practice, e.g., natural languages contain syntax and semantics, and proteins contain primary and secondary structures. To our best knowledge, this is the first rigorous result regarding a feature-level two-stage optimization process in transformers. Additionally, a corollary indicates that such a two-stage process is closely related to the spectral properties of the attention weights, which accords well with our empirical findings.
Authors: Kuang-Da Wang, Teng-Ruei Chen, Yu Heng Hung, Guo-Xun Ko, Shuoyang Ding, Yueh-Hua Wu, Yu-Chiang Frank Wang, Chao-Han Huck Yang, Wen-Chih Peng, Ping-Chun Hsieh
Abstract: Aligning Large Language Models (LLMs) with human preferences through finetuning is resource-intensive, motivating lightweight alternatives at test time. We address test-time alignment through the lens of sequential decision making, a perspective that reveals two fundamental challenges. When actions are defined at the token level, as in guided decoding, alignment suffers from the curse of horizon. Conversely, when actions are at the response level, as in traditional iterative refinement, the curse of dimensionality emerges. To resolve this trade-off, we draw inspiration from Model Predictive Control (MPC) in control theory to propose Textual Model Predictive Control (TMPC), a novel predictive planning framework adapted for aligning LLMs at inference time. A key limitation of standard MPC is its reliance on predefined, hard segment boundaries, which are often absent in text generation. TMPC overcomes this by introducing two principles inspired by hierarchical reinforcement learning: (1) Hindsight Subgoal Identification, where TMPC analyzes generation subgoals to retrospectively identify high-reward intermediate outputs as subgoals. This allows the framework to discover meaningful, task-specific planning steps (e.g., a sentence in machine translation or a bug fix in code generation.). (2) Subgoal-Conditioned Re-Generation, where these identified subgoals are used to guide subsequent planning iterations. By conditioning on these proven, high-quality subgoals, TMPC ensures stable improvement by building upon previously validated successes. TMPC is evaluated on three tasks with distinct segmentation properties: discourse-level translation, long-form response generation, and program synthesis. The results demonstrate that TMPC consistently improves performance, highlighting the generality.
Authors: Milad Alshomary, Nikhil Reddy Varimalla, Vishal Anand, Smaranda Muresan, Kathleen McKeown
Abstract: We propose a new approach for the authorship attribution task that leverages the various linguistic representations learned at different layers of pre-trained transformer-based models. We evaluate our approach on three datasets, comparing it to a state-of-the-art baseline in in-domain and out-of-domain scenarios. We found that utilizing various transformer layers improves the robustness of authorship attribution models when tested on out-of-domain data, resulting in new state-of-the-art results. Our analysis gives further insights into how our model's different layers get specialized in representing certain stylistic features that benefit the model when tested out of the domain.
Authors: Shiqi Chen, Tongyao Zhu, Ruochen Zhou, Jinghan Zhang, Siyang Gao, Juan Carlos Niebles, Mor Geva, Junxian He, Jiajun Wu, Manling Li
Abstract: Large Vision Language Models (VLMs) have long struggled with spatial reasoning tasks. Surprisingly, even simple spatial reasoning tasks, such as recognizing "under" or "behind" relationships between only two objects, pose significant challenges for current VLMs. In this work, we study the spatial reasoning challenge from the lens of mechanistic interpretability, diving into the model's internal states to examine the interactions between image and text tokens. By tracing attention distribution over the image through out intermediate layers, we observe that successful spatial reasoning correlates strongly with the model's ability to align its attention distribution with actual object locations, particularly differing between familiar and unfamiliar spatial relationships. Motivated by these findings, we propose ADAPTVIS based on inference-time confidence scores to sharpen the attention on highly relevant regions when confident, while smoothing and broadening the attention window to consider a wider context when confidence is lower. This training-free decoding method shows significant improvement (e.g., up to a 50 absolute point improvement) on spatial reasoning benchmarks such as WhatsUp and VSR with negligible cost. We make code and data publicly available for research purposes at https://github.com/shiqichen17/AdaptVis.
Authors: Zhuo Li, Yuhao Du, Xiaoqi Jiao, Yiwen Guo, Yuege Feng, Xiang Wan, Anningzhe Gao, Jinpeng Hu
Abstract: Selecting high-quality and diverse training samples from extensive datasets plays a crucial role in reducing training overhead and enhancing the performance of Large Language Models (LLMs). However, existing studies fall short in assessing the overall value of selected data, focusing primarily on individual quality, and struggle to strike an effective balance between ensuring diversity and minimizing data point traversals. Therefore, this paper introduces a novel choice-based sample selection framework that shifts the focus from evaluating individual sample quality to comparing the contribution value of different samples when incorporated into the subset. Thanks to the advanced language understanding capabilities of LLMs, we utilize LLMs to evaluate the value of each option during the selection process. Furthermore, we design a greedy sampling process where samples are incrementally added to the subset, thereby improving efficiency by eliminating the need for exhaustive traversal of the entire dataset with the limited budget. Extensive experiments demonstrate that selected data from our method not only surpasses the performance of the full dataset but also achieves competitive results with recent powerful studies, while requiring fewer selections. Moreover, we validate our approach on a larger medical dataset, highlighting its practical applicability in real-world applications. Our code and data are available at https://github.com/BIRlz/comperative_sample_selection.
URLs: https://github.com/BIRlz/comperative_sample_selection.
Authors: Jianghao Chen, Junhong Wu, Yangyifan Xu, Jiajun Zhang
Abstract: Long-context modeling has drawn more and more attention in the area of Large Language Models (LLMs). Continual training with long-context data becomes the de-facto method to equip LLMs with the ability to process long inputs. However, it still remains an open challenge to measure the quality of long-context training data. To address this issue, we propose a Long-context data selection framework with Attention-based Dependency Measurement (LADM), which can efficiently identify high-quality long-context data from a large-scale, multi-domain pre-training corpus. LADM leverages the retrieval capabilities of the attention mechanism to capture contextual dependencies, ensuring a comprehensive quality measurement of long-context data. Experimental results show that our LADM framework significantly boosts the performance of LLMs on multiple long-context tasks with only 1B tokens for continual training.
Authors: Hoang Hai Phan, Nguyen Duc Minh Vu, Nam Dang Phuong
Abstract: Neural Machine Translation (NMT) driven by Transformer architectures has advanced significantly, yet faces challenges with low-resource language pairs like Vietnamese-Japanese (Vi-Ja). Issues include sparse parallel data and handling linguistic/cultural nuances. Recent progress in Large Language Models (LLMs) with strong reasoning, often refined via Reinforcement Learning (RL), enables high-quality synthetic data generation. We introduce VNJPTranslate, a pipeline designed to systematically address the Vi-Ja translation task. It features a targeted data augmentation strategy using advanced LLMs with Chain-of-Thought prompting for challenging segments identified via corpus analysis. Subsequently, we employ efficient fine-tuning techniques (Unsloth with QLoRA) on a capable, low-parameter autoregressive model (specifically, a fine-tuned version of the 1.8B parameter Sailor model, which is based on the Qwen architecture) to create a practical and high-performing translation system. This integrated approach aims to improve Vi-Ja translation quality significantly over existing baselines.
Authors: Shuai Zhao, Yunqiu Xu, Linchao Zhu, Yi Yang
Abstract: Large language models~(LLMs) are expected to be helpful, harmless, and honest. In different alignment scenarios, such as safety, confidence, and general preference alignment, binary preference data collection and reward modeling are resource-intensive but play a central role in transferring human preferences. In this work, we explore using the similarity between sampled generations and reference answers as a supplementary reward function for alignment. When unary reference answers are available, such similarity-based rewards can circumvent the need for binary preference data and explicit reward modeling. We introduce \textit{RefAlign}, a versatile REINFORCE-style alignment algorithm that does not rely on reward or reference models. RefAlign utilizes language generation evaluation metrics, such as BERTScore, between sampled generations and reference answers as surrogate rewards. Beyond general preference optimization, RefAlign can be naturally extended to diverse scenarios, including safety and confidence alignment, by combining similarity-based rewards with task-specific objectives. Across multiple scenarios, RefAlign achieves performance comparable to prior alignment methods while operating without binary preference data or reward models. The code is available at https://github.com/mzhaoshuai/RefAlign.
Authors: Shahriar Noroozizadeh, Sayantan Kumar, Jeremy C. Weiss
Abstract: Clinical case reports encode temporal patient trajectories that are often underexploited by traditional machine learning methods relying on structured data. In this work, we introduce the forecasting problem from textual time series, where timestamped clinical findings -- extracted via an LLM-assisted annotation pipeline -- serve as the primary input for prediction. We systematically evaluate a diverse suite of models, including fine-tuned decoder-based large language models and encoder-based transformers, on tasks of event occurrence prediction, temporal ordering, and survival analysis. Our experiments reveal that encoder-based models consistently achieve higher F1 scores and superior temporal concordance for short- and long-horizon event forecasting, while fine-tuned masking approaches enhance ranking performance. In contrast, instruction-tuned decoder models demonstrate a relative advantage in survival analysis, especially in early prognosis settings. Our sensitivity analyses further demonstrate the importance of time ordering, which requires clinical time series construction, as compared to text ordering, the format of the text inputs that LLMs are classically trained on. This highlights the additional benefit that can be ascertained from time-ordered corpora, with implications for temporal tasks in the era of widespread LLM use.
Authors: William Bruns
Abstract: Humans understand new combinations of words encountered if they are combinations of words recognized from different contexts, an ability called Compositional Generalization. The COGS benchmark (Kim and Linzen, 2020) arXiv:2010.05465 reports 0% accuracy for Transformer models on some structural generalizations. We use (Weiss et al., 2021) arXiv:2106.06981's Restricted Access Sequence Processing (RASP), a Transformer-equivalent programming language, to demonstrate that a Transformer Encoder-Decoder can perform COGS and the semantically equivalent ReCOGS_pos (Wu et al., 2024) arXiv:2303.13716 systematically and compositionally: Our RASP models attain near perfect scores on structural generalization splits on COGS (exact match) and ReCOGS_pos (semantic exact match). Our RASP models show the (Re)COGS tasks do not require a hierarchical or tree-structured solution (contrary to (Kim and Linzen, 2020) arXiv:2010.05465, (Yao and Koller, 2022) arXiv:2210.13050, (Murty et al., 2022) arXiv:2305.18741, (Liu et al., 2021) arXiv:2107.06516): we use word-level tokens with an "embedding" layer that tags with possible part of speech, applying just once per encoder pass 19 attention-head compatible flat pattern-matching rules (easily identified with specific training examples), shown using grammar coverage (Zeller et al., 2023) to cover the non-recursive aspects of the input grammar, plus masking out prepositional phrases ("pp noun") and/or sentential complements (cp) when recognizing grammar patterns and extracting nouns related to the main verb in the sentence, and output the next logical form (LF) token (repeating until the LF is complete). The models do not apply recursive, tree-structured rules like "np_det pp np -> np_pp -> np", but score near perfect semantic and string exact match on both COGS and ReCOGS pp recursion, cp recursion using the decoder loop.
Authors: Zhibo Man, Yuanmeng Chen, Yujie Zhang, Jinan Xu
Abstract: Currently, Large Language Models (LLMs) have achieved remarkable results in machine translation. However, their performance in multi-domain translation (MDT) is less satisfactory, the meanings of words can vary across different domains, highlighting the significant ambiguity inherent in MDT. Therefore, evaluating the disambiguation ability of LLMs in MDT, remains an open problem. To this end, we present an evaluation and analysis of LLMs on disambiguation in multi-domain translation (DMDTEval), our systematic evaluation framework consisting of three critical aspects: (1) we construct a translation test set with multi-domain ambiguous word annotation, (2) we curate a diverse set of disambiguation prompt strategies, and (3) we design precise disambiguation metrics, and study the efficacy of various prompt strategies on multiple state-of-the-art LLMs. We conduct comprehensive experiments across 4 language pairs and 13 domains, our extensive experiments reveal a number of crucial findings that we believe will pave the way and also facilitate further research in the critical area of improving the disambiguation of LLMs.
Authors: Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yongkang Wu, Ji-Rong Wen, Yutao Zhu, Zhicheng Dou
Abstract: Large reasoning models (LRMs), such as OpenAI-o1 and DeepSeek-R1, demonstrate impressive long-horizon reasoning capabilities. However, their reliance on static internal knowledge limits their performance on complex, knowledge-intensive tasks and hinders their ability to produce comprehensive research reports requiring synthesis of diverse web information. To address this, we propose WebThinker, a deep research agent that empowers LRMs to autonomously search the web, navigate among web pages, and draft reports during the reasoning process. WebThinker integrates a Deep Web Explorer module, enabling LRMs to dynamically search, navigate, and extract information from the web when encountering knowledge gaps. It also employs an Autonomous Think-Search-and-Draft strategy, allowing the model to seamlessly interleave reasoning, information gathering, and report writing in real time. To further enhance research tool utilization, we introduce an RL-based training strategy via iterative online Direct Preference Optimization (DPO). Extensive experiments on complex reasoning benchmarks (GPQA, GAIA, WebWalkerQA, HLE) and scientific report generation tasks (Glaive) demonstrate that WebThinker significantly outperforms existing methods and strong proprietary systems. Our approach enhances LRM reliability and applicability in complex scenarios, paving the way for more capable and versatile deep research systems. The code is available at https://github.com/RUC-NLPIR/WebThinker.
Authors: Ran Zhang, Wei Zhao, Lieve Macken, Steffen Eger
Abstract: The impact of Large Language Models (LLMs) has extended into literary domains. However, existing evaluation metrics for literature prioritize mechanical accuracy over artistic expression and tend to overrate machine translation as being superior to human translation from experienced professionals. In the long run, this bias could result in an irreversible decline in translation quality and cultural authenticity. In response to the urgent need for a specialized literary evaluation metric, we introduce LITRANSPROQA, a novel, reference-free, LLM-based question-answering framework designed for literary translation evaluation. LITRANSPROQA integrates humans in the loop to incorporate insights from professional literary translators and researchers, focusing on critical elements in literary quality assessment such as literary devices, cultural understanding, and authorial voice. Our extensive evaluation shows that while literary-finetuned XCOMET-XL yields marginal gains, LITRANSPROQA substantially outperforms current metrics, achieving up to 0.07 gain in correlation and surpassing the best state-of-the-art metrics by over 15 points in adequacy assessments. Incorporating professional translator insights as weights further improves performance, highlighting the value of translator inputs. Notably, LITRANSPROQA reaches an adequacy performance comparable to trained linguistic student evaluators, though it still falls behind experienced professional translators. LITRANSPROQA shows broad applicability to open-source models like LLaMA3.3-70b and Qwen2.5-32b, indicating its potential as an accessible and training-free tool for evaluating literary translations that require local processing due to copyright or ethical considerations.
Authors: Xinyue Lou, You Li, Jinan Xu, Xiangyu Shi, Chi Chen, Kaiyu Huang
Abstract: The rapid development of Multimodal Large Reasoning Models (MLRMs) has demonstrated broad application potential, yet their safety and reliability remain critical concerns that require systematic exploration. To address this gap, we conduct a comprehensive and systematic safety evaluation of 11 MLRMs across 5 benchmarks and unveil prevalent safety degradation phenomena in most advanced models. Moreover, our analysis reveals distinct safety patterns across different benchmarks: significant safety degradation is observed across jailbreak robustness benchmarks, whereas safety-awareness benchmarks demonstrate less pronounced degradation. In particular, the long thought process in some scenarios even enhances safety performance. Therefore, it is a potential approach to address safety issues in MLRMs by leveraging the intrinsic reasoning capabilities of the model to detect unsafe intent. To operationalize this insight, we construct a multimodal tuning dataset that incorporates a safety-oriented thought process. Experimental results from fine-tuning existing MLRMs with this dataset effectively enhances the safety on both jailbreak robustness and safety-awareness benchmarks. This study provides a new perspective for developing safe MLRMs. Our dataset is available at https://github.com/xinyuelou/Think-in-Safety.
Authors: Doyoung Kim, Youngjun Lee, Joeun Kim, Jihwan Bang, Hwanjun Song, Susik Yoon, Jae-Gil Lee
Abstract: Conversational query reformulation (CQR) has become indispensable for improving retrieval in dialogue-based applications. However, existing approaches typically rely on reference passages for optimization, which are impractical to acquire in real-world scenarios. To address this limitation, we introduce a novel reference-free preference optimization framework DualReform that generates pseudo reference passages from commonly-encountered conversational datasets containing only queries and responses. DualReform attains this goal through two key innovations: (1) response-based inference, where responses serve as proxies to infer pseudo reference passages, and (2) response refinement via the dual-role of CQR, where a CQR model refines responses based on the shared objectives between response refinement and CQR. Despite not relying on reference passages, DualReform achieves 96.9--99.1% of the retrieval accuracy attainable only with reference passages and surpasses the state-of-the-art method by up to 31.6%.
Authors: Chenxi Whitehouse, Tianlu Wang, Ping Yu, Xian Li, Jason Weston, Ilia Kulikov, Swarnadeep Saha
Abstract: The progress of AI is bottlenecked by the quality of evaluation, making powerful LLM-as-a-Judge models a core solution. The efficacy of these judges depends on their chain-of-thought reasoning, creating a critical need for methods that can effectively optimize this reasoning process. In this work, we introduce J1, a reinforcement learning framework for teaching LLM judges to think before making decisions. Our core contribution lies in converting all judgment tasks for non-verifiable and verifiable prompts into a unified format with verifiable rewards, enabling direct optimization of evaluation quality while mitigating positional bias. We then use RL to train thinking-judges at scales of 8B, 32B, and 70B and show that they obtain state-of-the-art performance across multiple benchmarks. In particular, J1-Qwen-32B, our multitasked pointwise and pairwise judge also outperforms o1-mini, o3, and a much larger 671B DeepSeek-R1 on some benchmarks, while only training on synthetic data. Through comprehensive ablations of pairwise, pointwise, and multitask J1 variants, we demonstrate the effectiveness of our approach across seed prompts, reward strategies, and training recipes. Qualitative analysis reveals that J1 develops systematic evaluation strategies, including dynamic criteria generation, reference answer creation, iterative self-correction of initial assessments, and feedback generation for low-quality responses.
Authors: Huashan Sun, Shengyi Liao, Yansen Han, Yu Bai, Yang Gao, Cheng Fu, Weizhou Shen, Fanqi Wan, Ming Yan, Ji Zhang, Fei Huang
Abstract: Despite advances in pretraining with extended context lengths, large language models (LLMs) still face challenges in effectively utilizing real-world long-context information, primarily due to insufficient long-context alignment caused by data quality issues, training inefficiencies, and the lack of well-designed optimization objectives. To address these limitations, we propose a framework named $\textbf{S}$h$\textbf{o}$rt-to-$\textbf{Lo}$ng $\textbf{P}$reference $\textbf{O}$ptimization ($\textbf{SoLoPO}$), decoupling long-context preference optimization (PO) into two components: short-context PO and short-to-long reward alignment (SoLo-RA), supported by both theoretical and empirical evidence. Specifically, short-context PO leverages preference pairs sampled from short contexts to enhance the model's contextual knowledge utilization ability. Meanwhile, SoLo-RA explicitly encourages reward score consistency utilization for the responses when conditioned on both short and long contexts that contain identical task-relevant information. This facilitates transferring the model's ability to handle short contexts into long-context scenarios. SoLoPO is compatible with mainstream preference optimization algorithms, while substantially improving the efficiency of data construction and training processes. Experimental results show that SoLoPO enhances all these algorithms with respect to stronger length and domain generalization abilities across various long-context benchmarks, while achieving notable improvements in both computational and memory efficiency.
Authors: Jitai Hao, Qiang Huang, Hao Liu, Xinyan Xiao, Zhaochun Ren, Jun Yu
Abstract: Training high-performing Small Language Models (SLMs) remains costly, even with knowledge distillation and pruning from larger teacher models. Existing work often faces three key challenges: (1) information loss from hard pruning, (2) inefficient alignment of representations, and (3) underutilization of informative activations, particularly from Feed-Forward Networks (FFNs). To address these challenges, we introduce Low-Rank Clone (LRC), an efficient pre-training method that constructs SLMs aspiring to behavioral equivalence with strong teacher models. LRC trains a set of low-rank projection matrices that jointly enable soft pruning by compressing teacher weights, and activation clone by aligning student activations, including FFN signals, with those of the teacher. This unified design maximizes knowledge transfer while removing the need for explicit alignment modules. Extensive experiments with open-source teachers (e.g., Llama-3.2-3B-Instruct, Qwen2.5-3B/7B-Instruct) show that LRC matches or surpasses state-of-the-art models trained on trillions of tokens--while using only 20B tokens, achieving over 1,000x training efficiency. Our codes and model checkpoints are available at https://github.com/CURRENTF/LowRankClone and https://huggingface.co/collections/JitaiHao/low-rank-clone-lrc-6828389e96a93f1d4219dfaf.
URLs: https://github.com/CURRENTF/LowRankClone, https://huggingface.co/collections/JitaiHao/low-rank-clone-lrc-6828389e96a93f1d4219dfaf.
Authors: Jingyi Ren, Yekun Xu, Xiaolong Wang, Weitao Li, Weizhi Ma, Yang Liu
Abstract: Retrieval-Augmented Generation (RAG) delivers substantial value in knowledge-intensive applications. Many recent works use reinforcement learning (RL) to elicit strong reasoning in RAG generators. However, two key challenges remain unresolved: (1) Transparency: most prior methods do not explicitly indicate which references are actually used during the reasoning that leads to the final answer, limiting interpretability and visibility; (2) Stability: the KL divergence estimator used in existing RL-based approaches may cause gradient spikes, leading to unstable training. To address these challenges, we propose Adaptive-Rewarded Evidence Navigation Agent (ARENA), a transparent and robust RAG generator framework trained via RL with designed rewards. Based on our structured protocol, KL divergence stabilization, and adaptive reward calculation modules, ARENA enables the RAG generator to identify key evidence, perform structured reasoning, and generate answers with interpretable decision traces. Applied to Qwen2.5-7B-Instruct and Llama3.1-8B-Instruct, extensive experiments across multiple baselines show 10-30% accuracy improvements on three multi-hop QA datasets, comparable to advanced closed-source LLMs (e.g., OpenAI o1, DeepSeek R1). Further analyses show that ARENA generalizes well to unseen datasets and tasks. Our models and codes are publicly released.
Authors: Prithviraj Singh Shahani, Kaveh Eskandari Miandoab, Matthias Scheutz
Abstract: Safety guardrails in large language models (LLMs) are a critical component in preventing harmful outputs. Yet, their resilience under perturbation remains poorly understood. In this paper, we investigate the robustness of safety fine-tuning in LLMs by systematically injecting Gaussian noise into model activations. We show across multiple open-weight models that (1) Gaussian noise raises harmful-output rates (p < 0.001) by up to 27%, (2) that deeper safety fine-tuning affords no extra protection, and (3) that chain-of-thought reasoning remains largely intact. The findings reveal critical vulnerabilities in current safety alignment techniques and highlight the potential of reasoning-based and reinforcement learning approaches as promising direction for developing more robust AI safety systems. These results have important implications for real-world deployment of LLMs in safety-critical applications as these results imply that widely-deployed safety tuning methods can fail even without adversarial prompts.
Authors: Darpan Aswal, Siddharth D Jaiswal
Abstract: Recently released LLMs have strong multilingual \& multimodal capabilities. Model vulnerabilities are exposed using audits and red-teaming efforts. Existing efforts have focused primarily on the English language; thus, models continue to be susceptible to multilingual jailbreaking strategies, especially for multimodal contexts. In this study, we introduce a novel strategy that leverages code-mixing and phonetic perturbations to jailbreak LLMs for both text and image generation tasks. We also present an extension to a current jailbreak-template-based strategy and propose a novel template, showing higher effectiveness than baselines. Our work presents a method to effectively bypass safety filters in LLMs while maintaining interpretability by applying phonetic misspellings to sensitive words in code-mixed prompts. We achieve a 99\% Attack Success Rate for text generation and 78\% for image generation, with Attack Relevance Rate of 100\% for text generation and 96\% for image generation for the phonetically perturbed code-mixed prompts. Our interpretability experiments reveal that phonetic perturbations impact word tokenization, leading to jailbreak success. Our study motivates increasing the focus towards more generalizable safety alignment for multilingual multimodal models, especially in real-world settings wherein prompts can have misspelt words. \textit{\textbf{Warning: This paper contains examples of potentially harmful and offensive content.}}
Authors: Ryan Solgi, Kai Zhen, Rupak Vignesh Swaminathan, Nathan Susanj, Athanasios Mouchtaris, Siegfried Kunzmann, Zheng Zhang
Abstract: The efficient implementation of large language models (LLMs) is crucial for deployment on resource-constrained devices. Low-rank tensor compression techniques, such as tensor-train (TT) networks, have been widely studied for over-parameterized neural networks. However, their applications to compress pre-trained large language models (LLMs) for downstream tasks (post-training) remains challenging due to the high-rank nature of pre-trained LLMs and the lack of access to pretraining data. In this study, we investigate low-rank tensorized LLMs during fine-tuning and propose sparse augmented tensor networks (Saten) to enhance their performance. The proposed Saten framework enables full model compression. Experimental results demonstrate that Saten enhances both accuracy and compression efficiency in tensorized language models, achieving state-of-the-art performance.
Authors: Joonho Yang, Seunghyun Yoon, Hwan Chang, Byeongjeong Kim, Hwanhee Lee
Abstract: Large Language Models (LLMs) have significantly advanced text generation capabilities, including tasks like summarization, often producing coherent and fluent outputs. However, faithfulness to source material remains a significant challenge due to the generation of hallucinations. While extensive research focuses on detecting and reducing these inaccuracies, less attention has been paid to the positional distribution of hallucination within generated text, particularly in long outputs. In this work, we investigate where hallucinations occur in LLM-based long response generation, using long document summarization as a key case study. Focusing on the challenging setting of long context-aware long response generation, we find a consistent and concerning phenomenon: hallucinations tend to concentrate disproportionately in the latter parts of the generated long response. To understand this bias, we explore potential contributing factors related to the dynamics of attention and decoding over long sequences. Furthermore, we investigate methods to mitigate this positional hallucination, aiming to improve faithfulness specifically in the concluding segments of long outputs.
Authors: David Dinucu-Jianu, Jakub Macina, Nico Daheim, Ido Hakimi, Iryna Gurevych, Mrinmaya Sachan
Abstract: Large language models (LLMs) can transform education, but their optimization for direct question-answering often undermines effective pedagogy which requires strategically withholding answers. To mitigate this, we propose an online reinforcement learning (RL)-based alignment framework that can quickly adapt LLMs into effective tutors using simulated student-tutor interactions by emphasizing pedagogical quality and guided problem-solving over simply giving away answers. We use our method to train a 7B parameter tutor model without human annotations which reaches similar performance to larger proprietary models like LearnLM. We introduce a controllable reward weighting to balance pedagogical support and student solving accuracy, allowing us to trace the Pareto frontier between these two objectives. Our models better preserve reasoning capabilities than single-turn SFT baselines and can optionally enhance interpretability through thinking tags that expose the model's instructional planning.
Authors: Jinyang Wu, Chonghua Liao, Mingkuan Feng, Shuai Zhang, Zhengqi Wen, Haoran Luo, Ling Yang, Huazhe Xu, Jianhua Tao
Abstract: Reinforcement learning (RL) has emerged as an effective paradigm for enhancing model reasoning. However, existing RL methods like GRPO often rely on unstructured self-sampling to fit scalar rewards, often producing inefficient rollouts that fail to capture transferable problem-solving strategies. To address these limitations, we propose **TemplateRL**, a structured template-guided RL framework that augments policy optimization with explicit template guidance. Our approach first constructs a problem-solving template library via MCTS on a small seed set, then seamlessly integrates this high-level structured guidance into RL training. By guiding rollout generation to align with proven template structures, TemplateRL significantly improves high-quality trajectory hit rates while reducing ineffective exploration. This structure-guided design steers the policy toward validated strategic patterns, stabilizing training dynamics, and enhancing RL sampling efficiency. Notably, the explicit template library is interpretable, editable, and supports online updates-enabling continuous updates during both training and inference. Extensive experiments demonstrate that TemplateRL outperforms GRPO by 99% on AIME and 41% on AMC, with superior stability on weak models and remarkable cross-domain generalization, highlighting its potential for broader tasks.
Authors: Bohao Wu, Qingyun Wang, Yue Guo
Abstract: Personalizing jargon detection and explanation is essential for making technical documents accessible to readers with diverse disciplinary backgrounds. However, tailoring models to individual users typically requires substantial annotation efforts and computational resources due to user-specific finetuning. To address this, we present a systematic study of personalized jargon detection, focusing on methods that are both efficient and scalable for real-world deployment. We explore two personalization strategies: (1) lightweight finetuning using Low-Rank Adaptation (LoRA) on open-source models, and (2) personalized prompting, which tailors model behavior at inference time without retaining. To reflect realistic constraints, we also investigate semi-supervised approaches that combine limited annotated data with self-supervised learning from users' publications. Our personalized LoRA model outperforms GPT-4 with contextual prompting by 21.4% in F1 score and exceeds the best performing oracle baseline by 8.3%. Remarkably, our method achieves comparable performance using only 10% of the annotated training data, demonstrating its practicality for resource-constrained settings. Our study offers the first work to systematically explore efficient, low-resource personalization of jargon detection using open-source language models, offering a practical path toward scalable, user-adaptive NLP system.
Authors: Ruizhe Li, Chen Chen, Yuchen Hu, Yanjun Gao, Xi Wang, Emine Yilmaz
Abstract: Retrieval-Augmented Generation (RAG) leverages large language models (LLMs) combined with external contexts to enhance the accuracy and reliability of generated responses. However, reliably attributing generated content to specific context segments, context attribution, remains challenging due to the computationally intensive nature of current methods, which often require extensive fine-tuning or human annotation. In this work, we introduce a novel Jensen-Shannon Divergence driven method to Attribute Response to Context (ARC-JSD), enabling efficient and accurate identification of essential context sentences without additional fine-tuning, gradient-calculation or surrogate modelling. Evaluations on a wide range of RAG benchmarks, such as TyDi QA, Hotpot QA, and Musique, using instruction-tuned LLMs in different scales demonstrate superior accuracy and significant computational efficiency improvements compared to the previous surrogate-based method. Furthermore, our mechanistic analysis reveals specific attention heads and multilayer perceptron (MLP) layers responsible for context attribution, providing valuable insights into the internal workings of RAG models and how they affect RAG behaviours. Our code is available at https://github.com/ruizheliUOA/ARC_JSD.
Authors: Florentin Beck, William Rudman, Carsten Eickhoff
Abstract: Large Language Models (LLMs) present significant computational and memory challenges due to their extensive size, making pruning essential for their efficient deployment. Existing one-shot pruning methods often apply uniform sparsity constraints across layers or within each layer, resulting in suboptimal performance, especially at high sparsity ratios. This work introduces TRIM (Targeted Row-wise Iterative Metric-driven pruning), a novel approach that applies varying sparsity ratios to individual output dimensions (rows) within each layer. TRIM employs an iterative adjustment process guided by quality metrics to optimize dimension-wise sparsity allocation, focusing on reducing variance in quality retention across outputs to preserve critical information. TRIM can be seamlessly integrated with existing layer-wise pruning strategies. Our evaluations on perplexity and zero-shot tasks across diverse LLM families (Qwen2.5, LLaMA-2, and OPT) and sparsity levels demonstrate that TRIM achieves new state-of-the-art results and enhances stability. For instance, at 80% sparsity, TRIM reduces perplexity by 48% for Qwen2.5-14B and over 90% for OPT-13B compared to baseline methods. We conclude that fine-grained, dimension-wise sparsity adaptation is crucial for pushing the limits of extreme LLM compression. Code available at: https://github.com/flobk/TRIM
Authors: Md. Mithun Hossain, Md. Shakil Hossain, Sudipto Chaki, Md. Rajib Hossain
Abstract: Aspect-Based Sentiment Analysis (ABSA) is a fundamental task in natural language processing, offering fine-grained insights into opinions expressed in text. While existing research has largely focused on resource-rich languages like English which leveraging large annotated datasets, pre-trained models, and language-specific tools. These resources are often unavailable for low-resource languages such as Bengali. The ABSA task in Bengali remains poorly explored and is further complicated by its unique linguistic characteristics and a lack of annotated data, pre-trained models, and optimized hyperparameters. To address these challenges, this research propose CrosGrpsABS, a novel hybrid framework that leverages bidirectional cross-attention between syntactic and semantic graphs to enhance aspect-level sentiment classification. The CrosGrpsABS combines transformerbased contextual embeddings with graph convolutional networks, built upon rule-based syntactic dependency parsing and semantic similarity computations. By employing bidirectional crossattention, the model effectively fuses local syntactic structure with global semantic context, resulting in improved sentiment classification performance across both low- and high-resource settings. We evaluate CrosGrpsABS on four low-resource Bengali ABSA datasets and the high-resource English SemEval 2014 Task 4 dataset. The CrosGrpsABS consistently outperforms existing approaches, achieving notable improvements, including a 0.93% F1-score increase for the Restaurant domain and a 1.06% gain for the Laptop domain in the SemEval 2014 Task 4 benchmark.
Authors: Xuyang Liu, Zichen Wen, Shaobo Wang, Junjie Chen, Zhishan Tao, Yubo Wang, Tailai Chen, Xiangqi Jin, Chang Zou, Yiyu Wang, Chenfei Liao, Xu Zheng, Honggang Chen, Weijia Li, Xuming Hu, Conghui He, Linfeng Zhang
Abstract: The advancement of large language models (LLMs) and multi-modal LLMs (MLLMs) has historically relied on scaling model parameters. However, as hardware limits constrain further model growth, the primary computational bottleneck has shifted to the quadratic cost of self-attention over increasingly long sequences by ultra-long text contexts, high-resolution images, and extended videos. In this position paper, \textbf{we argue that the focus of research for efficient artificial intelligence (AI) is shifting from model-centric compression to data-centric compression}. We position data-centric compression as the emerging paradigm, which improves AI efficiency by directly compressing the volume of data processed during model training or inference. To formalize this shift, we establish a unified framework for existing efficiency strategies and demonstrate why it constitutes a crucial paradigm change for long-context AI. We then systematically review the landscape of data-centric compression methods, analyzing their benefits across diverse scenarios. Finally, we outline key challenges and promising future research directions. Our work aims to provide a novel perspective on AI efficiency, synthesize existing efforts, and catalyze innovation to address the challenges posed by ever-increasing context lengths.
Authors: Thang Nguyen, Peter Chin, Yu-Wing Tai
Abstract: We present MA-RAG, a Multi-Agent framework for Retrieval-Augmented Generation (RAG) that addresses the inherent ambiguities and reasoning challenges in complex information-seeking tasks. Unlike conventional RAG methods that rely on end-to-end fine-tuning or isolated component enhancements, MA-RAG orchestrates a collaborative set of specialized AI agents: Planner, Step Definer, Extractor, and QA Agents, each responsible for a distinct stage of the RAG pipeline. By decomposing tasks into subtasks such as query disambiguation, evidence extraction, and answer synthesis, and enabling agents to communicate intermediate reasoning via chain-of-thought prompting, MA-RAG progressively refines retrieval and synthesis while maintaining modular interpretability. Extensive experiments on multi-hop and ambiguous QA benchmarks, including NQ, HotpotQA, 2WikimQA, and TriviaQA, demonstrate that MA-RAG significantly outperforms standalone LLMs and existing RAG methods across all model scales. Notably, even a small LLaMA3-8B model equipped with MA-RAG surpasses larger standalone LLMs, while larger variants (LLaMA3-70B and GPT-4o-mini) set new state-of-the-art results on challenging multi-hop datasets. Ablation studies reveal that both the planner and extractor agents are critical for multi-hop reasoning, and that high-capacity models are especially important for the QA agent to synthesize answers effectively. Beyond general-domain QA, MA-RAG generalizes to specialized domains such as medical QA, achieving competitive performance against domain-specific models without any domain-specific fine-tuning. Our results highlight the effectiveness of collaborative, modular reasoning in retrieval-augmented systems: MA-RAG not only improves answer accuracy and robustness but also provides interpretable intermediate reasoning steps, establishing a new paradigm for efficient and reliable multi-agent RAG.
Authors: Siye Wu, Jian Xie, Yikai Zhang, Aili Chen, Kai Zhang, Yu Su, Yanghua Xiao
Abstract: While large reasoning models demonstrate strong performance on complex tasks, they lack the ability to adjust reasoning token usage based on task difficulty. This often leads to the "overthinking" problem -- excessive and unnecessary reasoning -- which, although potentially mitigated by human intervention to control the token budget, still fundamentally contradicts the goal of achieving fully autonomous AI. In this work, we propose Adaptive Reasoning Model (ARM), a reasoning model capable of adaptively selecting appropriate reasoning formats based on the task at hand. These formats include three efficient ones -- Direct Answer, Short CoT, and Code -- as well as a more elaborate format, Long CoT. To train ARM, we introduce Ada-GRPO, an adaptation of Group Relative Policy Optimization (GRPO), which addresses the format collapse issue in traditional GRPO. Ada-GRPO enables ARM to achieve high token efficiency, reducing tokens by an average of 30%, and up to 70%, while maintaining performance comparable to the model that relies solely on Long CoT. Furthermore, not only does it improve inference efficiency through reduced token generation, but it also brings a 2x speedup in training. In addition to the default Adaptive Mode, ARM supports two additional reasoning modes: 1) Instruction-Guided Mode, which allows users to explicitly specify the reasoning format via special tokens -- ideal when the appropriate format is known for a batch of tasks. 2) Consensus-Guided Mode, which aggregates the outputs of the three efficient formats and resorts to Long CoT in case of disagreement, prioritizing performance with higher token usage.
Authors: Yukun Zhang, Qi Dong
Abstract: We present Multi-Scale Manifold Alignment(MSMA), an information-geometric framework that decomposes LLM representations into local, intermediate, and global manifolds and learns cross-scale mappings that preserve geometry and information. Across GPT-2, BERT, RoBERTa, and T5, we observe consistent hierarchical patterns and find that MSMA improves alignment metrics under multiple estimators (e.g., relative KL reduction and MI gains with statistical significance across seeds). Controlled interventions at different scales yield distinct and architecture-dependent effects on lexical diversity, sentence structure, and discourse coherence. While our theoretical analysis relies on idealized assumptions, the empirical results suggest that multi-objective alignment offers a practical lens for analyzing cross-scale information flow and guiding representation-level control.
Authors: Yukun Zhang, Qi Dong
Abstract: This paper introduces the Dynamical Manifold Evolution Theory (DMET), a conceptual framework that models large language model (LLM) generation as a continuous trajectory evolving on a low-dimensional semantic manifold. The theory characterizes latent dynamics through three interpretable metrics-state continuity ($C$), attractor compactness ($Q$), and topological persistence ($P$)-which jointly capture the smoothness, stability, and structure of representation evolution. Empirical analyses across multiple Transformer architectures reveal consistent links between these latent dynamics and text quality: smoother trajectories correspond to greater fluency, and richer topological organization correlates with enhanced coherence. Different models exhibit distinct dynamical regimes, reflecting diverse strategies of semantic organization in latent space. Moreover, decoding parameters such as temperature and top-$p$ shape these trajectories in predictable ways, defining a balanced region that harmonizes fluency and creativity. As a phenomenological rather than first-principles framework, DMET provides a unified and testable perspective for interpreting, monitoring, and guiding LLM behavior, offering new insights into the interplay between internal representation dynamics and external text generation quality.
Authors: Keenan Samway, Max Kleiman-Weiner, David Guzman Piedrahita, Rada Mihalcea, Bernhard Sch\"olkopf, Zhijing Jin
Abstract: As AI systems increasingly navigate applications in healthcare, law, and governance, understanding how they handle ethically complex scenarios becomes critical. Previous work has mainly examined the moral judgments in large language models (LLMs), rather than their underlying moral reasoning process. In contrast, we focus on a large-scale analysis of the moral reasoning traces provided by LLMs. Furthermore, unlike prior work that attempted to draw inferences from only a handful of moral dilemmas, our study leverages over 600 distinct trolley problems as probes for revealing the reasoning patterns that emerge within different LLMs. We introduce and test a taxonomy of moral rationales to systematically classify reasoning traces according to two main normative ethical theories: consequentialism and deontology. Our analysis reveals that LLM chains-of-thought tend to favor deontological principles based on moral obligations, while post-hoc explanations shift notably toward consequentialist rationales that emphasize utility. Our framework provides a foundation for understanding how LLMs process and articulate ethical considerations, an important step toward safe and interpretable deployment of LLMs in high-stakes decision-making environments. Our code is available at https://github.com/keenansamway/moral-lens .
Authors: Hyeonbin Hwang, Byeongguk Jeon, Seungone Kim, Jiyeon Kim, Hoyeon Chang, Sohee Yang, Seungpil Won, Dohaeng Lee, Youbin Ahn, Minjoon Seo
Abstract: Autoregressive language models (LMs) generate one token at a time, yet human reasoning operates over higher-level abstractions - sentences, propositions, and concepts. This contrast raises a central question- Can LMs likewise learn to reason over structured semantic units rather than raw token sequences? In this work, we investigate whether pretrained LMs can be lifted into such abstract reasoning spaces by building on their learned representations. We present a framework that adapts a pretrained token-level LM to operate in sentence space by autoregressively predicting continuous embeddings of next sentences. We explore two embedding paradigms inspired by classical representation learning: 1) semantic embeddings, learned via autoencoding to preserve surface meaning; and 2) contextual embeddings, trained via next-sentence prediction to encode anticipatory structure. We evaluate both under two inference regimes: Discretized, which decodes each predicted embedding into text before re-encoding; and Continuous, which reasons entirely in embedding space for improved efficiency. Across four domains - mathematics, logic, commonsense, and planning - contextual embeddings under continuous inference show competitive performance with Chain-of-Thought (CoT) while reducing inference-time FLOPs on average by half. We also present early signs of scalability and modular adaptation. Finally, to visualize latent trajectories, we introduce SentenceLens, a diagnostic tool that decodes intermediate model states into interpretable sentences. Together, our results indicate that pretrained LMs can effectively transition to abstract, structured reasoning within latent embedding spaces.
Authors: Zicheng Xu, Guanchu Wang, Guangyao Zheng, Yu-Neng Chuang, Alexander Szalay, Xia Hu, Vladimir Braverman
Abstract: Although Large Language Models (LLMs) perform well in general fields, they exhibit a confidence distortion problem on multi-choice question-answering (MCQA), particularly as the number of answer choices increases. Specifically, on MCQA with many choices, LLMs suffer from under-confidence in correct predictions and over-confidence in incorrect ones, leading to a substantially degraded performance. To solve this problem, we propose Self-ensemble in this work. Our method splits the choices into several groups and ensembles LLM predictions across these groups to reach a final decision. The advantage of Self-ensemble is its plug-and-play nature, where it can be integrated into existing LLM architecture based on a designed attention mask and positional encoding, without requiring labeled datasets for parameter tuning. Experimental results on three LLMs and datasets demonstrate that Self-ensemble comprehensively addresses the confidence distortion problem of LLMs, outperforming standard inference as well as baseline methods.
Authors: Seokil Ham, Yubin Choi, Yujin Yang, Seungju Cho, Younghun Kim, Changick Kim
Abstract: Recently, major AI providers such as Google and OpenAI have introduced Finetuning-as-a-Service (FaaS), which allows users to customize Large Language Models (LLMs) using their own data. However, this service is vulnerable to safety degradation when user data includes harmful prompts, a threat known as harmful finetuning attacks. Prior works attempt to mitigate this issue by first constructing safety-aligned model and then finetuning the model on user data. However, we observe that the safety-aligned weights provide weak initialization for downstream task learning, leading to suboptimal safety-alignment and downstream task performance. To address this, we propose a Refusal-Teacher (Ref-Teacher)-guided finetuning framework. Instead of finetuning a safety-aligned model on user data, our approach directly finetunes the base model under the guidance of a safety-aligned Ref-Teacher, which filters harmful prompts from user data and distills safety-alignment knowledge into the base model. Extensive experiments demonstrate that our Ref-Teacher-guided finetuning strategy effectively minimizes harmful outputs and enhances finetuning accuracy for user-specific tasks, offering a practical solution for secure and reliable deployment of LLMs in FaaS.
Authors: Nadezhda Chirkova, Tunde Oluwaseyi Ajayi, Seth Aycock, Zain Muhammad Mujahid, Vladana Perli\'c, Ekaterina Borisova, Markarit Vartampetian
Abstract: Prompting large language models (LLMs) to evaluate generated text, known as LLM-as-a-judge, has become a standard evaluation approach in natural language generation (NLG), but is primarily used as a quantitative tool, i.e. with numerical scores as main outputs. In this work, we propose LLM-as-a-qualitative-judge, an LLM-based evaluation approach with the main output being a structured report of common issue types in the NLG system outputs. Our approach is targeted at providing developers with meaningful insights on what improvements can be done to a given NLG system and consists of two main steps, namely open-ended per-instance issue analysis and clustering of the discovered issues using an intuitive cumulative algorithm. We also introduce a strategy for evaluating the proposed approach, coupled with ~300 annotations of issues in instances from 12 NLG datasets. Our results show that instance-specific issues output by LLM-as-a-qualitative-judge match those annotated by humans in 2/3 cases, and that LLM-as-a-qualitative-judge is capable of producing error type reports resembling the reports composed by human annotators. We also demonstrate in a case study how the use of LLM-as-a-qualitative-judge can substantially improve NLG systems performance. Our code and data are publicly available at https://github.com/tunde-ajayi/llm-as-a-qualitative-judge.
URLs: https://github.com/tunde-ajayi/llm-as-a-qualitative-judge.
Authors: Greta Damo, Elena Cabrio, Serena Villata
Abstract: Counter-speech (CS) is a key strategy for mitigating online Hate Speech (HS), yet defining the criteria to assess its effectiveness remains an open challenge. We propose a novel computational framework for CS effectiveness classification, grounded in linguistics, communication and argumentation concepts. Our framework defines six core dimensions - Clarity, Evidence, Emotional Appeal, Rebuttal, Audience Adaptation, and Fairness - which we use to annotate 4,214 CS instances from two benchmark datasets, resulting in a novel linguistic resource released to the community. In addition, we propose two classification strategies, multi-task and dependency-based, achieving strong results (0.94 and 0.96 average F1 respectively on both expert- and user-written CS), outperforming standard baselines, and revealing strong interdependence among dimensions.
Authors: Joanito Agili Lopo, Muhammad Ravi Shulthan Habibi, Tack Hwa Wong, Muhammad Ilham Ghozali, Fajri Koto, Genta Indra Winata, Peerat Limkonchotiwat, Alham Fikri Aji, Samuel Cahyawijaya
Abstract: Large Language Models (LLMs) have demonstrated remarkable generalization capabilities across tasks and languages, revolutionizing natural language processing. This paper investigates the naturally emerging representation alignment in LLMs, particularly in the middle layers, and its implications for disentangling language-specific and language-agnostic information. We empirically confirm the existence of this alignment, analyze its behavior in comparison to explicitly designed alignment models, and demonstrate its potential for language-specific manipulation without semantic degradation. Building on these findings, we propose Inference-Time Language Control (ITLC), a novel method that leverages latent injection to enable precise cross-lingual language control and mitigate language confusion in LLMs. Our experiments highlight ITLC's strong cross-lingual control capabilities while preserving semantic integrity in target languages. Furthermore, we demonstrate its effectiveness in alleviating the cross-lingual language confusion problem, which persists even in current large-scale LLMs, leading to inconsistent language generation. This work advances our understanding of representation alignment in LLMs and introduces a practical solution for enhancing their monolingual and cross-lingual performance.
Authors: Xueqing Peng, Lingfei Qian, Yan Wang, Ruoyu Xiang, Yueru He, Yang Ren, Mingyang Jiang, Vincent Jim Zhang, Yuqing Guo, Jeff Zhao, Huan He, Yi Han, Yun Feng, Yuechen Jiang, Yupeng Cao, Haohang Li, Yangyang Yu, Xiaoyu Wang, Penglei Gao, Shengyuan Lin, Keyi Wang, Shanshan Yang, Yilun Zhao, Zhiwei Liu, Peng Lu, Jerry Huang, Suyuchen Wang, Triantafillos Papadopoulos, Polydoros Giannouris, Efstathia Soufleri, Nuo Chen, Zhiyang Deng, Heming Fu, Yijia Zhao, Mingquan Lin, Meikang Qiu, Kaleb E Smith, Arman Cohan, Xiao-Yang Liu, Jimin Huang, Guojun Xiong, Alejandro Lopez-Lira, Xi Chen, Junichi Tsujii, Jian-Yun Nie, Sophia Ananiadou, Qianqian Xie
Abstract: Real-world financial analysis involves information across multiple languages and modalities, from reports and news to scanned filings and meeting recordings. Yet most existing evaluations of LLMs in finance remain text-only, monolingual, and largely saturated by current models. To bridge these gaps, we present MultiFinBen, the first expert-annotated multilingual (five languages) and multimodal (text, vision, audio) benchmark for evaluating LLMs in realistic financial contexts. MultiFinBen introduces two new task families: multilingual financial reasoning, which tests cross-lingual evidence integration from filings and news, and financial OCR, which extracts structured text from scanned documents containing tables and charts. Rather than aggregating all available datasets, we apply a structured, difficulty-aware selection based on advanced model performance, ensuring balanced challenge and removing redundant tasks. Evaluating 21 leading LLMs shows that even frontier multimodal models like GPT-4o achieve only 46.01% overall, stronger on vision and audio but dropping sharply in multilingual settings. These findings expose persistent limitations in multilingual, multimodal, and expert-level financial reasoning. All datasets, evaluation scripts, and leaderboards are publicly released.
Authors: Xiang Cheng, Chengyan Pan, Minjun Zhao, Deyang Li, Fangchao Liu, Xinyu Zhang, Xiao Zhang, Yong Liu
Abstract: In-Context Learning (ICL) is an essential emergent ability of Large Language Models (LLMs), and recent studies introduce Chain-of-Thought (CoT) to exemplars of ICL to enhance the reasoning capability, especially in mathematics tasks. However, given the continuous advancement of model capabilities, it remains unclear whether CoT exemplars still benefit recent, stronger models in such tasks. Through systematic experiments, we find that for recent strong models such as the Qwen2.5 series, adding traditional CoT exemplars does not improve reasoning performance compared to Zero-Shot CoT. Instead, their primary function is to align the output format with human expectations. We further investigate the effectiveness of enhanced CoT exemplars, constructed using answers from advanced models such as \texttt{Qwen2.5-Max} and \texttt{DeepSeek-R1}. Experimental results indicate that these enhanced exemplars still fail to improve the model's reasoning performance. Further analysis reveals that models tend to ignore the exemplars and focus primarily on the instructions, leading to no observable gain in reasoning ability. Overall, our findings highlight the limitations of the current ICL+CoT framework in mathematical reasoning, calling for a re-examination of the ICL paradigm and the definition of exemplars.
Authors: Jackson Trager, Francielle Vargas, Diego Alves, Matteo Guida, Mikel K. Ngueajio, Ameeta Agrawal, Yalda Daryani, Farzan Karimi-Malekabadi, Flor Miriam Plaza-del-Arco
Abstract: Ensuring the moral reasoning capabilities of Large Language Models (LLMs) is a growing concern as these systems are used in socially sensitive tasks. Nevertheless, current evaluation benchmarks present two major shortcomings: a lack of annotations that justify moral classifications, which limits transparency and interpretability; and a predominant focus on English, which constrains the assessment of moral reasoning across diverse cultural settings. In this paper, we introduce MFTCXplain, a multilingual benchmark dataset for evaluating the moral reasoning of LLMs via multi-hop hate speech explanation using the Moral Foundations Theory. MFTCXplain comprises 3,000 tweets across Portuguese, Italian, Persian, and English, annotated with binary hate speech labels, moral categories, and text span-level rationales. Our results show a misalignment between LLM outputs and human annotations in moral reasoning tasks. While LLMs perform well in hate speech detection (F1 up to 0.836), their ability to predict moral sentiments is notably weak (F1 < 0.35). Furthermore, rationale alignment remains limited mainly in underrepresented languages. Our findings show the limited capacity of current LLMs to internalize and reflect human moral reasoning
Authors: Jiaxi Zhuang, Kangning Li, Jue Hou, Mingjun Xu, Zhifeng Gao, Hengxing Cai
Abstract: Extracting molecular structure-activity relationships (SARs) from scientific literature and patents is essential for drug discovery and materials research. However, this task remains challenging due to heterogeneous document formats and limitations of existing methods. Specifically, rule-based approaches relying on rigid templates fail to generalize across diverse document layouts, while general-purpose multimodal large language models (MLLMs) lack sufficient accuracy and reliability for specialized tasks, such as layout detection and optical chemical structure recognition (OCSR). To address these challenges, we introduce DocSAR-200, a rigorously annotated benchmark of 200 scientific documents designed specifically for evaluating SAR extraction methods. Additionally, we propose Doc2SAR, a novel synergistic framework that integrates domain-specific tools with MLLMs enhanced via supervised fine-tuning (SFT). Extensive experiments demonstrate that Doc2SAR achieves state-of-the-art performance across various document types, significantly outperforming leading end-to-end baselines. Specifically, Doc2SAR attains an overall Table Recall of 80.78% on DocSAR-200, exceeding end2end GPT-4o by 51.48%. Furthermore, Doc2SAR demonstrates practical usability through efficient inference and is accompanied by a web app.
Authors: Sana Kang, Myeongseok Gwon, Su Young Kwon, Jaewook Lee, Andrew Lan, Bhiksha Raj, Rita Singh
Abstract: Vocabulary acquisition poses a significant challenge for second-language (L2) learners, especially when learning typologically distant languages such as English and Korean, where phonological and structural mismatches complicate vocabulary learning. Recently, large language models (LLMs) have been used to generate keyword mnemonics by leveraging similar keywords from a learner's first language (L1) to aid in acquiring L2 vocabulary. However, most methods still rely on direct IPA-based phonetic matching or employ LLMs without phonological guidance. In this paper, we present PhoniTale, a novel cross-lingual mnemonic generation system that performs IPA-based phonological adaptation and syllable-aware alignment to retrieve L1 keyword sequence and uses LLMs to generate verbal cues. We evaluate PhoniTale through automated metrics and a short-term recall test with human participants, comparing its output to human-written and prior automated mnemonics. Our findings show that PhoniTale consistently outperforms previous automated approaches and achieves quality comparable to human-written mnemonics.
Authors: Ondrej Bohdal, Mete Ozay, Jijoong Moon, Kyeng-Hun Lee, Hyeonmok Ko, Umberto Michieli
Abstract: Adapter parameters provide a mechanism to modify the behavior of machine learning models and have gained significant popularity in the context of large language models (LLMs) and generative AI. These parameters can be merged to support multiple tasks via a process known as task merging. However, prior work on merging in LLMs, particularly in natural language processing, has been limited to scenarios where each test example addresses only a single task. In this paper, we focus on on-device settings and study the problem of text-based compositional multi-tasking, where each test example involves the simultaneous execution of multiple tasks. For instance, generating a translated summary of a long text requires solving both translation and summarization tasks concurrently. To facilitate research in this setting, we propose a benchmark comprising four practically relevant compositional tasks. We also present an efficient method (Learnable Calibration) tailored for on-device applications, where computational resources are limited, emphasizing the need for solutions that are both resource-efficient and high-performing. Our contributions lay the groundwork for advancing the capabilities of LLMs in real-world multi-tasking scenarios, expanding their applicability to complex, resource-constrained use cases.
Authors: Chuanhao Yan, Fengdi Che, Xuhan Huang, Xu Xu, Xin Li, Yizhi Li, Xingwei Qu, Jingzhe Shi, Chenghua Lin, Yaodong Yang, Binhang Yuan, Hang Zhao, Yu Qiao, Bowen Zhou, Jie Fu
Abstract: Existing informal language-based (e.g., human language) Large Language Models (LLMs) trained with Reinforcement Learning (RL) face a significant challenge: their verification processes, which provide crucial training signals, are neither reliable nor scalable. In fact, the prevalent large proprietary models could hardly generate verifiable programs. A promising yet largely uncharted alternative is formal language-based reasoning. Grounding LLMs in rigorous formal systems where generative models operate in formal language spaces (e.g., Dafny) enables the automatic and mathematically provable verification of their reasoning processes and outcomes. This capability is pivotal for achieving large-scale, reliable formal software verification. It is a common practice to employ human-annotated chain-of-thought and other human priors to induce the reasoning and coding capabilities of LLMs. Unfortunately, it becomes unacceptably all-consuming to provide such priors for supervising complex programming tasks. In this work, we systematically explore ways to reduce human priors with the formal language, Dafny, as the main environment for our pilot study. Our pipeline mainly relies on introducing an automatic and scalable data curation pipeline, and careful RL designs integrated with feedback from the formal language verifier. We introduce DafnyComp, a benchmark of compositional formal programs with auto-formalized specifications for specification reasoning. Our supervised fine-tuning (SFT) stage enables even small models (e.g., 0.5B) to generate syntactically valid and verifiable Dafny code, surpassing proprietary models. RL with regularization further improves performance, achieving stronger generalization to out-of-domain tasks and outperforming all strong baselines on the challenging DafnyComp benchmark.
Authors: Zhiyuan He, Yike Zhang, Chengruidong Zhang, Huiqiang Jiang, Yuqing Yang, Lili Qiu
Abstract: Large Language Models (LLMs) incur quadratic attention complexity with input length, creating a major time bottleneck in the prefilling stage. Existing acceleration methods largely exploit attention score sparsity by estimating blocks with high attention scores and applying dynamic sparse attention. In this work, we identify another untapped form of sparsity in the prefilling stage, namely decoding-time contribution sparsity, where many attention blocks exhibit nontrivial attention scores during prefilling yet contribute negligibly to subsequent decoding, as indicated by gradient-based analysis. Building on this observation, we propose TriangleMix, a training-free static attention pattern that uses dense attention in a subset of layers and switches to Triangle attention in the others. Extensive experiments show that TriangleMix preserves nearly lossless performance relative to dense attention while substantially reducing attention overhead in Triangle layers. For 128K inputs, Triangle attention achieves a 15.3x speedup in attention computation, significantly exceeding the acceleration of typical dynamic sparse methods (1.9x to 3.4x). Furthermore, TriangleMix can be seamlessly combined with dynamic sparsity approaches, delivering an additional 6% to 19% reduction in TTFT over using dynamic sparsity alone.
Authors: Yang Wang, Chenghao Xiao, Yizhi Li, Stuart E. Middleton, Noura Al Moubayed, Chenghua Lin
Abstract: Pre-trained language models (PLMs) have driven substantial progress in natural language processing but remain vulnerable to adversarial attacks, raising concerns about their robustness in real-world applications. Previous studies have sought to mitigate the impact of adversarial attacks by introducing adversarial perturbations into the training process, either implicitly or explicitly. While both strategies enhance robustness, they often incur high computational costs. In this work, we propose a simple yet effective add-on module that enhances the adversarial robustness of PLMs by removing instance-level principal components, without relying on conventional adversarial defences or perturbing the original training data. Our approach transforms the embedding space to approximate Gaussian properties, thereby reducing its susceptibility to adversarial perturbations while preserving semantic relationships. This transformation aligns embedding distributions in a way that minimises the impact of adversarial noise on decision boundaries, enhancing robustness without requiring adversarial examples or costly training-time augmentation. Evaluations on eight benchmark datasets show that our approach improves adversarial robustness while maintaining comparable before-attack accuracy to baselines, achieving a balanced trade-off between robustness and generalisation.
Authors: Sebastian Wind, Jeta Sopa, Daniel Truhn, Mahshad Lotfinia, Tri-Thien Nguyen, Keno Bressem, Lisa Adams, Mirabela Rusu, Harald K\"ostler, Gerhard Wellein, Andreas Maier, Soroosh Tayebi Arasteh
Abstract: Clinical decision-making in radiology increasingly benefits from artificial intelligence (AI), particularly through large language models (LLMs). However, traditional retrieval-augmented generation (RAG) systems for radiology question answering (QA) typically rely on single-step retrieval, limiting their ability to handle complex clinical reasoning tasks. Here we propose radiology Retrieval and Reasoning (RaR), a multi-step retrieval and reasoning framework designed to improve diagnostic accuracy, factual consistency, and clinical reliability of LLMs in radiology question answering. We evaluated 25 LLMs spanning diverse architectures, parameter scales (0.5B to >670B), and training paradigms (general-purpose, reasoning-optimized, clinically fine-tuned), using 104 expert-curated radiology questions from previously established RSNA-RadioQA and ExtendedQA datasets. To assess generalizability, we additionally tested on an unseen internal dataset of 65 real-world radiology board examination questions. RaR significantly improved mean diagnostic accuracy over zero-shot prompting and conventional online RAG. The greatest gains occurred in small-scale models, while very large models (>200B parameters) demonstrated minimal changes (<2% improvement). Additionally, RaR retrieval reduced hallucinations (mean 9.4%) and retrieved clinically relevant context in 46% of cases, substantially aiding factual grounding. Even clinically fine-tuned models showed gains from RaR (e.g., MedGemma-27B), indicating that retrieval remains beneficial despite embedded domain knowledge. These results highlight the potential of RaR to enhance factuality and diagnostic accuracy in radiology QA, warranting future studies to validate their clinical utility. All datasets, code, and the full RaR framework are publicly available to support open research and clinical translation.
Authors: Yinuo Xu, Veronica Derricks, Allison Earl, David Jurgens
Abstract: We present an approach to modeling annotator disagreement in subjective NLP tasks through both architectural and data-centric innovations. Our model, DEM-MoE (Demographic-Aware Mixture of Experts), routes inputs to expert subnetworks based on annotator demographics, enabling it to better represent structured, group-level variation compared to prior models. DEM-MoE consistently performs competitively across demographic groups, and shows especially strong results on datasets with high annotator disagreement. To address sparse demographic coverage, we test whether LLM-generated synthetic annotations via zero-shot persona prompting can be used for data imputation. We show these synthetic judgments align moderately well with human annotations on our data and offer a scalable way to potentially enrich training data. We then propose and evaluate approaches for blending real and synthetic data using strategies tailored to dataset structure. We find that the optimal strategies depend on dataset structure. Together, these contributions improve the representation of diverse perspectives.
Authors: Haotian Wu, Bo Xu, Yao Shu, Menglin Yang, Chengwei Qin
Abstract: Reasoning large language models (RLLMs) have recently demonstrated remarkable capabilities through structured and multi-step reasoning. While prior research has primarily focused on improving their training and inference strategies, their potential for in-context learning (ICL) remains largely underexplored. To fill this gap, we propose Thinking with Nothinking Calibration (JointThinking), a new ICL paradigm that prompts the model to generate two answers in parallel: one in Thinking mode and the other in Nothinking mode. A second round of Thinking is triggered only when the two initial responses are inconsistent, using a single prompt with two different answers. Extensive experiments across multiple reasoning benchmarks demonstrate that JointThinking significantly outperforms few-shot chain-of-thought (CoT), thinking twice and majority voting. Moreover, it achieves comparable in-distribution performance to training-based SOTA reasoning method, while substantially outperforming on out-of-distribution tasks. We further conduct a systematic analysis of the calibration mechanism, showing the importance of structural thinking diversity and the benefits of consistency check. Additionally, we observe that the performance gap between actual and ideal reasoning narrows as model size increases in the second thinking, indicating the strong scalability of our approach. Finally, we discuss current limitations and outline promising directions for future ICL research in RLLMs.
Authors: Khondoker Ittehadul Islam, Gabriele Sarti
Abstract: Language models have demonstrated remarkable performance on complex multi-step reasoning tasks. However, their evaluation has been predominantly confined to high-resource languages such as English. In this paper, we introduce a manually translated Bangla multi-step reasoning dataset derived from the English Reveal dataset, featuring both binary and non-binary question types. We conduct a controlled evaluation of English-centric and Bangla-centric multilingual small language models on the original dataset and our translated version to compare their ability to exploit relevant reasoning steps to produce correct answers. Our results show that, in comparable settings, reasoning context is beneficial for more challenging non-binary questions, but models struggle to employ relevant Bangla reasoning steps effectively. We conclude by exploring how reasoning steps contribute to models' predictions, highlighting different trends across models and languages.
Authors: Chiyu Zhang, Lu Zhou, Xiaogang Xu, Jiafei Wu, Liming Fang, Zhe Liu
Abstract: Jailbreaking commercial black-box models is one of the most challenging and serious security threats today. Existing attacks achieve certain success on non-reasoning models but perform limitedly on the latest reasoning models. We discover that carefully crafted developer messages can markedly boost jailbreak effectiveness. Building on this, we propose two developer-role-based attacks: D-Attack, which enhances contextual simulation, and DH-CoT, which strengthens attacks with deceptive chain-of-thought. In experiments, we further diccover that current red-teaming datasets often contain samples unsuited for measuring attack gains: prompts that fail to trigger defenses, prompts where malicious content is not the sole valid output, and benign prompts. Such data hinders accurate measurement of the true improvement brought by an attack method. To address this, we introduce MDH, a Malicious content Detection approach combining LLM-based screening with Human verification to balance accuracy and cost, with which we clean data and build the RTA dataset series. Experiments demonstrate that MDH reliably filters low-quality samples and that developer messages significantly improve jailbreak attack success. Codes, datasets, and other results will be released in https://github.com/AlienZhang1996/DH-CoT.
Authors: Xiangqi Jin, Yuxuan Wang, Yifeng Gao, Zichen Wen, Biqing Qi, Dongrui Liu, Linfeng Zhang
Abstract: Despite large language models (LLMs) have achieved remarkable success, their prefix-only prompting paradigm and sequential generation process offer limited flexibility for bidirectional information. Diffusion large language models (dLLMs) present new opportunities through their bidirectional attention mechanisms and iterative refinement processes, enabling more flexible in-place prompting strategies. We introduce ICE (In-Place Chain-of-Thought Prompting with Early Exit), a novel framework that transforms prefix-only prompting into in-place prompting specifically designed for dLLMs. ICE integrates in-place prompts directly within masked token positions during iterative refinement and employs a confidence-aware early exit mechanism to significantly reduce computational overhead. Extensive experiments demonstrate ICE's effectiveness, achieving up to 17.29% accuracy improvement with 4.12$\times$ speedup on GSM8K, and up to 276.67$\times$ acceleration on MMLU while maintaining competitive performance.
Authors: Yi Wang, Haoran Luo, Lu Meng, Ziyu Jia, Xinliang Zhou, Qingsong Wen
Abstract: With the widespread application of electroencephalography (EEG) in neuroscience and clinical practice, efficiently retrieving and semantically interpreting large-scale, multi-source, heterogeneous EEG data has become a pressing challenge. We propose EEG-MedRAG, a three-layer hypergraph-based retrieval-augmented generation framework that unifies EEG domain knowledge, individual patient cases, and a large-scale repository into a traversable n-ary relational hypergraph, enabling joint semantic-temporal retrieval and causal-chain diagnostic generation. Concurrently, we introduce the first cross-disease, cross-role EEG clinical QA benchmark, spanning seven disorders and five authentic clinical perspectives. This benchmark allows systematic evaluation of disease-agnostic generalization and role-aware contextual understanding. Experiments show that EEG-MedRAG significantly outperforms TimeRAG and HyperGraphRAG in answer accuracy and retrieval, highlighting its strong potential for real-world clinical decision support. Our data and code are publicly available at https://github.com/yi9206413-boop/EEG-MedRAG.
Authors: Guy Mor-Lan, Naama Rivlin-Angert, Yael R. Kaplan, Tamir Sheafer, Shaul R. Shenhav
Abstract: Political language is deeply intertwined with social identities. While social identities are often shaped by specific cultural contexts and expressed through particular uses of language, existing datasets for group and identity detection are predominantly English-centric, single-label and focus on coarse identity categories. We introduce HebID, the first multilabel Hebrew corpus for social identity detection: 5,536 sentences from Israeli politicians' Facebook posts (Dec 2018-Apr 2021), manually annotated for twelve nuanced social identities (e.g. Rightist, Ultra-Orthodox, Socially-oriented) grounded by survey data. We benchmark multilabel and single-label encoders alongside 2B-9B-parameter generative LLMs, finding that Hebrew-tuned LLMs provide the best results (macro-$F_1$ = 0.74). We apply our classifier to politicians' Facebook posts and parliamentary speeches, evaluating differences in popularity, temporal trends, clustering patterns, and gender-related variations in identity expression. We utilize identity choices from a national public survey, enabling a comparison between identities portrayed in elite discourse and the public's identity priorities. HebID provides a comprehensive foundation for studying social identities in Hebrew and can serve as a model for similar research in other non-English political contexts.
Authors: Naama Rivlin-Angert, Guy Mor-Lan
Abstract: We present the first large-scale computational study of political delegitimization discourse (PDD), defined as symbolic attacks on the normative validity of political entities. We curate and manually annotate a novel Hebrew-language corpus of 10,410 sentences drawn from Knesset speeches (1993-2023), Facebook posts (2018-2021), and leading news outlets, of which 1,812 instances (17.4\%) exhibit PDD and 642 carry additional annotations for intensity, incivility, target type, and affective framing. We introduce a two-stage classification pipeline combining finetuned encoder models and decoder LLMs. Our best model (DictaLM 2.0) attains an F$_1$ of 0.74 for binary PDD detection and a macro-F$_1$ of 0.67 for classification of delegitimization characteristics. Applying this classifier to longitudinal and cross-platform data, we see a marked rise in PDD over three decades, higher prevalence on social media versus parliamentary debate, greater use by male than female politicians, and stronger tendencies among right-leaning actors - with pronounced spikes during election campaigns and major political events. Our findings demonstrate the feasibility and value of automated PDD analysis for understanding democratic discourse.
Authors: Fei Lei, Yibo Yang, Wenxiu Sun, Dahua Lin
Abstract: Large Language Models (LLMs) are evolving from text generators into reasoning agents. This transition makes their ability to use external tools a critical capability. However, evaluating this skill presents a significant challenge. Existing benchmarks are often limited by their reliance on synthetic tools and severely constrained action spaces. To address these limitations, we introduce MCPVerse, an expansive, real-world benchmark for evaluating agentic tool use. MCPVerse integrates more than 550 real-world, executable tools to create an unprecedented action space exceeding 140k tokens, and employs outcome-based evaluation with real-time ground truth for time-sensitive tasks. We benchmarked the state-of-the-art LLMs across three modes (Oracle, Standard, and Max-Scale), revealing that while most models suffer performance degradation when confronted with larger tool sets, the agentic models, such as Claude-4-Sonnet, can effectively leverage expanded exploration spaces to improve accuracy. This finding not only exposes the limitations of state-of-the-art models in complex, real-world scenarios but also establishes MCPVerse as a critical benchmark for measuring and advancing agentic tool use capabilities.
Authors: Eliran Shem-Tov, Ella Rabinovich
Abstract: Music serves as a powerful reflection of individual identity, often aligning with deeper psychological traits. Prior research has established correlations between musical preferences and personality, while separate studies have demonstrated that personality is detectable through linguistic analysis. Our study bridges these two research domains by investigating whether individuals' musical preferences leave traces in their spontaneous language through the lens of the Big Five personality traits (Openness, Conscientiousness, Extroversion, Agreeableness, and Neuroticism). Using a carefully curated dataset of over 500,000 text samples from nearly 5,000 authors with reliably identified musical preferences, we build advanced models to assess personality characteristics. Our results reveal significant personality differences across fans of five musical genres. We release resources for future research at the intersection of computational linguistics, music psychology and personality analysis.
Authors: Xinglong Yang, Quan Feng, Zhongying Pan, Xiang Chen, Yu Tian, Wentong Li, Shuofei Qiao, Yuxia Geng, Xingyu Zhao, Sheng-Jun Huang
Abstract: The effectiveness of Multimodal Chain-of-Thought (MCoT) prompting is often limited by the use of randomly or manually selected examples. These examples fail to account for both model-specific knowledge distributions and the intrinsic complexity of the tasks, resulting in suboptimal and unstable model performance. To address this, we propose a novel framework inspired by the pedagogical principle of "tailored teaching with balanced difficulty". We reframe prompt selection as a prompt curriculum design problem: constructing a well ordered set of training examples that align with the model's current capabilities. Our approach integrates two complementary signals: (1) model-perceived difficulty, quantified through prediction disagreement in an active learning setup, capturing what the model itself finds challenging; and (2) intrinsic sample complexity, which measures the inherent difficulty of each question-image pair independently of any model. By jointly analyzing these signals, we develop a difficulty-balanced sampling strategy that ensures the selected prompt examples are diverse across both dimensions. Extensive experiments conducted on five challenging benchmarks and multiple popular Multimodal Large Language Models (MLLMs) demonstrate that our method yields substantial and consistent improvements and greatly reduces performance discrepancies caused by random sampling, providing a principled and robust approach for enhancing multimodal reasoning.
Authors: Byeongjeong Kim, Jeonghyun Park, Joonho Yang, Hwanhee Lee
Abstract: Long-context question answering over narrative tasks is challenging because correct answers often hinge on reconstructing a coherent timeline of events while preserving contextual f low in a limited context window. Retrievalaugmented generation (RAG) methods aim to address this challenge by selectively retrieving only necessary document segments. However, narrative texts possess unique characteristics that limit the effectiveness of these existing approaches. Specifically, understanding narrative texts requires more than isolated segments, as the broader context and sequential relationships between segments are crucial for comprehension. To address these limitations, we propose ChronoRAG, a novel RAG framework specialized for narrative texts. This approach focuses on two essential aspects: refining dispersed document information into coherent and structured passages and preserving narrative flow by explicitly capturing and maintaining the temporal order among retrieved passages. We empirically demonstrate the effectiveness of ChronoRAG through experiments on the NarrativeQA and GutenQAdataset, showing substantial improvements in tasks requiring both factual identification and comprehension of complex sequential relationships, underscoring that reasoning over temporal order is crucial in resolving narrative QA.
Authors: Rikuto Kotoge, Mai Nishimura, Jiaxin Ma
Abstract: Reinforcement Learning has emerged as a dominant post-training approach to elicit agentic RAG behaviors such as search and planning from language models. Despite its success with larger models, applying RL to compact models (e.g., 0.5--1B parameters) presents unique challenges. The compact models exhibit poor initial performance, resulting in sparse rewards and unstable training. To overcome these difficulties, we propose Distillation-Guided Policy Optimization (DGPO), which employs cold-start initialization from teacher demonstrations and continuous teacher guidance during policy optimization. To understand how compact models preserve agentic behavior, we introduce Agentic RAG Capabilities (ARC), a fine-grained metric analyzing reasoning, search coordination, and response synthesis. Comprehensive experiments demonstrate that DGPO enables compact models to achieve sophisticated agentic search behaviors, even outperforming the larger teacher model in some cases. DGPO makes agentic RAG feasible in computing resource-constrained environments.
Authors: Saksorn Ruangtanusak, Pittawat Taveekitworachai, Kunat Pipatanakul
Abstract: This report investigates approaches for prompting a tool-augmented large language model (LLM) to act as a role-playing dialogue agent in the API track of the Commonsense Persona-grounded Dialogue Challenge (CPDC) 2025. In this setting, dialogue agents often produce overly long in-character responses (over-speaking) while failing to use tools effectively according to the persona (under-acting), such as generating function calls that do not exist or making unnecessary tool calls before answering. We explore four prompting approaches to address these issues: 1) basic role prompting, 2) improved role prompting, 3) automatic prompt optimization (APO), and 4) rule-based role prompting. The rule-based role prompting (RRP) approach achieved the best performance through two novel techniques-character-card/scene-contract design and strict enforcement of function calling-which led to an overall score of 0.571, improving on the zero-shot baseline score of 0.519. These findings demonstrate that RRP design can substantially improve the effectiveness and reliability of role-playing dialogue agents compared with more elaborate methods such as APO. To support future efforts in developing persona prompts, we are open-sourcing all of our best-performing prompts and the APO tool Source code is available at https://github.com/scb-10x/apo
Authors: Hanqi Yan, Hainiu Xu, Siya Qi, Shu Yang, Yulan He
Abstract: With the growing accessibility and wide adoption of large language models, concerns about their safety and alignment with human values have become paramount. In this paper, we identify a concerning phenomenon: Reasoning-Induced Misalignment (RIM), in which misalignment emerges when reasoning capabilities strengthened-particularly when specific types of reasoning patterns are introduced during inference or training. Beyond reporting this vulnerability, we provide the first mechanistic account of its origins. Through representation analysis, we discover that specific attention heads facilitate refusal by reducing their attention to CoT tokens, a mechanism that modulates the model's rationalization process during inference. During training, we find significantly higher activation entanglement between reasoning and safety in safety-critical neurons than in control neurons, particularly after fine-tuning with those identified reasoning patterns. This entanglement strongly correlates with catastrophic forgetting, providing a neuron-level explanation for RIM.
Authors: Xiaoqiang Lin, Aritra Ghosh, Bryan Kian Hsiang Low, Anshumali Shrivastava, Vijai Mohan
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in leveraging extensive external knowledge to enhance responses in multi-turn and agentic applications, such as retrieval-augmented generation (RAG). However, processing long-context inputs introduces significant system latency and demands substantial memory for the key-value cache, resulting in reduced throughput and a fundamental trade-off between knowledge enrichment and system efficiency. While minimizing latency for long-context inputs is a primary objective for LLMs, we contend that RAG require specialized consideration. In RAG, much of the LLM context consists of concatenated passages from retrieval, with only a small subset directly relevant to the query. These passages often exhibit low semantic similarity due to diversity or deduplication during re-ranking, leading to block-diagonal attention patterns that differ from those in standard LLM generation tasks. Based on this observation, we argue that most computations over the RAG context during decoding are unnecessary and can be eliminated with minimal impact on performance. To this end, we propose REFRAG, an efficient decoding framework that compresses, senses, and expands to improve latency in RAG applications. By exploiting the sparsity structure, we demonstrate a 30.85 the time-to-first-token acceleration (3.75 improvement to previous work) without loss in perplexity. In addition, our optimization framework for large context enables REFRAG to extend the context size of LLMs by 16. We provide rigorous validation of REFRAG across diverse long-context tasks, including RAG, multi-turn conversations, and long document summarization, spanning a wide range of datasets. Experimental results confirm that REFRAG delivers substantial speedup with no loss in accuracy compared to LLaMA models and other state-of-the-art baselines across various context sizes.
Authors: Jusheng Zhang, Yijia Fan, Kaitong Cai, Zimeng Huang, Xiaofei Sun, Jian Wang, Chengpei Tang, Keze Wang
Abstract: This paper introduces DrDiff, a novel framework for long-text generation that overcomes the efficiency-quality trade-off through three core technologies. First, we design a dynamic expert scheduling mechanism that intelligently allocates computational resources during the diffusion process based on text complexity, enabling more efficient handling of text generation tasks of varying difficulty. Second, we introduce a Hierarchical Sparse Attention (HSA) mechanism that adaptively adjusts attention patterns according to a variety of input lengths, reducing computational complexity from O($n^2$) to O($n$) while maintaining model performance. Finally, we propose a soft absorption guidance optimization strategy that combines with DPM-solver++ to reduce diffusion steps, significantly improving generation speed. Comprehensive experiments on various long-text generation benchmarks demonstrate the superiority of our DrDiff over the existing SOTA methods.
Authors: Yuhao Zhang, Shaoming Duan, Jinhang Su, Chuanyi Liu, Peiyi Han
Abstract: Despite the significant advancements of self-play fine-tuning (SPIN), which can transform a weak large language model (LLM) into a strong one through competitive interactions between models of varying capabilities, it still faces challenges in the Text-to-SQL task. SPIN does not generate new information, and the large number of correct SQL queries produced by the opponent model during self-play reduces the main model's ability to generate accurate SQL queries. To address this challenge, we propose a new self-play fine-tuning method tailored for the Text-to-SQL task, called SPFT-SQL. Prior to self-play, we introduce a verification-based iterative fine-tuning approach, which synthesizes high-quality fine-tuning data iteratively based on the database schema and validation feedback to enhance model performance, while building a model base with varying capabilities. During the self-play fine-tuning phase, we propose an error-driven loss method that incentivizes incorrect outputs from the opponent model, enabling the main model to distinguish between correct SQL and erroneous SQL generated by the opponent model, thereby improving its ability to generate correct SQL. Extensive experiments and in-depth analyses on six open-source LLMs and five widely used benchmarks demonstrate that our approach outperforms existing state-of-the-art (SOTA) methods.
Authors: Andrea Wynn, Harsh Satija, Gillian Hadfield
Abstract: While multi-agent debate has been proposed as a promising strategy for improving AI reasoning ability, we find that debate can sometimes be harmful rather than helpful. Prior work has primarily focused on debates within homogeneous groups of agents, whereas we explore how diversity in model capabilities influences the dynamics and outcomes of multi-agent interactions. Through a series of experiments, we demonstrate that debate can lead to a decrease in accuracy over time - even in settings where stronger (i.e., more capable) models outnumber their weaker counterparts. Our analysis reveals that models frequently shift from correct to incorrect answers in response to peer reasoning, favoring agreement over challenging flawed reasoning. We perform additional experiments investigating various potential contributing factors to these harmful shifts - including sycophancy, social conformity, and model and task type. These results highlight important failure modes in the exchange of reasons during multi-agent debate, suggesting that naive applications of debate may cause performance degradation when agents are neither incentivised nor adequately equipped to resist persuasive but incorrect reasoning.
Authors: Nathaniel Imel, Noga Zaslavsky
Abstract: Converging evidence suggests that systems of semantic categories across human languages achieve near-optimal compression via the Information Bottleneck (IB) complexity-accuracy principle. Large language models (LLMs) are not trained for this objective, which raises the question: are LLMs capable of evolving efficient human-like semantic systems? To address this question, we focus on the domain of color as a key testbed of cognitive theories of categorization and replicate with LLMs (Gemini 2.0-flash and Llama 3.3-70B-Instruct) two influential human behavioral studies. First, we conduct an English color-naming study, showing that Gemini aligns well with the naming patterns of native English speakers and achieves a significantly high IB-efficiency score, while Llama exhibits an efficient but lower complexity system compared to English. Second, to test whether LLMs simply mimic patterns in their training data or actually exhibit a human-like inductive bias toward IB-efficiency, we simulate cultural evolution of pseudo color-naming systems in LLMs via iterated in-context language learning. We find that akin to humans, LLMs iteratively restructure initially random systems towards greater IB-efficiency and increased alignment with patterns observed across the world's languages. These findings demonstrate that LLMs are capable of evolving perceptually grounded, human-like semantic systems, driven by the same fundamental principle that governs semantic efficiency across human languages.
Authors: Ankan Mullick, Sombit Bose, Rounak Saha, Ayan Kumar Bhowmick, Aditya Vempaty, Prasenjit Dey, Ravi Kokku, Pawan Goyal, Niloy Ganguly
Abstract: In this paper, we introduce Spotlight, a novel paradigm for information extraction that produces concise, engaging narratives by highlighting the most compelling aspects of a document. Unlike traditional summaries, which prioritize comprehensive coverage, spotlights selectively emphasize intriguing content to foster deeper reader engagement with the source material. We formally differentiate spotlights from related constructs and support our analysis with a detailed benchmarking study using new datasets curated for this work. To generate high-quality spotlights, we propose a two-stage approach: fine-tuning a large language model on our benchmark data, followed by alignment via Direct Preference Optimization (DPO). Our comprehensive evaluation demonstrates that the resulting model not only identifies key elements with precision but also enhances readability and boosts the engagement value of the original document.
Authors: Md. Mithun Hossain, Sanjara, Md. Shakil Hossain, Sudipto Chaki
Abstract: Entity-level sentiment classification involves identifying the sentiment polarity linked to specific entities within text. This task poses several challenges: effectively modeling the subtle and complex interactions between entities and their surrounding sentiment expressions; capturing dependencies that may span across sentences; and ensuring consistent sentiment predictions for multiple mentions of the same entity through coreference resolution. Additionally, linguistic phenomena such as negation, ambiguity, and overlapping opinions further complicate the analysis. These complexities make entity-level sentiment classification a difficult problem, especially in real-world, noisy textual data. To address these issues, we propose SpanEIT, a novel framework integrating dynamic span interaction and graph-aware memory mechanisms for enhanced entity-sentiment relational modeling. SpanEIT builds span-based representations for entities and candidate sentiment phrases, employs bidirectional attention for fine-grained interactions, and uses a graph attention network to capture syntactic and co-occurrence relations. A coreference-aware memory module ensures entity-level consistency across documents. Experiments on FSAD, BARU, and IMDB datasets show SpanEIT outperforms state-of-the-art transformer and hybrid baselines in accuracy and F1 scores. Ablation and interpretability analyses validate the effectiveness of our approach, underscoring its potential for fine-grained sentiment analysis in applications like social media monitoring and customer feedback analysis.
Authors: Spandan Anaokar, Shrey Ganatra, Harshvivek Kashid, Swapnil Bhattacharyya, Shruti Nair, Reshma Sekhar, Siddharth Manohar, Rahul Hemrajani, Pushpak Bhattacharyya
Abstract: Large Language Models (LLMs) are widely used in industry but remain prone to hallucinations, limiting their reliability in critical applications. This work addresses hallucination reduction in consumer grievance chatbots built using LLaMA 3.1 8B Instruct, a compact model frequently used in industry. We develop HalluDetect, an LLM-based hallucination detection system that achieves an F1 score of 68.92% outperforming baseline detectors by 22.47%. Benchmarking five hallucination mitigation architectures, we find that out of them, AgentBot minimizes hallucinations to 0.4159 per turn while maintaining the highest token accuracy (96.13%), making it the most effective mitigation strategy. Our findings provide a scalable framework for hallucination mitigation, demonstrating that optimized inference strategies can significantly improve factual accuracy.
Authors: Senkang Hu, Xudong Han, Jinqi Jiang, Yihang Tao, Zihan Fang, Yong Dai, Sam Tak Wu Kwong, Yuguang Fang
Abstract: Adapting billion-parameter language models to a downstream task is still costly, even with parameter-efficient fine-tuning (PEFT). We re-cast task adaptation as output-distribution alignment: the objective is to steer the output distribution toward the task distribution directly during decoding rather than indirectly through weight updates. Building on this view, we introduce Steering Vector Decoding (SVDecode), a lightweight, PEFT-compatible, and theoretically grounded method. We start with a short warm-start fine-tune and extract a task-aware steering vector from the Kullback-Leibler (KL) divergence gradient between the output distribution of the warm-started and pre-trained models. This steering vector is then used to guide the decoding process to steer the model's output distribution towards the task distribution. We theoretically prove that SVDecode is first-order equivalent to the gradient step of full fine-tuning and derive a globally optimal solution for the strength of the steering vector. Across three tasks and nine benchmarks, SVDecode paired with four standard PEFT methods improves multiple-choice accuracy by up to 5 percentage points and open-ended truthfulness by 2 percentage points, with similar gains (1-2 percentage points) on commonsense datasets without adding trainable parameters beyond the PEFT adapter. SVDecode thus offers a lightweight, theoretically grounded path to stronger task adaptation for large language models.
Authors: Minsoo Kim, Arnav Kundu, Han-Byul Kim, Richa Dixit, Minsik Cho
Abstract: Modern large language models (LLMs) extend context lengths to millions of tokens, enabling coherent, personalized responses grounded in long conversational histories. This ability, however, hinges on Key-Value (KV) caching, whose memory grows linearly with dialogue length and quickly becomes the bottleneck in resource-constrained environments. An active line of research for reducing memory bottleneck is KV cache compression, which seeks to limit cache size while preserving accuracy. Yet existing methods face two major limitations: (i) evicting the KV cache after full-context prefill causes unbounded peak memory, and (ii) query-dependent eviction narrows the cache to a single query, leading to failure cases in multi-turn conversations. We introduce EpiCache, a training-free KV cache management framework for long conversational question answering (LongConvQA) under fixed memory budgets. EpiCache bounds cache growth through block-wise prefill and preserves topic-relevant context via episodic KV compression, which clusters conversation history into coherent episodes and applies episode-specific KV cache eviction. We further design an adaptive layer-wise budget allocation strategy that measures each layer's sensitivity to eviction and distributes the memory budget across layers accordingly. Across three LongConvQA benchmarks, EpiCache improves accuracy by up to 40%, maintains near-full KV accuracy under 4-6x compression, and reduces latency/memory by up to 2.4x/3.5x, enabling efficient multi-turn interaction under strict resource limits. Our code is available at https://github.com/apple/ml-epicache.
Authors: Alovi N Shohe, Chonglio Khiamungam, Teisovi Angami
Abstract: This paper investigates part-of-speech tagging, an important task in Natural Language Processing (NLP) for the Nagamese language. The Nagamese language, a.k.a. Naga Pidgin, is an Assamese-lexified Creole language developed primarily as a means of communication in trade between the Nagas and people from Assam in northeast India. A substantial amount of work in part-of-speech-tagging has been done for resource-rich languages like English, Hindi, etc. However, no work has been done in the Nagamese language. To the best of our knowledge, this is the first attempt at part-of-speech tagging for the Nagamese Language. The aim of this work is to identify the part-of-speech for a given sentence in the Nagamese language. An annotated corpus of 16,112 tokens is created and applied machine learning technique known as Conditional Random Fields (CRF). Using CRF, an overall tagging accuracy of 85.70%; precision, recall of 86%, and f1-score of 85% is achieved. Keywords. Nagamese, NLP, part-of-speech, machine learning, CRF.
Authors: Tiago Fernandes Tavares
Abstract: Unsupervised analysis of text corpora is challenging, especially in data-scarce domains where traditional topic models struggle. While these models offer a solution, they typically describe clusters with lists of keywords that require significant manual effort to interpret and often lack semantic coherence. To address this critical interpretability gap, we introduce Recursive Thematic Partitioning (RTP), a novel framework that leverages Large Language Models (LLMs) to interactively build a binary tree. Each node in the tree is a natural language question that semantically partitions the data, resulting in a fully interpretable taxonomy where the logic of each cluster is explicit. Our experiments demonstrate that RTP's question-driven hierarchy is more interpretable than the keyword-based topics from a strong baseline like BERTopic. Furthermore, we establish the quantitative utility of these clusters by showing they serve as powerful features in downstream classification tasks, particularly when the data's underlying themes correlate with the task labels. RTP introduces a new paradigm for data exploration, shifting the focus from statistical pattern discovery to knowledge-driven thematic analysis. Furthermore, we demonstrate that the thematic paths from the RTP tree can serve as structured, controllable prompts for generative models. This transforms our analytical framework into a powerful tool for synthesis, enabling the consistent imitation of specific characteristics discovered in the source corpus.
Authors: Junliang Li, Yucheng Wang, Yan Chen, Yu Ran, Ruiqing Zhang, Jing Liu, Hua Wu, Haifeng Wang
Abstract: Hallucination and factuality deficits remain key obstacles to the reliability of large language models (LLMs) in long-form generation. Existing reinforcement learning from human feedback (RLHF) frameworks primarily rely on preference rewards, yet they often overlook the model's internal knowledge boundaries, exacerbating the so-called "hallucination tax". To address this challenge, we propose Knowledge-Level Consistency Reinforcement Learning Framework (KLCF), a novel framework that focuses on the knowledge consistency between the policy model's expressed knowledge and the base model's parametric knowledge, and introduces a Dual-Fact Alignment mechanism to jointly optimize factual recall and precision. Specifically, KLCF leverages pretrained knowledge boundaries to construct fact checklist, guiding online reinforcement learning to improve factual coverage and recall; simultaneously, it trains a self-assessment module based on the base model's internal knowledge to enhance factual precision during generation. Unlike prior methods that rely on external retrieval or heavy verification, our reward design is fully external-knowledge-free and lightweight, making KLCF efficient and easily scalable to large-scale training. Experimental results demonstrate that KLCF substantially improves factuality metrics across multiple long-form benchmarks and effectively alleviates model hallucinations.
Authors: Guibin Zhang, Muxin Fu, Shuicheng Yan
Abstract: Agent memory shapes how Large Language Model (LLM)-powered agents, akin to the human brain, progressively refine themselves through environment interactions. Existing paradigms remain constrained: parametric memory forcibly adjusts model parameters, and retrieval-based memory externalizes experience into structured databases, yet neither captures the fluid interweaving of reasoning and memory that underlies human cognition. To address this gap, we propose MemGen, a dynamic generative memory framework that equips agents with a human-esque cognitive faculty. It consists of a \textit{memory trigger}, which monitors the agent's reasoning state to decide explicit memory invocation, and a \textit{memory weaver}, which takes the agent's current state as stimulus to construct a latent token sequence as machine-native memory to enrich its reasoning. In this way, MemGen enables agents to recall and augment latent memory throughout reasoning, producing a tightly interwoven cycle of memory and cognition. Extensive experiments across eight benchmarks show that MemGen surpasses leading external memory systems such as ExpeL and AWM by up to $38.22\%$, exceeds GRPO by up to $13.44\%$, and exhibits strong cross-domain generalization ability. More importantly, we find that without explicit supervision, MemGen spontaneously evolves distinct human-like memory faculties, including planning memory, procedural memory, and working memory, suggesting an emergent trajectory toward more naturalistic forms of machine cognition.
Authors: Rui Jia, Yuang Wei, Ruijia Li, Yuan-Hao Jiang, Xinyu Xie, Yaomin Shen, Min Zhang, Bo Jiang
Abstract: While cognitive diagnosis (CD) effectively assesses students' knowledge mastery from structured test data, applying it to real-world teacher-student dialogues presents two fundamental challenges. Traditional CD models lack a suitable framework for handling dynamic, unstructured dialogues, and it's difficult to accurately extract diagnostic semantics from lengthy dialogues. To overcome these hurdles, we propose DiaCDM, an innovative model. We've adapted the initiation-response-evaluation (IRE) framework from educational theory to design a diagnostic framework tailored for dialogue. We also developed a unique graph-based encoding method that integrates teacher questions with relevant knowledge components to capture key information more precisely. To our knowledge, this is the first exploration of cognitive diagnosis in a dialogue setting. Experiments on three real-world dialogue datasets confirm that DiaCDM not only significantly improves diagnostic accuracy but also enhances the results' interpretability, providing teachers with a powerful tool for assessing students' cognitive states. The code is available at https://github.com/Mind-Lab-ECNU/DiaCDM/tree/main.
Authors: Yanbin Fu, Hong Jiao, Tianyi Zhou, Nan Zhang, Ming Li, Qingshu Xu, Sydney Peters, Robert W. Lissitz
Abstract: Aligning test items to content standards is a critical step in test development to collect validity evidence based on content. Item alignment has typically been conducted by human experts. This judgmental process can be subjective and time-consuming. This study investigated the performance of fine-tuned small language models (SLMs) for automated item alignment using data from a large-scale standardized reading and writing test for college admissions. Different SLMs were trained for alignment at both domain and skill levels respectively with 10 skills mapped to 4 content domains. The model performance was evaluated in multiple criteria on two testing datasets. The impact of types and sizes of the input data for training was investigated. Results showed that including more item text data led to substantially better model performance, surpassing the improvements induced by sample size increase alone. For comparison, supervised machine learning models were trained using the embeddings from the multilingual-E5-large-instruct model. The study results showed that fine-tuned SLMs consistently outperformed the embedding-based supervised machine learning models, particularly for the more fine-grained skill alignment. To better understand model misclassifications, multiple semantic similarity analysis including pairwise cosine similarity, Kullback-Leibler divergence of embedding distributions, and two-dimension projections of item embeddings were conducted. These analyses consistently showed that certain skills in SAT and PSAT were semantically too close, providing evidence for the observed misclassification.
Authors: Zhuoran Zhuang, Ye Chen, Xia Zeng, Chao Luo, Luhui Liu, Yihan Chen
Abstract: We study deploying large language models (LLMs) as business development (BD) agents for persuasive price negotiation in online travel agencies (OTAs), where aligning traveler affordability and hotel profitability directly affects bookings, partner relationships, and access to travel. The agent must follow a Standard Operating Procedure (SOP) while conducting multi-turn persuasion, interpreting colloquial inputs, and adhering to guardrails (no over-promising, no hallucinations). Conventional post-training -- supervised fine-tuning (SFT) or single-source reward optimization -- overfits scripts, misses nuanced persuasive style, and fails to enforce verifiable business constraints. We propose Reward-Enhanced Policy Optimization (REPO), a reinforcement learning post-training framework that aligns an LLM with heterogeneous rewards: a preference-trained reward model (RM) for dense human alignment, a reward judge (RJ) for high-level persuasive behavior and SOP compliance, and programmatic reward functions (RF) for deterministic checks on numerics, formatting, and guardrails. A straightforward enhancement mechanism is proposed to combine the RM with RJ and RF signals to curb reward hacking and improve negotiation quality. In production-style evaluations -- approximately 150 turns from real dialogues and 225 turns from curated bad-case dialogues -- REPO lifts average dialogue rating to 4.63: +1.20 over base, +0.83 over Direct Preference Optimization (DPO); +0.33 over Group Relative Policy Optimization (GRPO), increases the share of conversations with at least one excellent response to 66.67% (+23.34 percentage points over GRPO), and achieves a 93.33% bad-case fix rate with 75.56% clean fixes, outperforming SFT, DPO, PPO, and GRPO. We also observe emergent capabilities -- proactive empathy, localized reasoning, calibrated tactics -- that surpass gold annotations.
Authors: Ayan Majumdar, Feihao Chen, Jinghui Li, Xiaozhen Wang
Abstract: Large-scale web-scraped text corpora used to train general-purpose AI models often contain harmful demographic-targeted social biases, creating a regulatory need for data auditing and developing scalable bias-detection methods. Although prior work has investigated biases in text datasets and related detection methods, these studies remain narrow in scope. They typically focus on a single content type (e.g., hate speech), cover limited demographic axes, overlook biases affecting multiple demographics simultaneously, and analyze limited techniques. Consequently, practitioners lack a holistic understanding of the strengths and limitations of recent large language models (LLMs) for automated bias detection. In this study, we present a comprehensive evaluation framework aimed at English texts to assess the ability of LLMs in detecting demographic-targeted social biases. To align with regulatory requirements, we frame bias detection as a multi-label task using a demographic-focused taxonomy. We then conduct a systematic evaluation with models across scales and techniques, including prompting, in-context learning, and fine-tuning. Using twelve datasets spanning diverse content types and demographics, our study demonstrates the promise of fine-tuned smaller models for scalable detection. However, our analyses also expose persistent gaps across demographic axes and multi-demographic targeted biases, underscoring the need for more effective and scalable auditing frameworks.
Authors: Mingjin Li, Yu Liu, Huayi Liu, Xiang Ye, Chao Jiang, Hongguang Zhang, Yu Ruan
Abstract: We propose MADS (Multi-Agent Dialogue Simulation), a scalable framework for generating persuasive multi-turn dialogues via agent self-play. MADS employs three coordinated agents: User Agents designed to simulate diverse persona-driven behaviors by leveraging personality signifiers such as Zodiac Signs and MBTI types, a Dialog Agent executing task-oriented persuasion strategies and an Optimization Agent evaluating and refining dialogue outcomes. We further validate its effectiveness through users' Chain-of-Attitude (CoA) modeling and dedicated LLMs' persuasion assessment. This approach enables low-cost generation of training data without human annotation, addressing key industry challenges such as lack of user data, cold-start evaluation difficulties, and prompt inefficiency. Applied to a real-world marketing scenario, MADS significantly improved the persuasion capacity of small LLMs, increasing the organic traffic conversion rate by 22.4% (from 1.83% to 2.24%) , demonstrating clear business value.
Authors: Qingshu Xu, Hong Jiao, Tianyi Zhou, Ming Li, Nan Zhang, Sydney Peters, Yanbin Fu
Abstract: Accurate alignment of items to content standards is critical for valid score interpretation in large-scale assessments. This study evaluates three automated paradigms for aligning items with four domain and nineteen skill labels. First, we extracted embeddings and trained multiple classical supervised machine learning models, and further investigated the impact of dimensionality reduction on model performance. Second, we fine-tuned eight BERT model and its variants for both domain and skill alignment. Third, we explored ensemble learning with majority voting and stacking with multiple meta-models. The DeBERTa-v3-base achieved the highest weighted-average F1 score of 0.950 for domain alignment while the RoBERTa-large yielded the highest F1 score of 0.869 for skill alignment. Ensemble models did not surpass the best-performing language models. Dimension reduction enhanced linear classifiers based on embeddings but did not perform better than language models. This study demonstrated different methods in automated item alignment to content standards.}
Authors: Maojia Song, Renhang Liu, Xinyu Wang, Yong Jiang, Pengjun Xie, Fei Huang, Soujanya Poria, Jingren Zhou
Abstract: RAG (Retrieval-Augmented Generation) systems and web agents are increasingly evaluated on multi-hop deep search tasks, yet current practice suffers from two major limitations. First, most benchmarks leak the reasoning path in the question text, allowing models to follow surface cues rather than discover reasoning chains autonomously. Second, evaluation is typically reduced to a single pass rate, which collapses diverse behaviours into one score and obscures whether failures stem from inadequate search, poor knowledge use, or inappropriate refusal. To address these issues, we present WebDetective, a benchmark of hint-free multi-hop questions paired with a controlled Wikipedia sandbox that ensures full traceability of model actions, and a holistic evaluation framework that separates search sufficiency, knowledge utilisation, and refusal behaviour. Our evaluation of 25 state-of-the-art models reveals systematic weaknesses across all architectures: models struggle with knowledge utilisation despite having sufficient evidence and demonstrate near-absent appropriate refusal when evidence is lacking. These patterns expose a fundamental gap: today's systems excel at executing given reasoning paths but fail when required to discover them. We develop an agentic workflow, EvidenceLoop, that explicitly targets the challenges our benchmark identifies, incorporating verification loops and systematic evidence tracking that improve both search and synthesis capabilities. This baseline demonstrates that WebDetective's diagnostic framework can guide concrete architectural improvements, establishing our benchmark as a critical tool for developing genuinely autonomous reasoning systems rather than pattern-following agents.
Authors: Yongmin Yoo, Xu Zhang, Longbing Cao
Abstract: Large language models (LLMs) increasingly generate natural language rationales to enhance interpretability, but these often contain logical errors, label mismatches, and domain-specific misalignments. Directly using such rationales as supervision risks propagating noise and undermining training stability. To address this challenge, we introduce Self-Filtered Distillation, a framework specifically tailored for patent classification, which treats LLM-generated rationales as trust signals rather than ground-truth supervision. The framework employs selective distillation guided by three unsupervised trust metrics: (1) Self-Consistency, which measures the stability of LLM-generated rationales across multiple generations; (2) Class Entailment Alignment, which assesses semantic coherence with patent-specific class definitions; and (3) LLM Agreement Scoring, which validates rationale-label plausibility. These metrics are integrated into a unified trust score that primarily weights training samples while optionally filtering out extremely low-trust cases, enabling reasoning-aware supervision. Experiments on the USPTO-2M dataset, a widely used benchmark for patent classification, show that our method outperforms label-based learning and conventional distillation in accuracy, stability, and interpretability, establishing a reliable paradigm for leveraging reasoning-aware trust indicators in patent analytics.
Authors: Sunbowen Lee, Qingyu Yin, Chak Tou Leong, Jialiang Zhang, Yicheng Gong, Shiwen Ni, Min Yang, Xiaoyu Shen
Abstract: Large language models (LLMs) are increasingly deployed on complex reasoning tasks, yet little is known about their ability to internally evaluate problem difficulty, which is an essential capability for adaptive reasoning and efficient resource allocation. In this work, we investigate whether LLMs implicitly encode problem difficulty in their internal representations. Using a linear probe on the final-token representations of LLMs, we demonstrate that the difficulty level of math problems can be linearly modeled. We further locate the specific attention heads of the final Transformer layer: these attention heads have opposite activation patterns for simple and difficult problems, thus achieving perception of difficulty. Our ablation experiments prove the accuracy of the location. Crucially, our experiments provide practical support for using LLMs as automatic difficulty annotators, potentially substantially reducing reliance on costly human labeling in benchmark construction and curriculum learning. We also uncover that there is a significant difference in entropy and difficulty perception at the token level. Our study reveals that difficulty perception in LLMs is not only present but also structurally organized, offering new theoretical insights and practical directions for future research. Our code is available at https://github.com/Aegis1863/Difficulty-Perception-of-LLMs.
URLs: https://github.com/Aegis1863/Difficulty-Perception-of-LLMs.
Authors: Toshiki Nakai, Ravi Kiran Chikkala, Lena Sophie Oberkircher, Nicholas Jennings, Natalia Skachkova, Tatiana Anikina, Jesujoba Oluwadara Alabi
Abstract: The 2025 Multimodal Models for Low-Resource Contexts and Social Impact (MMLoSo) Language Challenge addresses one of India's most pressing linguistic gaps: the lack of resources for its diverse low-resource languages (LRLs). In this study, we investigate whether enforcing cross-lingual similarity in specific internal layers of a decoder-only multilingual large language model (LLM) can improve translation quality from LRL to high-resource language (HRL). Specifically, we combine Centered Kernel Alignment (CKA), a similarity metric that encourages representations of different languages to align, with REPINA, a regularization method that constrains parameter updates to remain close to the pretrained model, into a joint method we call TRepLiNa. In this research project, we experiment with zero-shot, few-shot, and fine-tuning settings using Aya-23 8B with QLoRA across MMLoSo shared task language pairs (Mundari, Santali, Bhili) with Hindi/English pivots. Our results show that aligning mid-level layers using TRepLiNa (CKA+REPINA) is a low-cost, practical approach to improving LRL translation, especially in data-scarce settings.
Authors: Haotian Wu, Shufan Jiang, Mingyu Chen, Yiyang Feng, Hehai Lin, Heqing Zou, Yao Shu, Chengwei Qin
Abstract: As large language models (LLMs) advance in role-playing (RP) tasks, existing benchmarks quickly become obsolete due to their narrow scope, outdated interaction paradigms, and limited adaptability across diverse application scenarios. To address this gap, we introduce FURINA-Builder, a novel multi-agent collaboration pipeline that automatically constructs fully customizable RP benchmarks at any scale. It enables evaluation of arbitrary characters across diverse scenarios and prompt formats, as the first benchmark builder in RP area for adaptable assessment. FURINA-Builder simulates dialogues between a test character and other characters drawn from a well-constructed character-scene pool, while an LLM judge selects fine-grained evaluation dimensions and adjusts the test character's responses into final test utterances. Using this pipeline, we build FURINA-Bench, a new comprehensive role-playing benchmark featuring both established and synthesized test characters, each assessed with dimension-specific evaluation criteria. Human evaluation and preliminary separability analysis justify our pipeline and benchmark design. We conduct extensive evaluations of cutting-edge LLMs and find that o3 and DeepSeek-R1 achieve the best performance on English and Chinese RP tasks, respectively. Across all models, established characters consistently outperform synthesized ones, with reasoning capabilities further amplifying this disparity. Interestingly, we observe that model scale does not monotonically reduce hallucinations. More critically, for reasoning LLMs, we uncover a novel trade-off: reasoning improves RP performance but simultaneously increases RP hallucinations. This trade-off extends to a broader Pareto frontier between RP performance and reliability for all LLMs. These findings demonstrate the effectiveness of FURINA-Builder and the challenge posed by FURINA-Bench.
Authors: Jusen Du, Jiaxi Hu, Tao Zhang, Weigao Sun, Yu Cheng
Abstract: Transformers excel at sequence modeling but face quadratic complexity, while linear attention offers improved efficiency but often compromises recall accuracy over long contexts. In this work, we introduce Native Hybrid Attention (NHA), a novel hybrid architecture of linear and full attention that integrates both intra \& inter-layer hybridization into a unified layer design. NHA maintains long-term context in key-value slots updated by a linear RNN, and augments them with short-term tokens from a sliding window. A single \texttt{softmax attention} operation is then applied over all keys and values, enabling per-token and per-head context-dependent weighting without requiring additional fusion parameters. The inter-layer behavior is controlled through a single hyperparameter, the sliding window size, which allows smooth adjustment between purely linear and full attention while keeping all layers structurally uniform. Experimental results show that NHA surpasses Transformers and other hybrid baselines on recall-intensive and commonsense reasoning tasks. Furthermore, pretrained LLMs can be structurally hybridized with NHA, achieving competitive accuracy while delivering significant efficiency gains. Code is available at https://github.com/JusenD/NHA.
Authors: Xinliang Frederick Zhang, Anhad Mohananey, Alexandra Chronopoulou, Pinelopi Papalampidi, Somit Gupta, Tsendsuren Munkhdalai, Lu Wang, Shyam Upadhyay
Abstract: Models employing long chain-of-thought (CoT) reasoning have shown superior performance on complex reasoning tasks. Yet, this capability introduces a critical and often overlooked inefficiency -- overthinking -- models often engage in unnecessarily extensive reasoning even for simple queries, incurring significant computations without accuracy improvements. While prior work has explored solutions to mitigate overthinking, a fundamental gap remains in our understanding of its underlying causes. Most existing analyses are limited to superficial, profiling-based observations, failing to delve into LLMs' inner workings. This study introduces a systematic, fine-grained analyzer of LLMs' thought process to bridge the gap, TRACE. We first benchmark the overthinking issue, confirming that long-thinking models are five to twenty times slower on simple tasks with no substantial gains. We then use TRACE to first decompose the thought process into minimally complete sub-thoughts. Next, by inferring discourse relationships among sub-thoughts, we construct granular thought progression graphs and subsequently identify common thinking patterns for topically similar queries. Our analysis reveals two major patterns for open-weight thinking models -- Explorer and Late Landing. This finding provides evidence that over-verification and over-exploration are the primary drivers of overthinking in LLMs. Grounded in thought structures, we propose a utility-based definition of overthinking, which moves beyond length-based metrics. This revised definition offers a more insightful understanding of LLMs' thought progression, as well as practical guidelines for principled overthinking management.
Authors: Jialu Du, Guiyang Hou, Yihui Fu, Chen Wu, Wenqi Zhang, Yongliang Shen, Weiming Lu
Abstract: While large language models (LLMs) excel in mathematical and code reasoning, we observe they struggle with social reasoning tasks, exhibiting cognitive confusion, logical inconsistencies, and conflation between objective world states and subjective belief states. Through deteiled analysis of DeepSeek-R1's reasoning trajectories, we find that LLMs frequently encounter reasoning impasses and tend to output contradictory terms like "tricky" and "confused" when processing scenarios with multiple participants and timelines, leading to erroneous reasoning or infinite loops. The core issue is their inability to disentangle objective reality from agents' subjective beliefs. To address this, we propose an adaptive world model-enhanced reasoning mechanism that constructs a dynamic textual world model to track entity states and temporal sequences. It dynamically monitors reasoning trajectories for confusion indicators and promptly intervenes by providing clear world state descriptions, helping models navigate through cognitive dilemmas. The mechanism mimics how humans use implicit world models to distinguish between external events and internal beliefs. Evaluations on three social benchmarks demonstrate significant improvements in accuracy (e.g., +10% in Hi-ToM) while reducing computational costs (up to 33.8% token reduction), offering a simple yet effective solution for deploying LLMs in social contexts.
Authors: Gustave Cortal (ENS Paris Saclay, LISN), Alain Finkel (ENS Paris Saclay)
Abstract: Personal narratives are stories authors construct to make meaning of their experiences. Style, the distinctive way authors use language to express themselves, is fundamental to how these narratives convey subjective experiences. Yet there is a lack of a formal framework for systematically analyzing these stylistic choices. We present a novel approach that formalizes style in personal narratives as patterns in the linguistic choices authors make when communicating subjective experiences. Our framework integrates three domains: functional linguistics establishes language as a system of meaningful choices, computer science provides methods for automatically extracting and analyzing sequential patterns, and these patterns are linked to psychological observations. Using language models, we automatically extract linguistic features such as processes, participants, and circumstances. We apply our framework to hundreds of dream narratives, including a case study on a war veteran with post-traumatic stress disorder. Analysis of his narratives uncovers distinctive patterns, particularly how verbal processes dominate over mental ones, illustrating the relationship between linguistic choices and psychological states.
Authors: Yuxin Ma, Lun Du, Lanning Wei, Kun Chen, Qian Xu, Kangyu Wang, Guofeng Feng, Guoshan Lu, Lin Liu, Xiaojing Qi, Xinyuan Zhang, Zhen Tao, Haibo Feng, Ziyun Jiang, Ying Xu, Zenan Huang, Yihong Zhuang, Haokai Xu, Jiaqi Hu, Zhenzhong Lan, Junbo Zhao, Jianguo Li, Da Zheng
Abstract: Diffusion-based large language models (dLLMs) have emerged as a promising alternative to autoregressive (AR) LLMs, leveraging denoising-based generation to enable inherent parallelism. Even more and more open-sourced dLLM models emerge, yet their widespread adoption remains constrained by the lack of a standardized and efficient inference framework. We present dInfer, an efficient and extensible framework for dLLM inference. dInfer decomposes the inference pipeline into four modular components--model, diffusion iteration manager, decoding strategy, and KV-cache manager--and integrates novel algorithms for each component alongside system-level optimizations. Through this combination of algorithmic innovations and system enhancements, dInfer achieves substantial efficiency gains without compromising output quality on LLaDA-MoE. At batch size 1, it surpasses 1,100 tokens per second on HumanEval and averages over 800 tokens per second across six benchmarks on $8\times$ H800 GPUs. Compared to prior systems, dInfer delivers a $10\times$ speedup over Fast-dLLM while maintaining similar model performance. Even compared to the AR model (with a comparable number of activation parameters and performance) QWen2.5-3B, which is highly optimized with the latest vLLM inference engine, dInfer still delivers a $2$-$3\times$ speedup. The implementation of dInfer is open-sourced at https://github.com/inclusionAI/dInfer.
Authors: Xin Liu, Runsong Zhao, Pengcheng Huang, Xinyu Liu, Junyi Xiao, Chunyang Xiao, Tong Xiao, Shengxiang Gao, Zhengtao Yu, Jingbo Zhu
Abstract: Context compression presents a promising approach for accelerating large language model (LLM) inference by compressing long contexts into compact representations. Current context compression methods predominantly rely on autoencoding tasks to train context-agnostic compression tokens to compress contextual semantics. While autoencoding tasks enable compression tokens to acquire compression capabilities, compression via autoencoding tasks creates a fundamental mismatch: the models are optimized for reconstruction that diverge from actual downstream tasks, thereby weakening the features more beneficial for real-world usage. We propose Semantic-Anchor Compression (SAC), a novel method that shifts from autoencoding task based compression to an architecture that is equipped with this compression capability \textit{a priori}. Instead of training models to compress contexts through autoencoding tasks, SAC directly selects so-called anchor tokens from the original context and aggregates contextual information into their key-value (KV) representations. By deriving representations directly from the contextual tokens, SAC eliminates the need for autoencoding training. To ensure compression performance while directly leveraging anchor tokens, SAC incorporates two key designs: (1) anchor embeddings that enable the compressor to identify critical tokens, and (2) bidirectional attention modification that allows anchor tokens to capture information from the entire context. Experimental results demonstrate that SAC consistently outperforms existing context compression methods across various compression ratios. On out-of-distribution evaluation using MRQA, SAC achieves 1 EM improvement at 5x compression over strong baselines, with increasing advantages at higher compression ratios.
Authors: Enze Zhang, Jiaying Wang, Mengxi Xiao, Jifei Liu, Ziyan Kuang, Rui Dong, Eric Dong, Sophia Ananiadou, Min Peng, Qianqian Xie
Abstract: Large language models (LLMs) have substantially advanced machine translation (MT), yet their effectiveness in translating web novels remains unclear. Existing benchmarks rely on surface-level metrics that fail to capture the distinctive traits of this genre. To address these gaps, we introduce DITING, the first comprehensive evaluation framework for web novel translation, assessing narrative and cultural fidelity across six dimensions: idiom translation, lexical ambiguity, terminology localization, tense consistency, zero-pronoun resolution, and cultural safety, supported by over 18K expert-annotated Chinese-English sentence pairs. We further propose AgentEval, a reasoning-driven multi-agent evaluation framework that simulates expert deliberation to assess translation quality beyond lexical overlap, achieving the highest correlation with human judgments among seven tested automatic metrics. To enable metric comparison, we develop MetricAlign, a meta-evaluation dataset of 300 sentence pairs annotated with error labels and scalar quality scores. Comprehensive evaluation of fourteen open, closed, and commercial models reveals that Chinese-trained LLMs surpass larger foreign counterparts, and that DeepSeek-V3 delivers the most faithful and stylistically coherent translations. Our work establishes a new paradigm for exploring LLM-based web novel translation and provides public resources to advance future research.
Authors: Chenyang Gu, Yewen Pu, Bruce Yang, Xiaofan Li, Huan Gao
Abstract: Enhancing LLMs with the ability to actively search external knowledge is crucial for complex and real-world tasks. Current approaches either rely on prompting to elicit the model's innate agent capabilities, or suffer from performance ceilings and collapse when applying RL to complex interactive tasks, leaving their true agentic potential untapped. To address this, we introduce \textbf{D}ynamic-filter \textbf{S}equence-level \textbf{P}olicy \textbf{O}ptimization (DSPO), an improved RL algorithm designed for robust agent training through sequence-level optimization and dynamic sample filtering. We train our model purely through RL to interleave multi-turn search and reasoning, obviating the need for supervised demonstration data. Across multiple QA benchmarks, our DSPO-trained 7B model improves over a comparable previous work by \textbf{34.1\%}, and even outperforms the 14B model from previous work in complex multihop QA such as HotpotQA by nearly \textbf{9\% relative}, maintaining exceptional training stability.
Authors: Chenyu Wang, Paria Rashidinejad, DiJia Su, Song Jiang, Sid Wang, Siyan Zhao, Cai Zhou, Shannon Zejiang Shen, Feiyu Chen, Tommi Jaakkola, Yuandong Tian, Bo Liu
Abstract: Diffusion large language models (dLLMs) are emerging as an efficient alternative to autoregressive models due to their ability to decode multiple tokens in parallel. However, aligning dLLMs with human preferences or task-specific rewards via reinforcement learning (RL) is challenging because their intractable log-likelihood precludes the direct application of standard policy gradient methods. While prior work uses surrogates like the evidence lower bound (ELBO), these one-sided approximations can introduce significant policy gradient bias. To address this, we propose the Sandwiched Policy Gradient (SPG) that leverages both an upper and a lower bound of the true log-likelihood. Experiments show that SPG significantly outperforms baselines based on ELBO or one-step estimation. Specifically, SPG improves the accuracy over state-of-the-art RL methods for dLLMs by 3.6% in GSM8K, 2.6% in MATH500, 18.4% in Countdown and 27.0% in Sudoku.
Authors: James Y. Huang, Sailik Sengupta, Daniele Bonadiman, Yi-An Lai, Arshit Gupta, Nikolaos Pappas, Saab Mansour, Katrin Kirchhoff, Dan Roth
Abstract: Large Language Models (LLMs) are nowadays expected to generate content aligned with human preferences. Current work focuses on alignment at model training time, through techniques such as Reinforcement Learning with Human Feedback (RLHF). However, it is unclear if such methods are an effective choice to teach alignment objectives to the model. First, the inability to incorporate multiple, custom rewards and reliance on a model developer's view of universal and static principles are key limitations. Second, the reliability of such approaches is also questionable (e.g. susceptibility to jailbreaking even after safety training). To address these issues, we propose DeAL, a framework that allows the user to customize reward functions and enables Decoding-time Alignment of LLMs (DeAL). At its core, we view decoding as a heuristic-guided search process and facilitate the use of a wide variety of alignment objectives. Our experiments with programmatic constraints such as keyword and length constraints, and abstract objectives such as harmlessness and helpfulness, show that we can DeAL with fine-grained trade-offs and improve adherence to alignment objectives. Lastly, we demonstrate that DeAL is largely complementary to existing alignment strategies, and can be effectively paired with RLHF and prompting techniques to achieve better alignment.
Authors: Daking Rai, Yilun Zhou, Shi Feng, Abulhair Saparov, Ziyu Yao
Abstract: Mechanistic interpretability (MI) is an emerging sub-field of interpretability that seeks to understand a neural network model by reverse-engineering its internal computations. Recently, MI has garnered significant attention for interpreting transformer-based language models (LMs), resulting in many novel insights yet introducing new challenges. However, there has not been work that comprehensively reviews these insights and challenges, particularly as a guide for newcomers to this field. To fill this gap, we provide a comprehensive survey from a task-centric perspective, organizing the taxonomy of MI research around specific research questions or tasks. We outline the fundamental objects of study in MI, along with the techniques, evaluation methods, and key findings for each task in the taxonomy. In particular, we present a task-centric taxonomy as a roadmap for beginners to navigate the field by helping them quickly identify impactful problems in which they are most interested and leverage MI for their benefit. Finally, we discuss the current gaps in the field and suggest potential future directions for MI research.
Authors: Zafeirios Fountas, Martin A Benfeghoul, Adnan Oomerjee, Fenia Christopoulou, Gerasimos Lampouras, Haitham Bou-Ammar, Jun Wang
Abstract: Large language models (LLMs) have shown remarkable capabilities, but still struggle with processing extensive contexts, limiting their ability to maintain coherence and accuracy over long sequences. In contrast, the human brain excels at organising and retrieving episodic experiences across vast temporal scales, spanning a lifetime. In this work, we introduce EM-LLM, a novel approach that integrates key aspects of human episodic memory and event cognition into LLMs with no fine-tuning, enabling them to handle practically infinite context lengths while maintaining computational efficiency. EM-LLM organises sequences of tokens into coherent episodic events using a combination of Bayesian surprise and graph-theoretic boundary refinement in an online fashion. When needed, these events are retrieved through a two-stage memory process, combining similarity-based and temporally contiguous retrieval for efficient, human-inspired access to relevant information. Experiments on the LongBench and $\infty$-Bench benchmarks demonstrate EM-LLM's superior performance, consistently outperforming the state-of-the-art retrieval model InfLLM across various baseline LLMs. In addition, EM-LLM outperforms its popular counterpart, RAG, in a wide range of tasks, while requiring similar resources. Notably, EM-LLM's performance even surpasses full-context models in most tasks, while successfully performing retrieval across 10 million tokens -- a scale computationally infeasible for such models. Finally, our analysis reveals strong correlations between EM-LLM's event segmentation and human-perceived events, suggesting parallels between this artificial system and its biological counterpart, thereby offering a novel computational framework for exploring human memory mechanisms.
Authors: Yanxi Chen, Yaliang Li, Bolin Ding, Jingren Zhou
Abstract: This work presents an analytical framework for the design and analysis of LLM-based algorithms, i.e., algorithms that contain one or multiple calls of large language models (LLMs) as sub-routines and critically rely on the capabilities of LLMs. While such algorithms, ranging from basic LLM calls with prompt engineering to complicated LLM-powered agentic workflows and compound AI systems, have achieved remarkable empirical success, their design and optimization oftentimes require extensive trial-and-errors and case-by-case analysis. Our proposed framework serves as an attempt to mitigate such headaches, offering a formal and systematic approach for analyzing how the accuracy and efficiency of an LLM-based algorithm will be impacted by critical design choices, such as the pattern and granularity of task decomposition, or the prompt for each LLM call. Through a wide range of case studies covering diverse algorithm patterns (including parallel/hierarchical/recursive task decomposition and generic directed acyclic graphs), we demonstrate the proposed framework in action and derive interesting insights that generalize across scenarios, accompanied by systematic empirical validation in synthetic settings.
Authors: Shang Wang, Tianqing Zhu, Dayong Ye, Wanlei Zhou
Abstract: The deployment of large language models (LLMs) like ChatGPT and Gemini has shown their powerful natural language generation capabilities. However, these models can inadvertently learn and retain sensitive information and harmful content during training, raising significant ethical and legal concerns. To address these issues, machine unlearning has been introduced as a potential solution. While existing unlearning methods take into account the specific characteristics of LLMs, they often suffer from high computational demands, limited applicability, or the risk of catastrophic forgetting. To address these limitations, we propose a lightweight behavioral unlearning framework based on Retrieval-Augmented Generation (RAG) technology. By modifying the external knowledge base of RAG, we simulate the effects of forgetting without directly interacting with the unlearned LLM. We approach the construction of unlearned knowledge as a constrained optimization problem, deriving two key components that underpin the effectiveness of RAG-based unlearning. This RAG-based approach is particularly effective for closed-source LLMs, where existing unlearning methods often fail. We evaluate our framework through extensive experiments on both open-source and closed-source models, including ChatGPT, Gemini, Llama-2-7b-chat, and PaLM 2. The results demonstrate that our approach meets five key unlearning criteria: effectiveness, universality, harmlessness, simplicity, and robustness. Meanwhile, this approach can extend to multimodal large language models and LLM-based agents.
Authors: Kemal Kurniawan, Meladel Mistica, Timothy Baldwin, Jey Han Lau
Abstract: Human label variation (HLV) challenges the standard assumption that a labelled instance has a single ground truth, instead embracing the natural variation in human annotation to train and evaluate models. While various training methods and metrics for HLV have been proposed, it is still unclear which methods and metrics perform best in what settings. We propose new evaluation metrics for HLV leveraging fuzzy set theory. Since these new proposed metrics are differentiable, we then in turn experiment with employing these metrics as training objectives. We conduct an extensive study over 6 HLV datasets testing 14 training methods and 6 evaluation metrics. We find that training on either disaggregated annotations or soft labels performs best across metrics, outperforming training using the proposed training objectives with differentiable metrics. We also show that our proposed soft micro F1 score is one of the best metrics for HLV data.
Authors: Shashank Kirtania, Arun Iyer
Abstract: Recent advances in automated theorem proving use Large Language Models (LLMs) to translate informal mathematical statements into formal proofs. However, informal cues are often ambiguous or lack strict logical structure, making it hard for models to interpret them precisely. While existing methods achieve strong performance, little is known about how LLMs internally represent informal cues, or how these influence proof generation. To address this, we explore \textit{activation steering}, an inference-time intervention that identifies linear directions in residual activations associated with informal reasoning traces and adjusts them to improve proof construction without fine-tuning. This mechanism also yields interpretable information about how reasoning is internally encoded in the activation space of LLMs. We test our method for generating formal proofs from already-formalized theorems. Our contributions are twofold: (1) a novel activation-based intervention for guiding proof synthesis in LLMs; and (2) demonstration that this intervention improves performance under two decoding strategies (sampling and best-first search) without any further training.
Authors: Felix Stollenwerk
Abstract: Layer normalization (LN) is an essential component of modern neural networks. While many alternative techniques have been proposed, none of them have succeeded in replacing LN so far. The latest suggestion in this line of research is a dynamic activation function called Dynamic Tanh (DyT). Although it is empirically well-motivated and appealing from a practical point of view, it lacks a theoretical foundation. In this work, we shed light on the mathematical relationship between LN and dynamic activation functions. In particular, we derive DyT from the LN variant RMSNorm, and show that a well-defined decoupling in derivative space as well as an approximation are needed to do so. By applying the same decoupling procedure directly in function space, we are able to omit the approximation and obtain the exact element-wise counterpart of RMSNorm, which we call Dynamic Inverse Square Root Unit (DyISRU). We demonstrate numerically that DyISRU reproduces the normalization effect on outliers more accurately than DyT does.
Authors: Zhihan Zhou, Feng Hong, Jiaan Luo, Jiangchao Yao, Dongsheng Li, Bo Han, Ya Zhang, Yanfeng Wang
Abstract: We propose L2T, an advancement of visual instruction tuning (VIT). While VIT equips Multimodal LLMs (MLLMs) with promising multimodal capabilities, the current design choices for VIT often result in overfitting and shortcut learning, potentially degrading performance. This gap arises from an overemphasis on instruction-following abilities, while neglecting the proactive understanding of visual information. Inspired by this, L2T adopts a simple yet effective approach by incorporating the loss function into both the instruction and response sequences. It seamlessly expands the training data, and regularizes the MLLMs from overly relying on language priors. Based on this merit, L2T achieves a significant relative improvement of up to 9% on comprehensive multimodal benchmarks, requiring no additional training data and incurring negligible computational overhead. Surprisingly, L2T attains exceptional fundamental visual capabilities, yielding up to an 18% improvement in captioning performance, while simultaneously alleviating hallucination in MLLMs. Github code: https://github.com/Feng-Hong/L2T.
Authors: Shengao Wang, Arjun Chandra, Aoming Liu, Venkatesh Saligrama, Boqing Gong
Abstract: Human infants rapidly develop visual reasoning skills from minimal input, suggesting that developmentally inspired pretraining could significantly enhance the efficiency of vision-language models (VLMs). Although recent efforts have leveraged infant-inspired datasets like SAYCam, existing evaluation benchmarks remain misaligned--they are either too simplistic, narrowly scoped, or tailored for large-scale pretrained models. Additionally, training exclusively on infant data overlooks the broader, diverse input from which infants naturally learn. To address these limitations, we propose BabyVLM, a novel framework comprising comprehensive in-domain evaluation benchmarks and a synthetic training dataset created via child-directed transformations of existing datasets. We demonstrate that VLMs trained with our synthetic dataset achieve superior performance on BabyVLM tasks compared to models trained solely on SAYCam or general-purpose data of the SAYCam size. BabyVLM thus provides a robust, developmentally aligned evaluation tool and illustrates how compact models trained on carefully curated data can generalize effectively, opening pathways toward data-efficient vision-language learning paradigms.
Authors: Zhiqing Cui, Jiahao Yuan, Hanqing Wang, Yanshu Li, Chenxu Du, Zhenglong Ding
Abstract: Scientific diagrams are vital tools for communicating structured knowledge across disciplines. However, they are often published as static raster images, losing symbolic semantics and limiting reuse. While Multimodal Large Language Models (MLLMs) offer a pathway to bridging vision and structure, existing methods lack semantic control and structural interpretability, especially on complex diagrams. We propose Draw with Thought (DwT), a training-free framework that guides MLLMs to reconstruct diagrams into editable mxGraph XML code through cognitively-grounded Chain-of-Thought reasoning. DwT enables interpretable and controllable outputs without model fine-tuning by dividing the task into two stages: Coarse-to-Fine Planning, which handles perceptual structuring and semantic specification, and Structure-Aware Code Generation, enhanced by format-guided refinement. To support evaluation, we release Plot2XML, a benchmark of 247 real-world scientific diagrams with gold-standard XML annotations. Extensive experiments across eight MLLMs show that our approach yields high-fidelity, semantically aligned, and structurally valid reconstructions, with human evaluations confirming strong alignment in both accuracy and visual aesthetics, offering a scalable solution for converting static visuals into executable representations and advancing machine understanding of scientific graphics.
Authors: Zhenting Wang, Guofeng Cui, Yu-Jhe Li, Kun Wan, Wentian Zhao
Abstract: Recent advances in reinforcement learning (RL)-based post-training have led to notable improvements in large language models (LLMs), particularly in enhancing their reasoning capabilities to handle complex tasks. However, most existing methods treat the training data as a unified whole, overlooking the fact that modern LLM training often involves a mixture of data from diverse distributions-varying in both source and difficulty. This heterogeneity introduces a key challenge: how to adaptively schedule training across distributions to optimize learning efficiency. In this paper, we present a principled curriculum learning framework grounded in the notion of distribution-level learnability. Our core insight is that the magnitude of policy advantages reflects how much a model can still benefit from further training on a given distribution. Based on this, we propose a distribution-level curriculum learning framework for RL-based LLM post-training, which leverages the Upper Confidence Bound (UCB) principle to dynamically adjust sampling probabilities for different distrubutions. This approach prioritizes distributions with either high average advantage (exploitation) or low sample count (exploration), yielding an adaptive and theoretically grounded training schedule. We instantiate our curriculum learning framework with GRPO as the underlying RL algorithm and demonstrate its effectiveness on logic reasoning datasets with multiple difficulties and sources. Our experiments show that our framework significantly improves convergence speed and final performance, highlighting the value of distribution-aware curriculum strategies in LLM post-training. Code: https://github.com/ZhentingWang/DUMP.
Authors: Menglan Chen, Xianghe Pang, Jingjing Dong, WenHao Wang, Yaxin Du, Siheng Chen
Abstract: Aligning Vision-Language Models (VLMs) with safety standards is essential to mitigate risks arising from their multimodal complexity, where integrating vision and language unveils subtle threats beyond the reach of conventional safeguards. Inspired by the insight that reasoning across modalities is key to preempting intricate vulnerabilities, we propose a novel direction for VLM safety: multimodal reasoning-driven prompt rewriting. To this end, we introduce VLMGuard-R1, a proactive framework that refines user inputs through a reasoning-guided rewriter, dynamically interpreting text-image interactions to deliver refined prompts that bolster safety across diverse VLM architectures without altering their core parameters. To achieve this, we devise a three-stage reasoning pipeline to synthesize a dataset that trains the rewriter to infer subtle threats, enabling tailored, actionable responses over generic refusals. Extensive experiments across three benchmarks with five VLMs reveal that VLMGuard-R1 outperforms four baselines. In particular, VLMGuard-R1 achieves a remarkable 43.59\% increase in average safety across five models on the SIUO benchmark.
Authors: Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Yang Yue, Shiji Song, Gao Huang
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has recently demonstrated notable success in enhancing the reasoning performance of large language models (LLMs), particularly on mathematics and programming tasks. Similar to how traditional RL helps agents explore and learn new strategies, RLVR is believed to enable LLMs to continuously self-improve, thus acquiring novel reasoning abilities beyond those of the corresponding base models. In this study we critically examine the current state of RLVR by systematically probing the reasoning capability boundaries of RLVR-trained LLMs across various model families, RL algorithms, and math, coding, and visual reasoning benchmarks, using pass@k at large k values as the evaluation metric. Surprisingly, we find that the current training setup does not elicit fundamentally new reasoning patterns. While RLVR-trained models outperform their base models at small k (e.g., k = 1), the base models achieve a higher pass@k score when k is large. Coverage and perplexity analyses show that the observed reasoning abilities originate from and are bounded by the base model. Treating the base model as an upper bound, our quantitative analysis shows that six popular RLVR algorithms perform similarly and remain far from optimal in leveraging the potential of the base model. By contrast, we find that distillation can introduce new reasoning patterns from the teacher and genuinely expand the model's reasoning capabilities. Overall, our findings suggest that current RLVR methods have not yet realized the potential of RL to elicit truly novel reasoning abilities in LLMs. This highlights the need for improved RL paradigms, such as continual scaling and multi-turn agent-environment interaction, to unlock this potential.
Authors: Jiaqi Wei, Hao Zhou, Xiang Zhang, Di Zhang, Zijie Qiu, Wei Wei, Jinzhe Li, Wanli Ouyang, Siqi Sun
Abstract: Retrieval-augmented generation (RAG) has become a widely adopted paradigm for enabling knowledge-grounded large language models (LLMs). However, standard RAG pipelines often fail to ensure that model reasoning remains consistent with the evidence retrieved, leading to factual inconsistencies or unsupported conclusions. In this work, we reinterpret RAG as Retrieval-Augmented Reasoning and identify a central but underexplored problem: Reasoning Misalignment -- the divergence between an LLM's internal reasoning trajectory and the evidential constraints provided by retrieval. To address this issue, we propose AlignRAG, a novel iterative framework grounded in Critique-Driven Alignment (CDA). We further introduce AlignRAG-auto, an autonomous variant that dynamically terminates refinement, removing the need to pre-specify the number of critique iterations. At the heart of AlignRAG lies a contrastive critique synthesis mechanism that generates retrieval-sensitive critiques while mitigating self-bias. This mechanism trains a dedicated retrieval-augmented Critic Language Model (CLM) using labeled critiques that distinguish between evidence-aligned and misaligned reasoning. Empirical evaluations show that our approach significantly improves reasoning fidelity. Our 8B-parameter CLM improves performance over the Self-Refine baseline by 12.1% on out-of-domain tasks and outperforms a standard 72B-parameter CLM by 2.2%. Furthermore, AlignRAG-auto achieves this state-of-the-art performance while dynamically determining the optimal number of refinement steps, enhancing efficiency and usability. AlignRAG remains compatible with existing RAG architectures as a plug-and-play module and demonstrates strong robustness under both informative and noisy retrieval scenarios.
Authors: Adam Younsi, Ahmed Attia, Abdalgader Abubaker, Mohamed El Amine Seddik, Hakim Hacid, Salem Lahlou
Abstract: Achieving both accuracy and diverse reasoning remains challenging for Large Language Models (LLMs) in complex domains like mathematics. A key bottleneck is evaluating intermediate reasoning steps to guide generation without costly human annotations. To address this, we first introduce a novel Process Reward Model (PRM) trained automatically using Monte Carlo Tree Search coupled with a similarity-based data augmentation technique, effectively capturing step-level reasoning quality. Leveraging this PRM, we then adapt Generative Flow Networks (GFlowNets) to operate at the reasoning step level. Unlike traditional reinforcement learning focused on maximizing a single reward, GFlowNets naturally sample diverse, high-quality solutions proportional to their rewards, as measured by our PRM. Empirical evaluation shows strong improvements in both accuracy and solution diversity on challenging mathematical benchmarks (e.g., +2.59% absolute accuracy on MATH Level 5 for Llama3.2-3B), with effective generalization to unseen datasets (+9.4\% absolute on SAT MATH). Furthermore, we benchmark our PRM against existing open-source reward models, demonstrating superior alignment with reasoning quality and more consistent guidance for downstream generation. Our work demonstrates the potential of PRM-guided, step-level GFlowNets for developing more robust and versatile mathematical reasoning in LLMs.
Authors: Trinh T. L. Vuong, Jin Tae Kwak
Abstract: We present ViDRiP-LLaVA, the first large multimodal model (LMM) in computational pathology that integrates three distinct image scenarios, including single patch images, automatically segmented pathology video clips, and manually segmented pathology videos. This integration closely mirrors the natural diagnostic process of pathologists. By generating detailed histological descriptions and culminating in a definitive sign-out diagnosis, ViDRiP-LLaVA bridges visual narratives with diagnostic reasoning. Central to our approach is the ViDRiP-Instruct dataset, comprising 4278 video and diagnosis-specific chain-of-thought instructional pairs sourced from educational histopathology videos on YouTube. Although high-quality data is critical for enhancing diagnostic reasoning, its creation is time-intensive and limited in volume. To overcome this challenge, we transfer knowledge from existing single-image instruction datasets to train on weakly annotated, keyframe-extracted clips, followed by fine-tuning on manually segmented videos. ViDRiP-LLaVA establishes a new benchmark in pathology video analysis and offers a promising foundation for future AI systems that support clinical decision-making through integrated visual and diagnostic reasoning. Our code, data, and model are publicly available at: https://github.com/QuIIL/ViDRiP-LLaVA.
Authors: Neeloy Chakraborty, John Pohovey, Melkior Ornik, Katherine Driggs-Campbell
Abstract: Large language models (LLMs) have recently demonstrated success in generalizing across decision-making tasks including planning, control, and prediction, but their tendency to hallucinate unsafe and undesired outputs poses risks. We argue that detecting such failures is necessary, especially in safety-critical scenarios. Existing methods for black-box models often detect hallucinations by identifying inconsistencies across multiple samples. Many of these approaches typically introduce prompt perturbations like randomizing detail order or generating adversarial inputs, with the intuition that a confident model should produce stable outputs. We first perform a manual case study showing that other forms of perturbations (e.g., adding noise, removing sensor details) cause LLMs to hallucinate in a multi-agent driving environment. We then propose a novel method for efficiently searching the space of prompt perturbations using adaptive stress testing (AST) with Monte-Carlo tree search (MCTS). Our AST formulation enables discovery of scenarios and prompts that cause language models to act with high uncertainty or even crash. By generating MCTS prompt perturbation trees across diverse scenarios, we show through extensive experiments that offline analyses can be used at runtime to automatically generate prompts that influence model uncertainty, and to inform real-time trust assessments of an LLM. We further characterize LLMs deployed as planners in a single-agent lunar lander environment and in a multi-agent robot crowd navigation simulation. Overall, ours is one of the first hallucination intervention algorithms to pave a path towards rigorous characterization of black-box LLM planners.
Authors: Carl Edwards, Chi Han, Gawon Lee, Thao Nguyen, Sara Szymku\'c, Chetan Kumar Prasad, Bowen Jin, Jiawei Han, Ying Diao, Ge Liu, Hao Peng, Bartosz A. Grzybowski, Martin D. Burke, Heng Ji
Abstract: Despite their ability to understand chemical knowledge, large language models (LLMs) remain limited in their capacity to propose novel molecules with desired functions (e.g., drug-like properties). In addition, the molecules that LLMs propose can often be challenging to make, and are almost never compatible with automated synthesis approaches. To better enable the discovery of functional small molecules, LLMs need to learn a new molecular language that is more effective in predicting properties and inherently synced with automated synthesis technology. Current molecule LLMs are limited by representing molecules based on atoms. In this paper, we argue that just like tokenizing texts into meaning-bearing (sub-)word tokens instead of characters, molecules should be tokenized at the level of functional building blocks, i.e., parts of molecules that bring unique functions and serve as effective building blocks for real-world automated laboratory synthesis. This motivates us to propose mCLM, a modular Chemical-Language Model that comprises a bilingual language model that understands both natural language descriptions of functions and molecular blocks. mCLM front-loads synthesizability considerations while improving the predicted functions of molecules in a principled manner. mCLM, with only 3B parameters, achieves improvements in synthetic accessibility relative to 7 other leading generative AI methods including GPT-5. When tested on 122 out-of-distribution medicines using only building blocks/tokens that are compatible with automated modular synthesis, mCLM outperforms all baselines in property scores and synthetic accessibility. mCLM can also reason on multiple functions and iteratively self-improve to rescue drug candidates that failed late in clinical trials ("fallen angels").
Authors: Haokun Chen, Yueqi Zhang, Yuan Bi, Yao Zhang, Tong Liu, Jinhe Bi, Jian Lan, Jindong Gu, Claudia Grosser, Denis Krompass, Nassir Navab, Volker Tresp
Abstract: In recent years, Large Language Models (LLMs) have achieved remarkable advancements, drawing significant attention from the research community. Their capabilities are largely attributed to large-scale architectures, which require extensive training on massive datasets. However, such datasets often contain sensitive or copyrighted content sourced from the public internet, raising concerns about data privacy and ownership. Regulatory frameworks, such as the General Data Protection Regulation (GDPR), grant individuals the right to request the removal of such sensitive information. This has motivated the development of machine unlearning algorithms that aim to remove specific knowledge from models without the need for costly retraining. Despite these advancements, evaluating the efficacy of unlearning algorithms remains a challenge due to the inherent complexity and generative nature of LLMs. In this work, we introduce a comprehensive auditing framework for unlearning evaluation, comprising three benchmark datasets, six unlearning algorithms, and five prompt-based auditing methods. By using various auditing algorithms, we evaluate the effectiveness and robustness of different unlearning strategies. To explore alternatives beyond prompt-based auditing, we propose a novel technique that leverages intermediate activation perturbations, addressing the limitations of auditing methods that rely solely on model inputs and outputs.
Authors: Zifu Wang, Junyi Zhu, Bo Tang, Zhiyu Li, Feiyu Xiong, Jiaqian Yu, Matthew B. Blaschko
Abstract: The application of rule-based reinforcement learning (RL) to multimodal large language models (MLLMs) introduces unique challenges and potential deviations from findings in text-only domains, particularly for perception-heavy tasks. This paper provides a comprehensive study of rule-based visual RL, using jigsaw puzzles as a structured experimental framework. Jigsaw puzzles offer inherent ground truth, adjustable difficulty, and demand complex decision-making, making them ideal for this study. Our research reveals several key findings: \textit{Firstly,} we find that MLLMs, initially performing near to random guessing on the simplest jigsaw puzzles, achieve near-perfect accuracy and generalize to complex, unseen configurations through fine-tuning. \textit{Secondly,} training on jigsaw puzzles can induce generalization to other visual tasks, with effectiveness tied to specific task configurations. \textit{Thirdly,} MLLMs can learn and generalize with or without explicit reasoning, though open-source models often favor direct answering. Consequently, even when trained for step-by-step reasoning, they can ignore the thinking process in deriving the final answer. \textit{Fourthly,} we observe that complex reasoning patterns appear to be pre-existing rather than emergent, with their frequency increasing alongside training and task difficulty. \textit{Finally,} our results demonstrate that RL exhibits more effective generalization than Supervised Fine-Tuning (SFT), and an initial SFT cold start phase can hinder subsequent RL optimization. Although these observations are based on jigsaw puzzles and may vary across other visual tasks, this research contributes a valuable piece of jigsaw to the larger puzzle of collective understanding rule-based visual RL and its potential in multimodal learning. The code is available at: https://github.com/zifuwanggg/Jigsaw-R1
Authors: Ruida Wang, Yuxin Li, Yi R. Fung, Tong Zhang
Abstract: Enhancing the mathematical reasoning capabilities of LLMs has garnered significant attention in both the mathematical and computer science communities. Recent works have made substantial progress in both Natural Language (NL) reasoning and Formal Language (FL) reasoning by leveraging the potential of pure Reinforcement Learning (RL) methods on base models. However, RL approaches struggle to impart new capabilities not presented in the base model, highlighting the need to integrate more knowledge like FL into NL math reasoning effectively. Yet, this integration is challenging due to inherent disparities in problem structure and reasoning format between NL and FL. To address these challenges, we introduce **NL-FL HybridReasoning (NFL-HR)**, an end-to-end framework designed to incorporate the FL expert into NL math problem-solving. To bridge the NL and FL input format gap, we propose the NL-FL Problem Alignment method, which reformulates the Question-Answering (QA) problems in NL as existence theorems in FL. Subsequently, the Mixed Problem Input technique we provide enables the FL reasoner to handle both QA and existence problems concurrently. Lastly, we mitigate the NL and FL output format gap in reasoning through an LLM-based Answer Extraction mechanism. Comprehensive experiments demonstrate that the NFL-HR framework achieves **89.80**% and **84.34%** accuracy rates on the MATH-500 and the AMC benchmarks, surpassing the NL baseline by **4.60%** and **4.82%**, respectively. Notably, some problems resolved by our framework remain unsolved by the NL baseline model even under a larger number of trials.
Authors: Hairu Wang, Yuan Feng, Yukun Cao, Xike Xie, S Kevin Zhou
Abstract: Large language models excel at many tasks but often incur high inference costs during deployment. To mitigate hallucination, many systems use a knowledge graph to enhance retrieval-augmented generation (KG-RAG). However, the large amount of retrieved knowledge contexts increase these inference costs further. A promising solution to balance performance and cost is LLM routing, which directs simple queries to smaller LLMs and complex ones to larger LLMs. However, no dedicated routing methods currently exist for RAG, and existing training-based routers face challenges scaling to this domain due to the need for extensive training data. We observe that the score distributions produced by the retrieval scorer strongly correlate with query difficulty. Based on this, we propose an extremely simple yet effective routing framework, the first specifically designed for KG-RAG that efficiently balances performance and cost in a plug-and-play manner. It delivers over 3x higher routing effectiveness while reducing runtime to less than 0.001x compared to existing methods. Our code is available at https://github.com/hrwang00/SkewRoute.
Authors: James R. Golden
Abstract: Despite significant progress in transformer interpretability, an understanding of the computational mechanisms of large language models (LLMs) remains a fundamental challenge. Many approaches interpret a network's hidden representations but remain agnostic about how those representations are generated. We address this by mapping LLM inference for a given input sequence to an equivalent and interpretable linear system which reconstructs the predicted output embedding with relative error below $10^{-13}$ at double floating-point precision, requiring no additional model training. We exploit a property of transformers wherein every operation (gated activations, attention, and normalization) can be expressed as $A(x) \cdot x$, where $A(x)$ represents an input-dependent linear transform and $x$ preserves the linear pathway. To expose this linear structure, we strategically detach components of the gradient computation with respect to an input sequence, freezing the $A(x)$ terms at their values computed during inference, such that the Jacobian yields an equivalent linear mapping. This detached Jacobian of the model reconstructs the output with one linear operator per input token, which is shown for Qwen 3, Gemma 3 and Llama 3, up to Qwen 3 14B. These linear representations demonstrate that LLMs operate in extremely low-dimensional subspaces where the singular vectors can be decoded to interpretable semantic concepts. The computation for each intermediate output also has a linear equivalent, and we examine how the linear representations of individual layers and their attention and multilayer perceptron modules build predictions, and use these as steering operators to insert semantic concepts into unrelated text. Despite their global nonlinearity, LLMs can be interpreted through equivalent linear representations that reveal low-dimensional semantic structures in the next-token prediction process.
Authors: Xiao Yu, Baolin Peng, Ruize Xu, Michel Galley, Hao Cheng, Suman Nath, Jianfeng Gao, Zhou Yu
Abstract: Recent progress in reasoning with large language models (LLMs), such as DeepSeek-R1, demonstrates impressive capabilities in domains like mathematics and coding, by exhibiting complex cognitive behaviors such as verification, goal decomposition, and self-reflection. However, it is unclear what behavior is effective and what behavior is missing for long-horizon AI agents tasks. In this work, we propose Dyna-Think, a thinking framework that integrates planning with an internal world model with reasoning and acting to enhance AI agent performance. To enable Dyna-Think, we propose Dyna-Think Imitation Learning (DIT) and Dyna-Think Dyna Training (DDT). To initialize a policy with Dyna-Think, DIT reconstructs the thinking process of R1 to focus on performing world model simulation relevant to the proposed (and planned) action, and trains the policy using this reconstructed data. To enhance Dyna-Think, DDT uses a two-stage training process to first improve the agent's world modeling ability via objectives such as state prediction or critique generation, and then improve the agent's action via policy training. We evaluate our methods on OSWorld and WindowsAgentArena, and demonstrate that Dyna-Think improves the agent's in-domain and out-of-domain performance, achieving similar best-of-n performance compared to R1 while generating 2x less tokens on average. Our extensive empirical studies reveal that 1) using critique generation for world model training is effective to improve policy performance; and 2) AI agents with better performance correlate with better world modeling abilities. We believe our results suggest a promising research direction to integrate world model simulation into AI agents to enhance their reasoning, planning, and acting capabilities.
Authors: Beomsik Cho, Jaehyung Kim
Abstract: Large Vision-Language Models (LVLMs) achieve strong performance across multimodal tasks by integrating visual perception with language understanding. However, how vision information contributes to the model's decoding process remains under-explored, as reflected in frequent hallucinations. Through a series of analyses, we found that (i) vision tokens provide meaningful visual information even when hallucinations occur, and (ii) their semantics are encoded in the textual space and become explicit under appropriate vocabulary constraints. Building on these observations, we propose ReVisiT, a simple training-free decoding method that references vision tokens to guide text generation. Our approach leverages the semantic information embedded within vision tokens by projecting them into the text token distribution. Specifically, ReVisiT dynamically selects the most relevant vision token at each decoding step via context-aware constrained divergence minimization, and using its constrained projection to refine the output distribution to better incorporate visual semantics. Across five benchmarks on recent LVLMs, ReVisiT consistently enhances visual grounding with minimal computational overhead, and achieves competitive or superior results to state-of-the-art decoding baselines while reducing computational cost by up to $2\times$.
Authors: Moussa Koulako Bala Doumbouya, Dan Jurafsky, Christopher D. Manning
Abstract: Work in psychology has highlighted that the geometric model of similarity standard in deep learning is not psychologically plausible because its metric properties such as symmetry do not align with human perception of similarity. In contrast, Tversky (1977) proposed an axiomatic theory of similarity with psychological plausibility based on a representation of objects as sets of features, and their similarity as a function of their common and distinctive features. This model of similarity has not been used in deep learning before, in part because of the challenge of incorporating discrete set operations. In this paper, we develop a differentiable parameterization of Tversky's similarity that is learnable through gradient descent, and derive basic neural network building blocks such as the Tversky projection layer, which unlike the linear projection layer can model non-linear functions such as XOR. Through experiments with image recognition and language modeling neural networks, we show that the Tversky projection layer is a beneficial replacement for the linear projection layer. For instance, on the NABirds image classification task, a frozen ResNet-50 adapted with a Tversky projection layer achieves a 24.7% relative accuracy improvement over the linear layer adapter baseline. With Tversky projection layers, GPT-2's perplexity on PTB decreases by 7.8%, and its parameter count by 34.8%. Finally, we propose a unified interpretation of both types of projection layers as computing similarities of input stimuli to learned prototypes for which we also propose a novel visualization technique highlighting the interpretability of Tversky projection layers. Our work offers a new paradigm for thinking about the similarity model implicit in modern deep learning, and designing neural networks that are interpretable under an established theory of psychological similarity.
Authors: Dongjie Yang, Chengqiang Lu, Qimeng Wang, Xinbei Ma, Yan Gao, Yao Hu, Hai Zhao
Abstract: Unlike reasoning, which often entails a deep sequence of deductive steps, complex real-world planning is characterized by the need to synthesize a broad spectrum of parallel and potentially conflicting information and constraints. For example, in travel planning scenarios, it requires the integration of diverse real-world information and user preferences. While LLMs show promise, existing methods with long-horizon thinking struggle with handling multifaceted constraints, leading to suboptimal solutions. Motivated by the challenges of real-world travel planning, this paper introduces the Multiple Aspects of Planning (MAoP), empowering LLMs with "wide-horizon thinking" to solve planning problems with multifaceted constraints. Instead of direct planning, MAoP leverages the strategist to conduct pre-planning from various aspects and provide the planning blueprint for planners, enabling strong inference-time scalability by scaling aspects to consider various constraints. In addition, existing benchmarks for multi-constraint planning are flawed because they assess constraints in isolation, ignoring causal dependencies within the constraints, e.g, travel planning, where past activities dictate future itinerary. To address this, we propose Travel-Sim, an agent-based benchmark assessing plans via real-world simulation, thereby inherently resolving these causal dependencies. This paper advances LLM capabilities in complex planning and offers novel insights for evaluating sophisticated scenarios through simulation.
Authors: Andrey Goncharov, Daniil Vyazhev, Petr Sychev, Edvard Khalafyan, Alexey Zaytsev
Abstract: General-purpose Large Language Models (LLMs) are frequently fine-tuned through supervised fine-tuning (SFT) to enhance performance in specific domains. Better results can be achieved by distilling the chain-of-thought of a larger model at the cost of numerous expensive calls and a much greater amount of data. We propose a novel blueprint for efficient fine-tuning that uses reasoning only for complex data identified by entropy. Specifically, across two small open models ($~3B$) we split the training data into complexity categories by a single token answer entropy (ROC AUC $0.73$), fine-tune large language models (LLMs) via SFT and distillation, and show that our pipeline significantly outperforms the standard SFT approach ($0.58$ vs $0.45$ average accuracy) and outperforms the distillation approach ($0.58$ vs $0.56$ average accuracy) while using $81%$ less data.
Authors: Xiangrui Liu, Man Luo, Agneet Chatterjee, Hua Wei, Chitta Baral, Yezhou Yang
Abstract: Hallucination is a long-standing problem that has been actively investigated in Vision-Language Models (VLMs). Existing research commonly attributes hallucinations to technical limitations or sycophancy bias, where the latter means the models tend to generate incorrect answers to align with user expectations. However, these explanations primarily focus on technical or externally driven factors, and may have neglected the possibility that hallucination behaviours might mirror cognitive biases observed in human psychology. In this work, we introduce a psychological taxonomy, categorizing VLMs' cognitive biases that lead to hallucinations, including sycophancy, logical inconsistency, and a newly identified VLMs behaviour: appeal to authority. To systematically analyze these behaviours, we design AIpsych, a scalable benchmark that reveals psychological tendencies in model response patterns. Leveraging this benchmark, we investigate how variations in model architecture and parameter size influence model behaviour when responding to strategically manipulated questions. Our experiments reveal that as model size increases, VLMs exhibit stronger sycophantic tendencies but reduced authority bias, suggesting increasing competence but a potential erosion of response integrity. A human subject study further validates our hypotheses and highlights key behavioural differences between VLMs and human respondents. This work suggests a new perspective for understanding hallucination in VLMs and highlights the importance of integrating psychological principles into model evaluation.
Authors: Anita Kriz, Elizabeth Laura Janes, Xing Shen, Tal Arbel
Abstract: Multimodal large language models (MLLMs) hold considerable promise for applications in healthcare. However, their deployment in safety-critical settings is hindered by two key limitations: (i) sensitivity to prompt design, and (ii) a tendency to generate incorrect responses with high confidence. As clinicians may rely on a model's stated confidence to gauge the reliability of its predictions, it is especially important that when a model expresses high confidence, it is also highly accurate. We introduce Prompt4Trust, the first reinforcement learning (RL) framework for prompt augmentation targeting confidence calibration in MLLMs. A lightweight LLM is trained to produce context-aware auxiliary prompts that guide a downstream task MLLM to generate responses in which the expressed confidence more accurately reflects predictive accuracy. Unlike conventional calibration techniques, Prompt4Trust specifically prioritizes aspects of calibration most critical for safe and trustworthy clinical decision-making. Beyond improvements driven by this clinically motivated calibration objective, our proposed method also improves task accuracy, achieving state-of-the-art medical visual question answering (VQA) performance on the PMC-VQA benchmark, which is composed of multiple-choice questions spanning diverse medical imaging modalities. Moreover, our framework trained with a small downstream task MLLM showed promising zero-shot generalization to larger MLLMs in our experiments, suggesting the potential for scalable calibration without the associated computational costs. This work demonstrates the potential of automated yet human-aligned prompt engineering for improving the the trustworthiness of MLLMs in safety critical settings. Our codebase can be found at https://github.com/xingbpshen/prompt4trust.
Authors: Yinghao Zhu, Yifan Qi, Zixiang Wang, Lei Gu, Dehao Sui, Haoran Hu, Xichen Zhang, Ziyi He, Junjun He, Liantao Ma, Lequan Yu
Abstract: The rapid proliferation of scientific knowledge presents a grand challenge: transforming this vast repository of information into an active engine for discovery, especially in high-stakes domains like healthcare. Current AI agents, however, are constrained by static, predefined strategies, limiting their ability to navigate the complex, evolving ecosystem of scientific research. This paper introduces HealthFlow, a self-evolving AI agent that overcomes this limitation through a novel meta-level evolution mechanism. HealthFlow autonomously refines its high-level problem-solving policies by distilling procedural successes and failures into a durable, structured knowledge base, enabling it to learn not just how to use tools, but how to strategize. To anchor our research and provide a community resource, we introduce EHRFlowBench, a new benchmark featuring complex health data analysis tasks systematically derived from peer-reviewed scientific literature. Our experiments demonstrate that HealthFlow's self-evolving approach significantly outperforms state-of-the-art agent frameworks. This work offers a new paradigm for intelligent systems that can learn to operationalize the procedural knowledge embedded in scientific content, marking a critical step toward more autonomous and effective AI for healthcare scientific discovery.
Authors: Alexander Golubev, Maria Trofimova, Sergei Polezhaev, Ibragim Badertdinov, Maksim Nekrashevich, Anton Shevtsov, Simon Karasik, Sergey Abramov, Andrei Andriushchenko, Filipp Fisin, Sergei Skvortsov, Boris Yangel
Abstract: Research on applications of reinforcement learning (RL) to large language models has mostly been focused on single-turn problems, such as mathematical reasoning or single-shot code generation. While these problems can be viewed as token-level multi-turn Markov decision processes (MDPs), this view corresponds to a degenerate case of multi-turn interaction where the environment provides no feedback. This contrasts with many real-world domains, such as software engineering (SWE), which require rich multi-turn interactions with a stateful environment that responds to each action with a non-trivial observation. To bridge this gap, we demonstrate the successful application of RL to this general regime. Our methodology begins with rejection fine-tuning (RFT) using execution feedback to train a policy to follow instructions and formatting effectively, followed by a synchronous RL pipeline using DAPO for iterative improvement. Applying this pipeline to Qwen2.5-72B-Instruct, we increase its Pass@1 on the SWE-bench Verified benchmark from 11% to 39%, substantially improving upon the 20% RFT baseline. On the May and June splits of SWE-rebench, the resulting agent achieves Pass@1 of 35% and 31% respectively, competitive with even larger models such as DeepSeek-V3-0324 or Qwen3-235B-A22B, demonstrating that our methodology offers a practical approach for training capable agents for multi-turn interactive tasks using open-weight models.
Authors: Xiaohan Chen, Zhongying Pan, Quan Feng, Yu Tian, Shuqun Yang, Mengru Wang, Lina Gong, Yuxia Geng, Piji Li, Xiang Chen
Abstract: Despite Retrieval-Augmented Generation improving code completion, traditional retrieval methods struggle with information redundancy and a lack of diversity within limited context windows. To solve this, we propose a resource-optimized retrieval augmentation method, SaraCoder. It maximizes information diversity and representativeness in a limited context window, significantly boosting the accuracy and reliability of repository-level code completion. Its core Hierarchical Feature Optimization module systematically refines candidates by distilling deep semantic relationships, pruning exact duplicates, assessing structural similarity with a novel graph-based metric that weighs edits by their topological importance, and reranking results to maximize both relevance and diversity. Furthermore, an External-Aware Identifier Disambiguator module accurately resolves cross-file symbol ambiguity via dependency analysis. Extensive experiments on the challenging CrossCodeEval and RepoEval-Updated benchmarks demonstrate that SaraCoder outperforms existing baselines across multiple programming languages and models. Our work proves that systematically refining retrieval results across multiple dimensions provides a new paradigm for building more accurate and resource-optimized repository-level code completion systems.
Authors: Zhongang Cai, Yubo Wang, Qingping Sun, Ruisi Wang, Chenyang Gu, Wanqi Yin, Zhiqian Lin, Zhitao Yang, Chen Wei, Xuanke Shi, Kewang Deng, Xiaoyang Han, Zukai Chen, Jiaqi Li, Xiangyu Fan, Hanming Deng, Lewei Lu, Bo Li, Ziwei Liu, Quan Wang, Dahua Lin, Lei Yang
Abstract: Multimodal models have achieved remarkable progress in recent years. Nevertheless, they continue to exhibit notable limitations in spatial understanding and reasoning, the very capability that anchors artificial general intelligence in the physical world. With the recent release of GPT-5, allegedly the most powerful AI model to date, it is timely to examine where the leading models (GPT, Gemini, Grok, Seed, Qwen, and Intern) stand on the path toward spatial intelligence. We first propose a holistic taxonomy of spatial tasks that unifies existing benchmarks and a standardized protocol for the fair evaluation of state-of-the-art proprietary and open-source models across eight key benchmarks, at a cost exceeding ten billion total tokens. Our empirical study then reveals that (1) GPT-5 demonstrates unprecedented strength in spatial intelligence (SI), yet (2) still falls short of human performance significantly across a broad spectrum of SI-tasks. Moreover, we (3) show that SI-tasks expose greater model capability deficiency than non-SI tasks, to the extent that (4) proprietary models do not exhibit a decisive advantage when facing the most difficult ones. In addition, we conduct a qualitative evaluation across a diverse set of scenarios that are intuitive for humans, yet fail even the most advanced multimodal models.
Authors: Zifeng Ding, Shenyang Huang, Zeyu Cao, Emma Kondrup, Zachary Yang, Xingyue Huang, Yuan Sui, Zhangdie Yuan, Yuqicheng Zhu, Xianglong Hu, Yuan He, Farimah Poursafaei, Michael Bronstein, Andreas Vlachos
Abstract: Forecasting future links is a central task in temporal graph (TG) reasoning, requiring models to leverage historical interactions to predict upcoming ones. Traditional neural approaches, such as temporal graph neural networks, achieve strong performance but lack explainability and cannot be applied to unseen graphs without retraining. Recent studies have begun to explore using large language models (LLMs) for graph reasoning, but most of them are constrained to static graphs or small synthetic TGs and lack the evaluation of the quality of reasoning traces generated by LLMs. In this work, we present Reasoning-Enhanced Learning for Temporal Graphs (ReaL-TG), a reinforcement learning framework that fine-tunes LLMs to perform explainable link forecasting on real-world TGs. ReaL-TG uses outcome-based reward to encourage models to self-explore reasoning strategies from graph structure and to produce explanations that directly justify their predictions. To enable evaluation on LLM-generated reasoning traces, we propose a new evaluation protocol combining ranking metrics with an LLM-as-a-Judge system that assesses both the quality of reasoning and the impact of hallucinations. Experiments with ReaL-TG-4B, obtained by fine-tuning Qwen3-4B under our framework, show that it outperforms much larger frontier LLMs, including GPT-5 mini, on ranking metrics, while producing high-quality explanations confirmed by both the LLM judge and human evaluation.
Authors: Hanyi Mao, Quanjia Xiao, Lei Pang, Haixiao Liu
Abstract: We propose FSPO (Fair Sequence Policy Optimization), a sequence-level reinforcement learning method for LLMs that enforces length-fair clipping on the importance-sampling (IS) weight. We study RL methods with sequence-level IS and identify a mismatch when PPO/GRPO-style clipping is transplanted to sequences: a fixed clip range systematically reweights short vs. long responses, distorting the optimization direction. FSPO introduces a simple remedy: we clip the sequence log-IS ratio with a band that scales as $\sqrt{L}$. Theoretically, we formalize length fairness via a Length Reweighting Error (LRE) and prove that small LRE yields a cosine directional guarantee between the clipped and true updates. Empirically, FSPO flattens clip rates across length bins, stabilizes training, and outperforms baselines across model sizes and evaluation datasets, with the largest gains on the Qwen3-8B-Base model.
Authors: Xinyu Liu, Bei Li, Jiahao Liu, Junhao Ruan, Kechen Jiao, Hongyin Tang, Jingang Wang, Xiao Tong, Jingbo Zhu
Abstract: High-order numerical methods enhance Transformer performance in tasks like NLP and CV, but introduce a performance-efficiency trade-off due to increased computational overhead. Our analysis reveals that conventional efficiency techniques, such as distillation, can be detrimental to the performance of these models, exemplified by PCformer. To explore more optimizable ODE-based Transformer architectures, we propose the Iterative Implicit Euler Transformer (IIET), which simplifies high-order methods using an iterative implicit Euler approach. This simplification not only leads to superior performance but also facilitates model compression compared to PCformer. To enhance inference efficiency, we introduce Iteration Influence-Aware Distillation (IIAD). Through a flexible threshold, IIAD allows users to effectively balance the performance-efficiency trade-off. On lm-evaluation-harness, IIET boosts average accuracy by 2.65% over vanilla Transformers and 0.8% over PCformer. Its efficient variant, E-IIET, significantly cuts inference overhead by 55% while retaining 99.4% of the original task accuracy. Moreover, the most efficient IIET variant achieves an average performance gain exceeding 1.6% over vanilla Transformer with comparable speed.
Authors: Zi-Yuan Hu, Shuo Liang, Duo Zheng, Yanyang Li, Yeyao Tao, Shijia Huang, Wei Feng, Jia Qin, Jianguang Yu, Jing Huang, Meng Fang, Yin Li, Liwei Wang
Abstract: Recent advances in video large language models (VideoLLMs) call for new evaluation protocols and benchmarks for complex temporal reasoning in video-language understanding. Inspired by the needle in a haystack test widely used by LLMs, we introduce a novel task of Needle in a Montage (NeMo), designed to assess VideoLLMs' critical reasoning capabilities, including long-context recall and temporal grounding. To generate video question answering data for our task, we develop a scalable automated data generation pipeline that facilitates high-quality data synthesis. Built upon the proposed pipeline, we present NeMoBench, a video-language benchmark centered on our task. Specifically, our full set of NeMoBench features 31,378 automatically generated question-answer (QA) pairs from 13,486 videos with various durations ranging from seconds to hours. Experiments demonstrate that our pipeline can reliably and automatically generate high-quality evaluation data, enabling NeMoBench to be continuously updated with the latest videos. We evaluate 20 state-of-the-art models on our benchmark, providing extensive results and key insights into their capabilities and limitations. Our project page is available at: https://lavi-lab.github.io/NeMoBench.
Authors: Junwei Lan, Jianlyu Chen, Zheng Liu, Chaofan Li, Siqi Bao, Defu Lian
Abstract: With the growing popularity of LLM agents and RAG, it has become increasingly important to retrieve documents that are essential for solving a task, even when their connection to the task is indirect or implicit. Addressing this problem requires fine-grained reasoning to accurately assess the relevance between the task and each candidate document. This capability, however, poses a significant challenge for existing IR techniques. Despite recent progress in reasoning-enhanced IR, existing approaches still face significant challenges in applicability, scalability, and efficiency. In this work, we propose Retro*, a novel approach for reasoning-intensive document retrieval. Our method introduces a rubric-based relevance scoring mechanism, enabling the model to reason about the relationship between a task and a document based on explicitly defined criteria, whereby producing a fine-grained, interpretable relevance score. Retro* also supports test-time scaling by combining multiple reasoning trajectories via score integration, which produces more reliable relevance estimates. To optimize Retro*'s reasoning capabilities, we introduce a novel reinforcement learning algorithm tailored for its relevance scoring mechanism, which employs two composite rewards to fully exploit the trajectories of each training sample. Our experiments show that Retro* outperforms existing document retrieval methods with notable advantages, leading to state-of-the-art performance on the BRIGHT benchmark.
Authors: Lifan Yuan, Weize Chen, Yuchen Zhang, Ganqu Cui, Hanbin Wang, Ziming You, Ning Ding, Zhiyuan Liu, Maosong Sun, Hao Peng
Abstract: Does RL teach LLMs genuinely new skills, or does it merely activate existing ones? This question lies at the core of ongoing debates about the role of RL in LLM post-training. On one side, strong empirical results can be achieved with RL even without preceding supervised finetuning; on the other, critics argue that RL contributes little beyond reweighting existing reasoning strategies. This work provides concrete evidence that LLMs can acquire genuinely new skills during RL by composing existing ones, mirroring one of the central mechanisms by which humans acquire new cognitive skills. To mitigate data contamination and other confounding factors, and to allow precise control over task complexity, we develop a synthetic framework for our investigation. Specifically, we define a skill as the ability to infer the output of a string transformation function f(x) given x. When an LLM has already learned f and g prior to RL, our experiments reveal that RL enables it to learn unseen compositions of them h(x)=g(f(x)). Further, this compositional ability generalizes to more difficult problems such as compositions of >2 functions unseen during RL training. Surprisingly, our experiments show that compositional skill acquired on a source task transfers to a different target task. This transfer happens even without compositional training on the target, requiring only prior knowledge of the target's atomic skills. Our qualitative analysis shows that RL fundamentally changes the reasoning behaviors of the models. In contrast, next-token training with the same data yields none of these findings. Our systematic experiments provide fresh insights into LLM learning, suggesting the value of first building base models with basic skills, then using RL to incentivize advanced, generalizable skills for complex problems.
Authors: Shenao Zhang, Donghan Yu, Yihao Feng, Bowen Jin, Zhaoran Wang, John Peebles, Zirui Wang
Abstract: Large language models excel with reinforcement learning (RL), but fully unlocking this potential requires a mid-training stage. An effective mid-training phase should identify a compact set of useful actions and enable fast selection among them through online RL. We formalize this intuition by presenting the first theoretical result on how mid-training shapes post-training: it characterizes an action subspace that minimizes both the value approximation error from pruning and the RL error during subsequent planning. Our analysis reveals two key determinants of mid-training effectiveness: pruning efficiency, which shapes the prior of the initial RL policy, and its impact on RL convergence, which governs the extent to which that policy can be improved via online interactions. These results suggest that mid-training is most effective when the decision space is compact and the effective horizon is short, highlighting the importance of operating in the space of action abstractions rather than primitive actions. Building on these insights, we propose Reasoning as Action Abstractions (RA3), a scalable mid-training algorithm. Specifically, we derive a sequential variational lower bound and optimize it by iteratively discovering temporally-consistent latent structures via RL, followed by fine-tuning on the bootstrapped data. Experiments on code generation tasks demonstrate the effectiveness of our approach. Across multiple base models, RA3 improves the average performance on HumanEval and MBPP by 8 and 4 points over the base model and the next-token prediction baseline. Furthermore, RA3 achieves faster convergence and higher asymptotic performance in RLVR on HumanEval+, MBPP+, LiveCodeBench, and Codeforces.
Authors: Tudor Achim, Alex Best, Alberto Bietti, Kevin Der, Math\"is F\'ed\'erico, Sergei Gukov, Daniel Halpern-Leistner, Kirsten Henningsgard, Yury Kudryashov, Alexander Meiburg, Martin Michelsen, Riley Patterson, Eric Rodriguez, Laura Scharff, Vikram Shanker, Vladmir Sicca, Hari Sowrirajan, Aidan Swope, Matyas Tamas, Vlad Tenev, Jonathan Thomm, Harold Williams, Lawrence Wu
Abstract: We introduce Aristotle, an AI system that combines formal verification with informal reasoning, achieving gold-medal-equivalent performance on the 2025 International Mathematical Olympiad problems. Aristotle integrates three main components: a Lean proof search system, an informal reasoning system that generates and formalizes lemmas, and a dedicated geometry solver. Our system demonstrates state-of-the-art performance with favorable scaling properties for automated theorem proving.
Authors: Wenlong Deng, Yi Ren, Yushu Li, Boying Gong, Danica J. Sutherland, Xiaoxiao Li, Christos Thrampoulidis
Abstract: Reinforcement learning with verifiable rewards has significantly advanced the reasoning capabilities of large language models, yet how to explicitly steer training toward exploration or exploitation remains an open problem. We introduce Token Hidden Reward (THR), a token-level metric that quantifies each token's influence on the likelihood of correct responses under Group Relative Policy Optimization (GRPO). We find that training dynamics are dominated by a small subset of tokens with high absolute THR values. Most interestingly, tokens with positive THR strengthen confidence in correct outputs, thus favoring exploitation, while tokens with negative THR preserve probability mass for alternative outputs, enabling exploration. This insight suggests a natural intervention: a THR-guided reweighting algorithm that modulates GRPO's learning signals to explicitly bias training toward exploitation or exploration. We validate the efficacy of this algorithm on diverse math reasoning benchmarks. By amplifying tokens with positive THR value and weakening negative ones, our algorithm improves greedy-decoding accuracy, favoring exploitation. The reverse strategy yields consistent gains in Pass@K accuracy, favoring exploration. We further demonstrate that our algorithm integrates seamlessly with other RL objectives such as GSPO and generalizes across architectures including Llama. These findings establish THR as a principled and fine-grained mechanism for dynamically controlling exploration and exploitation in RL-tuned LLMs, providing new tools for targeted fine-tuning in reasoning-intensive applications.
Authors: Haoqiang Kang, Yizhe Zhang, Nikki Lijing Kuang, Nicklas Majamaki, Navdeep Jaitly, Yi-An Ma, Lianhui Qin
Abstract: Large Language Models (LLMs) demonstrate their reasoning ability through chain-of-thought (CoT) generation. However, LLM's autoregressive decoding may limit the ability to revisit and refine earlier tokens in a holistic manner, which can also lead to inefficient exploration for diverse solutions. In this paper, we propose LaDiR (Latent Diffusion Reasoner), a novel reasoning framework that unifies the expressiveness of continuous latent representation with the iterative refinement capabilities of latent diffusion models for an existing LLM. We first construct a structured latent reasoning space using a Variational Autoencoder (VAE) that encodes text reasoning steps into blocks of thought tokens, preserving semantic information and interpretability while offering compact but expressive representations. Subsequently, we utilize a latent diffusion model that learns to denoise a block of latent thought tokens with a blockwise bidirectional attention mask, enabling longer horizon and iterative refinement with adaptive test-time compute. This design allows efficient parallel generation of diverse reasoning trajectories, allowing the model to plan and revise the reasoning process holistically. We conduct evaluations on a suite of mathematical reasoning and planning benchmarks. Empirical results show that LaDiR consistently improves accuracy, diversity, and interpretability over existing autoregressive, diffusion-based, and latent reasoning methods, revealing a new paradigm for text reasoning with latent diffusion.
Authors: Ruiling Xu, Yifan Zhang, Qingyun Wang, Carl Edwards, Heng Ji
Abstract: Organic reaction mechanisms are the stepwise elementary reactions by which reactants form intermediates and products, and are fundamental to understanding chemical reactivity and designing new molecules and reactions. Although large language models (LLMs) have shown promise in understanding chemical tasks such as synthesis design, it is unclear to what extent this reflects genuine chemical reasoning capabilities, i.e., the ability to generate valid intermediates, maintain chemical consistency, and follow logically coherent multi-step pathways. We address this by introducing oMeBench, the first large-scale, expert-curated benchmark for organic mechanism reasoning in organic chemistry. It comprises over 10,000 annotated mechanistic steps with intermediates, type labels, and difficulty ratings. Furthermore, to evaluate LLM capability more precisely and enable fine-grained scoring, we propose oMeS, a dynamic evaluation framework that combines step-level logic and chemical similarity. We analyze the performance of state-of-the-art LLMs, and our results show that although current models display promising chemical intuition, they struggle with correct and consistent multi-step reasoning. Notably, we find that using prompting strategy and fine-tuning a specialist model on our proposed dataset increases performance by 50% over the leading closed-source model. We hope that oMeBench will serve as a rigorous foundation for advancing AI systems toward genuine chemical reasoning.
Authors: Onur Kele\c{s}, Asl{\i} \"Ozy\"urek, Gerardo Ortega, Kadir G\"okg\"oz, Esam Ghaleb
Abstract: Iconicity, the resemblance between linguistic form and meaning, is pervasive in signed languages, offering a natural testbed for visual grounding. For vision-language models (VLMs), the challenge is to recover such essential mappings from dynamic human motion rather than static context. We introduce the Visual Iconicity Challenge, a novel video-based benchmark that adapts psycholinguistic measures to evaluate VLMs on three tasks: (i) phonological sign-form prediction (e.g., handshape, location), (ii) transparency (inferring meaning from visual form), and (iii) graded iconicity ratings. We assess 13 state-of-the-art VLMs in zero- and few-shot settings on Sign Language of the Netherlands and compare them to human baselines. On phonological form prediction, VLMs recover some handshape and location detail but remain below human performance; on transparency, they are far from human baselines; and only top models correlate moderately with human iconicity ratings. Interestingly, models with stronger phonological form prediction correlate better with human iconicity judgment, indicating shared sensitivity to visually grounded structure. Our findings validate these diagnostic tasks and motivate human-centric signals and embodied learning methods for modelling iconicity and improving visual grounding in multimodal models.
Authors: Md Kowsher, Ali O. Polat, Ehsan Mohammady Ardehaly, Mehrdad Salehi, Zia Ghiasi, Prasanth Murali, Chen Chen
Abstract: This paper presents a theoretical framework explaining why fine tuning small, randomly selected subnetworks (slices) within pre trained models can be sufficient for downstream adaptation. We prove that pretrained networks exhibit a universal winning slice property arising from two phenomena: (1) spectral balance the eigenspectra of different weight matrix slices are remarkably similar; and (2) high task energy their backbone representations retain rich, task relevant features. This leads to the Universal Winning Slice Hypothesis, which provides a theoretical foundation for parameter efficient fine tuning (PEFT) in large scale models. Inspired by this, we propose SliceFine, a PEFT method that exploits this inherent redundancy by updating only selected slices of the original weights introducing zero new parameters, unlike adapter-based approaches. Empirically, SliceFine matches the performance of state of the art PEFT methods across language and vision tasks, while significantly improving training speed, memory efficiency, and model compactness. Our work bridges theory and practice, offering a theoretically grounded alternative to existing PEFT techniques.