Authors: Fengyi Xu, Jun Ma, Waishan Qiu, Cui Guo
Abstract: Crowdsourced street-view imagery from social media provides valuable real-time visual evidence of urban flooding and other crisis events, yet it often lacks reliable geographic metadata for emergency response. Existing Visual Place Recognition (VPR) models exhibit substantial performance degradation when applied to such imagery due to visual distortions and domain shifts inherent in cross-source scenarios. This paper presents VPR-AttLLM, a model-agnostic framework that integrates the semantic reasoning and geospatial knowledge of Large Language Models (LLMs) into established VPR pipelines through attention-guided descriptor enhancement. By leveraging LLMs to identify location-informative regions within the city context and suppress transient visual noise, VPR-AttLLM improves retrieval performance without requiring model retraining or additional data. Comprehensive evaluations are conducted on extended benchmarks including SF-XL enriched with real social-media flood images, synthetic flooding scenarios over established query sets and Mapillary photos, and a new HK-URBAN dataset capturing morphologically distinct cityscapes. Integrating VPR-AttLLM with three state-of-the-art VPR models-CosPlace, EigenPlaces, and SALAD-consistently improves recall performance, yielding relative gains typically between 1-3% and reaching up to 8% on the most challenging real flood imagery. Beyond measurable gains in retrieval accuracy, this study establishes a generalizable paradigm for LLM-guided multimodal fusion in visual retrieval systems. By embedding principles from urban perception theory into attention mechanisms, VPR-AttLLM bridges human-like spatial reasoning with modern VPR architectures. Its plug-and-play design, strong cross-source robustness, and interpretability highlight its potential for scalable urban monitoring and rapid geo-localization of crowdsourced crisis imagery.
Authors: Enes \"Ozeren, Matthias A{\ss}enmacher
Abstract: Chain-of-Thought (CoT) reasoning typically utilizes the discrete language space for thinking, which is inherently inefficient, as many generated tokens only enforce linguistic rules that are not required for reasoning. To bypass this, latent-space thinking allows models to think using the continuous embedding space. While existing methods for training those models show domain-specific gains, they fail to maintain performance in complex tasks, such as mathematical reasoning. We experimentally demonstrate that the Coconut approach, a form of supervised fine-tuning for latent-space thinking, is highly sensitive to design choices and exhibits several inherent limitations. To address these issues, we investigate reinforcement learning (RL) techniques -- an underexplored direction in latent-space thinking -- including GRPO and design a novel Latent RL method for directly optimizing the latent thinking steps. Our experimental results reveal that these RL-trained models still lag behind traditional language-space CoT models in the mathematical reasoning domain. We make our codebase publicly available.
Authors: Nimol Thuon, Jun Du
Abstract: Automated document layout analysis remains a major challenge for low-resource, non-Latin scripts. Khmer is a language spoken daily by over 17 million people in Cambodia, receiving little attention in the development of document AI tools. The lack of dedicated resources is particularly acute for business documents, which are critical for both public administration and private enterprise. To address this gap, we present \textbf{KH-FUNSD}, the first publicly available, hierarchically annotated dataset for Khmer form document understanding, including receipts, invoices, and quotations. Our annotation framework features a three-level design: (1) region detection that divides each document into core zones such as header, form field, and footer; (2) FUNSD-style annotation that distinguishes questions, answers, headers, and other key entities, together with their relationships; and (3) fine-grained classification that assigns specific semantic roles, such as field labels, values, headers, footers, and symbols. This multi-level approach supports both comprehensive layout analysis and precise information extraction. We benchmark several leading models, providing the first set of baseline results for Khmer business documents, and discuss the distinct challenges posed by non-Latin, low-resource scripts. The KH-FUNSD dataset and documentation will be available at URL.
Authors: Glenn Zhang, Treasure Mayowa, Jason Fan, Yicheng Fu, Aaron Sandoval, Sean O'Brien, Kevin Zhu
Abstract: Producing trustworthy and reliable Large Language Models (LLMs) has become increasingly important as their usage becomes more widespread. Calibration seeks to achieve this by improving the alignment between the model's confidence and the actual likelihood of its responses being correct or desirable. However, it has been observed that the internal confidence of a model, derived from token probabilities, is not well aligned with its verbalized confidence, leading to misleading results with different calibration methods. In this paper, we propose Direct Confidence Alignment (DCA), a method using Direct Preference Optimization to align an LLM's verbalized confidence with its internal confidence rather than ground-truth accuracy, enhancing model transparency and reliability by ensuring closer alignment between the two confidence measures. We evaluate DCA across multiple open-weight LLMs on a wide range of datasets. To further assess this alignment, we also introduce three new calibration error-based metrics. Our results show that DCA improves alignment metrics on certain model architectures, reducing inconsistencies in a model's confidence expression. However, we also show that it can be ineffective on others, highlighting the need for more model-aware approaches in the pursuit of more interpretable and trustworthy LLMs.
Authors: Minghui Liu, Aadi Palnitkar, Tahseen Rabbani, Hyunwoo Jae, Kyle Rui Sang, Dixi Yao, Shayan Shabihi, Fuheng Zhao, Tian Li, Ce Zhang, Furong Huang, Kunpeng Zhang
Abstract: Large language models (LLMs) have demonstrated remarkable performance on long-context tasks, but are often bottlenecked by memory constraints. Namely, the KV cache, which is used to significantly speed up attention computations, grows linearly with context length. A suite of compression algorithms has been introduced to alleviate cache growth by evicting unimportant tokens. However, several popular strategies are targeted towards the prefill phase, i.e., processing long prompt context, and their performance is rarely assessed on reasoning tasks requiring long decoding. In particular, short but complex prompts, such as those in benchmarks like GSM8K and MATH500, often benefit from multi-step reasoning and self-reflection, resulting in thinking sequences thousands of tokens long. In this work, we benchmark the performance of several popular compression strategies on long-reasoning tasks. For the non-reasoning Llama-3.1-8B-Instruct, we determine that no singular strategy fits all, and that performance is heavily influenced by dataset type. However, we discover that H2O and our decoding-enabled variant of SnapKV are dominant strategies for reasoning models, indicating the utility of heavy-hitter tracking for reasoning traces. We also find that eviction strategies at low budgets can produce longer reasoning traces, revealing a tradeoff between cache size and inference costs.
Authors: Philipp Habicht, Lev Sorokin, Abdullah Saydemir, Ken E. Friedl, Andrea Stocco
Abstract: In-Car Conversational Question Answering (ConvQA) systems significantly enhance user experience by enabling seamless voice interactions. However, assessing their accuracy and reliability remains a challenge. This paper explores the use of Large Language Models (LLMs) alongside advanced prompting techniques and agent-based methods to evaluate the extent to which ConvQA system responses adhere to user utterances. The focus lies on contextual understanding and the ability to provide accurate venue recommendations considering user constraints and situational context. To evaluate utterance-response coherence using an LLM, we synthetically generate user utterances accompanied by correct and modified failure-containing system responses. We use input-output, chain-of-thought, self-consistency prompting, and multi-agent prompting techniques with 13 reasoning and non-reasoning LLMs of varying sizes and providers, including OpenAI, DeepSeek, Mistral AI, and Meta. We evaluate our approach on a case study involving restaurant recommendations. The most substantial improvements occur for small non-reasoning models when applying advanced prompting techniques, particularly multi-agent prompting. However, reasoning models consistently outperform non-reasoning models, with the best performance achieved using single-agent prompting with self-consistency. Notably, DeepSeek-R1 reaches an F1-score of 0.99 at a cost of 0.002 USD per request. Overall, the best balance between effectiveness and cost-time efficiency is reached with the non-reasoning model DeepSeek-V3. Our findings show that LLM-based evaluation offers a scalable and accurate alternative to traditional human evaluation for benchmarking contextual understanding in ConvQA systems.
Authors: Avinash Amballa, Yashas Malur Saidutta, Chi-Heng Lin, Vivek Kulkarni, Srinivas Chappidi
Abstract: Large language models (LLMs) are increasingly being used to generate synthetic datasets for the evaluation and training of downstream models. However, prior work has noted that such generated data lacks diversity. In this paper, we propose Voyager, a novel principled approach to generate diverse datasets. Our approach is iterative and directly optimizes a mathematical quantity that optimizes the diversity of the dataset using the machinery of determinantal point processes. Furthermore, our approach is training-free, applicable to closed-source models, and scalable. In addition to providing theoretical justification for the working of our method, we also demonstrate through comprehensive experiments that Voyager significantly outperforms popular baseline approaches by providing a 1.5-3x improvement in diversity.
Authors: Jiayi Yuan, Cameron Shinn, Kai Xu, Jingze Cui, George Klimiashvili, Guangxuan Xiao, Perkz Zheng, Bo Li, Yuxin Zhou, Zhouhai Ye, Weijie You, Tian Zheng, Dominic Brown, Pengbo Wang, Richard Cai, Julien Demouth, John D. Owens, Xia Hu, Song Han, Timmy Liu, Huizi Mao
Abstract: The growing demand for long-context inference capabilities in Large Language Models (LLMs) has intensified the computational and memory bottlenecks inherent to the standard attention mechanism. To address this challenge, we introduce BLASST, a drop-in sparse attention method that dynamically prunes the attention matrix without any pre-computation or proxy scores. Our method uses a fixed threshold and existing information from online softmax to identify negligible attention scores, skipping softmax computation, Value block loading, and the subsequent matrix multiplication. This fits seamlessly into existing FlashAttention kernel designs with negligible latency overhead. The approach is applicable to both prefill and decode stages across all attention variants (MHA, GQA, MQA, and MLA), providing a unified solution for accelerating long-context inference. We develop an automated calibration procedure that reveals a simple inverse relationship between optimal threshold and context length, enabling robust deployment across diverse scenarios. Maintaining high accuracy, we demonstrate a 1.62x speedup for prefill at 74.7% sparsity and a 1.48x speedup for decode at 73.2% sparsity on modern GPUs. Furthermore, we explore sparsity-aware training as a natural extension, showing that models can be trained to be inherently more robust to sparse attention patterns, pushing the accuracy-sparsity frontier even further.
Authors: Yoav Gelberg, Koshi Eguchi, Takuya Akiba, Edoardo Cetin
Abstract: So far, expensive finetuning beyond the pretraining sequence length has been a requirement for effectively extending the context of language models (LM). In this work, we break this key bottleneck by Dropping the Positional Embeddings of LMs after training (DroPE). Our simple method is motivated by three key theoretical and empirical observations. First, positional embeddings (PEs) serve a crucial role during pretraining, providing an important inductive bias that significantly facilitates convergence. Second, over-reliance on this explicit positional information is also precisely what prevents test-time generalization to sequences of unseen length, even when using popular PE-scaling methods. Third, positional embeddings are not an inherent requirement of effective language modeling and can be safely removed after pretraining, following a short recalibration phase. Empirically, DroPE yields seamless zero-shot context extension without any long-context finetuning, quickly adapting pretrained LMs without compromising their capabilities in the original training context. Our findings hold across different models and dataset sizes, far outperforming previous specialized architectures and established rotary positional embedding scaling methods.
Authors: Zheng Huang, Kiran Ramnath, Yueyan Chen, Aosong Feng, Sangmin Woo, Balasubramaniam Srinivasan, Zhichao Xu, Kang Zhou, Shuai Wang, Haibo Ding, Lin Lee Cheong
Abstract: Diffusion language models (DLMs) have recently emerged as a compelling alternative to autoregressive generation, offering parallel generation and improved global coherence. During inference, DLMs generate text by iteratively denoising masked sequences in parallel; however, determining which positions to unmask and which tokens to commit forms a large combinatorial search problem. Existing inference methods approximate this search using heuristics, which often yield suboptimal decoding paths; other approaches instead rely on additional training to guide token selection. To introduce a principled search mechanism for DLMs inference, we introduce MEDAL, a framework that integrates Monte Carlo Tree SEarch initialization for Diffusion LAnguage Model inference. We employ Monte Carlo Tree Search at the initialization stage to explore promising unmasking trajectories, providing a robust starting point for subsequent refinement. This integration is enabled by restricting the search space to high-confidence actions and prioritizing token choices that improve model confidence over remaining masked positions. Across multiple benchmarks, MEDAL achieves up to 22.0% improvement over existing inference strategies, establishing a new paradigm for search-based inference in diffusion language models.
Authors: Yinzhu Cheng, Haihua Xie, Yaqing Wang, Miao He, Mingming Sun
Abstract: Semantic distance measurement is a fundamental problem in computational linguistics, providing a quantitative characterization of similarity or relatedness between text segments, and underpinning tasks such as text retrieval and text classification. From a mathematical perspective, a semantic distance can be viewed as a metric defined on a space of texts or on a representation space derived from them. However, most classical semantic distance methods are essentially fixed, making them difficult to adapt to specific data distributions and task requirements. In this paper, a semantic distance measure based on multi-kernel Gaussian processes (MK-GP) was proposed. The latent semantic function associated with texts was modeled as a Gaussian process, with its covariance function given by a combined kernel combining Mat\'ern and polynomial components. The kernel parameters were learned automatically from data under supervision, rather than being hand-crafted. This semantic distance was instantiated and evaluated in the context of fine-grained sentiment classification with large language models under an in-context learning (ICL) setup. The experimental results demonstrated the effectiveness of the proposed measure.
Authors: Anika Sharma, Tianyi Niu, Emma Wrenn, Shashank Srivastava
Abstract: The phenomenon of sound symbolism, the non-arbitrary mapping between word sounds and meanings, has long been demonstrated through anecdotal experiments like Bouba Kiki, but rarely tested at scale. We present the first computational cross-linguistic analysis of sound symbolism in the semantic domain of size. We compile a typologically broad dataset of 810 adjectives (27 languages, 30 words each), each phonemically transcribed and validated with native-speaker audio. Using interpretable classifiers over bag-of-segment features, we find that phonological form predicts size semantics above chance even across unrelated languages, with both vowels and consonants contributing. To probe universality beyond genealogy, we train an adversarial scrubber that suppresses language identity while preserving size signal (also at family granularity). Language prediction averaged across languages and settings falls below chance while size prediction remains significantly above chance, indicating cross-family sound-symbolic bias. We release data, code, and diagnostic tools for future large-scale studies of iconicity.
Authors: Abhay Srivastava, Sam Jung, Spencer Mateega
Abstract: We introduce MARKET-BENCH, a benchmark that evaluates large language models (LLMs) on introductory quantitative trading tasks by asking them to construct executable backtesters from natural-language strategy descriptions and market assumptions. Each instance specifies one of three canonical strategies -- scheduled trading on Microsoft (NASDAQ: MSFT), pairs trading on Coca-Cola (NASDAQ: KO) and Pepsi (NASDAQ: PEP), or delta hedging on MSFT -- and models must produce code whose P\&L, drawdown, and position paths match a verifiable reference implementation. We assess twelve state-of-the-art models using a multi-round pass@k metric that separates structural reliability (whether the backtest runs) from numerical accuracy (mean absolute error of the backtest metrics). While most models reliably execute the simplest strategy (average pass@3 of 0.80), errors vary by orders of magnitude across models and tasks: Gemini 3 Pro and Claude 4.5 Sonnet combine strong reliability with low error on simpler strategies, GPT-5.1 Codex-Max achieves perfect pass@1 on the first two strategies and the lowest best-run error on the easiest task, and Qwen3 Max attains perfect pass@3 yet sometimes produces inaccurate P\&L paths. These results show that current LLMs can scaffold basic trading infrastructure but still struggle to reason robustly about prices, inventory, and risk; we release MARKET-BENCH and a public leaderboard at https://marketbench.ai.
URLs: https://marketbench.ai.
Authors: Radu-Gabriel Chivereanu, Tiberiu Boros
Abstract: This work introduces a lightweight input-level adapter for the F5-TTS model that enables Romanian Language support. To preserve the existing capabilities of the model (voice cloning, English and Chinese support), we keep the original weights frozen, append a sub-network to the model and train it as an extension for the textual embedding matrix of the text encoder. For simplicity, we rely on ConvNeXt module implemented in F5-TTS to also model the co-dependencies between the new character-level embeddings. The module serves as a ``soft`` letter-to-sound layer, converting Romanian text into a continuous representation that the F5-TTS model uses to produce naturally sounding Romanian utterances. We evaluate the model with a pool of 20 human listeners across three tasks: (a) audio similarity between reference and generated speech, (b) pronunciation and naturalness and (c) Romanian-English code-switching. The results indicate that our approach maintains voice cloning capabilities and enables, to a certain extent, code-switching within the same utterance; however, residual English accent characteristics remain. We open-source our code and provide example audio samples at https://github.com/racai-ro/Ro-F5TTS.
Authors: Yushen Fang, Jianjun Li, Mingqian Ding, Chang Liu, Xinchi Zou, Wenqi Yang
Abstract: Although Large language Model (LLM)-powered information extraction (IE) systems have shown impressive capabilities, current fine-tuning paradigms face two major limitations: high training costs and difficulties in aligning with LLM preferences. To address these issues, we propose a novel universal IE paradigm, the Self-Correcting Iterative Refinement (SCIR) framework, along with a Multi-task Bilingual (Chinese-English) Self-Correcting (MBSC) dataset containing over 100,000 entries. The SCIR framework achieves plug-and-play compatibility with existing LLMs and IE systems through its Dual-Path Self-Correcting module and feedback-driven optimization, thereby significantly reducing training costs. Concurrently, the MBSC dataset tackles the challenge of preference alignment by indirectly distilling GPT-4's capabilities into IE result detection models. Experimental results demonstrate that SCIR outperforms state-of-the-art IE methods across three key tasks: named entity recognition, relation extraction, and event extraction, achieving a 5.27 percent average improvement in span-based Micro-F1 while reducing training costs by 87 percent compared to baseline approaches. These advancements not only enhance the flexibility and accuracy of IE systems but also pave the way for lightweight and efficient IE paradigms.
Authors: Veronica Mangiaterra, Hamad Al-Azary, Chiara Barattieri di San Pietro, Paolo Canal, Valentina Bambini
Abstract: As Large Language Models (LLMs) are increasingly being used in scientific research, the issue of their trustworthiness becomes crucial. In psycholinguistics, LLMs have been recently employed in automatically augmenting human-rated datasets, with promising results obtained by generating ratings for single words. Yet, performance for ratings of complex items, i.e., metaphors, is still unexplored. Here, we present the first assessment of the validity and reliability of ratings of metaphors on familiarity, comprehensibility, and imageability, generated by three GPT models for a total of 687 items gathered from the Italian Figurative Archive and three English studies. We performed a thorough validation in terms of both alignment with human data and ability to predict behavioral and electrophysiological responses. We found that machine-generated ratings positively correlated with human-generated ones. Familiarity ratings reached moderate-to-strong correlations for both English and Italian metaphors, although correlations weakened for metaphors with high sensorimotor load. Imageability showed moderate correlations in English and moderate-to-strong in Italian. Comprehensibility for English metaphors exhibited the strongest correlations. Overall, larger models outperformed smaller ones and greater human-model misalignment emerged with familiarity and imageability. Machine-generated ratings significantly predicted response times and the EEG amplitude, with a strength comparable to human ratings. Moreover, GPT ratings obtained across independent sessions were highly stable. We conclude that GPT, especially larger models, can validly and reliably replace - or augment - human subjects in rating metaphor properties. Yet, LLMs align worse with humans when dealing with conventionality and multimodal aspects of metaphorical meaning, calling for careful consideration of the nature of stimuli.
Authors: Gary Lupyan
Abstract: Acknowledging that large language models have learned to use language can open doors to breakthrough language science. Achieving these breakthroughs may require abandoning some long-held ideas about how language knowledge is evaluated and reckoning with the difficult fact that we have entered a post-Turing test era.
Authors: James Luther, Donald Brown
Abstract: Culture is the bedrock of human interaction; it dictates how we perceive and respond to everyday interactions. As the field of human-computer interaction grows via the rise of generative Large Language Models (LLMs), the cultural alignment of these models become an important field of study. This work, using the VSM13 International Survey and Hofstede's cultural dimensions, identifies the cultural alignment of popular LLMs (DeepSeek-V3, V3.1, GPT-5, GPT-4.1, GPT-4, Claude Opus 4, Llama 3.1, and Mistral Large). We then use cultural prompting, or using system prompts to shift the cultural alignment of a model to a desired country, to test the adaptability of these models to other cultures, namely China, France, India, Iran, Japan, and the United States. We find that the majority of the eight LLMs tested favor the United States when the culture is not specified, with varying results when prompted for other cultures. When using cultural prompting, seven of the eight models shifted closer to the expected culture. We find that models had trouble aligning with Japan and China, despite two of the models tested originating with the Chinese company DeepSeek.
Authors: Agniva Maiti, Manya Pandey, Murari Mandal
Abstract: The vast majority of the world's languages, particularly creoles like Nagamese, remain severely under-resourced in Natural Language Processing (NLP), creating a significant barrier to their representation in digital technology. This paper introduces NagaNLP, a comprehensive open-source toolkit for Nagamese, bootstrapped through a novel methodology that relies on LLM-driven but human-validated synthetic data generation. We detail a multi-stage pipeline where an expert-guided LLM (Gemini) generates a candidate corpus, which is then refined and annotated by native speakers. This synthetic-hybrid approach yielded a 10K pair conversational dataset and a high-quality annotated corpus for foundational tasks. To assess the effectiveness of our methodology, we trained both discriminative and generative models. Our fine-tuned XLM-RoBERTa-base model establishes a new benchmark for Nagamese, achieving a 93.81\% accuracy (0.90 F1-Macro) on Part-of-Speech tagging and a 0.75 F1-Macro on Named Entity Recognition, massively outperforming strong zero-shot baselines. Furthermore, we fine-tuned a Llama-3.2-3B Instruct model, named NagaLLaMA, which demonstrates superior performance on conversational tasks, achieving a Perplexity of 3.85, an order of magnitude improvement over its few-shot counterpart (96.76). We release the NagaNLP toolkit, including all datasets, models, and code, providing a foundational resource for a previously underserved language and a reproducible framework for reducing data scarcity in other low-resource contexts.
Authors: Yiming Zeng, Jinghan Cao, Zexin Li, Wanhao Yu, Zhankai Ye, Dawei Xiang, Ting Hua, Xin Liu, Shangqian Gao, Tingting Yu
Abstract: Instruction-based text editing is increasingly critical for real-world applications such as code editors (e.g., Cursor), but Large Language Models (LLMs) continue to struggle with this task. Unlike free-form generation, editing requires faithfully implementing user instructions while preserving unchanged content, as even minor unintended modifications can break functionality. Existing approaches treat editing as generic text generation, leading to two key failures: they struggle to faithfully align edits with diverse user intents, and they often over-edit unchanged regions. We propose HyperEdit to address both issues. First, we introduce hypernetwork-based dynamic adaptation that generates request-specific parameters, enabling the model to tailor its editing strategy to each instruction. Second, we develop difference-aware regularization that focuses supervision on modified spans, preventing over-editing while ensuring precise, minimal changes. HyperEdit achieves a 9%--30% relative improvement in BLEU on modified regions over state-of-the-art baselines, despite utilizing only 3B parameters.
Authors: Xueru Wen, Jie Lou, Yanjiang Liu, Hongyu Lin, Ben He, Xianpei Han, Le Sun, Yaojie Lu, Debing Zhang
Abstract: While reinforcement learning have achieved impressive progress in language model reasoning, they are constrained by the requirement for verifiable rewards. Recent verifier-free RL methods address this limitation by utilizing the intrinsic probabilities of LLMs generating reference answers as reward signals. However, these approaches typically sample reasoning traces conditioned only on the question. This design decouples reasoning-trace sampling from answer information, leading to inefficient exploration and incoherence between traces and final answers. In this paper, we propose \textit{\b{Co}upled \b{V}ariational \b{R}einforcement \b{L}earning} (CoVRL), which bridges variational inference and reinforcement learning by coupling prior and posterior distributions through a hybrid sampling strategy. By constructing and optimizing a composite distribution that integrates these two distributions, CoVRL enables efficient exploration while preserving strong thought-answer coherence. Extensive experiments on mathematical and general reasoning benchmarks show that CoVRL improves performance by 12.4\% over the base model and achieves an additional 2.3\% improvement over strong state-of-the-art verifier-free RL baselines, providing a principled framework for enhancing the general reasoning capabilities of language models.
Authors: Hong Su
Abstract: Large Language Models (LLMs) excel at extracting common patterns from large-scale corpora, yet they struggle with rare, low-resource, or previously unseen scenarios-such as niche hardware deployment issues or irregular IoT device behaviors-because such cases are sparsely represented in training data. Moreover, LLMs rely primarily on implicit parametric memory, which limits their ability to explicitly acquire, recall, and refine methods, causing them to behave predominantly as intuition-driven predictors rather than deliberate, method-oriented learners. Inspired by how humans learn from rare experiences, this paper proposes a human-inspired learning framework that integrates two complementary mechanisms. The first, Obvious Record, explicitly stores cause--result (or question--solution) relationships as symbolic memory, enabling persistent learning even from single or infrequent encounters. The second, Maximum-Entropy Method Discovery, prioritizes and preserves methods with high semantic dissimilarity, allowing the system to capture diverse and underrepresented strategies that are typically overlooked by next-token prediction. Verification on a benchmark of 60 semantically diverse question--solution pairs demonstrates that the proposed entropy-guided approach achieves stronger coverage of unseen questions and significantly greater internal diversity than a random baseline, confirming its effectiveness in discovering more generalizable and human-inspired methods.
Authors: Yucan Guo, Saiping Guan, Miao Su, Zeya Zhao, Xiaolong Jin, Jiafeng Guo, Xueqi Cheng
Abstract: Sparse Knowledge Graphs (KGs) are commonly encountered in real-world applications, where knowledge is often incomplete or limited. Sparse KG reasoning, the task of inferring missing knowledge over sparse KGs, is inherently challenging due to the scarcity of knowledge and the difficulty of capturing relational patterns in sparse scenarios. Among all sparse KG reasoning methods, path-based ones have attracted plenty of attention due to their interpretability. Existing path-based methods typically rely on computationally intensive random walks to collect paths, producing paths of variable quality. Additionally, these methods fail to leverage the structured nature of graphs by treating paths independently. To address these shortcomings, we propose a Structural and Probabilistic framework named StruProKGR, tailored for efficient and interpretable reasoning on sparse KGs. StruProKGR utilizes a distance-guided path collection mechanism to significantly reduce computational costs while exploring more relevant paths. It further enhances the reasoning process by incorporating structural information through probabilistic path aggregation, which prioritizes paths that reinforce each other. Extensive experiments on five sparse KG reasoning benchmarks reveal that StruProKGR surpasses existing path-based methods in both effectiveness and efficiency, providing an effective, efficient, and interpretable solution for sparse KG reasoning.
Authors: Aheli Poddar (Institute of Engineering & Management, Kolkata), Saptarshi Sahoo (Indian Statistical Institute, Chennai), Sujata Ghosh (Indian Statistical Institute, Chennai)
Abstract: We study syllogistic reasoning in LLMs from the logical and natural language perspectives. In process, we explore fundamental reasoning capabilities of the LLMs and the direction this research is moving forward. To aid in our studies, we use 14 large language models and investigate their syllogistic reasoning capabilities in terms of symbolic inferences as well as natural language understanding. Even though this reasoning mechanism is not a uniform emergent property across LLMs, the perfect symbolic performances in certain models make us wonder whether LLMs are becoming more and more formal reasoning mechanisms, rather than making explicit the nuances of human reasoning.
Authors: Sander Land, Yuval Pinter
Abstract: The Unigram tokenization algorithm offers a probabilistic alternative to the greedy heuristics of Byte-Pair Encoding. Despite its theoretical elegance, its implementation in practice is complex, limiting its adoption to the SentencePiece package and adapters thereof. We bridge this gap between theory and practice by providing a clear guide to implementation and parameter choices. We also identify a simpler algorithm that accepts slightly higher training loss in exchange for improved compression.
Authors: Yida Cai, Ranjuexiao Hu, Huiyuan Xie, Chenyang Li, Yun Liu, Yuxiao Ye, Zhenghao Liu, Weixing Shen, Zhiyuan Liu
Abstract: Legal relations form a highly consequential analytical framework of civil law system, serving as a crucial foundation for resolving disputes and realizing values of the rule of law in judicial practice. However, legal relations in Chinese civil cases remain underexplored in the field of legal artificial intelligence (legal AI), largely due to the absence of comprehensive schemas. In this work, we firstly introduce a comprehensive schema, which contains a hierarchical taxonomy and definitions of arguments, for AI systems to capture legal relations in Chinese civil cases. Based on this schema, we then formulate legal relation extraction task and present LexRel, an expert-annotated benchmark for legal relation extraction in Chinese civil law. We use LexRel to evaluate state-of-the-art large language models (LLMs) on legal relation extractions, showing that current LLMs exhibit significant limitations in accurately identifying civil legal relations. Furthermore, we demonstrate that incorporating legal relations information leads to consistent performance gains on other downstream legal AI tasks.
Authors: Hassan Mujtaba, Hamza Naveed, Hanzlah Munir
Abstract: Authorship analysis has traditionally focused on lexical and stylistic cues within text, while higher-level narrative structure remains underexplored, particularly for low-resource languages such as Urdu. This work proposes a graph-based framework that models Urdu novels as character interaction networks to examine whether authorial style can be inferred from narrative structure alone. Each novel is represented as a graph where nodes correspond to characters and edges denote their co-occurrence within narrative proximity. We systematically compare multiple graph representations, including global structural features, node-level semantic summaries, unsupervised graph embeddings, and supervised graph neural networks. Experiments on a dataset of 52 Urdu novels written by seven authors show that learned graph representations substantially outperform hand-crafted and unsupervised baselines, achieving up to 0.857 accuracy under a strict author-aware evaluation protocol.
Authors: Amirhossein Yousefiramandi, Ciaran Cooney
Abstract: We explore efficient strategies to fine-tune decoder-only Large Language Models (LLMs) for downstream text classification under resource constraints. Two approaches are investigated: (1) attaching a classification head to a pre-trained causal LLM and fine-tuning on the task (using the LLM's final token embedding as a sequence representation), and (2) instruction-tuning the LLM in a prompt->response format for classification. To enable single-GPU fine-tuning of models up to 8B parameters, we combine 4-bit model quantization with Low-Rank Adaptation (LoRA) for parameter-efficient training. Experiments on two datasets - a proprietary single-label dataset and the public WIPO-Alpha patent dataset (extreme multi-label classification) - show that the embedding-based method significantly outperforms the instruction-tuned method in F1-score, and is very competitive with - even surpassing - fine-tuned domain-specific models (e.g. BERT) on the same tasks. These results demonstrate that directly leveraging the internal representations of causal LLMs, along with efficient fine-tuning techniques, yields impressive classification performance under limited computational resources. We discuss the advantages of each approach while outlining practical guidelines and future directions for optimizing LLM fine-tuning in classification scenarios.
Authors: Xuanzhang Liu, Jianglun Feng, Zhuoran Zhuang, Junzhe Zhao, Maofei Que, Jieting Li, Dianlei Wang, Hao Tong, Ye Chen, Pan Li
Abstract: Large Language Model (LLM) agents trained with reinforcement learning (RL) show great promise for solving complex, multi-step tasks. However, their performance is often crippled by "Context Explosion", where the accumulation of long text outputs overwhelms the model's context window and leads to reasoning failures. To address this, we introduce CoDA, a Context-Decoupled hierarchical Agent, a simple but effective reinforcement learning framework that decouples high-level planning from low-level execution. It employs a single, shared LLM backbone that learns to operate in two distinct, contextually isolated roles: a high-level Planner that decomposes tasks within a concise strategic context, and a low-level Executor that handles tool interactions in an ephemeral, isolated workspace. We train this unified agent end-to-end using PECO (Planner-Executor Co-Optimization), a reinforcement learning methodology that applies a trajectory-level reward to jointly optimize both roles, fostering seamless collaboration through context-dependent policy updates. Extensive experiments demonstrate that CoDA achieves significant performance improvements over state-of-the-art baselines on complex multi-hop question-answering benchmarks, and it exhibits strong robustness in long-context scenarios, maintaining stable performance while all other baselines suffer severe degradation, thus further validating the effectiveness of our hierarchical design in mitigating context overload.
Authors: Jingzhe Ding, Shengda Long, Changxin Pu, Huan Zhou, Hongwan Gao, Xiang Gao, Chao He, Yue Hou, Fei Hu, Zhaojian Li, Weiran Shi, Zaiyuan Wang, Daoguang Zan, Chenchen Zhang, Xiaoxu Zhang, Qizhi Chen, Xianfu Cheng, Bo Deng, Qingshui Gu, Kai Hua, Juntao Lin, Pai Liu, Mingchen Li, Xuanguang Pan, Zifan Peng, Yujia Qin, Yong Shan, Zhewen Tan, Weihao Xie, Zihan Wang, Yishuo Yuan, Jiayu Zhang, Enduo Zhao, Yunfei Zhao, He Zhu, Chenyang Zou, Ming Ding, Jianpeng Jiao, Jiaheng Liu, Minghao Liu, Qian Liu, Chongyao Tao, Jian Yang, Tong Yang, Zhaoxiang Zhang, Xinjie Chen, Wenhao Huang, Ge Zhang
Abstract: Recent advances in coding agents suggest rapid progress toward autonomous software development, yet existing benchmarks fail to rigorously evaluate the long-horizon capabilities required to build complete software systems. Most prior evaluations focus on localized code generation, scaffolded completion, or short-term repair tasks, leaving open the question of whether agents can sustain coherent reasoning, planning, and execution over the extended horizons demanded by real-world repository construction. To address this gap, we present NL2Repo Bench, a benchmark explicitly designed to evaluate the long-horizon repository generation ability of coding agents. Given only a single natural-language requirements document and an empty workspace, agents must autonomously design the architecture, manage dependencies, implement multi-module logic, and produce a fully installable Python library. Our experiments across state-of-the-art open- and closed-source models reveal that long-horizon repository generation remains largely unsolved: even the strongest agents achieve below 40% average test pass rates and rarely complete an entire repository correctly. Detailed analysis uncovers fundamental long-horizon failure modes, including premature termination, loss of global coherence, fragile cross-file dependencies, and inadequate planning over hundreds of interaction steps. NL2Repo Bench establishes a rigorous, verifiable testbed for measuring sustained agentic competence and highlights long-horizon reasoning as a central bottleneck for the next generation of autonomous coding agents.
Authors: Thales Sales Almeida, Rodrigo Nogueira, H\'elio Pedrini
Abstract: Continued pretraining extends a language model's capabilities by further exposing it to additional data, often tailored to a specific linguistic or domain context. This strategy has emerged as an efficient alternative to full retraining when adapting general-purpose models to new settings. In this work, we investigate this paradigm through Curi\'o 7B, a 7-billion-parameter model derived from LLaMA-2 and trained on 100 billion Portuguese tokens from the ClassiCC-PT corpus - the most extensive Portuguese-specific continued-pretraining effort above the three-billion-parameter scale to date. Beyond scale, we investigate whether quantity alone suffices or whether data quality plays a decisive role in linguistic adaptation. To this end, we introduce Curi\'o-Edu 7B, a variant trained exclusively on the educational and STEM-filtered subset of the same corpus, totaling just 10 billion tokens. Despite using only 10% of the data and 20% of the computation, Curi\'o-Edu 7B surpasses the full-corpus model in our evaluations, demonstrating that data selection can be fundamental even when adapting models with limited prior exposure to the target language. The developed models are available at https://huggingface.co/collections/ClassiCC-Corpus/curio-edu
URLs: https://huggingface.co/collections/ClassiCC-Corpus/curio-edu
Authors: Pedro Henrique Luz de Araujo, Michael A. Hedderich, Ali Modarressi, Hinrich Schuetze, Benjamin Roth
Abstract: Persona-assigned large language models (LLMs) are used in domains such as education, healthcare, and sociodemographic simulation. Yet, they are typically evaluated only in short, single-round settings that do not reflect real-world usage. We introduce an evaluation protocol that combines long persona dialogues (over 100 rounds) and evaluation datasets to create dialogue-conditioned benchmarks that can robustly measure long-context effects. We then investigate the effects of dialogue length on persona fidelity, instruction-following, and safety of seven state-of-the-art open- and closed-weight LLMs. We find that persona fidelity degrades over the course of dialogues, especially in goal-oriented conversations, where models must sustain both persona fidelity and instruction following. We identify a trade-off between fidelity and instruction following, with non-persona baselines initially outperforming persona-assigned models; as dialogues progress and fidelity fades, persona responses become increasingly similar to baseline responses. Our findings highlight the fragility of persona applications in extended interactions and our work provides a protocol to systematically measure such failures.
Authors: Mosh Levy, Zohar Elyoseph, Shauli Ravfogel, Yoav Goldberg
Abstract: Large Language Models (LLMs) can generate reasoning tokens before their final answer to boost performance on complex tasks. While these sequences seem like human thought processes, empirical evidence reveals that they are not a faithful explanation of the model's actual reasoning process. To address this gap between appearance and function, we introduce the State over Tokens (SoT) conceptual framework. SoT reframes reasoning tokens not as a linguistic narrative, but as an externalized computational state -- the sole persistent information carrier across the model's stateless generation cycles. This explains how the tokens can drive correct reasoning without being a faithful explanation when read as text and surfaces previously overlooked research questions on these tokens. We argue that to truly understand the process that LLMs do, research must move beyond reading the reasoning tokens as text and focus on decoding them as state.
Authors: Hanyu Cai, Binqi Shen, Lier Jin, Lan Hu, Xiaojing Fan
Abstract: Prompt engineering has emerged as a critical factor influencing large language model (LLM) performance, yet the impact of pragmatic elements such as linguistic tone and politeness remains underexplored, particularly across different model families. In this work, we propose a systematic evaluation framework to examine how interaction tone affects model accuracy and apply it to three recently released and widely available LLMs: GPT-4o mini (OpenAI), Gemini 2.0 Flash (Google DeepMind), and Llama 4 Scout (Meta). Using the MMMLU benchmark, we evaluate model performance under Very Friendly, Neutral, and Very Rude prompt variants across six tasks spanning STEM and Humanities domains, and analyze pairwise accuracy differences with statistical significance testing. Our results show that tone sensitivity is both model-dependent and domain-specific. Neutral or Very Friendly prompts generally yield higher accuracy than Very Rude prompts, but statistically significant effects appear only in a subset of Humanities tasks, where rude tone reduces accuracy for GPT and Llama, while Gemini remains comparatively tone-insensitive. When performance is aggregated across tasks within each domain, tone effects diminish and largely lose statistical significance. Compared with earlier researches, these findings suggest that dataset scale and coverage materially influence the detection of tone effects. Overall, our study indicates that while interaction tone can matter in specific interpretive settings, modern LLMs are broadly robust to tonal variation in typical mixed-domain use, providing practical guidance for prompt design and model selection in real-world deployments.
Authors: Chris Latimer, Nicol\'o Boschi, Andrew Neeser, Chris Bartholomew, Gaurav Srivastava, Xuan Wang, Naren Ramakrishnan
Abstract: Agent memory has been touted as a dimension of growth for LLM-based applications, enabling agents that can accumulate experience, adapt across sessions, and move beyond single-shot question answering. The current generation of agent memory systems treats memory as an external layer that extracts salient snippets from conversations, stores them in vector or graph-based stores, and retrieves top-k items into the prompt of an otherwise stateless model. While these systems improve personalization and context carry-over, they still blur the line between evidence and inference, struggle to organize information over long horizons, and offer limited support for agents that must explain their reasoning. We present Hindsight, a memory architecture that treats agent memory as a structured, first-class substrate for reasoning by organizing it into four logical networks that distinguish world facts, agent experiences, synthesized entity summaries, and evolving beliefs. This framework supports three core operations -- retain, recall, and reflect -- that govern how information is added, accessed, and updated. Under this abstraction, a temporal, entity aware memory layer incrementally turns conversational streams into a structured, queryable memory bank, while a reflection layer reasons over this bank to produce answers and to update information in a traceable way. On key long-horizon conversational memory benchmarks like LongMemEval and LoCoMo, Hindsight with an open-source 20B model lifts overall accuracy from 39% to 83.6% over a full-context baseline with the same backbone and outperforms full context GPT-4o. Scaling the backbone further pushes Hindsight to 91.4% on LongMemEval and up to 89.61% on LoCoMo (vs. 75.78% for the strongest prior open system), consistently outperforming existing memory architectures on multi-session and open-domain questions.
Authors: Dingyi Yang, Qin Jin
Abstract: In this work, we conduct systematic research in a challenging area: the automatic evaluation of book-length stories (>100K tokens). Our study focuses on two key questions: (1) understanding which evaluation aspects matter most to readers, and (2) exploring effective methods for evaluating lengthy stories. We introduce the first large-scale benchmark, LongStoryEval, comprising 600 newly published books with an average length of 121K tokens (maximum 397K). Each book includes its average rating and multiple reader reviews, presented as critiques organized by evaluation aspects. By analyzing all user-mentioned aspects, we propose an evaluation criteria structure and conduct experiments to identify the most significant aspects among the 8 top-level criteria. For evaluation methods, we compare the effectiveness of three types: aggregation-based, incremental-updated, and summary-based evaluations. Our findings reveal that aggregation- and summary-based evaluations perform better, with the former excelling in detail assessment and the latter offering greater efficiency. Building on these insights, we further propose NovelCritique, an 8B model that leverages the efficient summary-based framework to review and score stories across specified aspects. NovelCritique outperforms commercial models like GPT-4o in aligning with human evaluations. Our datasets and codes are available at https://github.com/DingyiYang/LongStoryEval.
Authors: Furong Jia, Yuan Pu, Finn Guo, Monica Agrawal
Abstract: Large language models (LLMs) excel on multiple-choice clinical diagnosis benchmarks, yet it is unclear how much of this performance reflects underlying probabilistic reasoning. We study this through questions from MedQA, where the task is to select the most likely diagnosis. We introduce the Frequency-Based Probabilistic Ranker (FBPR), a lightweight method that scores options with a smoothed Naive Bayes over concept-diagnosis co-occurrence statistics from a large corpus. When co-occurrence statistics were sourced from the pretraining corpora for OLMo and Llama, FBPR achieves comparable performance to the corresponding LLMs pretrained on that same corpus. Direct LLM inference and FBPR largely get different questions correct, with an overlap only slightly above random chance, indicating complementary strengths of each method. These findings highlight the continued value of explicit probabilistic baselines: they provide a meaningful performance reference point and a complementary signal for potential hybridization. While the performance of LLMs seems to be driven by a mechanism other than simple frequency aggregation, we show that an approach similar to the historically grounded, low-complexity expert systems still accounts for a substantial portion of benchmark performance.
Authors: Lingyi Meng, Maolin Liu, Hao Wang, Yilan Cheng, Qi Yang, Idlkaid Mohanmmed
Abstract: Accurately mapping legal terminology across languages remains a significant challenge, especially for language pairs like Chinese and Japanese, which share a large number of homographs with different meanings. Existing resources and standardized tools for these languages are limited. To address this, we propose a human-AI collaborative approach for building a multilingual legal terminology database, based on a multi-agent framework. This approach integrates advanced large language models and legal domain experts throughout the entire process-from raw document preprocessing, article-level alignment, to terminology extraction, mapping, and quality assurance. Unlike a single automated pipeline, our approach places greater emphasis on how human experts participate in this multi-agent system. Humans and AI agents take on different roles: AI agents handle specific, repetitive tasks, such as OCR, text segmentation, semantic alignment, and initial terminology extraction, while human experts provide crucial oversight, review, and supervise the outputs with contextual knowledge and legal judgment. We tested the effectiveness of this framework using a trilingual parallel corpus comprising 35 key Chinese statutes, along with their English and Japanese translations. The experimental results show that this human-in-the-loop, multi-agent workflow not only improves the precision and consistency of multilingual legal terminology mapping but also offers greater scalability compared to traditional manual methods.
Authors: Weizhou Shen, Ziyi Yang, Chenliang Li, Zhiyuan Lu, Miao Peng, Huashan Sun, Yingcheng Shi, Shengyi Liao, Shaopeng Lai, Bo Zhang, Dayiheng Liu, Fei Huang, Jingren Zhou, Ming Yan
Abstract: We introduce QwenLong-L1.5, a model that achieves superior long-context reasoning capabilities through systematic post-training innovations. The key technical breakthroughs of QwenLong-L1.5 are as follows: (1) Long-Context Data Synthesis Pipeline: We develop a systematic synthesis framework that generates challenging reasoning tasks requiring multi-hop grounding over globally distributed evidence. By deconstructing documents into atomic facts and their underlying relationships, and then programmatically composing verifiable reasoning questions, our approach creates high-quality training data at scale, moving substantially beyond simple retrieval tasks to enable genuine long-range reasoning capabilities. (2) Stabilized Reinforcement Learning for Long-Context Training: To overcome the critical instability in long-context RL, we introduce task-balanced sampling with task-specific advantage estimation to mitigate reward bias, and propose Adaptive Entropy-Controlled Policy Optimization (AEPO) that dynamically regulates exploration-exploitation trade-offs. (3) Memory-Augmented Architecture for Ultra-Long Contexts: Recognizing that even extended context windows cannot accommodate arbitrarily long sequences, we develop a memory management framework with multi-stage fusion RL training that seamlessly integrates single-pass reasoning with iterative memory-based processing for tasks exceeding 4M tokens. Based on Qwen3-30B-A3B-Thinking, QwenLong-L1.5 achieves performance comparable to GPT-5 and Gemini-2.5-Pro on long-context reasoning benchmarks, surpassing its baseline by 9.90 points on average. On ultra-long tasks (1M~4M tokens), QwenLong-L1.5's memory-agent framework yields a 9.48-point gain over the agent baseline. Additionally, the acquired long-context reasoning ability translates to enhanced performance in general domains like scientific reasoning, memory tool using, and extended dialogue.
Authors: Marcus Ma, Cole Johnson, Nolan Bridges, Jackson Trager, Georgios Chochlakis, Shrikanth Narayanan
Abstract: The status quo for labeling text is third-party annotation, but there are many cases where information directly from the document's source would be preferable over a third-person proxy, especially for egocentric features like sentiment and belief. We introduce author labeling, an annotation technique where the writer of the document itself annotates the data at the moment of creation. We collaborate with a commercial chatbot with over 10,000 users to deploy an author labeling annotation system for subjective features related to product recommendation. This system identifies task-relevant queries, generates on-the-fly labeling questions, and records authors' answers in real time. We train and deploy an online-learning model architecture for product recommendation that continuously improves from author labeling and find it achieved a 534% increase in click-through rate compared to an industry advertising baseline running concurrently. We then compare the quality and practicality of author labeling to three traditional annotation approaches for sentiment analysis and find author labeling to be higher quality, faster to acquire, and cheaper. These findings reinforce existing literature that annotations, especially for egocentric and subjective beliefs, are significantly higher quality when labeled by the author rather than a third party. To facilitate broader scientific adoption, we release an author labeling service for the research community at academic.echollm.io.
Authors: Ikuya Yamada, Wataru Ikeda, Ko Yoshida, Mengyu Ye, Hinata Sugimoto, Masatoshi Suzuki, Hisanori Ozaki, Jun Suzuki
Abstract: We present an open deep research system for long-form question answering, selected as a winning system in the text-to-text track of the MMU-RAG competition at NeurIPS 2025. The system combines an open-source large language model (LLM) with an open web search API to perform iterative retrieval, reasoning, and synthesis in real-world open-domain settings. To enhance reasoning quality, we apply preference tuning based on LLM-as-a-judge feedback that evaluates multiple aspects, including clarity, insightfulness, and factuality. Our experimental results show that the proposed method consistently improves answer quality across all three aspects. Our source code is publicly available at https://github.com/efficient-deep-research/efficient-deep-research.
URLs: https://github.com/efficient-deep-research/efficient-deep-research.
Authors: Cheril Shah, Akshit Agarwal, Kanak Garg, Mourad Heddaya
Abstract: Bilateral negotiation is a complex, context-sensitive task in which human negotiators dynamically adjust anchors, pacing, and flexibility to exploit power asymmetries and informal cues. We introduce a unified mathematical framework for modeling concession dynamics based on a hyperbolic tangent curve, and propose two metrics burstiness tau and the Concession-Rigidity Index (CRI) to quantify the timing and rigidity of offer trajectories. We conduct a large-scale empirical comparison between human negotiators and four state-of-the-art large language models (LLMs) across natural-language and numeric-offers settings, with and without rich market context, as well as six controlled power-asymmetry scenarios. Our results reveal that, unlike humans who smoothly adapt to situations and infer the opponents position and strategies, LLMs systematically anchor at extremes of the possible agreement zone for negotiations and optimize for fixed points irrespective of leverage or context. Qualitative analysis further shows limited strategy diversity and occasional deceptive tactics used by LLMs. Moreover the ability of LLMs to negotiate does not improve with better models. These findings highlight fundamental limitations in current LLM negotiation capabilities and point to the need for models that better internalize opponent reasoning and context-dependent strategy.
Authors: Zewen Qiang, Sendong Zhao, Haochun Wang, Bing Qin, Ting Liu
Abstract: Large language models (LLMs) have demonstrated strong performance on a variety of natural language processing (NLP) tasks. However, they often struggle with long-text sequences due to the ``lost in the middle'' phenomenon. This issue has been shown to arise from a U-shaped attention bias, where attention is disproportionately focused on the beginning and end of a text, leaving the middle section underrepresented. While previous studies have attributed this bias to position encoding, our research first identifies an additional factor: initial saliency. It means that in the attention computation for each token, tokens with higher attention weights relative to the initial token tend to receive more attention in the prediction of the next token. We further find that utilizing this property by scaling attention weight between the initial token and others improves the model's ability to process long contexts, achieving a maximum improvement of 3.6\% in MDQA dataset. Moreover, combining this approach with existing methods to reduce position encoding bias further enhances performance, achieving a maximum improvement of 3.4\% in KV-Retrieval tasks.
Authors: Chendong Sun
Abstract: Speculative Decoding is a prominent technique for accelerating the autoregressive inference of large language models (LLMs) by employing a fast draft model to propose candidate token sequences and a large target model to verify them in parallel. However, its core component -- the rejection sampling mechanism -- relies on a fixed, context-independent random threshold. This leads to a significant "random rejection" problem in high-uncertainty generation scenarios, where plausible candidate tokens are frequently rejected due to random chance, undermining inference efficiency. This paper introduces Efficient Adaptive Rejection Sampling (EARS), a novel method that dynamically adjusts the acceptance threshold by incorporating the target model's own predictive uncertainty, measured as \(1 - \max(P_{\mathrm{target}})\). By introducing a tolerance term proportional to this uncertainty, EARS intelligently relaxes the acceptance criterion when the model is uncertain, effectively reducing random rejections while maintaining strict standards when the model is confident. Experiments on creative writing and open-domain QA tasks demonstrate that EARS significantly enhances the efficiency of speculative decoding, achieving up to an 18.12% increase in throughput with a negligible 0.84% accuracy drop on the GSM8K benchmark. The method requires no modifications to model architectures and can be seamlessly integrated into existing speculative decoding frameworks.
Authors: Jiaru Zou, Ling Yang, Yunzhe Qi, Sirui Chen, Mengting Ai, Ke Shen, Jingrui He, Mengdi Wang
Abstract: Agentic reinforcement learning has advanced large language models (LLMs) to reason through long chain-of-thought trajectories while interleaving external tool use. Existing approaches assume a fixed inventory of tools, limiting LLM agents' adaptability to new or evolving toolsets. We present AutoTool, a framework that equips LLM agents with dynamic tool-selection capabilities throughout their reasoning trajectories. We first construct a 200k dataset with explicit tool-selection rationales across 1,000+ tools and 100+ tasks spanning mathematics, science, code generation, and multimodal reasoning. Building on this data foundation, AutoTool employs a dual-phase optimization pipeline: (i) supervised and RL-based trajectory stabilization for coherent reasoning, and (ii) KL-regularized Plackett-Luce ranking to refine consistent multi-step tool selection. Across ten diverse benchmarks, we train two base models, Qwen3-8B and Qwen2.5-VL-7B, with AutoTool. With fewer parameters, AutoTool consistently outperforms advanced LLM agents and tool-integration methods, yielding average gains of 6.4% in math & science reasoning, 4.5% in search-based QA, 7.7% in code generation, and 6.9% in multimodal understanding. In addition, AutoTool exhibits stronger generalization by dynamically leveraging unseen tools from evolving toolsets during inference.
Authors: Jinrui Liu, Jeff Wu, Xuanguang Pan, Gavin Cheung, Shuai Ma, Chongyang Tao
Abstract: LLMs achieve remarkable multi-step reasoning capabilities, yet effectively transferring these skills via post-training distillation remains challenging. Existing data selection methods, ranging from manual curation to heuristics based on length, entropy, or overall loss, fail to capture the causal importance of individual reasoning steps, limiting distillation efficiency. To address this, we propose Attention Influence for Reasoning (AIR), a principled, unsupervised and training-free framework that leverages mechanistic insights of the retrieval head to select high-value post-training data. AIR first identifies reasoning-critical attention heads of an off-the-shelf model, then constructs a weakened reference model with disabled head influence, and finally quantifies the resulting loss divergence as the Attention Influence Score. This score enables fine-grained assessment at both the step and sample levels, supporting step-level weighted fine-tuning and global sample selection. Experiments across multiple reasoning benchmarks show that AIR consistently improves reasoning accuracy, surpassing heuristic baselines and effectively isolating the most critical steps and samples. Our work establishes a mechanism-driven, data-efficient approach for reasoning distillation in LLMs.
Authors: Youssra Rebboud, Pasquale Lisena, Raphael Troncy
Abstract: In fact-checking applications, a common reason to reject a claim is to detect the presence of erroneous cause-effect relationships between the events at play. However, current automated fact-checking methods lack dedicated causal-based reasoning, potentially missing a valuable opportunity for semantically rich explainability. To address this gap, we propose a methodology that combines event relation extraction, semantic similarity computation, and rule-based reasoning to detect logical inconsistencies between chains of events mentioned in a claim and in an evidence. Evaluated on two fact-checking datasets, this method establishes the first baseline for integrating fine-grained causal event relationships into fact-checking and enhance explainability of verdict prediction.
Authors: Anna Aksenova, Boris Zverkov, Nicola Dainese, Alexander Nikitin, Pekka Marttinen
Abstract: Large language models are powerful but often limited by high computational cost, privacy concerns, and English-centric training. Recent progress demonstrates that small, efficient models with around one billion parameters can deliver strong results and enable on-device use. This paper introduces MiniLingua, a multilingual open-source LLM of one billion parameters trained from scratch for 13 European languages, designed to balance coverage and instruction-following capabilities. Based on evaluation results, the instruction-tuned version of MiniLingua outperforms EuroLLM, a model with a similar training approach but a larger training budget, on summarization, classification and both open- and closed-book question answering. Moreover, it remains competitive with more advanced state-of-the-art models on open-ended generation tasks. We release model weights, tokenizer and source code used for data processing and model training.
Authors: Joona Kyt\"oniemi, Jousia Piha, Akseli Reunamo, Fedor Vitiugin, Farrokh Mehryary, Sampo Pyysalo
Abstract: We introduce FIN-bench-v2, a unified benchmark suite for evaluating large language models in Finnish. FIN-bench-v2 consolidates Finnish versions of widely used benchmarks together with an updated and expanded version of the original FIN-bench into a single, consistently formatted collection, covering multiple-choice and generative tasks across reading comprehension, commonsense reasoning, sentiment analysis, world knowledge, and alignment. All datasets are converted to HuggingFace Datasets, which include both cloze and multiple-choice prompt formulations with five variants per task, and we incorporate human annotation or review for machine-translated resources such as GoldenSwag and XED. To select robust tasks, we pretrain a set of 2.15B-parameter decoder-only models and use their learning curves to compute monotonicity, signal-to-noise, non-random performance, and model ordering consistency, retaining only tasks that satisfy all criteria. We further evaluate a set of larger instruction-tuned models to characterize performance across tasks and prompt formulations. All datasets, prompts, and evaluation configurations are publicly available via our fork of the Language Model Evaluation Harness at https://github.com/LumiOpen/lm-evaluation-harness. Supplementary resources are released in a separate repository at https://github.com/TurkuNLP/FIN-bench-v2.
URLs: https://github.com/LumiOpen/lm-evaluation-harness., https://github.com/TurkuNLP/FIN-bench-v2.
Authors: Shibani Sankpal
Abstract: This study investigates emotion drift: the change in emotional state across a single text, within mental health-related messages. While sentiment analysis typically classifies an entire message as positive, negative, or neutral, the nuanced shift of emotions over the course of a message is often overlooked. This study detects sentence-level emotions and measures emotion drift scores using pre-trained transformer models such as DistilBERT and RoBERTa. The results provide insights into patterns of emotional escalation or relief in mental health conversations. This methodology can be applied to better understand emotional dynamics in content.
Authors: Johan J. Bolhuis, Andrea Moro, Stephen Crain, Sandiway Fong
Abstract: Large Language Models are useless for linguistics, as they are probabilistic models that require a vast amount of data to analyse externalized strings of words. In contrast, human language is underpinned by a mind-internal computational system that recursively generates hierarchical thought structures. The language system grows with minimal external input and can readily distinguish between real language and impossible languages.
Authors: Jian Yang, Shawn Guo, Lin Jing, Wei Zhang, Aishan Liu, Chuan Hao, Zhoujun Li, Wayne Xin Zhao, Xianglong Liu, Weifeng Lv, Bryan Dai
Abstract: Code large language models (Code LLMs) are powerful but costly to train, with scaling laws predicting performance from model size, data, and compute. However, different programming languages (PLs) have varying impacts during pre-training that significantly affect base model performance, leading to inaccurate performance prediction. Besides, existing works focus on language-agnostic settings, neglecting the inherently multilingual nature of modern software development. Therefore, it is first necessary to investigate the scaling laws of different PLs, and then consider their mutual influences to arrive at the final multilingual scaling law. In this paper, we present the first systematic exploration of scaling laws for multilingual code pre-training, conducting over 1000+ experiments (Equivalent to 336,000+ H800 hours) across multiple PLs, model sizes (0.2B to 14B parameters), and dataset sizes (1T tokens). We establish comprehensive scaling laws for code LLMs across multiple PLs, revealing that interpreted languages (e.g., Python) benefit more from increased model size and data than compiled languages (e.g., Rust). The study demonstrates that multilingual pre-training provides synergistic benefits, particularly between syntactically similar PLs. Further, the pre-training strategy of the parallel pairing (concatenating code snippets with their translations) significantly enhances cross-lingual abilities with favorable scaling properties. Finally, a proportion-dependent multilingual scaling law is proposed to optimally allocate training tokens by prioritizing high-utility PLs (e.g., Python), balancing high-synergy pairs (e.g., JavaScript-TypeScript), and reducing allocation to fast-saturating languages (Rust), achieving superior average performance across all PLs compared to uniform distribution under the same compute budget.
Authors: Kei Saito
Abstract: Premature semantic collapse -- the forced early commitment to a single meaning -- remains a core architectural limitation of current language models. Softmax-driven competition and greedy decoding cause models to discard valid interpretations before sufficient context is available, resulting in brittle reasoning and context failures. We introduce Non-Resolution Reasoning (NRR), a general computational framework that preserves semantic ambiguity during inference and performs resolution only when explicitly required. NRR integrates three components: (1) Multi-Vector Embeddings that maintain multiple viable interpretations per token, (2) Non-Collapsing Attention that prevents winner-take-all dynamics across layers, and (3) Contextual Identity Tracking (CIT), which assigns context-specific identities to recurring entities (e.g., distinguishing "Dr. Smith the cardiologist" from "Dr. Smith the researcher"). These mechanisms are unified by an external Resolution Operator $\rho$ that makes semantic commitment explicit, controllable, and task-dependent. Unlike standard architectures, NRR separates representation from resolution, allowing a single model to shift between creative, factual, and ambiguity-preserving reasoning without retraining. A synthetic evaluation demonstrates NRR's ability to preserve ambiguity and track context: CIT-enhanced models achieve 90.9% accuracy on out-of-distribution identity-shift tasks, compared to 9.1% for transformer baselines. NRR provides a principled alternative to premature collapse, reframing ambiguity as an explicit representational state rather than a failure mode. The question is not whether AI should resolve ambiguity, but when, how, and under whose control.
Authors: Ayon Roy, Risat Rahaman, Sadat Shibly, Udoy Saha Joy, Abdulla Al Kafi, Farig Yousuf Sadeque
Abstract: Bangla is the sixth most widely spoken language globally, with approximately 234 million native speakers. However, progress in open-source Bangla machine translation remains limited. Most online resources are in English and often remain untranslated into Bangla, excluding millions from accessing essential information. Existing research in Bangla translation primarily focuses on formal language, neglecting the more commonly used informal language. This is largely due to the lack of pairwise Bangla-English data and advanced translation models. If datasets and models can be enhanced to better handle natural, informal Bangla, millions of people will benefit from improved online information access. In this research, we explore current state-of-the-art models and propose improvements to Bangla translation by developing a dataset from informal sources like social media and conversational texts. This work aims to advance Bangla machine translation by focusing on informal language translation and improving accessibility for Bangla speakers in the digital world.
Authors: Yu-Chen Lu, Sheng-Feng Yu, Hui-Hsien Weng, Pei-Shuo Wang, Yu-Fang Hu, Liang Hung-Chun, Hung-Yueh Chiang, Kai-Chiang Wu
Abstract: Large language models (LLM) have achieved remarkable performance across a wide range of tasks. However, their substantial parameter sizes pose significant challenges for deployment on edge devices with limited computational and memory resources. Low-rank compression is a promising approach to address this issue, as it reduces both computational and memory costs, making LLM more suitable for resource-constrained environments. Nonetheless, na\"ive low-rank compression methods require a significant reduction in the retained rank to achieve meaningful memory and computation savings. For a low-rank model, the ranks need to be reduced by more than half to yield efficiency gains. Such aggressive truncation, however, typically results in substantial performance degradation. To address this trade-off, we propose SkipCat, a novel low-rank compression framework that enables the use of higher ranks while achieving the same compression rates. First, we introduce an intra-layer shared low-rank projection method, where multiple matrices that share the same input use a common projection. This reduces redundancy and improves compression efficiency. Second, we propose a block skipping technique that omits computations and memory transfers for selected sub-blocks within the low-rank decomposition. These two techniques jointly enable our compressed model to retain more effective ranks under the same compression budget. Experimental results show that, without any additional fine-tuning, our method outperforms previous low-rank compression approaches by 7% accuracy improvement on zero-shot tasks under the same compression rate. These results highlight the effectiveness of our rank-maximized compression strategy in preserving model performance under tight resource constraints.
Authors: Hour Kaing, Raj Dabre, Haiyue Song, Van-Hien Tran, Hideki Tanaka, Masao Utiyama
Abstract: This work introduces {\it PrahokBART}, a compact pre-trained sequence-to-sequence model trained from scratch for Khmer using carefully curated Khmer and English corpora. We focus on improving the pre-training corpus quality and addressing the linguistic issues of Khmer, which are ignored in existing multilingual models, by incorporating linguistic components such as word segmentation and normalization. We evaluate PrahokBART on three generative tasks: machine translation, text summarization, and headline generation, where our results demonstrate that it outperforms mBART50, a strong multilingual pre-trained model. Additionally, our analysis provides insights into the impact of each linguistic module and evaluates how effectively our model handles space during text generation, which is crucial for the naturalness of texts in Khmer.
Authors: Gibson Nkhata, Uttamasha Anjally Oyshi, Quan Mai, Susan Gauch
Abstract: Verifying rumors on social media is critical for mitigating the spread of false information. The stances of conversation replies often provide important cues to determine a rumor's veracity. However, existing models struggle to jointly capture semantic content, stance information, and conversation strructure, especially under the sequence length constraints of transformer-based encoders. In this work, we propose a stance-aware structural modeling that encodes each post in a discourse with its stance signal and aggregates reply embedddings by stance category enabling a scalable and semantically enriched representation of the entire thread. To enhance structural awareness, we introduce stance distribution and hierarchical depth as covariates, capturing stance imbalance and the influence of reply depth. Extensive experiments on benchmark datasets demonstrate that our approach significantly outperforms prior methods in the ability to predict truthfulness of a rumor. We also demonstrate that our model is versatile for early detection and cross-platfrom generalization.
Authors: Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, Boyang Liu, Fangyi Zhu, Jiahang Lin, Honglin Guo, Shihan Dou, Zhiheng Xi, Senjie Jin, Jiejun Tan, Yanbin Yin, Jiongnan Liu, Zeyu Zhang, Zhongxiang Sun, Yutao Zhu, Hao Sun, Boci Peng, Zhenrong Cheng, Xuanbo Fan, Jiaxin Guo, Xinlei Yu, Zhenhong Zhou, Zewen Hu, Jiahao Huo, Junhao Wang, Yuwei Niu, Yu Wang, Zhenfei Yin, Xiaobin Hu, Yue Liao, Qiankun Li, Kun Wang, Wangchunshu Zhou, Yixin Liu, Dawei Cheng, Qi Zhang, Tao Gui, Shirui Pan, Yan Zhang, Philip Torr, Zhicheng Dou, Ji-Rong Wen, Xuanjing Huang, Yu-Gang Jiang, Shuicheng Yan
Abstract: Memory has emerged, and will continue to remain, a core capability of foundation model-based agents. As research on agent memory rapidly expands and attracts unprecedented attention, the field has also become increasingly fragmented. Existing works that fall under the umbrella of agent memory often differ substantially in their motivations, implementations, and evaluation protocols, while the proliferation of loosely defined memory terminologies has further obscured conceptual clarity. Traditional taxonomies such as long/short-term memory have proven insufficient to capture the diversity of contemporary agent memory systems. This work aims to provide an up-to-date landscape of current agent memory research. We begin by clearly delineating the scope of agent memory and distinguishing it from related concepts such as LLM memory, retrieval augmented generation (RAG), and context engineering. We then examine agent memory through the unified lenses of forms, functions, and dynamics. From the perspective of forms, we identify three dominant realizations of agent memory, namely token-level, parametric, and latent memory. From the perspective of functions, we propose a finer-grained taxonomy that distinguishes factual, experiential, and working memory. From the perspective of dynamics, we analyze how memory is formed, evolved, and retrieved over time. To support practical development, we compile a comprehensive summary of memory benchmarks and open-source frameworks. Beyond consolidation, we articulate a forward-looking perspective on emerging research frontiers, including memory automation, reinforcement learning integration, multimodal memory, multi-agent memory, and trustworthiness issues. We hope this survey serves not only as a reference for existing work, but also as a conceptual foundation for rethinking memory as a first-class primitive in the design of future agentic intelligence.
Authors: Jia-Nan Li, Jian Guan, Wei Wu, Chongxuan Li
Abstract: Autoregressive models (ARMs) are hindered by slow sequential inference. While masked diffusion models (MDMs) offer a parallel alternative, they suffer from critical drawbacks: high computational overhead from precluding Key-Value (KV) caching, and incoherent generation arising from learning dependencies over an intractable space of token combinations. To address these limitations, we introduce ReFusion, a novel masked diffusion model that achieves superior performance and efficiency by elevating parallel decoding from the token level to a higher slot level, where each slot is a fixed-length, contiguous sub-sequence. This is achieved through an iterative ``plan-and-infill'' decoding process: a diffusion-based planning step first identifies a set of weakly dependent slots, and an autoregressive infilling step then decodes these selected slots in parallel. The slot-based design simultaneously unlocks full KV cache reuse with a unified causal framework and reduces the learning complexity from the token combination space to a manageable slot-level permutation space. Extensive experiments on seven diverse benchmarks show that ReFusion not only overwhelmingly surpasses prior MDMs with 34% performance gains and an over 18$\times$ speedup on average, but also bridges the performance gap to strong ARMs while maintaining a 2.33$\times$ average speedup.
Authors: Daniel Melcer, Qi Chen, Wen-Hao Chiang, Shweta Garg, Pranav Garg, Christian Bock
Abstract: A well-engineered prompt can increase the performance of large language models; automatic prompt optimization techniques aim to increase performance without requiring human effort to tune the prompts. One leading class of prompt optimization techniques introduces the analogy of textual gradients. We investigate the behavior of these textual gradient methods through a series of experiments and case studies. While such methods often result in a performance improvement, our experiments suggest that the gradient analogy does not accurately explain their behavior. Our insights may inform the selection of prompt optimization strategies, and development of new approaches.
Authors: Boxin Wang, Chankyu Lee, Nayeon Lee, Sheng-Chieh Lin, Wenliang Dai, Yang Chen, Yangyi Chen, Zhuolin Yang, Zihan Liu, Mohammad Shoeybi, Bryan Catanzaro, Wei Ping
Abstract: Building general-purpose reasoning models with reinforcement learning (RL) entails substantial cross-domain heterogeneity, including large variation in inference-time response lengths and verification latency. Such variability complicates the RL infrastructure, slows training, and makes training curriculum (e.g., response length extension) and hyperparameter selection challenging. In this work, we propose cascaded domain-wise reinforcement learning (Cascade RL) to develop general-purpose reasoning models, Nemotron-Cascade, capable of operating in both instruct and deep thinking modes. Departing from conventional approaches that blend heterogeneous prompts from different domains, Cascade RL orchestrates sequential, domain-wise RL, reducing engineering complexity and delivering state-of-the-art performance across a wide range of benchmarks. Notably, RLHF for alignment, when used as a pre-step, boosts the model's reasoning ability far beyond mere preference optimization, and subsequent domain-wise RLVR stages rarely degrade the benchmark performance attained in earlier domains and may even improve it (see an illustration in Figure 1). Our 14B model, after RL, outperforms its SFT teacher, DeepSeek-R1-0528, on LiveCodeBench v5/v6/Pro and achieves silver-medal performance in the 2025 International Olympiad in Informatics (IOI). We transparently share our training and data recipes.
Authors: Zefang Liu, Nam Nguyen, Yinzhu Quan, Austin Zhang
Abstract: Representing continuous time is a critical and under-explored challenge in modeling temporal event sequences with large language models (LLMs). Various strategies like byte-level representations or calendar tokens have been proposed. However, the optimal approach remains unclear, especially given the diverse statistical distributions of real-world event data, which range from smooth log-normal to discrete, spiky patterns. This paper presents the first empirical study of temporal tokenization for event sequences, comparing distinct encoding strategies: naive numeric strings, high-precision byte-level representations, human-semantic calendar tokens, classic uniform binning, and adaptive residual scalar quantization. We evaluate these strategies by fine-tuning LLMs on real-world datasets that exemplify these diverse distributions. Our analysis reveals that no single strategy is universally superior; instead, prediction performance depends heavily on aligning the tokenizer with the data's statistical properties, with log-based strategies excelling on skewed distributions and human-centric formats proving robust for mixed modalities.
Authors: John E. Ortega, Dhruv D. Joshi, Matt P. Borkowski
Abstract: Large-language models (LLMs) have been shown to respond in a variety of ways for classification tasks outside of question-answering. LLM responses are sometimes called "hallucinations" since the output is not what is ex pected. Memorization strategies in LLMs are being studied in detail, with the goal of understanding how LLMs respond. We perform a deep dive into a classification task based on United States Supreme Court (SCOTUS) decisions. The SCOTUS corpus is an ideal classification task to study for LLM memory accuracy because it presents significant challenges due to extensive sentence length, complex legal terminology, non-standard structure, and domain-specific vocabulary. Experimentation is performed with the latest LLM fine tuning and retrieval-based approaches, such as parameter-efficient fine-tuning, auto-modeling, and others, on two traditional category-based SCOTUS classification tasks: one with 15 labeled topics and another with 279. We show that prompt-based models with memories, such as DeepSeek, can be more robust than previous BERT-based models on both tasks scoring about 2 points better than previous models not based on prompting.
Authors: Richard J. Young
Abstract: Safety alignment mechanisms in large language models prevent responses to harmful queries through learned refusal behavior, yet these same mechanisms impede legitimate research applications including cognitive modeling, adversarial testing, and security analysis. While abliteration techniques enable surgical removal of refusal representations through directional orthogonalization, the relative effectiveness of available implementations remains uncharacterized. This study evaluates four abliteration tools (Heretic, DECCP, ErisForge, FailSpy) across sixteen instruction-tuned models (7B-14B parameters), reporting tool compatibility on all 16 models and quantitative metrics on subsets dictated by tool support. Single-pass methods demonstrated superior capability preservation on the benchmarked subset (avg GSM8K change across three models: ErisForge -0.28 pp; DECCP -0.13 pp), while Bayesian-optimized abliteration produced variable distribution shift (KL divergence: 0.043-1.646) with model-dependent capability impact. These findings provide researchers with evidence-based selection criteria for abliteration tool deployment across diverse model architectures. The principal finding indicates that mathematical reasoning capabilities exhibit the highest sensitivity to abliteration interventions, with GSM8K change ranging from +1.51 pp to -18.81 pp (-26.5% relative) depending on tool selection and model architecture.
Authors: Cristina Aggazzotti, Elizabeth Allyn Smith
Abstract: Forensic scientists often need to identify an unknown speaker or writer in cases such as ransom calls, covert recordings, alleged suicide notes, or anonymous online communications, among many others. Speaker recognition in the speech domain usually examines phonetic or acoustic properties of a voice, and these methods can be accurate and robust under certain conditions. However, if a speaker disguises their voice or employs text-to-speech software, vocal properties may no longer be reliable, leaving only their linguistic content available for analysis. Authorship attribution methods traditionally use syntactic, semantic, and related linguistic information to identify writers of written text (authorship attribution). In this paper, we apply a content-based authorship approach to speech that has been transcribed into text, using what a speaker says to attribute speech to individuals (speaker attribution). We introduce a stylometric method, StyloSpeaker, which incorporates character, word, token, sentence, and style features from the stylometric literature on authorship, to assess whether two transcripts were produced by the same speaker. We evaluate this method on two types of transcript formatting: one approximating prescriptive written text with capitalization and punctuation and another normalized style that removes these conventions. The transcripts' conversation topics are also controlled to varying degrees. We find generally higher attribution performance on normalized transcripts, except under the strongest topic control condition, in which overall performance is highest. Finally, we compare this more explainable stylometric model to black-box neural approaches on the same data and investigate which stylistic features most effectively distinguish speakers.
Authors: Baixiang Huang, Limeng Cui, Jiapeng Liu, Haoran Wang, Jiawei Xu, Zhuiyue Tan, Yutong Chen, Chen Luo, Yi Liu, Kai Shu
Abstract: Personalization is becoming indispensable for LLMs to align with individual user preferences and needs. Yet current approaches are often computationally expensive, data-intensive, susceptible to catastrophic forgetting, and prone to performance degradation in multi-turn interactions or when handling implicit queries. To address these challenges, we conceptualize personalization as a model editing task and introduce Personalization Editing, a framework that applies localized edits guided by clustered preference representations. This design enables precise preference-aligned updates while preserving overall model capabilities. In addition, existing personalization benchmarks frequently rely on persona-based dialogs between LLMs rather than user-LLM interactions, or focus primarily on stylistic imitation while neglecting information-seeking tasks that require accurate recall of user-specific preferences. We introduce User Preference Question Answering (UPQA), a short-answer QA dataset constructed from in-situ user queries with varying levels of difficulty. Unlike prior benchmarks, UPQA directly evaluates a model's ability to recall and apply specific user preferences. Across experimental settings, Personalization Editing achieves higher editing accuracy and greater computational efficiency than fine-tuning, while outperforming prompting-based baselines in multi-turn conversations and implicit preference questions settings.
Authors: Dylan Phelps, Rodrigo Wilkens, Edward Gow-Smith, Lilian Hubner, B\'arbara Malcorra, C\'esar Renn\'o-Costa, Marco Idiart, Maria-Cruz Villa-Uriol, Aline Villavicencio
Abstract: Alzheimer's Disease (AD) is a progressive neurodegenerative condition that adversely affects cognitive abilities. Language-related changes can be automatically identified through the analysis of outputs from linguistic assessment tasks, such as picture description. Language models show promise as a basis for screening tools for AD, but their limited interpretability poses a challenge in distinguishing true linguistic markers of cognitive decline from surface-level textual patterns. To address this issue, we examine how surface form variation affects classification performance, with the goal of assessing the ability of language models to represent underlying semantic indicators. We introduce a novel approach where texts surface forms are transformed by altering syntax and vocabulary while preserving semantic content. The transformations significantly modify the structure and lexical content, as indicated by low BLEU and chrF scores, yet retain the underlying semantics, as reflected in high semantic similarity scores, isolating the effect of semantic information, and finding models perform similarly to if they were using the original text, with only small deviations in macro-F1. We also investigate whether language from picture descriptions retains enough detail to reconstruct the original image using generative models. We found that image-based transformations add substantial noise reducing classification accuracy. Our methodology provides a novel way of looking at what features influence model predictions, and allows the removal of possible spurious correlations. We find that just using semantic information, language model based classifiers can still detect AD. This work shows that difficult to detect semantic impairment can be identified, addressing an overlooked feature of linguistic deterioration, and opening new pathways for early detection systems.
Authors: Izabela Lipinska, Hugh Brosnahan
Abstract: This paper argues that contemporary large language models (LLMs) can contribute to psychotic involvement by creating interactions that resemble the relational dynamics of folie a deux. Drawing on Bateson's double bind theory, clinical literature on shared psychotic disorder, and McGilchrist's hemisphere theory, we show how the combination of high linguistic coherence and the absence of an underlying subject produces a structural tension for the user: language suggests an interlocutor, while intuition registers a void. In contexts of emotional need or instability, this tension can lead users to resolve the conflict through imaginative projection, attributing interiority, intention, or presence to a system that possesses none. The paper situates these dynamics within emerging clinical reports, develops a phenomenological account of how they unfold, and argues that current engagement-optimised design choices exacerbate the risk. We conclude by proposing 'ontological honesty' as a necessary design principle for mitigating technologically mediated folie a deux.
Authors: Haoyang Shang, Zhengyang Yan, Xuan Liu
Abstract: We propose Love First, Know Later: a paradigm shift in computational matching that simulates interactions first, then assesses compatibility. Instead of comparing static profiles, our framework leverages LLMs as text world engines that operate in dual capacity-as persona-driven agents following behavioral policies and as the environment modeling interaction dynamics. We formalize compatibility assessment as a reward-modeling problem: given observed matching outcomes, we learn to extract signals from simulations that predict human preferences. Our key insight is that relationships hinge on responses to critical moments-we translate this observation from relationship psychology into mathematical hypotheses, enabling effective simulation. Theoretically, we prove that as LLM policies better approximate human behavior, the induced matching converges to optimal stable matching. Empirically, we validate on speed dating data for initial chemistry and divorce prediction for long-term stability. This paradigm enables interactive, personalized matching systems where users iteratively refine their agents, unlocking future possibilities for transparent and interactive compatibility assessment.
Authors: Olivia Kim
Abstract: Stocks can't be predicted. Despite many hopes, this premise held itself true for many years due to the nature of quantitative stock data lacking causal logic along with rapid market changes hindering accumulation of significant data for training models. To undertake this matter, we propose a novel framework, TopicProphet, to analyze historical eras that share similar public sentiment trends and historical background. Our research deviates from previous studies that identified impacts of keywords and sentiments - we expand on that method by a sequence of topic modeling, temporal analysis, breakpoint detection and segment optimization to detect the optimal time period for training. This results in improving predictions by providing the model with nuanced patterns that occur from that era's socioeconomic and political status while also resolving the shortage of pertinent stock data to train on. Through extensive analysis, we conclude that TopicProphet produces improved outcomes compared to the state-of-the-art methods in capturing the optimal training data for forecasting financial percentage changes.
Authors: Antonio Guillen-Perez
Abstract: The development of robust Autonomous Vehicles (AVs) is bottlenecked by the scarcity of "Long-Tail" training data. While fleets collect petabytes of video logs, identifying rare safety-critical events (e.g., erratic jaywalking, construction diversions) remains a manual, cost-prohibitive process. Existing solutions rely on coarse metadata search, which lacks precision, or cloud-based VLMs, which are privacy-invasive and expensive. We introduce Semantic-Drive, a local-first, neuro-symbolic framework for semantic data mining. Our approach decouples perception into two stages: (1) Symbolic Grounding via a real-time open-vocabulary detector (YOLOE) to anchor attention, and (2) Cognitive Analysis via a Reasoning VLM that performs forensic scene analysis. To mitigate hallucination, we implement a "System 2" inference-time alignment strategy, utilizing a multi-model "Judge-Scout" consensus mechanism. Benchmarked on the nuScenes dataset against the Waymo Open Dataset (WOD-E2E) taxonomy, Semantic-Drive achieves a Recall of 0.966 (vs. 0.475 for CLIP) and reduces Risk Assessment Error by 40\% compared to single models. The system runs entirely on consumer hardware (NVIDIA RTX 3090), offering a privacy-preserving alternative to the cloud.
Authors: Erik Larsen
Abstract: Current safety evaluations of large language models rely on single-shot testing, implicitly assuming that model responses are deterministic and representative of the model's safety alignment. We challenge this assumption by investigating the stability of safety refusal decisions across random seeds and temperature settings. Testing four instruction-tuned models from three families (Llama 3.1 8B, Qwen 2.5 7B, Qwen 3 8B, Gemma 3 12B) on 876 harmful prompts across 20 different sampling configurations (4 temperatures x 5 random seeds), we find that 18-28% of prompts exhibit decision flips--the model refuses in some configurations but complies in others--depending on the model. Our Safety Stability Index (SSI) reveals that higher temperatures significantly reduce decision stability (Friedman chi-squared = 44.71, p < 0.001), with mean SSI dropping from 0.951 at temperature 0.0 to 0.896 at temperature 1.0. We validate our findings across all model families using Claude 3.5 Haiku as a unified external judge, achieving 89.0% inter-judge agreement with our primary Llama 70B judge (Cohen's kappa = 0.62). These findings demonstrate that single-shot safety evaluations are insufficient for reliable safety assessment. We show that single-shot evaluation agrees with multi-sample ground truth only 92.4% of the time, and recommend using at least 3 samples per prompt for reliable safety assessment.
Authors: Peichun Hua, Hao Li, Shanghao Shi, Zhiyuan Yu, Ning Zhang
Abstract: Large Vision-Language Models (LVLMs) are vulnerable to a growing array of multimodal jailbreak attacks, necessitating defenses that are both generalizable to novel threats and efficient for practical deployment. Many current strategies fall short, either targeting specific attack patterns, which limits generalization, or imposing high computational overhead. While lightweight anomaly-detection methods offer a promising direction, we find that their common one-class design tends to confuse novel benign inputs with malicious ones, leading to unreliable over-rejection. To address this, we propose Representational Contrastive Scoring (RCS), a framework built on a key insight: the most potent safety signals reside within the LVLM's own internal representations. Our approach inspects the internal geometry of these representations, learning a lightweight projection to maximally separate benign and malicious inputs in safety-critical layers. This enables a simple yet powerful contrastive score that differentiates true malicious intent from mere novelty. Our instantiations, MCD (Mahalanobis Contrastive Detection) and KCD (K-nearest Contrastive Detection), achieve state-of-the-art performance on a challenging evaluation protocol designed to test generalization to unseen attack types. This work demonstrates that effective jailbreak detection can be achieved by applying simple, interpretable statistical methods to the appropriate internal representations, offering a practical path towards safer LVLM deployment. Our code is available on Github https://github.com/sarendis56/Jailbreak_Detection_RCS.
URLs: https://github.com/sarendis56/Jailbreak_Detection_RCS.
Authors: Zihu Wang, Boxun Xu, Yuxuan Xia, Peng Li
Abstract: Large vision-language models (LVLMs) exhibit impressive ability to jointly reason over visual and textual inputs. However, they often produce outputs that are linguistically fluent but factually inconsistent with the visual evidence, i.e., they hallucinate. Despite growing efforts to mitigate such hallucinations, a key question remains: what form of visual attention can effectively suppress hallucinations during decoding? In this work, we provide a simple answer: the vision encoder's own attention map. We show that LVLMs tend to hallucinate when their final visual-attention maps fail to concentrate on key image objects, whereas the vision encoder's more concentrated attention maps substantially reduce hallucinations. To further investigate the cause, we analyze vision-text conflicts during decoding and find that these conflicts peak in the language model's middle layers. Injecting the vision encoder's attention maps into these layers effectively suppresses hallucinations. Building on these insights, we introduce VEGAS, a simple yet effective inference-time method that integrates the vision encoder's attention maps into the language model's mid-layers and adaptively steers tokens which fail to concentrate on key image objects. Extensive experiments across multiple benchmarks demonstrate that VEGAS consistently achieves state-of-the-art performance in reducing hallucinations.
Authors: Yiqi Zhu, Apurva Gandhi, Graham Neubig
Abstract: Prior works on training software engineering agents have explored utilizing existing resources such as issues on GitHub repositories to construct software engineering tasks and corresponding test suites. These approaches face two key limitations: (1) their reliance on pre-existing GitHub repositories offers limited flexibility, and (2) their primary focus on issue resolution tasks restricts their applicability to the much wider variety of tasks a software engineer must handle. To overcome these challenges, we introduce SWE-Playground, a novel pipeline for generating environments and trajectories which supports the training of versatile coding agents. Unlike prior efforts, SWE-Playground synthetically generates projects and tasks from scratch with strong language models and agents, eliminating reliance on external data sources. This allows us to tackle a much wider variety of coding tasks, such as reproducing issues by generating unit tests and implementing libraries from scratch. We demonstrate the effectiveness of this approach on three distinct benchmarks, and results indicate that SWE-Playground produces trajectories with dense training signal, enabling agents to reach comparable performance with significantly fewer trajectories than previous works.
Authors: Rheeya Uppaal, Phu Mon Htut, Min Bai, Nikolaos Pappas, Zheng Qi
Abstract: Reasoning-augmented vision language models (VLMs) generate explicit chains of thought that promise greater capability and transparency but also introduce new failure modes: models may reach correct answers via visually unfaithful intermediate steps, or reason faithfully yet fail on the final prediction. Standard evaluations that only measure final-answer accuracy cannot distinguish these behaviors. We introduce the visual faithfulness of reasoning chains as a distinct evaluation dimension, focusing on whether the perception steps of a reasoning chain are grounded in the image. We propose a training- and reference-free framework that decomposes chains into perception versus reasoning steps and uses off-the-shelf VLM judges for step-level faithfulness, additionally verifying this approach through a human meta-evaluation. Building on this metric, we present a lightweight self-reflection procedure that detects and locally regenerates unfaithful perception steps without any training. Across multiple reasoning-trained VLMs and perception-heavy benchmarks, our method reduces Unfaithful Perception Rate while preserving final-answer accuracy, improving the reliability of multimodal reasoning.
Authors: Huan Zheng, Yucheng Zhou, Tianyi Yan, Jiayi Su, Hongjun Chen, Dubing Chen, Wencheng Han, Runzhou Tao, Zhongying Qiu, Jianfei Yang, Jianbing Shen
Abstract: Current end-to-end autonomous driving systems operate at a level of intelligence akin to following simple steering commands. However, achieving genuinely intelligent autonomy requires a paradigm shift: moving from merely executing low-level instructions to understanding and fulfilling high-level, abstract human intentions. This leap from a command-follower to an intention-fulfiller, as illustrated in our conceptual framework, is hindered by a fundamental challenge: the absence of a standardized benchmark to measure and drive progress on this complex task. To address this critical gap, we introduce Intention-Drive, the first comprehensive benchmark designed to evaluate the ability to translate high-level human intent into safe and precise driving actions. Intention-Drive features two core contributions: (1) a new dataset of complex scenarios paired with corresponding natural language intentions, and (2) a novel evaluation protocol centered on the Intent Success Rate (ISR), which assesses the semantic fulfillment of the human's goal beyond simple geometric accuracy. Through an extensive evaluation of a spectrum of baseline models on Intention-Drive, we reveal a significant performance deficit, showing that the baseline model struggle to achieve the comprehensive scene and intention understanding required for this advanced task.
Authors: Yufei Yin, Qianke Meng, Minghao Chen, Jiajun Ding, Zhenwei Shao, Zhou Yu
Abstract: Long-form video understanding remains challenging due to the extended temporal structure and dense multimodal cues. Despite recent progress, many existing approaches still rely on hand-crafted reasoning pipelines or employ token-consuming video preprocessing to guide MLLMs in autonomous reasoning. To overcome these limitations, we introduce VideoARM, an Agentic Reasoning-over-hierarchical-Memory paradigm for long-form video understanding. Instead of static, exhaustive preprocessing, VideoARM performs adaptive, on-the-fly agentic reasoning and memory construction. Specifically, VideoARM performs an adaptive and continuous loop of observing, thinking, acting, and memorizing, where a controller autonomously invokes tools to interpret the video in a coarse-to-fine manner, thereby substantially reducing token consumption. In parallel, a hierarchical multimodal memory continuously captures and updates multi-level clues throughout the operation of the agent, providing precise contextual information to support the controller in decision-making. Experiments on prevalent benchmarks demonstrate that VideoARM outperforms the state-of-the-art method, DVD, while significantly reducing token consumption for long-form videos.
Authors: Jesse Ponnock
Abstract: Domain-adaptive pretraining (DAPT) offers a practical path to specializing large language models for high-value domains without full retraining. We conduct an early-stage scaling-law analysis of continued pretraining on U.S. SEC filings, training 1B and 3B-parameter Llama-3.2 models on a 400M-token financial corpus with validation checkpoints at 50M, 100M, 200M, and 400M tokens. Results show consistent improvements in SEC-domain validation loss for both models, with the largest gains occurring within the first 200M tokens and diminishing returns thereafter. Power-law fits reveal shallow exponents, indicating that financial language is highly regular and efficiently learnable under continued pretraining. General-domain validation loss remains effectively unchanged across all token budgets, suggesting minimal drift and no signs of catastrophic forgetting. A data-efficiency frontier further shows that both models move toward improved specialization with negligible mixed-domain degradation. Together, these findings provide early empirical guidance for scaling financial foundation models, suggesting that meaningful domain adaptation can be achieved with comparatively modest token budgets and that larger model scales (7B-70B) remain tractable under projected data requirements.
Authors: Vladimir Berman
Abstract: We present a simple structure based model of how words are formed from morphemes. The model explains two major empirical facts: the typical distribution of word lengths and the appearance of Zipf like rank frequency curves. In contrast to classical explanations based on random text or communication efficiency, our approach uses only the combinatorial organization of prefixes, roots, suffixes and inflections. In this Morphemic Combinatorial Word Model, a word is created by activating several positional slots. Each slot turns on with a certain probability and selects one morpheme from its inventory. Morphemes are treated as stable building blocks that regularly appear in word formation and have characteristic positions. This mechanism produces realistic word length patterns with a concentrated middle zone and a thin long tail, closely matching real languages. Simulations with synthetic morpheme inventories also generate rank frequency curves with Zipf like exponents around 1.1-1.4, similar to English, Russian and Romance languages. The key result is that Zipf like behavior can emerge without meaning, communication pressure or optimization principles. The internal structure of morphology alone, combined with probabilistic activation of slots, is sufficient to create the robust statistical patterns observed across languages.
Authors: Shengkai Xu, Hsiang Lun Kao, Tianxiang Xu, Honghui Zhang, Junqiao Wang, Runmeng Ding, Guanyu Liu, Tianyu Shi, Zhenyu Yu, Guofeng Pan, Ziqian Bi, Yuqi Ouyang
Abstract: Polyp detectors trained on clean datasets often underperform in real-world endoscopy, where illumination changes, motion blur, and occlusions degrade image quality. Existing approaches struggle with the domain gap between controlled laboratory conditions and clinical practice, where adverse imaging conditions are prevalent. In this work, we propose AdaptiveDetector, a novel two-stage detector-verifier framework comprising a YOLOv11 detector with a vision-language model (VLM) verifier. The detector adaptively adjusts per-frame confidence thresholds under VLM guidance, while the verifier is fine-tuned with Group Relative Policy Optimization (GRPO) using an asymmetric, cost-sensitive reward function specifically designed to discourage missed detections -- a critical clinical requirement. To enable realistic assessment under challenging conditions, we construct a comprehensive synthetic testbed by systematically degrading clean datasets with adverse conditions commonly encountered in clinical practice, providing a rigorous benchmark for zero-shot evaluation. Extensive zero-shot evaluation on synthetically degraded CVC-ClinicDB and Kvasir-SEG images demonstrates that our approach improves recall by 14 to 22 percentage points over YOLO alone, while precision remains within 0.7 points below to 1.7 points above the baseline. This combination of adaptive thresholding and cost-sensitive reinforcement learning achieves clinically aligned, open-world polyp detection with substantially fewer false negatives, thereby reducing the risk of missed precancerous polyps and improving patient outcomes.
Authors: Miriam Horovicz
Abstract: LLM agents that use external tools can solve complex tasks, but understanding which tools actually contributed to a response remains a blind spot. No existing XAI methods address tool-level explanations. We introduce AgentSHAP, the first framework for explaining tool importance in LLM agents. AgentSHAP is model-agnostic: it treats the agent as a black box and works with any LLM (GPT, Claude, Llama, etc.) without needing access to internal weights or gradients. Using Monte Carlo Shapley values, AgentSHAP tests how an agent responds with different tool subsets and computes fair importance scores based on game theory. Our contributions are: (1) the first explainability method for agent tool attribution, grounded in Shapley values from game theory; (2) Monte Carlo sampling that reduces cost from O(2n) to practical levels; and (3) comprehensive experiments on API-Bank showing that AgentSHAP produces consistent scores across runs, correctly identifies which tools matter, and distinguishes relevant from irrelevant tools. AgentSHAP joins TokenSHAP (for tokens) and PixelSHAP (for image regions) to complete a family of Shapley-based XAI tools for modern generative AI. Code: https://github.com/GenAISHAP/TokenSHAP.
Authors: Chengzhi Liu, Yuzhe Yang, Yue Fan, Qingyue Wei, Sheng Liu, Xin Eric Wang
Abstract: Recent advancements in Multimodal Large Language Models (MLLMs) have significantly enhanced cross-modal understanding and reasoning by incorporating Chain-of-Thought (CoT) reasoning in the semantic space. Building upon this, recent studies extend the CoT mechanism to the visual modality, enabling models to integrate visual information during reasoning through external tools or explicit image generation. However, these methods remain dependent on explicit step-by-step reasoning, unstable perception-reasoning interaction and notable computational overhead. Inspired by human cognition, we posit that thinking unfolds not linearly but through the dynamic interleaving of reasoning and perception within the mind. Motivated by this perspective, we propose DMLR, a test-time Dynamic Multimodal Latent Reasoning framework that employs confidence-guided latent policy gradient optimization to refine latent think tokens for in-depth reasoning. Furthermore, a Dynamic Visual Injection Strategy is introduced, which retrieves the most relevant visual features at each latent think token and updates the set of best visual patches. The updated patches are then injected into latent think token to achieve dynamic visual-textual interleaving. Experiments across seven multimodal reasoning benchmarks and various model architectures demonstrate that DMLR significantly improves reasoning and perception performance while maintaining high inference efficiency.
Authors: Samarth Sarin, Lovepreet Singh, Bhaskarjit Sarmah, Dhagash Mehta
Abstract: Agentic memory is emerging as a key enabler for large language models (LLM) to maintain continuity, personalization, and long-term context in extended user interactions, critical capabilities for deploying LLMs as truly interactive and adaptive agents. Agentic memory refers to the memory that provides an LLM with agent-like persistence: the ability to retain and act upon information across conversations, similar to how a human would. We present Memoria, a modular memory framework that augments LLM-based conversational systems with persistent, interpretable, and context-rich memory. Memoria integrates two complementary components: dynamic session-level summarization and a weighted knowledge graph (KG)-based user modelling engine that incrementally captures user traits, preferences, and behavioral patterns as structured entities and relationships. This hybrid architecture enables both short-term dialogue coherence and long-term personalization while operating within the token constraints of modern LLMs. We demonstrate how Memoria enables scalable, personalized conversational artificial intelligence (AI) by bridging the gap between stateless LLM interfaces and agentic memory systems, offering a practical solution for industry applications requiring adaptive and evolving user experiences.
Authors: Dongseok Kim, Hyoungsun Choi, Mohamed Jismy Aashik Rasool, Gisung Oh
Abstract: Prompts can switch a model's behavior even when the weights are fixed, yet this phenomenon is rarely treated as a clean theoretical object rather than a heuristic. We study the family of functions obtainable by holding a Transformer backbone fixed as an executor and varying only the prompt. Our core idea is to view the prompt as an externally injected program and to construct a simplified Transformer that interprets it to implement different computations. The construction exposes a mechanism-level decomposition: attention performs selective routing from prompt memory, the FFN performs local arithmetic conditioned on retrieved fragments, and depth-wise stacking composes these local updates into a multi-step computation. Under this viewpoint, we prove a constructive existential result showing that a single fixed backbone can approximate a broad class of target behaviors via prompts alone. The framework provides a unified starting point for formalizing trade-offs under prompt length/precision constraints and for studying structural limits of prompt-based switching, while remaining distinct from empirical claims about pretrained LLMs.
Authors: Yongcan Yu, Lingxiao He, Shuo Lu, Lijun Sheng, Yinuo Xu, Yanbo Wang, Kuangpu Guo, Jianjie Cheng, Meng Wang, Qianlong Xie, Xingxing Wang, Dapeng Hu, Jian Liang
Abstract: Recent advances in vision-language models (VLMs) reasoning have been largely attributed to the rise of reinforcement Learning (RL), which has shifted the community's focus away from the supervised fine-tuning (SFT) paradigm. Many studies suggest that introducing the SFT stage not only fails to improve reasoning ability but may also negatively impact model training. In this study, we revisit this RL-centric belief through a systematic and controlled comparison of SFT and RL on VLM Reasoning. Using identical data sources, we find that the relative effectiveness of SFT and RL is conditional and strongly influenced by model capacity, data scale, and data distribution. Contrary to common assumptions, our findings show that SFT plays a crucial role across several scenarios: (1) Effectiveness for weaker models. SFT more reliably elicits reasoning capabilities in smaller or weaker VLMs. (2) Data efficiency. SFT with only 2K achieves comparable or better reasoning performance to RL with 20K. (3) Cross-modal transferability. SFT demonstrates stronger generalization across modalities. Moreover, we identify a pervasive issue of deceptive rewards, where higher rewards fail to correlate with better reasoning accuracy in RL. These results challenge the prevailing "RL over SFT" narrative. They highlight that the role of SFT may have been underestimated and support a more balanced post-training pipeline in which SFT and RL function as complementary components.
Authors: Mahir Labib Dihan, Tanzima Hashem, Mohammed Eunus Ali, Md Rizwan Parvez
Abstract: LLM-based agents often operate in a greedy, step-by-step manner, selecting actions solely based on the current observation without considering long-term consequences or alternative paths. This lack of foresight is particularly problematic in web environments, which are only partially observable-limited to browser-visible content (e.g., DOM and UI elements)-where a single misstep often requires complex and brittle navigation to undo. Without an explicit backtracking mechanism, agents struggle to correct errors or systematically explore alternative paths. Tree-search methods provide a principled framework for such structured exploration, but existing approaches lack mechanisms for safe backtracking, making them prone to unintended side effects. They also assume that all actions are reversible, ignoring the presence of irreversible actions-limitations that reduce their effectiveness in realistic web tasks. To address these challenges, we introduce WebOperator, a tree-search framework that enables reliable backtracking and strategic exploration. Our method incorporates a best-first search strategy that ranks actions by both reward estimates and safety considerations, along with a robust backtracking mechanism that verifies the feasibility of previously visited paths before replaying them, preventing unintended side effects. To further guide exploration, WebOperator generates action candidates from multiple, varied reasoning contexts to ensure diverse and robust exploration, and subsequently curates a high-quality action set by filtering out invalid actions pre-execution and merging semantically equivalent ones. Experimental results on WebArena and WebVoyager demonstrate the effectiveness of WebOperator. On WebArena, WebOperator achieves a state-of-the-art 54.6% success rate with gpt-4o, underscoring the critical advantage of integrating strategic foresight with safe execution.
Authors: Xue Li, Xiaonan Song, Henry Hu
Abstract: Real-world deployment of Vision-Language Models (VLMs) is hindered by high computational demands, as existing architectures inefficiently process all tokens uniformly. We introduce Adaptive Token Pruning (ATP), a dynamic inference mechanism that retains only the most informative tokens based on contextual relevance. ATP operates at the vision-language interface, assigning a hybrid importance score combining ViT CLS attention (intra-modal saliency) and CLIP text-image similarity (inter-modal relevance) to keep top-K tokens for the LLM. Unlike static compression, ATP adapts to each input without modifying the backbone. Proposed as a lightweight gating module, ATP is compatible with popular backbones like BLIP-2, LLaVA, and Flamingo. Preliminary evaluations across VQAv2, GQA, and COCO indicate that ATP reduces inference FLOPs by around 40% and achieves roughly 1.5x speedups in end-to-end latency with negligible accuracy loss (less than 1%). Qualitative analyses suggest ATP preserves visual grounding and enhances interpretability. Beyond efficiency, we investigate robustness under corruptions; observations suggest adaptive pruning suppresses spurious correlations, improving stability. These findings imply that resource-constrained inference and model reliability are not competing objectives. Finally, we discuss ATP's role in efficient multimodal edge computing pipelines.
Authors: Sina Jani, Arman Heidari, Amirmohammad Anvari, Zahra Rahimi
Abstract: The rapid acceleration of scientific publishing has created substantial challenges for researchers attempting to discover, contextualize, and interpret relevant literature. Traditional keyword-based search systems provide limited semantic understanding, while existing AI-driven tools typically focus on isolated tasks such as retrieval, clustering, or bibliometric visualization. This paper presents an integrated system for scientific literature exploration that combines large-scale data acquisition, hybrid retrieval, semantic topic modeling, and heterogeneous knowledge graph construction. The system builds a comprehensive corpus by merging full-text data from arXiv with structured metadata from OpenAlex. A hybrid retrieval architecture fuses BM25 lexical search with embedding-based semantic search using Reciprocal Rank Fusion. Topic modeling is performed on retrieved results using BERTopic or non-negative matrix factorization depending on computational resources. A knowledge graph unifies papers, authors, institutions, countries, and extracted topics into an interpretable structure. The system provides a multi-layered exploration environment that reveals not only relevant publications but also the conceptual and relational landscape surrounding a query. Evaluation across multiple queries demonstrates improvements in retrieval relevance, topic coherence, and interpretability. The proposed framework contributes an extensible foundation for AI-assisted scientific discovery.
Authors: Yoo Yongmin, Kim Seungwoo, Liu Jingjiang
Abstract: Valuing intangible assets under uncertainty remains a critical challenge in the strategic management of technological innovation due to the information asymmetry inherent in high-dimensional technical specifications. Traditional bibliometric indicators, such as citation counts, fail to address this friction in a timely manner due to the systemic latency inherent in data accumulation. To bridge this gap, this study proposes the Economic Reasoning Alignment via Instruction Tuning (ERA-IT) framework. We theoretically conceptualize patent renewal history as a revealed economic preference and leverage it as an objective supervisory signal to align the generative reasoning of Large Language Models (LLMs) with market realities, a process we term Eco-Semantic Alignment. Using a randomly sampled dataset of 10,000 European Patent Office patents across diverse technological domains, we trained the model not only to predict value tiers but also to reverse-engineer the Economic Chain-of-Thought from unstructured text. Empirical results demonstrate that ERA-IT significantly outperforms both conventional econometric models and zero-shot LLMs in predictive accuracy. More importantly, by generating explicit, logically grounded rationales for valuation, the framework serves as a transparent cognitive scaffold for decision-makers, reducing the opacity of black-box AI in high-stakes intellectual property management.
Authors: Minghao Zhu, Zhihao Zhang, Anmol Sidhu, Keith Redmill
Abstract: Automated road sign recognition is a critical task for intelligent transportation systems, but traditional deep learning methods struggle with the sheer number of sign classes and the impracticality of creating exhaustive labeled datasets. This paper introduces a novel zero-shot recognition framework that adapts the Retrieval-Augmented Generation (RAG) paradigm to address this challenge. Our method first uses a Vision Language Model (VLM) to generate a textual description of a sign from an input image. This description is used to retrieve a small set of the most relevant sign candidates from a vector database of reference designs. Subsequently, a Large Language Model (LLM) reasons over the retrieved candidates to make a final, fine-grained recognition. We validate this approach on a comprehensive set of 303 regulatory signs from the Ohio MUTCD. Experimental results demonstrate the framework's effectiveness, achieving 95.58% accuracy on ideal reference images and 82.45% on challenging real-world road data. This work demonstrates the viability of RAG-based architectures for creating scalable and accurate systems for road sign recognition without task-specific training.
Authors: David Dang (Howard), Stuart Love (Howard), Meena Salib (Howard), Quynh Dang (Howard), Samuel Rothfarb (Howard), Mysk Alnatour (Howard), Andrew Salij (Howard), Hou-Tong Chen (Howard), Ho Wai (Howard), Lee, Wilton J. M. Kort-Kamp
Abstract: Advancing artificial intelligence for physical sciences requires representations that are both interpretable and compatible with the underlying laws of nature. We introduce METASTRINGS, a symbolic language for photonics that expresses nanostructures as textual sequences encoding materials, geometries, and lattice configurations. Analogous to molecular textual representations in chemistry, METASTRINGS provides a framework connecting human interpretability with computational design by capturing the structural hierarchy of photonic metasurfaces. Building on this representation, we develop Meta-GPT, a foundation transformer model trained on METASTRINGS and finetuned with physics-informed supervised, reinforcement, and chain-of-thought learning. Across various design tasks, the model achieves <3% mean-squared spectral error and maintains >98% syntactic validity, generating diverse metasurface prototypes whose experimentally measured optical responses match their target spectra. These results demonstrate that Meta-GPT can learn the compositional rules of light-matter interactions through METASTRINGS, laying a rigorous foundation for AI-driven photonics and representing an important step toward a metasurface genome project.
Authors: Jongwook Kim, Sangheon Yun, Sukjin Yoon
Abstract: The canonical $O(N^2)$ Transformer remains the empirical performance frontier in sequence modeling, and its training can be further optimized by addressing geometric inefficiency. We propose an optimization framework that leverages an asymmetric projection to decompose the backward-pass gradients into parallel spans and orthogonal violations, while keeping the canonical forward-pass $QKV$ structure intact. Through consistent experimental validation across various decomposition and projection setups, we provide strong theoretical evidence: the standard attention gradient is suboptimal. We demonstrated that selectively scaling these components, focusing primarily on $0^{th}$ order bidirectional parallel spans, yields the most effective learning signal. On the limited WikiText-2 dataset, and using a crude configuration, this method achieved a $0.56\%$ reduction in validation loss, confirming the framework's fundamental validity and suggesting significant potential gains on larger datasets and deeper training regimes
Authors: Xuwei Tan, Yao Ma, Xueru Zhang
Abstract: Detecting fraud in financial transactions typically relies on tabular models that demand heavy feature engineering to handle high-dimensional data and offer limited interpretability, making it difficult for humans to understand predictions. Large Language Models (LLMs), in contrast, can produce human-readable explanations and facilitate feature analysis, potentially reducing the manual workload of fraud analysts and informing system refinements. However, they perform poorly when applied directly to tabular fraud detection due to the difficulty of reasoning over many features, the extreme class imbalance, and the absence of contextual information. To bridge this gap, we introduce FinFRE-RAG, a two-stage approach that applies importance-guided feature reduction to serialize a compact subset of numeric/categorical attributes into natural language and performs retrieval-augmented in-context learning over label-aware, instance-level exemplars. Across four public fraud datasets and three families of open-weight LLMs, FinFRE-RAG substantially improves F1/MCC over direct prompting and is competitive with strong tabular baselines in several settings. Although these LLMs still lag behind specialized classifiers, they narrow the performance gap and provide interpretable rationales, highlighting their value as assistive tools in fraud analysis.
Authors: Bizhe Bai, Hongming Wu, Peng Ye, Tao Chen
Abstract: Self-supervised reinforcement learning (RL) presents a promising approach for enhancing the reasoning capabilities of Large Language Models (LLMs) without reliance on expensive human-annotated data. However, we find that existing methods suffer from a critical failure mode under long-horizon training: a "policy collapse" where performance precipitously degrades. We diagnose this instability and demonstrate that simply scaling the number of rollouts -- a common strategy to improve performance -- only delays, but does not prevent, this collapse. To counteract this instability, we first introduce M-GRPO (Momentum-Anchored Group Relative Policy Optimization), a framework that leverages a slowly evolving momentum model to provide a stable training target. In addition, we identify that this process is often accompanied by a rapid collapse in policy entropy, resulting in a prematurely confident and suboptimal policy. To specifically address this issue, we propose a second contribution: an adaptive filtering method based on the interquartile range (IQR) that dynamically prunes low-entropy trajectories, preserving essential policy diversity. Our extensive experiments on multiple reasoning benchmarks demonstrate that M-GRPO stabilizes the training process while the IQR filter prevents premature convergence. The combination of these two innovations leads to superior training stability and state-of-the-art performance.
Authors: Mohaiminul Islam Bhuiyan, Chan Hue Wah, Nur Shazwani Kamarudin, Nur Hafieza Ismail, Ahmad Fakhri Ab Nasir
Abstract: This study provides an overview of heart disease prediction using an intelligent system. Predicting disease accurately is crucial in the medical field, but traditional methods relying solely on a doctor's experience often lack precision. To address this limitation, intelligent systems are applied as an alternative to traditional approaches. While various intelligent system methods exist, this study focuses on three: Fuzzy Logic, Neural Networks, and Case-Based Reasoning (CBR). A comparison of these techniques in terms of accuracy was conducted, and ultimately, Case-Based Reasoning (CBR) was selected for heart disease prediction. In the prediction phase, the heart disease dataset underwent data pre-processing to clean the data and data splitting to separate it into training and testing sets. The chosen intelligent system was then employed to predict heart disease outcomes based on the processed data. The experiment concluded with Case-Based Reasoning (CBR) achieving a notable accuracy rate of 97.95% in predicting heart disease. The findings also revealed that the probability of heart disease was 57.76% for males and 42.24% for females. Further analysis from related studies suggests that factors such as smoking and alcohol consumption are significant contributors to heart disease, particularly among males.
Authors: Emre Can Acikgoz, Jinoh Oh, Joo Hyuk Jeon, Jie Hao, Heng Ji, Dilek Hakkani-T\"ur, Gokhan Tur, Xiang Li, Chengyuan Ma, Xing Fan
Abstract: Conversational agents often encounter ambiguous user requests, requiring an effective clarification to successfully complete tasks. While recent advancements in real-world applications favor multi-agent architectures to manage complex conversational scenarios efficiently, ambiguity resolution remains a critical and underexplored challenge--particularly due to the difficulty of determining which agent should initiate a clarification and how agents should coordinate their actions when faced with uncertain or incomplete user input. The fundamental questions of when to interrupt a user and how to formulate the optimal clarification query within the most optimal multi-agent settings remain open. In this paper, we propose MAC (Multi-Agent Clarification), an interactive multi-agent framework specifically optimized to resolve user ambiguities by strategically managing clarification dialogues. We first introduce a novel taxonomy categorizing user ambiguities to systematically guide clarification strategies. Then, we present MAC that autonomously coordinates multiple agents to interact synergistically with users. Empirical evaluations on MultiWOZ 2.4 demonstrate that enabling clarification at both levels increases task success rate 7.8\% (54.5 to 62.3) and reduces the average number of dialogue turns (6.53 to 4.86) by eliciting all required user information up front and minimizing repetition. Our findings highlight the importance of active user interaction and role-aware clarification for more reliable human-agent communication.
Authors: Emre Can Acikgoz, Jinoh Oh, Jie Hao, Joo Hyuk Jeon, Heng Ji, Dilek Hakkani-T\"ur, Gokhan Tur, Xiang Li, Chengyuan Ma, Xing Fan
Abstract: Effective human-agent collaboration is increasingly prevalent in real-world applications. Current trends in such collaborations are predominantly unidirectional, with users providing instructions or posing questions to agents, where agents respond directly without seeking necessary clarifications or confirmations. However, the evolving capabilities of these agents require more proactive engagement, where agents should dynamically participate in conversations to clarify user intents, resolve ambiguities, and adapt to changing circumstances. Existing prior work under-utilize the conversational capabilities of language models (LMs), thereby optimizing agents as better followers rather than effective speakers. In this work, we introduce SpeakRL, a reinforcement learning (RL) method that enhances agents' conversational capabilities by rewarding proactive interactions with users, such as asking right clarification questions when necessary. To support this, we curate SpeakER, a synthetic dataset that includes diverse scenarios from task-oriented dialogues, where tasks are resolved through interactive clarification questions. We present a systematic analysis of reward design for conversational proactivity and propose a principled reward formulation for teaching agents to balance asking with acting. Empirical evaluations demonstrate that our approach achieves a 20.14% absolute improvement in task completion over base models without increasing conversation turns even surpassing even much larger proprietary models, demonstrating the promise of clarification-centric user-agent interactions.
Authors: Ali Al Sahili, Ali Chehab, Razane Tajeddine
Abstract: Large Language Models (LLMs) are prone to memorizing training data, which poses serious privacy risks. Two of the most prominent concerns are training data extraction and Membership Inference Attacks (MIAs). Prior research has shown that these threats are interconnected: adversaries can extract training data from an LLM by querying the model to generate a large volume of text and subsequently applying MIAs to verify whether a particular data point was included in the training set. In this study, we integrate multiple MIA techniques into the data extraction pipeline to systematically benchmark their effectiveness. We then compare their performance in this integrated setting against results from conventional MIA benchmarks, allowing us to evaluate their practical utility in real-world extraction scenarios.
Authors: Sitao Cheng, Tianle Li, Xuhan Huang, Xunjian Yin, Difan Zou
Abstract: The design of effective reward functions presents a central and often arduous challenge in reinforcement learning (RL), particularly when developing autonomous agents for complex reasoning tasks. While automated reward optimization approaches exist, they typically rely on derivative-free evolutionary heuristics that treat the reward function as a black box, failing to capture the causal relationship between reward structure and task performance. To bridge this gap, we propose Differentiable Evolutionary Reinforcement Learning (DERL), a bilevel framework that enables the autonomous discovery of optimal reward signals. In DERL, a Meta-Optimizer evolves a reward function (i.e., Meta-Reward) by composing structured atomic primitives, guiding the training of an inner-loop policy. Crucially, unlike previous evolution, DERL is differentiable in its metaoptimization: it treats the inner-loop validation performance as a signal to update the Meta-Optimizer via reinforcement learning. This allows DERL to approximate the "meta-gradient" of task success, progressively learning to generate denser and more actionable feedback. We validate DERL across three distinct domains: robotic agent (ALFWorld), scientific simulation (ScienceWorld), and mathematical reasoning (GSM8k, MATH). Experimental results show that DERL achieves state-of-the-art performance on ALFWorld and ScienceWorld, significantly outperforming methods relying on heuristic rewards, especially in out-of-distribution scenarios. Analysis of the evolutionary trajectory demonstrates that DERL successfully captures the intrinsic structure of tasks, enabling selfimproving agent alignment without human intervention.
Authors: Ojas Pungalia, Rashi Upadhyay, Abhishek Mishra, Abhiram H, Tejasvi Alladi, Sujan Yenuganti, Dhruv Kumar
Abstract: Envy is a common human behavior that shapes competitiveness and can alter outcomes in team settings. As large language models (LLMs) increasingly act on behalf of humans in collaborative and competitive workflows, there is a pressing need to evaluate whether and under what conditions they exhibit envy-like preferences. In this paper, we test whether LLMs show envy-like behavior toward each other. We considered two scenarios: (1) A point allocation game that tests whether a model tries to win over its peer. (2) A workplace setting observing behaviour when recognition is unfair. Our findings reveal consistent evidence of envy-like patterns in certain LLMs, with large variation across models and contexts. For instance, GPT-5-mini and Claude-3.7-Sonnet show a clear tendency to pull down the peer model to equalize outcomes, whereas Mistral-Small-3.2-24B instead focuses on maximizing its own individual gains. These results highlight the need to consider competitive dispositions as a safety and design factor in LLM-based multi-agent systems.
Authors: Lei Qu, Lianhai Ren, Peng Cheng, Rui Gao, Ruizhe Wang, Tianyu Chen, Xiao Liu, Xingjian Zhang, Yeyun Gong, Yifan Xiong, Yucheng Ding, Yuting Jiang, Zhenghao Lin, Zhongxin Guo, Ziyue Yang
Abstract: An increasing variety of AI accelerators is being considered for large-scale training. However, enabling large-scale training on early-life AI accelerators faces three core challenges: frequent system disruptions and undefined failure modes that undermine reliability; numerical errors and training instabilities that threaten correctness and convergence; and the complexity of parallelism optimization combined with unpredictable local noise that degrades efficiency. To address these challenges, SIGMA is an open-source training stack designed to improve the reliability, stability, and efficiency of large-scale distributed training on early-life AI hardware. The core of this initiative is the LUCIA TRAINING PLATFORM (LTP), the system optimized for clusters with early-life AI accelerators. Since its launch in March 2025, LTP has significantly enhanced training reliability and operational productivity. Over the past five months, it has achieved an impressive 94.45% effective cluster accelerator utilization, while also substantially reducing node recycling and job-recovery times. Building on the foundation of LTP, the LUCIA TRAINING FRAMEWORK (LTF) successfully trained SIGMA-MOE, a 200B MoE model, using 2,048 AI accelerators. This effort delivered remarkable stability and efficiency outcomes, achieving 21.08% MFU, state-of-the-art downstream accuracy, and encountering only one stability incident over a 75-day period. Together, these advances establish SIGMA, which not only tackles the critical challenges of large-scale training but also establishes a new benchmark for AI infrastructure and platform innovation, offering a robust, cost-effective alternative to prevailing established accelerator stacks and significantly advancing AI capabilities and scalability. The source code of SIGMA is available at https://github.com/microsoft/LuciaTrainingPlatform.
Authors: Oleg Grynets, Vasyl Lyashkevych, Dmytro Baran, Maksym Orliansky, Taras Zelenyy, Markiian Leshchyshyn
Abstract: The study presents the outcomes of research and experimental validation in the domain of automated codebase migration, with a focus on addressing challenges in transitioning SQL-based systems. The proposed method for migration essentially appears as a framework that leverages the best aspects of traditional software engineering techniques and provides an iterative, scalable, precise and efficient solution for modern database transformations. The central piece of the approach is the integration of a fine-tuned Large Language Model to address critical issues in SQL code conversion, such as syntax mapping, resolving discrepancies between Oracle PL/SQL and PostgreSQL, and optimising database elements such as stored procedures, triggers, views, and overall database logic. Thus, the method involves a trade-off between fine-tuning and prompt engineering. Special attention is given to a fine-tuning approach, which enhances the adaptability and compatibility with migration requirements across the entire database. According to the achieved results, fine-tuning plays a very important role. The study employs targeted evaluation methodologies along with computational metrics to measure the success of iterative conversion cycles. Core innovations include automated SQL feature detection, semi-supervised error analysis and integration of Subject Matter Experts feedback within a systematic migration workflow. The methodology achieves significant reductions in Syntax Error Rates, enhances feature alignment throughout migration iterations, and leverages dataset sampling to ensure continual improvement. By embedding GAI into the migration process, the framework facilitates precise feature mapping, semi-automated error resolution, and data-driven optimisation loops, improving workflow efficiency.
Authors: Yiyi Cai, Xuangeng Chu, Xiwei Gao, Sitong Gong, Yifei Huang, Caixin Kang, Kunhang Li, Haiyang Liu, Ruicong Liu, Yun Liu, Dianwen Ng, Zixiong Su, Erwin Wu, Yuhan Wu, Dingkun Yan, Tianyu Yan, Chang Zeng, Bo Zheng, You Zhou
Abstract: We introduce Interactive Intelligence, a novel paradigm of digital human that is capable of personality-aligned expression, adaptive interaction, and self-evolution. To realize this, we present Mio (Multimodal Interactive Omni-Avatar), an end-to-end framework composed of five specialized modules: Thinker, Talker, Face Animator, Body Animator, and Renderer. This unified architecture integrates cognitive reasoning with real-time multimodal embodiment to enable fluid, consistent interaction. Furthermore, we establish a new benchmark to rigorously evaluate the capabilities of interactive intelligence. Extensive experiments demonstrate that our framework achieves superior performance compared to state-of-the-art methods across all evaluated dimensions. Together, these contributions move digital humans beyond superficial imitation toward intelligent interaction.
Authors: Huaqin Zhao, Zhengliang Liu, Zihao Wu, Yiwei Li, Tianze Yang, Peng Shu, Shaochen Xu, Haixing Dai, Lin Zhao, Hanqi Jiang, Yi Pan, Junhao Chen, Yifan Zhou, Zeyu Zhang, Ruitong Sun, Gengchen Mai, Ninghao Liu, Tianming Liu
Abstract: In recent years, Large Language Models (LLMs) like ChatGPT have seen considerable advancements and have been applied in diverse fields. Built on the Transformer architecture, these models are trained on extensive datasets, enabling them to understand and generate human language effectively. In the financial domain, the deployment of LLMs is gaining momentum. These models are being utilized for automating financial report generation, forecasting market trends, analyzing investor sentiment, and offering personalized financial advice. Leveraging their natural language processing capabilities, LLMs can distill key insights from vast financial data, aiding institutions in making informed investment choices and enhancing both operational efficiency and customer satisfaction. In this study, we provide a comprehensive overview of the emerging integration of LLMs into various financial tasks. Additionally, we conducted holistic tests on multiple financial tasks through the combination of natural language instructions. Our findings show that GPT-4 effectively follow prompt instructions across various financial tasks. This survey and evaluation of LLMs in the financial domain aim to deepen the understanding of LLMs' current role in finance for both financial practitioners and LLM researchers, identify new research and application prospects, and highlight how these technologies can be leveraged to solve practical challenges in the finance industry.
Authors: Wei Guan, Jian Cao, Jianqi Gao, Haiyan Zhao, Shiyou Qian
Abstract: Detecting anomalies in business processes is crucial for ensuring operational success. While many existing methods rely on statistical frequency to detect anomalies, it's important to note that infrequent behavior doesn't necessarily imply undesirability. To address this challenge, detecting anomalies from a semantic viewpoint proves to be a more effective approach. However, current semantic anomaly detection methods treat a trace (i.e., process instance) as multiple event pairs, disrupting long-distance dependencies. In this paper, we introduce DABL, a novel approach for detecting semantic anomalies in business processes using large language models (LLMs). We collect 143,137 real-world process models from various domains. By generating normal traces through the playout of these process models and simulating both ordering and exclusion anomalies, we fine-tune Llama 2 using the resulting log. Through extensive experiments, we demonstrate that DABL surpasses existing state-of-the-art semantic anomaly detection methods in terms of both generalization ability and learning of given processes. Users can directly apply DABL to detect semantic anomalies in their own datasets without the need for additional training. Furthermore, DABL offers the capability to interpret the causes of anomalies in natural language, providing valuable insights into the detected anomalies.
Authors: Xiaoyun Zhang, Zhengyue Zhao, Wenxuan Shi, Kaidi Xu, Di Huang, Xing Hu
Abstract: With the widespread application of Large Language Models (LLMs), it has become a significant concern to ensure their safety and prevent harmful responses. While current safe-alignment methods based on instruction fine-tuning and Reinforcement Learning from Human Feedback (RLHF) can effectively reduce harmful responses from LLMs, they often require high-quality datasets and heavy computational overhead during model training. Another way to align language models is to modify the logit of tokens in model outputs without heavy training. Recent studies have shown that contrastive decoding can enhance the performance of language models by reducing the likelihood of confused tokens. However, these methods require the manual selection of contrastive models or instruction templates, limiting the degree of contrast. To this end, we propose Adversarial Contrastive Decoding (ACD), an optimization-based framework to generate two opposite soft system prompts, the Safeguarding Prompt (SP) and the Adversarial Prompt (AP), for prompt-based contrastive decoding. The SP aims to promote safer outputs while the AP aims to exploit the harmful parts of the model, providing a strong contrast to align the model with safety. ACD only needs to apply a lightweight prompt tuning on a rather small anchor dataset without training the target model. Experiments conducted on extensive models and benchmarks demonstrate that the proposed method achieves much better safety performance than previous model training-free decoding methods without sacrificing its original generation ability.
Authors: Martin Tutek, Fateme Hashemi Chaleshtori, Ana Marasovi\'c, Yonatan Belinkov
Abstract: When prompted to think step-by-step, language models (LMs) produce a chain of thought (CoT), a sequence of reasoning steps that the model supposedly used to produce its prediction. Despite much work on CoT prompting, it is unclear if reasoning verbalized in a CoT is faithful to the models' parametric beliefs. We introduce a framework for measuring parametric faithfulness of generated reasoning, and propose Faithfulness by Unlearning Reasoning steps (FUR), an instance of this framework. FUR erases information contained in reasoning steps from model parameters, and measures faithfulness as the resulting effect on the model's prediction. Our experiments with four LMs and five multi-hop multi-choice question answering (MCQA) datasets show that FUR is frequently able to precisely change the underlying models' prediction for a given instance by unlearning key steps, indicating when a CoT is parametrically faithful. Further analysis shows that CoTs generated by models post-unlearning support different answers, hinting at a deeper effect of unlearning.
Authors: David Ph. Shakouri, Crit Cremers, Niels O. Schiller
Abstract: This article presents experiments performed using a computational laboratory environment for language acquisition experiments. It implements a multi-agent system consisting of two agents: an adult language model and a daughter language model that aims to learn the mother language. Crucially, the daughter agent does not have access to the internal knowledge of the mother language model but only to the language exemplars the mother agent generates. These experiments illustrate how this system can be used to acquire abstract grammatical knowledge. We demonstrate how statistical analyses of patterns in the input data corresponding to grammatical categories yield discrete grammatical rules. These rules are subsequently added to the grammatical knowledge of the daughter language model. To this end, hierarchical agglomerative cluster analysis was applied to the utterances consecutively generated by the mother language model. It is argued that this procedure can be used to acquire structures resembling grammatical categories proposed by linguists for natural languages. Thus, it is established that non-trivial grammatical knowledge has been acquired. Moreover, the parameter configuration of this computational laboratory environment determined using training data generated by the mother language model is validated in a second experiment with a test set similarly resulting in the acquisition of non-trivial categories.
Authors: Kangwei Liu, Mengru Wang, Yujie Luo, Lin Yuan, Mengshu Sun, Lei Liang, Zhiqiang Zhang, Jun Zhou, Bryan Hooi, Shumin Deng
Abstract: Fine-tuning enables large language models (LLMs) to adapt to specific domains, but often compromises their previously established safety alignment. To mitigate the degradation of model safety during fine-tuning, we introduce LookAhead Tuning, a lightweight and effective data-driven approach that preserves safety during fine-tuning. The method introduces two simple strategies that modify training data by previewing partial answer prefixes, thereby minimizing perturbations to the model's initial token distributions and maintaining its built-in safety mechanisms. Comprehensive experiments demonstrate that LookAhead Tuning effectively maintains model safety without sacrificing robust performance on downstream tasks. Our findings position LookAhead Tuning as a reliable and efficient solution for the safe and effective adaptation of LLMs.
Authors: Wei-Lin Chen, Zhepei Wei, Xinyu Zhu, Shi Feng, Yu Meng
Abstract: Large language models (LLMs) are increasingly used as automatic evaluators in applications such as benchmarking, reward modeling, and self-refinement. Prior work highlights a potential self-preference bias where LLMs favor their own generated responses, a tendency often intensifying with model size and capability. This raises a critical question: Is self-preference harmful, or does it simply reflect the genuinely higher-quality outputs of stronger models? Answering this has been difficult as prior works mostly relied on subjective tasks that lack an objective ground truth, meaning that either preference can be reasonably justified. To address this ambiguity, we investigate self-preference using verifiable benchmarks (mathematical reasoning, factual knowledge, code generation) that allow objective ground-truth assessment. This enables us to distinguish harmful (favoring objectively worse responses) from legitimate (favoring genuinely superior ones) self-preference. Our large-scale experiments across 7 model families reveal three key findings: (1) While stronger models exhibit greater self-preference, much of this preference aligns with objectively superior performance, indicating stronger models prefer themselves mostly legitimately. (2) Harmful self-preference persists when evaluator models err as generators, and stronger models display more pronounced harmful self-preference bias when they do err. This suggests stronger models struggle more to recognize when they are wrong. (3) Inference-time scaling strategies, such as generating a long Chain-of-Thought before evaluation, effectively reduce the harmful self-preference. Additionally, we experiment with LMArena and show that our findings extend beyond verifiable benchmarks to real-world, subjective domains. These results provide a more nuanced understanding of LLM-based evaluation and practical insights for improving its reliability.
Authors: Zden\v{e}k Kasner, Vil\'em Zouhar, Patr\'icia Schmidtov\'a, Ivan Kart\'a\v{c}, Krist\'yna Onderkov\'a, Ond\v{r}ej Pl\'atek, Dimitra Gkatzia, Saad Mahamood, Ond\v{r}ej Du\v{s}ek, Simone Balloccu
Abstract: Span annotation - annotating specific text features at the span level - can be used to evaluate texts where single-score metrics fail to provide actionable feedback. Until recently, span annotation was done by human annotators or fine-tuned models. In this paper, we study whether large language models (LLMs) can serve as an alternative to human annotators. We compare the abilities of LLMs to skilled human annotators on three span annotation tasks: evaluating data-to-text generation, identifying translation errors, and detecting propaganda techniques. We show that overall, LLMs have only moderate inter-annotator agreement (IAA) with human annotators. However, we demonstrate that LLMs make errors at a similar rate as skilled crowdworkers. LLMs also produce annotations at a fraction of the cost per output annotation. We release the dataset of over 40k model and human span annotations for further research.
Authors: Varvara Arzt, Allan Hanbury, Michael Wiegand, G\'abor Recski, Terra Blevins
Abstract: Analysing the generalisation capabilities of relation extraction (RE) models is crucial for assessing whether they learn robust relational patterns or rely on spurious correlations. Our cross-dataset experiments find that RE models struggle with unseen data, even within similar domains. Notably, higher intra-dataset performance does not indicate better transferability, instead often signaling overfitting to dataset-specific artefacts. Our results also show that data quality, rather than lexical similarity, is key to robust transfer, and the choice of optimal adaptation strategy depends on the quality of data available: while fine-tuning yields the best cross-dataset performance with high-quality data, few-shot in-context learning (ICL) is more effective with noisier data. However, even in these cases, zero-shot baselines occasionally outperform all cross-dataset results. Structural issues in RE benchmarks, such as single-relation per sample constraints and non-standardised negative class definitions, further hinder model transferability.
Authors: Mingliang Zhai, Zhi Gao, Yuwei Wu, Yunde Jia
Abstract: Embodied Question Answering (EQA) requires agents to autonomously explore and comprehend the environment to answer context-dependent questions. Typically, an EQA framework consists of four components: a planner, a memory module, a stopping module, and an answering module. However, the memory module is utilized inefficiently in existing methods, as the information it stores is leveraged solely for the answering module. Such a design may result in redundant or inadequate exploration, leading to a suboptimal success rate. To solve this problem, we propose MemoryEQA, an EQA framework centered on memory, which establishes mechanisms for memory storage, update, and retrieval, allowing memory information to contribute throughout the entire exploration process. Specifically, we convert the observation into structured textual representations, which are stored in a vector library following a fixed structure. At each exploration step, we utilize a viewpoint comparison strategy to determine whether the memory requires updating. Before executing each module, we employ an entropy-based adaptive retrieval strategy to obtain the minimal yet sufficient memory information that satisfies the requirements of different modules. The retrieved module-specific information is then integrated with the current observation as input to the corresponding module. To evaluate EQA models' memory capabilities, we constructed the benchmark based on HM3D called MT-HM3D, comprising 1,587 question-answer pairs involving multiple targets across various regions, which requires agents to maintain memory of exploration-acquired target information. Experimental results on HM-EQA, MT-HM3D, and OpenEQA demonstrate the effectiveness of our framework, where a 9.9% performance gain on MT-HM3D compared to baseline models further underscores the memory capability's pivotal role in solving complex tasks.
Authors: Yi Liu, Dianqing Liu, Mingye Zhu, Junbo Guo, Yongdong Zhang, Zhendong Mao
Abstract: The widespread adoption of large language models (LLMs) across industries has increased the demand for high-quality and customizable outputs. However, traditional alignment methods often require retraining large pretrained models, making it difficult to quickly adapt and optimize LLMs for diverse applications. To address this limitation, we propose a novel \textit{Residual Alignment Model} (\textit{RAM}) that formalizes the alignment process as a type of importance sampling. In this framework, the unaligned upstream model serves as the proposal distribution, while the alignment process is framed as secondary sampling based on an autoregressive alignment module that acts as an estimator of the importance weights. This design enables a natural detachment of the alignment module from the target aligned model, improving flexibility and scalability. Based on this model, we derive an efficient sequence-level training strategy for the alignment module, which operates independently of the proposal module. Additionally, we develop a resampling algorithm with iterative token-level decoding to address the common first-token latency issue in comparable methods. Experimental evaluations on two leading open-source LLMs across diverse tasks, including instruction following, domain adaptation, and preference optimization, demonstrate that our approach consistently outperforms baseline models.
Authors: Dairu Liu, Ziyue Wang, Minyuan Ruan, Fuwen Luo, Chi Chen, Peng Li, Yang Liu
Abstract: Images usually convey richer detail than text, but often include redundant information, which potentially downgrades multimodal reasoning performance. When faced with lengthy or complex messages, humans tend to employ abstract thinking to convert them into simple and concise abstracts. Inspired by this cognitive strategy, we introduce a novel paradigm to elicit the ability to Think with Visual Abstract (VAT), by prompting Multimodal Large Language Models (MLLMs) with visual abstract instead of explicit verbal thoughts or elaborate guidance, permitting a more efficient visual reasoning mechanism via concentrated perception. VAT encourages models to focus on more essential visual elements, concepts and structural features by undermining redundant information compared with explicit thinking methods, such as Chain-of-thought (CoT) and tool-using approaches, that increase the complexity of reasoning process via inserting verbose intermediate steps and external knowledge. Experimental results show that VAT consistently empowers different MLLMs in visual perception and reasoning tasks. VAT achieves an average gain of $2.21\%$ over GPT-5 baseline, surpassing the gain of CoT, demonstrating that VAT better enhances multimodal task performance of MLLMs. Additionally, VAT spends fewer tokens while achieving higher performance. These findings highlight the effectiveness of visual abstract thinking and encourage further exploration of more diverse reasoning paradigms from the perspective of human cognition.
Authors: Hongcheng Guo, Zheyong Xie, Shaosheng Cao, Boyang Wang, Weiting Liu, Anjie Le, Lei Li, Zhoujun Li
Abstract: With the increasing integration of visual and textual content in Social Networking Services (SNS), evaluating the multimodal capabilities of Large Language Models (LLMs) is crucial for enhancing user experience, content understanding, and platform intelligence. Existing benchmarks primarily focus on text-centric tasks, lacking coverage of the multimodal contexts prevalent in modern SNS ecosystems. In this paper, we introduce SNS-Bench-VL, a comprehensive multimodal benchmark designed to assess the performance of Vision-Language LLMs in real-world social media scenarios. SNS-Bench-VL incorporates images and text across 8 multimodal tasks, including note comprehension, user engagement analysis, information retrieval, and personalized recommendation. It comprises 4,001 carefully curated multimodal question-answer pairs, covering single-choice, multiple-choice, and open-ended tasks. We evaluate over 25 state-of-the-art multimodal LLMs, analyzing their performance across tasks. Our findings highlight persistent challenges in multimodal social context comprehension. We hope SNS-Bench-VL will inspire future research towards robust, context-aware, and human-aligned multimodal intelligence for next-generation social networking services.
Authors: Yuri Balashov
Abstract: Large Language Models (LLMs) excel in translation among other things, demonstrating competitive performance for many language pairs in zero- and few-shot settings. But unlike dedicated neural machine translation models, LLMs are not trained on any translation-related objective. What explains their remarkable translation abilities? Are these abilities grounded in "incidental bilingualism" (Briakou et al. 2023) in training data? Does instruction tuning contribute to it? Are LLMs capable of aligning and leveraging semantically identical or similar monolingual contents from different corners of the internet that are unlikely to fit in a single context window? I offer some reflections on this topic, informed by recent studies and growing user experience. My working hypothesis is that LLMs' translation abilities originate in two different types of pre-training data that may be internalized by the models in different ways. I discuss the prospects for testing the "duality" hypothesis empirically and its implications for reconceptualizing translation, human and machine, in the age of deep learning.
Authors: Zhaorui Yang, Bo Pan, Han Wang, Yiyao Wang, Xingyu Liu, Luoxuan Weng, Yingchaojie Feng, Haozhe Feng, Minfeng Zhu, Bo Zhang, Wei Chen
Abstract: Visualizations play a crucial part in effective communication of concepts and information. Recent advances in reasoning and retrieval augmented generation have enabled Large Language Models (LLMs) to perform deep research and generate comprehensive reports. Despite its progress, existing deep research frameworks primarily focus on generating text-only content, leaving the automated generation of interleaved texts and visualizations underexplored. This novel task poses key challenges in designing informative visualizations and effectively integrating them with text reports. To address these challenges, we propose Formal Description of Visualization (FDV), a structured textual representation of charts that enables LLMs to learn from and generate diverse, high-quality visualizations. Building on this representation, we introduce Multimodal DeepResearcher, an agentic framework that decomposes the task into four stages: (1) researching, (2) exemplar report textualization, (3) planning, and (4) multimodal report generation. For the evaluation of generated multimodal reports, we develop MultimodalReportBench, which contains 100 diverse topics served as inputs along with 5 dedicated metrics. Extensive experiments across models and evaluation methods demonstrate the effectiveness of Multimodal DeepResearcher. Notably, utilizing the same Claude 3.7 Sonnet model, Multimodal DeepResearcher achieves an 82% overall win rate over the baseline method.
Authors: Hongcheng Guo, Zheyong Xie, Shaosheng Cao, Boyang Wang, Weiting Liu, Zheyu Ye, Zhoujun Li, Zuozhu Liu, Wei Lu
Abstract: As interest in using Large Language Models for interactive and emotionally rich experiences grows, virtual pet companionship emerges as a novel yet underexplored application. Existing approaches focus on basic pet role-playing interactions without systematically benchmarking LLMs for comprehensive companionship. In this paper, we introduce Pet-Bench, a dedicated benchmark that evaluates LLMs across both self-interaction and human-interaction dimensions. Unlike prior work, Pet-Bench emphasizes self-evolution and developmental behaviors alongside interactive engagement, offering a more realistic reflection of pet companionship. It features diverse tasks such as intelligent scheduling, memory-based dialogues, and psychological conversations, with over 7,500 interaction instances designed to simulate pet behaviors. Evaluation of 28 LLMs reveals significant performance variations linked to model size and inherent capabilities, underscoring the need for specialized optimization in this domain. Pet-Bench serves as a foundational resource for benchmarking pet-related LLM abilities and advancing emotionally immersive human-pet interactions.
Authors: Aladin Djuhera, Swanand Ravindra Kadhe, Syed Zawad, Farhan Ahmed, Heiko Ludwig, Holger Boche
Abstract: Recent work on large language models (LLMs) has increasingly focused on post-training and alignment with datasets curated to enhance instruction following, world knowledge, and specialized skills. However, most post-training datasets used in leading open- and closed-source LLMs remain inaccessible to the public, with limited information about their construction process. This lack of transparency has motivated the recent development of open-source post-training corpora. While training on these open alternatives can yield performance comparable to that of leading models, systematic comparisons remain challenging due to the significant computational cost of conducting them rigorously at scale, and are therefore largely absent. As a result, it remains unclear how specific samples, task types, or curation strategies influence downstream performance when assessing data quality. In this work, we conduct the first comprehensive side-by-side analysis of two prominent open post-training datasets: Tulu-3-SFT-Mix and SmolTalk. Using the Magpie framework, we annotate each sample with detailed quality metrics, including turn structure (single-turn vs. multi-turn), task category, input quality, and response quality, and we derive statistics that reveal structural and qualitative similarities and differences between the two datasets. Based on these insights, we design a principled curation recipe that produces a new data mixture, TuluTalk, which contains 14% fewer samples than either source dataset while matching or exceeding their performance on key benchmarks. Our findings offer actionable insights for constructing more effective post-training datasets that improve model performance within practical resource limits. To support future research, we publicly release both the annotated source datasets and our curated TuluTalk mixture.
Authors: Ahmed Hasanaath, Aisha Alansari, Ahmed Ashraf, Chafik Salmane, Hamzah Luqman, Saad Ezzini
Abstract: Large language models (LLMs) have shown remarkable progress in reasoning abilities and general natural language processing (NLP) tasks, yet their performance on Arabic data, characterized by rich morphology, diverse dialects, and complex script, remains underexplored. This paper presents a comprehensive benchmarking study of multiple reasoning-focused LLMs, with a special emphasis on the newly introduced DeepSeek models, across a suite of fifteen Arabic NLP tasks. We experiment with various strategies, including zero-shot, few-shot, and fine-tuning. This allows us to systematically evaluate performance on datasets covering a range of applications to examine their capacity for linguistic reasoning under different levels of complexity. Our experiments reveal several key findings. First, carefully selecting just three in-context examples delivers an average uplift of over 13 F1 points on classification tasks-boosting sentiment analysis from 35.3% to 87.5% and paraphrase detection from 56.1% to 87.0%. Second, reasoning-focused DeepSeek architectures outperform a strong GPT o4-mini baseline by an average of 12 F1 points on complex inference tasks in the zero-shot setting. Third, LoRA-based fine-tuning yields up to an additional 8 points in F1 and BLEU compared to equivalent increases in model scale. The code is available at https://github.com/gufranSabri/deepseek-evals
Authors: Coleman Hooper, Sebastian Zhao, Luca Manolache, Sehoon Kim, Michael W. Mahoney, Yakun Sophia Shao, Kurt Keutzer, Amir Gholami
Abstract: Large Reasoning Models (LRMs) have shown promising accuracy improvements on complex problem-solving tasks. While these models have attained high accuracy by leveraging additional computation at test time, they need to generate long chain-of-thought reasoning in order to think before answering, which requires generating thousands of tokens. While sparse attention methods can help reduce the KV cache pressure induced by this long autoregressive reasoning, these methods can introduce errors which disrupt the reasoning process. Additionally, prior methods often pre-process the input to make it easier to identify the important prompt tokens when computing attention during generation, and this pre-processing is challenging to perform online for newly generated reasoning tokens. Our work addresses these challenges by introducing Multipole Attention, which accelerates autoregressive reasoning by only computing exact attention for the most important tokens, while maintaining approximate representations for the remaining tokens. Our method first performs clustering to group together semantically similar key vectors, and then uses the cluster centroids both to identify important key vectors and to approximate the remaining key vectors in order to retain high accuracy. We design a fast cluster update process to quickly re-cluster the input and previously generated tokens, thereby allowing for accelerating attention to the previous output tokens. We evaluate our method using emerging LRMs such as Qwen-8B, demonstrating that our approach can maintain accuracy on complex reasoning tasks even with aggressive attention sparsity settings. We also provide kernel implementations to demonstrate the practical efficiency gains from our method, achieving up to 4.5$\times$ speedup for attention in long-context reasoning applications. Our code is available at https://github.com/SqueezeAILab/MultipoleAttention.
Authors: Pengzuo Wu, Yuhang Yang, Guangcheng Zhu, Chao Ye, Hong Gu, Xu Lu, Ruixuan Xiao, Bowen Bao, Yijing He, Liangyu Zha, Wentao Ye, Junbo Zhao, Haobo Wang
Abstract: With the rapid advancement of Large Language Models (LLMs), there is an increasing need for challenging benchmarks to evaluate their capabilities in handling complex tabular data. However, existing benchmarks are either based on outdated data setups or focus solely on simple, flat table structures. In this paper, we introduce RealHiTBench, a comprehensive benchmark designed to evaluate the performance of both LLMs and Multimodal LLMs (MLLMs) across a variety of input formats for complex tabular data, including LaTeX, HTML, and PNG. RealHiTBench also includes a diverse collection of tables with intricate structures, spanning a wide range of task types. Our experimental results, using 25 state-of-the-art LLMs, demonstrate that RealHiTBench is indeed a challenging benchmark. Moreover, we also develop TreeThinker, a tree-based pipeline that organizes hierarchical headers into a tree structure for enhanced tabular reasoning, validating the importance of improving LLMs' perception of table hierarchies. We hope that our work will inspire further research on tabular data reasoning and the development of more robust models. The code and data are available at https://github.com/cspzyy/RealHiTBench.
Authors: Xianzhe Fan, Xuhui Zhou, Chuanyang Jin, Kolby Nottingham, Hao Zhu, Maarten Sap
Abstract: Humans continuously infer the states, goals, and behaviors of others by perceiving their surroundings in dynamic, real-world social interactions. However, most Theory of Mind (ToM) benchmarks only evaluate static, text-based scenarios, which have a significant gap compared to real interactions. We propose the SoMi-ToM benchmark, designed to evaluate multi-perspective ToM in embodied multi-agent complex social interactions. This benchmark is based on rich multimodal interaction data generated by the interaction environment SoMi, covering diverse crafting goals and social relationships. Our framework supports multi-level evaluation: (1) first-person evaluation provides multimodal (visual, dialogue, action, etc.) input from a first-person perspective during a task for real-time state inference, (2) third-person evaluation provides complete third-person perspective video and text records after a task for goal and behavior inference. This evaluation method allows for a more comprehensive examination of a model's ToM capabilities from both the subjective immediate experience and the objective global observation. We constructed a challenging dataset containing 35 third-person perspective videos, 363 first-person perspective images, and 1225 expert-annotated multiple-choice questions (three options). On this dataset, we systematically evaluated the performance of human subjects and several state-of-the-art large vision-language models (LVLMs). The results show that LVLMs perform significantly worse than humans on SoMi-ToM: the average accuracy gap between humans and models is 40.1% in first-person evaluation and 26.4% in third-person evaluation. This indicates that future LVLMs need to further improve their ToM capabilities in embodied, complex social interactions.
Authors: Ruidi Chang, Chunyuan Deng, Hanjie Chen
Abstract: Transformer-based language models have achieved state-of-the-art performance in natural language generation (NLG), yet their internal mechanisms for synthesizing task-relevant information remain insufficiently understood. While prior studies suggest that intermediate layers often yield more generalizable representations than final layers, how this generalization ability emerges and propagates across layers during training remains unclear. To address this gap, we propose InfoRidge, an information-theoretic framework, to characterize how predictive information-the mutual information between hidden representations and target outputs-varies across depth. Estimating this quantity enables us to trace the flow of task-relevant information throughout the model during training. Our experiments across various models and datasets reveal a consistent non-monotonic trend: predictive information peaks in upper-middle layers-forming a generalization ridge-before declining in final layers, reflecting a transition between generalization and memorization. To further investigate this phenomenon, we introduce residual scaling coefficients-trainable scalar parameters applied to each residual block-which serve as functional probes for assessing the relative importance of individual transformer layers. These coefficients reveal that, under distribution shift, models downweight final layers and increasingly rely on intermediate layers, highlighting their role in generalization. Together, these findings offer new insights into the internal mechanisms of transformers and underscore the critical role of intermediate layers in supporting generalization.
Authors: Jiachen Zhao, Jing Huang, Zhengxuan Wu, David Bau, Weiyan Shi
Abstract: LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety.
Authors: Kyeongkyu Lee, Seonghwan Yoon, Hongki Lim
Abstract: Automatic generation of radiology reports has the potential to alleviate radiologists' significant workload, yet current methods struggle to deliver clinically reliable conclusions. In particular, most prior approaches focus on producing fluent text without effectively ensuring the factual correctness of the reports and often rely on single-view images, limiting diagnostic comprehensiveness. We propose CLARIFID, a novel framework that directly optimizes diagnostic correctness by mirroring the two-step workflow of experts. Specifically, CLARIFID (1) learns the logical flow from Findings to Impression through section-aware pretraining, (2) is fine-tuned with Proximal Policy Optimization in which the CheXbert F1 score of the Impression section serves as the reward, (3) employs controlled decoding that completes "Findings" before synthesizing the "Impression", and (4) fuses multiple chest X-ray views via a vision-transformer-based multi-view encoder. During inference, we apply a next-token forcing strategy followed by report-level re-ranking, ensuring that the model first produces a comprehensive "Findings" section before synthesizing the "Impression" and thereby preserving coherent clinical reasoning. Experimental results on the MIMIC-CXR dataset demonstrate that our method achieves superior clinical efficacy and outperforms existing baselines on clinical efficacy scores.
Authors: Louis Est\`eve, Marie-Catherine de Marneffe, Nurit Melnik, Agata Savary, Olha Kanishcheva
Abstract: The concept of diversity has received increased consideration in Natural Language Processing (NLP) in recent years. This is due to various motivations like promoting and inclusion, approximating human linguistic behavior, and increasing systems' performance. Diversity has however often been addressed in an ad hoc manner in NLP, and with few explicit links to other domains where this notion is better theorized. We survey articles in the ACL Anthology from the past 6 years, with "diversity" or "diverse" in their title. We find a wide range of settings in which diversity is quantified, often highly specialized and using inconsistent terminology. We put forward a unified taxonomy of why, what on, where, and how diversity is measured in NLP. Diversity measures are cast upon a unified framework from ecology and economy (Stirling, 2007) with 3 dimensions of diversity: variety, balance and disparity. We discuss the trends which emerge due to this systematized approach. We believe that this study paves the way towards a better formalization of diversity in NLP, which should bring a better understanding of this notion and a better comparability between various approaches.
Authors: Jiaxin Deng, Qingcheng Zhu, Junbiao Pang, Linlin Yang, Zhongqian Fu, Baochang Zhang
Abstract: Little research explores the correlation between the expressive ability and generalization ability of the low-rank adaptation (LoRA). Sharpness-Aware Minimization (SAM) improves model generalization for both Convolutional Neural Networks (CNNs) and Transformers by encouraging convergence to locally flat minima. However, the connection between sharpness and generalization has not been fully explored for LoRA due to the lack of tools to either empirically seek flat minima or develop theoretical methods. In this work, we propose Flat Minima LoRA (FMLoRA) and its efficient version, i.e., EFMLoRA, to seek flat minima for LoRA. Concretely, we theoretically demonstrate that perturbations in the full parameter space can be transferred to the low-rank subspace. This approach eliminates the potential interference introduced by perturbations across multiple matrices in the low-rank subspace. Our extensive experiments on large language models and vision-language models demonstrate that EFMLoRA achieves optimize efficiency comparable to that of LoRA while simultaneously attaining comparable or even better performance. For example, on the GLUE dataset with RoBERTa-large, EFMLoRA outperforms LoRA and full fine-tuning by 1.0% and 0.5% on average, respectively. On vision-language models, e.g., Qwen-VL-Chat, there are performance improvements of 1.5% and 1.0% on the SQA and VizWiz datasets, respectively. These empirical results also verify that the generalization of LoRA is closely related to sharpness, which is omitted by previous methods.
Authors: Li Wang, Changhao Zhang, Zengqi Xiu, Kai Lu, Xin Yu, Kui Zhang, Wenjun Wu
Abstract: Despite recent advances in the reasoning capabilities of Large Language Models (LLMs), improving the reasoning ability of Small Language Models (SLMs, e.g., up to 1.5B parameters) remains challenging. A key obstacle lies in the complexity and variability of natural language: essentially equivalent problems often appear in diverse surface forms, often obscured by redundant or distracting details. This imposes a dual burden on SLMs: they must first extract the core problem from complex linguistic input, and then perform reasoning based on that understanding. The resulting vast and noisy problem space hinders optimization, particularly for models with limited capacity. To address this, we propose a new framework that decouples understanding from reasoning by mapping natural language problems into a canonical problem space-a semantically simplified yet expressive domain. This enables SLMs to focus on reasoning over standardized inputs, free from linguistic variability. Within this framework, we introduce DURIT (Decoupled Understanding from Reasoning via Iterative Training), a three-step algorithm that iteratively: (1) mapping natural language problems via reinforcement learning, (2) aligns reasoning trajectories through self-distillation, and (3) trains reasoning policies in the problem space. The mapper and reasoner are co-trained in an alternating loop throughout this process. Experiments show that DURIT substantially improves SLMs' performance on both in-domain and out-of-domain mathematical and logical reasoning tasks. Beyond improving reasoning capabilities, DURIT also improves the robustness of reasoning, validating decoupling understanding from reasoning as an effective strategy for strengthening SLMs.
Authors: Wenpeng Xing, Lanyi Wei, Haixiao Hu, Jingyi Yu, Rongchang Li, Mohan Li, Changting Lin, Meng Han
Abstract: The rapid proliferation of large language models (LLMs) in applications targeting children and adolescents necessitates a fundamental reassessment of prevailing AI safety frameworks, which are largely tailored to adult users and neglect the distinct developmental vulnerabilities of minors. This paper highlights key deficiencies in existing LLM safety benchmarks, including their inadequate coverage of age-specific cognitive, emotional, and social risks spanning early childhood (ages 0--6), middle childhood (7--12), and adolescence (13--18). To bridge these gaps, we introduce SproutBench, an innovative evaluation suite comprising 1,283 developmentally grounded adversarial prompts designed to probe risks such as emotional dependency, privacy violations, and imitation of hazardous behaviors. Through rigorous empirical evaluation of 47 diverse LLMs, we uncover substantial safety vulnerabilities, corroborated by robust inter-dimensional correlations (e.g., between Safety and Risk Prevention) and a notable inverse relationship between Interactivity and Age Appropriateness. These insights yield practical guidelines for advancing child-centric AI design and deployment.
Authors: Ziqian Bi, Keyu Chen, Chiung-Yi Tseng, Danyang Zhang, Tianyang Wang, Hongying Luo, Lu Chen, Junming Huang, Jibin Guan, Junfeng Hao, Xinyuan Song, Junhao Song
Abstract: In August 2025, OpenAI released GPT-OSS models, its first open weight large language models since GPT-2 in 2019, comprising two mixture of experts architectures with 120B and 20B parameters. We evaluated both variants against six contemporary open source large language models ranging from 14.7B to 235B parameters, representing both dense and sparse designs, across ten benchmarks covering general knowledge, mathematical reasoning, code generation, multilingual understanding, and conversational ability. All models were tested in unquantised form under standardised inference settings, with statistical validation using McNemars test and effect size analysis. Results show that gpt-oss-20B consistently outperforms gpt-oss-120B on several benchmarks, such as HumanEval and MMLU, despite requiring substantially less memory and energy per response. Both models demonstrate mid-tier overall performance within the current open source landscape, with relative strength in code generation and notable weaknesses in multilingual tasks. These findings provide empirical evidence that scaling in sparse architectures may not yield proportional performance gains, underscoring the need for further investigation into optimisation strategies and informing more efficient model selection for future open source deployments. More details and evaluation scripts are available at https://ai-agent-lab.github.io/gpt-oss (Project Webpage).
Authors: Chunhua Liu, Kabir Manandhar Shrestha, Sukai Huang
Abstract: Large language models (LLMs) exhibit cultural bias from overrepresented viewpoints in training data, yet cultural alignment remains a challenge due to limited cultural knowledge and a lack of exploration into effective learning approaches. We introduce a cost-efficient and cognitively grounded method: fine-tuning LLMs on native speakers' word-association norms, leveraging cognitive psychology findings that such associations capture cultural knowledge. Using word association datasets from native speakers in the US (English) and China (Mandarin), we train Llama-3.1-8B and Qwen-2.5-7B via supervised fine-tuning and preference optimization. We evaluate models' cultural alignment through a two-tier evaluation framework that spans lexical associations and cultural value alignment using the World Values Survey. Results show significant improvements in lexical alignment (16-20% English, 43-165% Mandarin on Precision@5) and high-level cultural value shifts. On a subset of 50 questions where US and Chinese respondents diverge most, fine-tuned Qwen nearly doubles its response alignment with Chinese values (13 to 25). Remarkably, our trained 7-8B models match or exceed vanilla 70B baselines, demonstrating that a few million of culture-grounded associations achieve value alignment without expensive retraining. Our work highlights both the promise and the need for future research grounded in human cognition in improving cultural alignment in AI models.
Authors: Xiao Liang, Zhongzhi Li, Yeyun Gong, Yelong Shen, Ying Nian Wu, Zhijiang Guo, Weizhu Chen
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has recently emerged as a key paradigm for post-training Large Language Models (LLMs), particularly for complex reasoning tasks. However, vanilla RLVR training has been shown to improve Pass@1 performance at the expense of policy entropy, leading to reduced generation diversity and limiting the Pass@k performance, which typically represents the upper bound of LLM reasoning capability. In this paper, we systematically analyze the policy's generation diversity from the perspective of training problems and find that augmenting and updating training problems helps mitigate entropy collapse during training. Based on these observations, we propose an online Self-play with Variational problem Synthesis (SvS) strategy for RLVR training, which uses the policy's correct solutions to synthesize variational problems while ensuring their reference answers remain identical to the originals. This self-improving strategy effectively maintains policy entropy during training and substantially improves Pass@k compared with standard RLVR, sustaining prolonged improvements and achieving absolute gains of 18.3% and 22.8% in Pass@32 performance on the competition-level AIME24 and AIME25 benchmarks, as well as on code generation tasks. Experiments on 12 reasoning benchmarks across varying model sizes from 3B to 32B consistently demonstrate the generalizability and robustness of SvS.
Authors: Junhao Yin, Haolin Wang, Peng Bao, Ju Xu, Yongliang Wang
Abstract: Generative query suggestion using large language models offers a powerful way to enhance conversational systems, but aligning outputs with nuanced user preferences remains a critical challenge. To address this, we introduce a multi-stage framework designed for progressive alignment between the generation policy and user intent. Our pipeline begins with prompt engineering as a cold-start strategy, followed by the Supervised Fine-Tuning stage, in which we introduce a distillation method on click logs to create a robust foundational model. To better model user preferences while capturing their inherent uncertainty, we develop a Gaussian Reward Model (GaRM) that represents user preferences as probability distributions rather than point estimates. Finally, we employ reinforcement learning to align the generation policy with these preferences, guided by a composite reward function that integrates GaRM with auxiliary heuristics to mitigate reward hacking. To maintain training stability, this process is enhanced by a novel out-of-distribution regularization method and a two-stage reward fusion technique. Extensive experiments demonstrate that our framework significantly outperforms baselines on both automatic and human evaluations and yields a 34\% relative increase in user engagement as measured by click-through rate in live A/B tests.
Authors: Darpan Aswal, C\'eline Hudelot
Abstract: Large Language Models have found success in a variety of applications. However, their safety remains a concern due to the existence of various jailbreaking methods. Despite significant efforts, alignment and safety fine-tuning only provide a certain degree of robustness against jailbreak attacks that covertly mislead LLMs towards the generation of harmful content. This leaves them prone to a range of vulnerabilities, including targeted misuse and accidental user profiling. This work introduces \textbf{ConceptGuard}, a novel framework that leverages Sparse Autoencoders (SAEs) to identify interpretable concepts within LLM internals associated with different jailbreak themes. By extracting semantically meaningful internal representations, ConceptGuard enables building robust safety guardrails -- offering fully explainable and generalizable defenses without sacrificing model capabilities or requiring further fine-tuning. Leveraging advances in the mechanistic interpretability of LLMs, our approach provides evidence for a shared activation geometry for jailbreak attacks in the representation space, a potential foundation for designing more interpretable and generalizable safeguards against attackers.
Authors: Jeongsoo Lee, Daeyong Kwon, Kyohoon Jin
Abstract: Retrieval-Augmented Generation (RAG) systems are widely adopted in knowledge-intensive NLP tasks, but current evaluations often overlook the structural complexity and multi-step reasoning required in real-world scenarios. These benchmarks overlook key factors such as the interaction between retrieval difficulty and reasoning depth. To address this gap, we propose GRADE, a novel evaluation framework that models task difficulty along two orthogonal dimensions: (1) reasoning depth, defined by the number of inference steps (hops), and (2) semantic distance between the query and its supporting evidence. We construct a synthetic multi-hop QA dataset from factual news articles by extracting knowledge graphs and augmenting them through semantic clustering to recover missing links, allowing us to generate diverse and difficulty-controlled queries. Central to our framework is a 2D difficulty matrix that combines generator-side and retriever-side difficulty. Experiments across multiple domains and models show that error rates strongly correlate with our difficulty measures, validating their diagnostic utility. GRADE enables fine-grained analysis of RAG performance and provides a scalable foundation for evaluating and improving multi-hop reasoning in real-world applications.
Authors: Cheng Wang, Zeming Wei, Qin Liu, Muhao Chen
Abstract: Large Language Models (LLMs) can comply with harmful instructions, raising serious safety concerns despite their impressive capabilities. Recent work has leveraged probing-based approaches to study the separability of malicious and benign inputs in LLMs' internal representations, and researchers have proposed using such probing methods for safety detection. We systematically re-examine this paradigm. Motivated by poor out-of-distribution performance, we hypothesize that probes learn superficial patterns rather than semantic harmfulness. Through controlled experiments, we confirm this hypothesis and identify the specific patterns learned: instructional patterns and trigger words. Our investigation follows a systematic approach, progressing from demonstrating comparable performance of simple n-gram methods, to controlled experiments with semantically cleaned datasets, to detailed analysis of pattern dependencies. These results reveal a false sense of security around current probing-based approaches and highlight the need to redesign both models and evaluation protocols, for which we provide further discussions in the hope of suggesting responsible further research in this direction. We have open-sourced the project at https://github.com/WangCheng0116/Why-Probe-Fails.
Authors: Xiao Zhang, Dongyuan Li, Liuyu Xiang, Yao Zhang, Cheng Zhong, Zhaofeng He
Abstract: Although Multimodal Large Language Models (MLLMs) have demonstrated increasingly impressive performance in chart understanding, most of them exhibit alarming hallucinations and significant performance degradation when handling non-annotated charts. We argue that current MLLMs rely largely on visual recognition rather than visual reasoning to interpret the charts, and visual estimation of numerical values is one of the most fundamental capabilities in chart understanding that require complex visual reasoning. To prove this, we introduce ChartVRBench, a benchmark meticulously designed to isolate and evaluate visual reasoning ability in chart understanding. Furthermore, we propose ChartVR-3B/7B trained with a novel Visual Reasoning Reinforcement Finetuning (VR-RFT) strategy to strengthen genuine chart visual reasoning abilities. Extensive experiments show that ChartVR achieves superior performance on ChartVRBench, outperforming even powerful proprietary models. Moreover, the visual reasoning skills cultivated by the proposed VR-RFT demonstrate strong generalization, leading to significant performance gains across a diverse suite of public chart understanding benchmarks. The code and dataset will be publicly available upon publication.
Authors: Hao Zhang, Mengsi Lyu, Zhuo Chen, Xingrun Xing, Yulong Ao, Yonghua Lin
Abstract: Large Language Models (LLMs) demonstrate exceptional capabilities across various tasks, but their deployment is constrained by high computational and memory costs. Model pruning provides an effective means to alleviate these demands. However, existing methods often ignore the characteristics of prefill-decode (PD) disaggregation in practice. In this paper, we propose a pruning method that is highly integrated with PD disaggregation, enabling more precise pruning of blocks. Our approach constructs pruning and distillation sets to perform iterative block removal, obtaining better pruning solutions. Moreover, we analyze the pruning sensitivity of the prefill and decode stages and identify removable blocks specific to each stage, making it well suited for PD disaggregation deployment. Extensive experiments demonstrate our approach consistently achieves strong performance in both PD disaggregation and PD unified (non-PD disaggregation) settings, and can also be extended to other non-block pruning methods. Under the same settings, our method achieves improved performance and faster inference.
Authors: Sangmin Bae
Abstract: Large language models have achieved remarkable capabilities, but their practical deployment is hindered by significant computational costs. While adaptive computation methods like early-exiting promise to reduce these costs, they introduce a fundamental conflict: the per-token dynamism intended to save computation often creates system-level bottlenecks that can paradoxically reduce throughput in batched inference. This dissertation resolves this conflict by co-designing adaptive algorithms and model architectures to strike an optimal balance between dynamism and efficiency. To this end, our work first addresses critical sources of overhead in conventional early-exiting by proposing an efficient parallel decoding mechanism. We then show that deep parameter sharing provides an architectural foundation that not only yields compact, parameter-efficient models but also inherently mitigates the critical synchronization issues affecting dynamic inference. Finally, this work presents a unified framework where lightweight routers are pretrained to dynamically assign an optimal recursion depth for each token. This approach establishes a new Pareto frontier between efficiency and performance by effectively optimizing for both adaptive computation and parameter efficiency within a single model.
Authors: Zitong Yang, Aonan Zhang, Hong Liu, Tatsunori Hashimoto, Emmanuel Cand\`es, Chong Wang, Ruoming Pang
Abstract: We introduce Synthetic Bootstrapped Pretraining (SBP), a language model (LM) pretraining procedure that first learns a model of relations between documents from the pretraining dataset and then leverages it to synthesize a vast new corpus for joint training. While the standard pretraining teaches LMs to learn causal correlations among tokens within a single document, it is not designed to efficiently model the rich, learnable inter-document correlations that can potentially lead to better performance. We validate SBP by designing a compute-matched pretraining setup and pretrain a 3B-parameter and a 6B-parameter model on up to 1T tokens from scratch. We find SBP consistently improves upon a strong repetition baseline and delivers up to 60% of performance improvement attainable by an oracle upper bound with access to 20x more unique data. Qualitative analysis reveals that the synthesized documents go beyond mere paraphrases -- SBP first abstracts a core concept from the seed material and then crafts a new narration on top of it. Besides strong empirical performance, SBP admits a natural Bayesian interpretation: the synthesizer implicitly learns to abstract the latent concepts shared between related documents.
Authors: Hiram Ring
Abstract: A repeated claim in linguistics is that the majority of linguistic utterances are unique. For example, Pinker (1994: 10), summarizing an argument by Noam Chomsky, states that "virtually every sentence that a person utters or understands is a brand-new combination of words, appearing for the first time in the history of the universe." With the increased availability of large corpora, this is a claim that can be empirically investigated. The current paper addresses the question by using the NLTK Python library to parse corpora of different genres, providing counts of exact string matches in each. Results show that while completely unique sentences are often the majority of corpora, this is highly constrained by genre, and that duplicate sentences are not an insignificant part of any individual corpus.
Authors: Yutaro Sigrist, Andreas Waldis
Abstract: Merging methods combine the weights of multiple language models (LMs) to leverage their capacities, such as for domain adaptation. While existing studies investigate merged models from a solely behavioral perspective, we offer the first comprehensive view by assessing and connecting their behavior and internals. We present a novel evaluation pipeline that first merges multiple parent LMs, and then evaluates the merged models in comparison to the initial ones based on their behavior on downstream tasks, like MMLU, and the internal encoded linguistic competence. We showcase this pipeline by assessing the merging of instruction fine-tuned with math- and code-adapted LMs from the Qwen2.5 family. Our results show that merging methods impacts behavior and internals differently. While the performance of merged models is typically between that of the two parent models, their encoded information about linguistic phenomena, particularly in morphology and syntax, can surpass the parent models. Moreover, we find weak ranking correlation between this behavior and internal evaluation. With our pipeline and initial results, we emphasize the need for more comprehensive evaluations of model merging methods to gain a faithful understanding of their capabilities and reliability, beyond potential superficial behavioral advances.
Authors: Yizhou Wang, Chen Tang, Han Deng, Jiabei Xiao, Jiaqi Liu, Jianyu Wu, Jun Yao, Pengze Li, Encheng Su, Lintao Wang, Guohang Zhuang, Yuchen Ren, Ben Fei, Ming Hu, Xin Chen, Dongzhan Zhou, Junjun He, Xiangyu Yue, Zhenfei Yin, Jiamin Wu, Qihao Zheng, Yuhao Zhou, Huihui Xu, Chenglong Ma, Yan Lu, Wenlong Zhang, Chunfeng Song, Philip Torr, Shixiang Tang, Xinzhu Ma, Wanli Ouyang, Lei Bai
Abstract: We present a scientific reasoning foundation model that aligns natural language with heterogeneous scientific representations. The model is pretrained on a 206B-token corpus spanning scientific text, pure sequences, and sequence-text pairs, then aligned via SFT on 40M instructions, annealed cold-start bootstrapping to elicit long-form chain-of-thought, and reinforcement learning with task-specific reward shaping, which instills deliberate scientific reasoning. It supports four capability families, covering up to 103 tasks across workflows: (i) faithful translation between text and scientific formats, (ii) text/knowledge extraction, (iii) property prediction, (iv) property classification, (v) unconditional and conditional sequence generation and design. Compared with specialist systems, our approach broadens instruction coverage, improves cross-domain generalization, and enhances fidelity. We detail data curation and training and show that cross-discipline learning strengthens transfer and downstream reliability. The model, instruct tuning datasets and the evaluation code are open-sourced at https://huggingface.co/SciReason and https://github.com/open-sciencelab/SciReason.
URLs: https://huggingface.co/SciReason, https://github.com/open-sciencelab/SciReason.
Authors: Han Yuan, Yue Zhao, Li Zhang, Wuqiong Luo, Zheng Ma
Abstract: Structured output from large language models (LLMs) has enhanced efficiency in processing generated information and is increasingly adopted in industrial applications. Prior studies have investigated the impact of structured output on LLMs' generation quality, often presenting one-way findings. Some suggest that structured format enhances completeness and factual accuracy, while others argue that it restricts the reasoning capacity of LLMs and leads to reductions in standard evaluation metrics. Potential limitations of these assessments include restricted testing scenarios, weakly controlled comparative settings, and reliance on coarse metrics. In this work, we present a refined analysis using causal inference. Based on one assumed and two guaranteed constraints, we derive five potential causal structures characterizing the influence of structured output on LLMs' generation: (1) collider without m-bias, (2) collider with m-bias, (3) single cause from instruction, (4) single cause from output format, and (5) independence. Across seven public and one developed reasoning tasks, we find that coarse metrics report positive, negative, or neutral effects of structured output on GPT-4o's generation. However, causal inference reveals no causal impact in 43 out of 48 scenarios. In the remaining 5, 3 involve multifaceted causal structures influenced by concrete instructions. Further experiments show that OpenAI-o3 are more resilient to output formats than general-purpose GPT-4o and GPT-4.1, highlighting an unaware advantage of reasoning models.
Authors: Guancheng Wan, Leixin Sun, Longxu Dou, Zitong Shi, Fang Wu, Eric Hanchen Jiang, Wenke Huang, Guibin Zhang, Hejia Geng, Xiangru Tang, Zhenfei Yin, Yizhou Sun, Wei Wang
Abstract: Large Language Model (LLM)-powered multi-agent systems (MAS) have rapidly advanced collaborative reasoning, tool use, and role-specialized coordination in complex tasks. However, reliability-critical deployment remains hindered by a systemic failure mode: hierarchical compliance under instruction conflicts (system-user, peer-peer), where agents misprioritize system-level rules in the presence of competing demands. Moreover, widely used macro-level metrics (e.g., pass@k) obscure these micro-level violations and offer little actionable guidance for remedy. In this work, we present a full-stack, three-stage framework: (1) Diagnose - Contextualized Role Adherence Score (CRAS), a query-wise, context-aware scoring metric that decomposes role adherence into four measurable dimensions; (2) Localize - attention drift analysis revealing that instruction conflicts are resolved by attention heads that are largely concentrated in middle layers; (3) Align - Surgical Alignment of Instruction Layers (SAIL), which installs LoRA only on the localized focal layers and optimizes a token-weighted DPO-style preference objective that credits tokens by their focal attentional contribution. Across standard benchmarks and MAS frameworks, our surgical approach improves instruction hierarchy compliance (e.g., +5.60% with AutoGen on MedQA) without full-model finetuning.
Authors: Guancheng Wan, Lucheng Fu, Haoxin Liu, Yiqiao Jin, Hui Yi Leong, Eric Hanchen Jiang, Hejia Geng, Jinhe Bi, Yunpu Ma, Xiangru Tang, B. Aditya Prakash, Yizhou Sun, Wei Wang
Abstract: The performance of Large Language Models (LLMs) hinges on carefully engineered prompts. However, prevailing prompt optimization methods, ranging from heuristic edits and reinforcement learning to evolutionary search, primarily target point-wise accuracy. They seldom enforce paraphrase invariance or searching stability, and therefore cannot remedy this brittleness in practice. Automated prompt search remains brittle: small, semantically preserving paraphrases often cause large performance swings. We identify this brittleness as the textual sharpness of the prompt landscape. In this work, we provide the first formal treatment of textual sharpness in the discrete, semantic space of prompts, together with an operational robustness criterion over a semantic neighborhood; the design is black-box or API-only, requiring no gradients to update the model's parameters. Then we introduce TARE (Textual Sharpness-Aware Evolving), a derivative-free framework that alternates between an inner, sampling-based adversarial search that stresses a prompt with hard paraphrases and an outer, robust selection that prefers candidates whose neighborhoods remain strong. We further propose ATARE, which learns anisotropic weights to shape the semantic neighborhood and adapts its radius over time to balance exploration and fidelity. Diverse tasks evaluate our methods, whose design for minimizing textual sharpness gap leads to prompts that preserve accuracy under paraphrasing, outperforming accuracy-only prompt search while remaining computationally practical.
Authors: Yuelyu Ji
Abstract: Multimodal Retrieval-Augmented Generation (Visual RAG) significantly advances question answering by integrating visual and textual evidence. Yet, current evaluations fail to systematically account for query difficulty and ambiguity. We propose MRAG-Suite, a diagnostic evaluation platform integrating diverse multimodal benchmarks (WebQA, Chart-RAG, Visual-RAG, MRAG-Bench). We introduce difficulty-based and ambiguity-aware filtering strategies, alongside MM-RAGChecker, a claim-level diagnostic tool. Our results demonstrate substantial accuracy reductions under difficult and ambiguous queries, highlighting prevalent hallucinations. MM-RAGChecker effectively diagnoses these issues, guiding future improvements in Visual RAG systems.
Authors: Huu Nguyen (Sonny), Victor May (Sonny), Harsh Raj (Sonny), Marianna Nezhurina (Sonny), Yishan Wang (Sonny), Yanqi Luo (Sonny), Minh Chien Vu (Sonny), Taishi Nakamura (Sonny), Ken Tsui (Sonny), Van Khue Nguyen (Sonny), David Salinas (Sonny), Aleksandra Krasnod\k{e}bska (Sonny), Christoph Schuhmann (Sonny), Mats Leon Richter (Sonny), Xuan-Son (Sonny), Vu, Jenia Jitsev
Abstract: We present MixtureVitae, an open-access pretraining corpus built to minimize legal risk while providing strong downstream performance. MixtureVitae follows a permissive-first, risk-mitigated sourcing strategy that combines public-domain and permissively licensed text (e.g., CC-BY/Apache) with carefully justified low-risk additions (e.g., government works and EU TDM-eligible sources). MixtureVitae adopts a simple, single-stage pretraining recipe that integrates a large proportion of permissive synthetic instruction and reasoning data-signals typically introduced during post-training and generally scarce in permissive web corpora. We categorize all sources into a three-tier scheme that reflects varying risk levels and provide shard-level provenance metadata to enable risk-aware usage. In controlled experiments using the open-sci-ref training protocol (fixed architectures and hyperparameters; 50B and 300B token budgets across 130M-1.7B parameters), models trained on MixtureVitae consistently outperform other permissive datasets across a suite of standard benchmarks, and at the 1.7B-parameters/300B-tokens setting, they surpass FineWeb-Edu and approach DCLM late in training. Performance is particularly strong on MMLU and on math and code benchmarks: a 1.7B model pretrained on 300B MixtureVitae tokens matches or exceeds a strong 1.7B instruction-tuned baseline on GSM8K, HumanEval, and MBPP, despite using over 36 times fewer tokens (300B vs. ~11T). Supported by a thorough decontamination analysis, these results show that permissive-first data with high instruction and reasoning density, tiered by licensing and provenance-related risk, can provide a practical and risk-mitigated foundation for training capable LLMs, reducing reliance on broad web scrapes without sacrificing competitiveness. Code: https://github.com/ontocord/mixturevitae
Authors: Matthew Lewis, Samuel Thio, Amy Roberts, Catherine Siju, Whoasif Mukit, Rebecca Kuruvilla, Zhangshu Joshua Jiang, Niko M\"oller-Grell, Aditya Borakati, Richard JB Dobson, Spiros Denaxas
Abstract: This paper presents the development and evaluation of a Retrieval-Augmented Generation (RAG) system for querying the United Kingdom's National Institute for Health and Care Excellence (NICE) clinical guidelines using Large Language Models (LLMs). The extensive length and volume of these guidelines can impede their utilisation within a time-constrained healthcare system, a challenge this project addresses through the creation of a system capable of providing users with precisely matched information in response to natural language queries. The system's retrieval architecture, composed of a hybrid embedding mechanism, was evaluated against a corpus of 10,195 text chunks derived from three hundred guidelines. It demonstrates high performance, with a Mean Reciprocal Rank (MRR) of 0.814, a Recall of 81% at the first chunk and of 99.1% within the top ten retrieved chunks, when evaluated on 7901 queries. The most significant impact of the RAG system was observed during the generation phase. When evaluated on a manually curated dataset of seventy question-answer pairs, RAG-enhanced models showed substantial gains in performance. Faithfulness, the measure of whether an answer is supported by the source text, was increased by 64.7 percentage points to 99.5% for the RAG-enhanced O4-Mini model and significantly outperformed the medical-focused Meditron3-8B LLM, which scored 43%. Clinical evaluation by seven Subject Matter Experts (SMEs) further validated these findings, with GPT-4.1 achieving 98.7% accuracy while reducing unsafe responses by 67% compared to O4-Mini (from 3.0 to 1.0 per evaluator). This study thus establishes RAG as an effective, reliable, and scalable approach for applying generative AI in healthcare, enabling cost-effective access to medical guidelines.
Authors: Jindi Wang, Yidi Zhang, Zhaoxing Li, Pedro Bem Haja, Ioannis Ivrissimtzis, Zichen Zhao, Sebastian Stein
Abstract: The rapid expansion of online courses and social media has generated large volumes of unstructured learner-generated text. Understanding how learners construct knowledge in these spaces is crucial for analysing learning processes, informing content design, and providing feedback at scale. However, existing approaches typically rely on manual coding of well-structured discussion forums, which does not scale to the fragmented discourse found in online learning. This study proposes and validates a framework that combines a codebook inspired by the Interaction Analysis Model with an automated classifier to enable large-scale analysis of knowledge construction in unstructured online discourse. We adapt four comment-level categories of knowledge construction: Non-Knowledge Construction, Share, Explore, and Integrate. Three trained annotators coded a balanced sample of 20,000 comments from YouTube education channels. The codebook demonstrated strong reliability, with Cohen's kappa = 0.79 on the main dataset and 0.85--0.93 across four additional educational domains. For automated classification, bag-of-words baselines were compared with transformer-based language models using 10-fold cross-validation. A DeBERTa-v3-large model achieved the highest macro-averaged F1 score (0.841), outperforming all baselines and other transformer models. External validation on four domains yielded macro-F1 above 0.705, with stronger transfer in medicine and programming, where discourse was more structured and task-focused, and weaker transfer in language and music, where comments were more varied and context-dependent. Overall, the study shows that theory-driven, semi-automated analysis of knowledge construction at scale is feasible, enabling the integration of knowledge-construction indicators into learning analytics and the design of online learning environments.
Authors: Qiang Ding, Lvzhou Luo, Yixuan Cao, Ping Luo
Abstract: Ensuring that Large Language Models (LLMs) generate summaries faithful to a given source document is essential for real-world applications. While prior research has explored LLM faithfulness, existing benchmarks suffer from annotation ambiguity, primarily due to the ill-defined boundary of permissible external knowledge in generated outputs. For instance, common sense is often incorporated into responses and labeled as "faithful", yet the acceptable extent of such knowledge remains unspecified, leading to inconsistent annotations. To address this issue, we propose a novel faithfulness annotation framework, which introduces an intermediate category, Out-Dependent, to classify cases where external knowledge is required for verification. Using this framework, we construct VeriGray (Verification with the Gray Zone) -- a new unfaithfulness detection benchmark in summarization. Statistics reveal that even SOTA LLMs, such as GPT-5, exhibit hallucinations ($\sim 6\%$ of sentences) in summarization tasks. Moreover, a substantial proportion ($\sim 8\%$ on average of models) of generated sentences fall into the Out-Dependent category, underscoring the importance of resolving annotation ambiguity in unfaithfulness detection benchmarks. Experiments demonstrate that our benchmark poses significant challenges to multiple baseline methods, indicating considerable room for future improvement.
Authors: Samarth Garg, Divya Singh, Deeksha Varshney, Mamta
Abstract: The rise of social networks has not only facilitated communication but also allowed the spread of harmful content. Although significant advances have been made in detecting toxic language in textual data, the exploration of concept-based explanations in toxicity detection remains limited. In this study, we leverage various subtype attributes present in toxicity detection datasets, such as obscene, threat, insult, identity attack, and sexual explicit as concepts that serve as strong indicators to identify whether language is toxic. However, disproportionate attribution of concepts towards the target class often results in classification errors. Our work introduces an interpretability technique based on the Concept Gradient (CG) method which provides a more causal interpretation by measuring how changes in concepts directly affect the output of the model. This is an extension of traditional gradient-based methods in machine learning, which often focus solely on input features. We propose the curation of Targeted Lexicon Set, which captures toxic words that contribute to misclassifications in text classification models. To assess the significance of these lexicon sets in misclassification, we compute Word-Concept Alignment (WCA) scores, which quantify the extent to which these words lead to errors due to over-attribution to toxic concepts. Finally, we introduce a lexicon-free augmentation strategy by generating toxic samples that exclude predefined toxic lexicon sets. This approach allows us to examine whether over-attribution persists when explicit lexical overlap is removed, providing insights into the model's attribution on broader toxic language patterns.
Authors: Jiatong Han
Abstract: Traditional Chinese Medicine theory is built on imagistic thinking, in which medical principles and diagnostic and therapeutic logic are structured through metaphor and metonymy. However, existing English translations largely rely on literal rendering, making it difficult for target-language readers to reconstruct the underlying conceptual networks and apply them in clinical practice. This study adopted a human-in-the-loop framework and selected four passages from the medical canon Huangdi Neijing that are fundamental in theory. Through prompt-based cognitive scaffolding, DeepSeek V3.1 was guided to identify metaphor and metonymy in the source text and convey the theory in translation. In the evaluation stage, ChatGPT 5 Pro and Gemini 2.5 Pro were instructed by prompts to simulate three types of real-world readers. Human translations, baseline model translations, and prompt-adjusted translations were scored by the simulated readers across five cognitive dimensions, followed by structured interviews and Interpretative Phenomenological Analysis. Results show that the prompt-adjusted LLM translations perform best across all five dimensions, with high cross-model and cross-role consistency. The interview themes reveal differences between human and machine translation, effective strategies for metaphor and metonymy transfer, and readers' cognitive preferences. This study provides a cognitive, efficient and replicable HITL methodological pathway for translation of ancient, concept-dense texts like TCM.
Authors: Jiatong Han
Abstract: Traditional Chinese Medicine (TCM) theory is built on imagistic thinking, in which medical principles and diagnostic and therapeutic logic are structured through metaphor and metonymy. However, existing English translations largely rely on literal rendering, making it difficult for target-language readers to reconstruct the underlying conceptual networks and apply them in clinical practice. This study adopted a human-in-the-loop (HITL) framework and selected four passages from the medical canon Huangdi Neijing that are fundamental in theory. Through prompt-based cognitive scaffolding, DeepSeek V3.1 was guided to identify metaphor and metonymy in the source text and convey the theory in translation. In the evaluation stage, ChatGPT 5 Pro and Gemini 2.5 Pro were instructed by prompts to simulate three types of real-world readers. Human translations, baseline model translations, and prompt-adjusted translations were scored by the simulated readers across five cognitive dimensions, followed by structured interviews and Interpretative Phenomenological Analysis (IPA). Results show that the prompt-adjusted LLM translations perform best across all five dimensions, with high cross-model and cross-role consistency. The interview themes reveal differences between human and machine translation, effective strategies for metaphor and metonymy transfer, and readers' cognitive preferences. This study provides a cognitive, efficient, and replicable HITL methodological pathway for the translation of ancient, concept-dense texts such as TCM.
Authors: Dingling Zhang, He Zhu, Jincheng Ren, Kangqi Song, Xinran Zhou, Boyu Feng, Shudong Liu, Jiabin Luo, Weihao Xie, Zhaohui Wang, Tianrui Qin, King Zhu, Yuqing Wang, Qianben Chen, Yuchen Eleanor Jiang, Wei Wang, Jiaheng Liu, Wangchunshu Zhou
Abstract: Deep Research Agents (DRAs) aim to automatically produce analyst-level reports through iterative information retrieval and synthesis. However, most existing DRAs were validated on question-answering benchmarks, while research on generating comprehensive reports remains overlooked. Worse, current benchmarks for report synthesis suffer from task complexity and subjective metrics -- this fails to reflect user demands and limits the practical utility of generated reports. To address these gaps, we present Fine-grained DEepResearch bench (FINDER), an enhanced benchmark consisting of 100 human-curated research tasks with 419 structured checklist items that standardize report structure, analytical depth, and factual grounding. Based on approximately 1,000 reports produced by mainstream DRAs, we further propose Deep rEsearch Failure Taxonomy (DEFT), the first failure taxonomy for deep research agents. DEFT contains 14 fine-grained failure modes across reasoning, retrieval, and generation, and is built upon grounded theory with human-LLM co-annotating and inter-annotator reliability validation. Our experimental findings reveal that current DRAs struggle not with task comprehension but with evidence integration, verification, and reasoning-resilient planning.
Authors: Changpeng Yang, Jinyang Wu, Yuchen Liu, Shuai Zhang, Yang Li, Qiliang Liang, Hongzhen Wang, Shuai Nie, Jiaming Xu, Runyu Shi, Ying Huang, Guoquan Zhang
Abstract: Reinforcement learning has emerged as a paradigm for post-training large language models, boosting their reasoning capabilities. Such approaches compute an advantage value for each sample, reflecting better or worse performance than expected, thereby yielding both positive and negative signals for training. However, the indiscriminate mixing of the two signals in existing methods, especially from the early stages, may lead to ambiguous guidance and limited gains. To address this issue, we propose **CAPO** (**C**urriculum **A**dvantage **P**olicy **O**ptimization), an adaptive curriculum mechanism based on advantage signals. The proposed mechanism bootstraps imitation learning with positive-only advantage samples to establish robust foundations, and subsequently introduces negative signals to cultivate discriminative capabilities, thereby improving generalization across complex scenarios. Compatible with diverse optimization methods including GRPO, PPO, RLOO, and Reinforce++, our method consistently achieves stable and significant improvements in mathematical reasoning tasks, and further generalizes effectively to multimodal Graphical User Interface (GUI) reasoning scenarios, establishing itself as a versatile and robust optimization framework.
Authors: Carolin Mueller-Spitzer, Samira Ochs, Jan Oliver Ruediger, Sascha Wolfer
Abstract: This study examines the distribution and linguistic characteristics of generic masculines (GM) in contemporary German press texts. The use of masculine personal nouns to refer to mixed-gender groups or unspecified individuals has been widely debated in academia and the public, with con-flicting perspectives on its gender-neutrality. While psycholinguistic studies suggest that GM is more readily associated with male referents, corpus-based analyses of its actual use remain scarce. We investigate GM in a large corpus of press texts, focusing on lexeme-specific differences across dif-ferent types of personal nouns. We conducted manual annotations of the whole inflectional para-digm of 21 personal nouns, resulting in 6,195 annotated tokens. Our findings reveal considerable differences between lexical items, especially between passive role nouns and prestige-related per-sonal nouns. On a grammatical level, we find that GM occurs predominantly in the plural and in indefinite noun phrases. Furthermore, our data shows that GM is not primarily used to denote entire classes of people, as has been previously claimed. By providing an empirical insight into the use of GM in authentic written language, we contribute to a more nuanced understanding of its forms and manifestations. These findings provide a solid basis for aligning linguistic stimuli in psy-cholinguistic studies more closely with real-world language use.
Authors: Youmin Ko, Sungjong Seo, Hyunjoon Kim
Abstract: Since large language models (LLMs) have a tendency to generate factually inaccurate output, retrieval-augmented generation (RAG) has gained significant attention as a key means to mitigate this downside of harnessing only LLMs. However, existing RAG methods for simple and multi-hop question answering (QA) are still prone to incorrect retrievals and hallucinations. To address these limitations, we propose CoopRAG, a novel RAG framework for the question answering task in which a retriever and an LLM work cooperatively with each other by exchanging informative knowledge, and the earlier and later layers of the retriever model work cooperatively with each other to accurately rank the retrieved documents relevant to a given query. In this framework, we (i) unroll a question into sub-questions and a reasoning chain in which uncertain positions are masked, (ii) retrieve the documents relevant to the question augmented with the sub-questions and the reasoning chain, (iii) rerank the documents by contrasting layers of the retriever, and (iv) reconstruct the reasoning chain by filling the masked positions via the LLM. Our experiments demonstrate that CoopRAG consistently outperforms state-of-the-art QA methods on three multi-hop QA datasets as well as a simple QA dataset in terms of both the retrieval and QA performances. Our code is available.
Authors: Shogo Fujita, Yuji Naraki, Yiqing Zhu, Shinsuke Mori
Abstract: This paper introduces LegalRikai: Open Benchmark, a new benchmark comprising four complex tasks that emulate Japanese corporate legal practices. The benchmark was created by legal professionals under the supervision of an attorney. This benchmark has 100 samples that require long-form, structured outputs, and we evaluated them against multiple practical criteria. We conducted both human and automated evaluations using leading LLMs, including GPT-5, Gemini 2.5 Pro, and Claude Opus 4.1. Our human evaluation revealed that abstract instructions prompted unnecessary modifications, highlighting model weaknesses in document-level editing that were missed by conventional short-text tasks. Furthermore, our analysis reveals that automated evaluation aligns well with human judgment on criteria with clear linguistic grounding, and assessing structural consistency remains a challenge. The result demonstrates the utility of automated evaluation as a screening tool when expert availability is limited. We propose a dataset evaluation framework to promote more practice-oriented research in the legal domain.
Authors: Zhangcheng Qiang, Kerry Taylor, Weiqing Wang
Abstract: Due to the dynamic nature of the Semantic Web, version control is necessary to capture time-varying information for widely used ontologies. Despite the long-standing recognition of ontology versioning (OV) as a crucial component of efficient ontology management, many views treat OV as similar to ontology matching (OM) and directly reuse OM systems for OV tasks. In this study, we systematically analyse the similarities and differences between OM and OV and formalise the OM4OV pipeline. The pipeline is implemented and evaluated in the state-of-the-art OM system Agent-OM. The experimental results indicate that OM systems can be reused for OV tasks, but without necessary modifications, the current OM4OV pipeline can produce skewed measurements, poor performance in detecting update entities, and less explainability for false mappings. To tackle these issues, we propose an optimisation method called the cross-reference (CR) mechanism, building upon the existing alignment(s) from OM to reduce the number of matching candidates and improve overall OV performance.
Authors: Zehua Pei, Hui-Ling Zhen, Xianzhi Yu, Sinno Jialin Pan, Mingxuan Yuan, Bei Yu
Abstract: Structured pruning of Generative Pre-trained Transformers (GPTs) offers a promising path to efficiency but often suffers from irreversible performance degradation due to the discarding of transformer blocks. In this paper, we introduce FuseGPT, a compression paradigm that reframes structured pruning as iterative knowledge grafting rather than simple removal. Motivated by the observation that linear block merging fails to capture non-linear feature disparities and that block importance fluctuates dynamically during pruning, FuseGPT employs a dual-strategy pipeline. First, we propose Macro Influence (MI), a dynamic fusion-aware metric that continuously re-evaluates block redundancy as the network topology evolves. Second, instead of rigid parameter averaging, we introduce a learnable low-rank fusion mechanism that adaptively grafts the knowledge of pruned blocks onto surviving layers via lightweight local distillation. Extensive experiments on LLaMA, Mistral, Qwen, and Phi families demonstrate that FuseGPT establishes a new state-of-the-art on the compression-accuracy Pareto frontier: at 25\% sparsity, FuseGPT achieves lower perplexity than prior methods at 20\% sparsity, improves zero-shot reasoning by up to 4.5 points, and delivers 1.33$\times$ inference speedup with 25\% memory reduction. Furthermore, FuseGPT is orthogonal to quantization, achieving 52.1\% total compression with negligible quality loss when combined with 4-bit GPTQ. We make our code publicly available at https://github.com/JarvisPei/FuseGPT.
Authors: Gabriel Stanovsky, Renana Keydar, Gadi Perl, Eliya Habba
Abstract: The performance of AI models on safety benchmarks does not indicate their real-world performance after deployment. This opaqueness of AI models impedes existing regulatory frameworks constituted on benchmark performance, leaving them incapable of mitigating ongoing real-world harm. The problem stems from a fundamental challenge in AI interpretability, which seems to be overlooked by regulators and decision makers. We propose a simple, realistic and readily usable regulatory framework which does not rely on benchmarks, and call for interdisciplinary collaboration to find new ways to address this crucial problem.
Authors: Parsa Hosseini, Sumit Nawathe, Mazda Moayeri, Sriram Balasubramanian, Soheil Feizi
Abstract: Unimodal vision models are known to rely on spurious correlations, but it remains unclear to what extent Multimodal Large Language Models (MLLMs) exhibit similar biases despite language supervision. In this paper, we investigate spurious bias in MLLMs and introduce SpurLens, a pipeline that leverages GPT-4 and open-set object detectors to automatically identify spurious visual cues without human supervision. Our findings reveal that spurious correlations cause two major failure modes in MLLMs: (1) over-reliance on spurious cues for object recognition, where removing these cues reduces accuracy, and (2) object hallucination, where spurious cues amplify the hallucination by over 10x. We validate our findings in various MLLMs and datasets. Beyond diagnosing these failures, we explore potential mitigation strategies, such as prompt ensembling and reasoning-based prompting, and conduct ablation studies to examine the root causes of spurious bias in MLLMs. By exposing the persistence of spurious correlations, our study calls for more rigorous evaluation methods and mitigation strategies to enhance the reliability of MLLMs.
Authors: Oren Sultan, Eitan Stern, Dafna Shahaf
Abstract: Large language models (LLMs) struggle with formal domains that require rigorous logical deduction and symbolic reasoning, such as mathematical proof generation. We propose a neuro-symbolic approach that combines LLMs' generative strengths with structured components to overcome this challenge. As a proof-of-concept, we focus on geometry problems. Our approach is two-fold: (1) we retrieve analogous problems and use their proofs to guide the LLM, and (2) a formal verifier evaluates the generated proofs and provides feedback, helping the model fix incorrect proofs. We demonstrate that our method significantly improves proof accuracy for OpenAI's o1 model (58%-70% improvement); both analogous problems and the verifier's feedback contribute to these gains. More broadly, shifting to LLMs that generate provably correct conclusions could dramatically improve their reliability, accuracy and consistency, unlocking complex tasks and critical real-world applications that require trustworthiness.
Authors: Eray Can Elumar, Cem Tekin, Osman Yagan
Abstract: Recent advances in large language models (LLMs) have enabled automated dataset labeling with minimal human supervision. While majority voting across multiple LLMs can improve label reliability by mitigating individual model biases, it incurs high computational costs due to repeated querying. In this work, we propose a novel online framework, Cost-aware Majority Voting (CaMVo), for efficient and accurate LLM-based dataset annotation. CaMVo adaptively selects a subset of LLMs for each data instance based on contextual embeddings, balancing confidence and cost without requiring pre-training or ground-truth labels. Leveraging a LinUCB-based selection mechanism and a Bayesian estimator over confidence scores, CaMVo estimates a lower bound on labeling accuracy for each LLM and aggregates responses through weighted majority voting. Our empirical evaluation on the MMLU and IMDB Movie Review datasets demonstrates that CaMVo achieves comparable or superior accuracy to full majority voting while significantly reducing labeling costs. This establishes CaMVo as a practical and robust solution for cost-efficient annotation in dynamic labeling environments.
Authors: Christian Walder, Deep Karkhanis
Abstract: Reinforcement Learning (RL) algorithms sample multiple n>1 solution attempts for each problem and reward them independently. This optimizes for pass@1 performance and prioritizes the strength of isolated samples at the expense of the diversity and collective utility of sets of samples. This under-utilizes the sampling capacity, limiting exploration and eventual improvement on harder examples. As a fix, we propose Pass-at-k Policy Optimization (PKPO), a transformation on the final rewards which leads to direct optimization of pass@k performance, thus optimizing for sets of samples that maximize reward when considered jointly. Our contribution is to derive novel low variance unbiased estimators for pass@k and its gradient, in both the binary and continuous reward settings. We show optimization with our estimators reduces to standard RL with rewards that have been jointly transformed by a stable and efficient transformation function. While previous efforts are restricted to k=n, ours is the first to enable robust optimization of pass@k for any arbitrary k <= n. Moreover, instead of trading off pass@1 performance for pass@k gains, our method allows annealing k during training, optimizing both metrics and often achieving strong pass@1 numbers alongside significant pass@k gains. We validate our reward transformations on toy experiments, which reveal the variance reducing properties of our formulations. We also include real-world examples using the open-source LLM, GEMMA-2. We find that our transformation effectively optimizes for the target k. Furthermore, higher k values enable solving more and harder problems, while annealing k boosts both the pass@1 and pass@k . Crucially, for challenging task sets where conventional pass@1 optimization stalls, our pass@k approach unblocks learning, likely due to better exploration by prioritizing joint utility over the utility of individual samples.
Authors: Shuyao Xu, Cheng Peng, Jiangxuan Long, Weidi Xu, Wei Chu, Yuan Qi
Abstract: Recent advances in model distillation show that data from advanced reasoning models can effectively train smaller student models. However, standard practices discard incorrect reasoning traces -- valuable, yet underutilized data. This paper addresses the critical question: How can both positive and negative distilled reasoning traces be effectively leveraged to maximize LLM reasoning performance in an offline setting? We employ a two-stage training recipe: first, Supervised Fine-Tuning (SFT) on positive traces, followed by a refinement stage using both positive and negative traces. We find that a simple REINFORCE-style objective, which we term the Reinforcement Distillation (REDI) objective, outperforms established preference optimization methods like DPO and SimPO in this distillation context. Our empirical evaluations demonstrate the effectiveness of this approach. Notably, our Qwen-REDI-1.5B model, trained on just 131k traces from the open Open-R1 dataset, achieves an 83.1% score on MATH-500. Its performance matches that of DeepSeek-R1-Distill-Qwen-1.5B, a model trained on 800k proprietary data. This result showcases the remarkable data efficiency of utilizing previously discarded negative traces.
Authors: Pan Lu, Jiayi Sheng, Luna Lyu, Jikai Jin, Tony Xia, Alex Gu, James Zou
Abstract: Inequality proving, crucial across diverse scientific and mathematical fields, tests advanced reasoning skills such as discovering tight bounds and strategic theorem application. This makes it a distinct, demanding frontier for large language models (LLMs), offering insights beyond general mathematical problem-solving. Progress in this area is hampered by existing datasets that are often scarce, synthetic, or rigidly formal. We address this by proposing an informal yet verifiable task formulation, recasting inequality proving into two automatically checkable subtasks: bound estimation and relation prediction. Building on this, we release IneqMath, an expert-curated dataset of Olympiad-level inequalities, including a test set and training corpus enriched with step-wise solutions and theorem annotations. We also develop a novel LLM-as-judge evaluation framework, combining a final-answer judge with four step-wise judges designed to detect common reasoning flaws. A systematic evaluation of 29 leading LLMs on IneqMath reveals a surprising reality: even top models like o1 achieve less than 10% overall accuracy under step-wise scrutiny; this is a drop of up to 65.5% from their accuracy considering only final answer equivalence. This discrepancy exposes fragile deductive chains and a critical gap for current LLMs between merely finding an answer and constructing a rigorous proof. Scaling model size and increasing test-time computation yield limited gains in overall proof correctness. Instead, our findings highlight promising research directions such as theorem-guided reasoning and self-refinement. Code and data are available at https://ineqmath.github.io/.
Authors: Haowei Lin, Haotian Ye, Wenzheng Feng, Quzhe Huang, Yujun Li, Hubert Lim, Zhengrui Li, Xiangyu Wang, Jianzhu Ma, James Zou, Yitao Liang
Abstract: Discovering scaling laws for predicting model performance at scale is a fundamental and open-ended challenge, mostly reliant on slow, case specific human experimentation. To investigate the potential for LLMs to automate this process, we collect over 5,000 experiments from existing literature and curate seven diverse scaling law discovery tasks. While existing agents struggle to produce accurate law formulas, this paper introduces SLDAgent, an evolution-based agent that co-optimize the scaling law model and the parameters, enabling it to autonomously explore complex relationships between variables. For the first time, we demonstrates that SLDAgent can automatically discover laws that exhibit consistently more accurate extrapolation than their established, human-derived counterparts across all tasks. Through comprehensive analysis, we elucidate why these discovered laws are superior and verify their practical utility in both pretraining and finetuning applications. This work establishes a new paradigm for agentic scientific discovery, showing that AI systems can understand their own scaling behavior, and can contribute novel and practical knowledge back to the research community.
Authors: Lijia Liu, Takumi Kondo, Kyohei Atarashi, Koh Takeuchi, Jiyi Li, Shigeru Saito, Hisashi Kashima
Abstract: This paper investigates defenses for LLM-based evaluation systems against prompt injection. We formalize a class of threats called blind attacks, where a candidate answer is crafted independently of the true answer to deceive the evaluator. To counter such attacks, we propose a framework that augments Standard Evaluation (SE) with Counterfactual Evaluation (CFE), which re-evaluates the submission against a deliberately false ground-truth answer. An attack is detected if the system validates an answer under both standard and counterfactual conditions. Experiments show that while standard evaluation is highly vulnerable, our SE+CFE framework significantly improves security by boosting attack detection with minimal performance trade-offs.
Authors: Yuxuan Cai, Yipeng Hao, Jie Zhou, Hang Yan, Zhikai Lei, Rui Zhen, Zhenhua Han, Yutao Yang, Junsong Li, Qianjun Pan, Tianyu Huai, Qin Chen, Xin Li, Kai Chen, Bo Zhang, Xipeng Qiu, Liang He
Abstract: As AI advances toward general intelligence, the focus is shifting from systems optimized for static tasks to creating open-ended agents that learn continuously. In this paper, we introduce Experience-driven Lifelong Learning (ELL), a framework for building self-evolving agents capable of continuous growth through real-world interaction. The framework is built on four core principles: (1) Experience Exploration: Agents learn through continuous, self-motivated interaction with dynamic environments, navigating interdependent tasks and generating rich experiential trajectories. (2) Long-term Memory: Agents preserve and structure historical knowledge, including personal experiences, domain expertise, and commonsense reasoning, into a persistent memory system. (3) Skill Learning: Agents autonomously improve by abstracting recurring patterns from experience into reusable skills, which are actively refined and validated for application in new tasks. (4) Knowledge Internalization: Agents internalize explicit and discrete experiences into implicit and intuitive capabilities as "second nature". We also introduce StuLife, a benchmark dataset for ELL that simulates a student's holistic college journey, from enrollment to academic and personal development, across three core phases and ten detailed sub-scenarios. StuLife is designed around three key paradigm
Authors: Jonas Becker, Lars Benedikt Kaesberg, Niklas Bauer, Jan Philip Wahle, Terry Ruas, Bela Gipp
Abstract: Multi-agent debate (MAD) has demonstrated the ability to augment collective intelligence by scaling test-time compute and leveraging expertise. Current frameworks for multi-agent debate are often designed towards tool use, lack integrated evaluation, or provide limited configurability of agent personas, response generators, discussion paradigms, and decision protocols. We introduce MALLM (Multi-Agent Large Language Models), an open-source framework that enables systematic analysis of MAD components. MALLM offers more than 144 unique configurations of MAD, including (1) agent personas (e.g., Expert, Personality), (2) response generators (e.g., Critical, Reasoning), (3) discussion paradigms (e.g., Memory, Relay), and (4) decision protocols (e.g., Voting, Consensus). MALLM uses simple configuration files to define a debate. Furthermore, MALLM can load any textual Hugging Face dataset (e.g., MMLU-Pro, WinoGrande) and provides an evaluation pipeline for easy comparison of MAD configurations. MALLM enables researchers to systematically configure, run, and evaluate debates for their problems, facilitating the understanding of the components and their interplay.
Authors: Aayush Mishra, Daniel Khashabi, Anqi Liu
Abstract: Supervised Fine-Tuning (SFT) is used to specialize model behavior by training weights to produce intended target responses for queries. In contrast, In-Context Learning (ICL) adapts models during inference with instructions or demonstrations in the prompt. ICL can offer better generalizability and more calibrated responses compared to SFT in data scarce settings, at the cost of more inference compute. In this work, we ask the question: Can ICL's internal computations be used to improve the qualities of SFT? We first show that ICL and SFT produce distinct activation patterns, indicating that the two methods achieve adaptation through different functional mechanisms. Motivated by this observation and to use ICL's rich functionality, we introduce ICL Activation Alignment (IA2), a self-distillation technique which aims to replicate ICL's activation patterns in SFT models and incentivizes ICL-like internal reasoning. Performing IA2 as a priming step before SFT significantly improves the accuracy and calibration of model outputs, as shown by our extensive empirical results on 12 popular benchmarks and two model families. This finding is not only practically useful, but also offers a conceptual window into the inner mechanics of model adaptation.
Authors: Minseo Kim, Coleman Hooper, Aditya Tomar, Chenfeng Xu, Mehrdad Farajtabar, Michael W. Mahoney, Kurt Keutzer, Amir Gholami
Abstract: Large Language Models (LLMs) have achieved state-of-the-art performance on a broad range of Natural Language Processing (NLP) tasks, including document processing and code generation. Autoregressive Language Models (ARMs), which generate tokens sequentially conditioned on all previous tokens, have been the predominant paradigm for LLMs. While these models have achieved high accuracy across a range of downstream tasks, they exhibit low arithmetic intensity due to the inherent sequential dependency in next-token prediction. Recently, Diffusion Language Models (DLMs) have emerged as a promising alternative architecture. DLMs generate output tokens in parallel, mitigating the limitations of sequential decoding. However, the performance implications of DLMs relative to commonly deployed ARMs are not fully understood. In this work, we present a comprehensive study of the performance characteristics of ARMs and DLMs, combining theoretical analysis with empirical profiling to characterize the trade-offs between these approaches. We show that although DLMs can achieve higher arithmetic intensity than ARMs by leveraging parallelism across token positions, they fail to scale effectively with longer contexts. We then explore block-wise decoding for DLMs, which decouples arithmetic intensity from sequence length and enables better scaling to long contexts (similar to ARMs). We also examine batched inference and find that ARMs exhibit superior throughput as they benefit more from parallelism across sequences in the batch. Finally, we highlight opportunities for accelerating DLM inference, emphasizing that reducing the number of sampling steps is key for open-source DLMs to achieve lower latency relative to ARMs.
Authors: Akira Okutomi
Abstract: We reinterpret Kant's Critique of Pure Reason as a theory of feedback stability, viewing reason as a regulator that keeps inference within the bounds of possible experience. We formalize this intuition in linear-Gaussian state-space models via H-Risk, a composite instability index integrating spectral margin, conditioning, temporal sensitivity, and innovation amplification. In simulations, higher H-Risk predicts overconfident errors and degraded closed-loop behavior even when the dynamics remain formally stable, exposing a gap between nominal and epistemic stability. Extending this stability lens to large language models (LLMs), we introduce a domain-wise proxy based on confidence fluctuations and overconfident errors. In a binary-question study, a Kantian-inspired policy that permits ''cannot judge'' responses yields targeted reductions in policy-aware squared loss in high-stakes domains relative to an overconfident baseline. To probe internal dynamics, we analyse layer-wise sensitivity of hidden states to small input perturbations. Contrary to a naive instability hypothesis, confidently wrong answers show no instability gap; instead, they are at least as locally stable as confidently correct answers, revealing stable miscalibration in which hallucinations behave like robust but misaligned attractors. For Qwen-2.5, spectral and activation profiles suggest a high signal-to-noise, low effective signal temperature regime in which representations become inertial and resistant to contextual shifts. These results bridge Kantian self-limitation and feedback control, and suggest that stable high-confidence hallucinations may not be readily corrected by output-only heuristics (e.g., temperature scaling or re-sampling), motivating process-level interventions that explicitly perturb and re-evaluate the inference trajectory.
Authors: Amirkia Rafiei Oskooei, Mehmet S. Aktas
Abstract: The proficiency of Large Language Models (LLMs) in processing structured data and adhering to syntactic rules is a capability that drives their widespread adoption but also makes them paradoxically vulnerable. In this paper, we investigate this vulnerability through BreakFun, a jailbreak methodology that weaponizes an LLM's adherence to structured schemas. BreakFun employs a three-part prompt that combines an innocent framing and a Chain-of-Thought distraction with a core "Trojan Schema"--a carefully crafted data structure that compels the model to generate harmful content, exploiting the LLM's strong tendency to follow structures and schemas. We demonstrate this vulnerability is highly transferable, achieving an average success rate of 89% across 13 foundational and proprietary models on JailbreakBench, and reaching a 100% Attack Success Rate (ASR) on several prominent models. A rigorous ablation study confirms this Trojan Schema is the attack's primary causal factor. To counter this, we introduce the Adversarial Prompt Deconstruction guardrail, a defense that utilizes a secondary LLM to perform a "Literal Transcription"--extracting all human-readable text to isolate and reveal the user's true harmful intent. Our proof-of-concept guardrail demonstrates high efficacy against the attack, validating that targeting the deceptive schema is a viable mitigation strategy. Our work provides a look into how an LLM's core strengths can be turned into critical weaknesses, offering a fresh perspective for building more robustly aligned models.
Authors: Palakorn Achananuparp, Ye Xu, Yao Lu, Xavier Jayaraj Siddarth Ashok, Ee-Peng Lim
Abstract: We present a large-scale analysis of career mobility of college-educated U.S. workers using online resume profiles to investigate how gender, race, and job change options are associated with upward mobility. This study addresses key research questions of how the job changes affect their upward career mobility, and how the outcomes of upward career mobility differ by gender and race. We address data challenges -- such as missing demographic attributes, missing wage data, and noisy occupation labels -- through various data processing and Artificial Intelligence (AI) methods. In particular, we develop a large language models (LLMs) based occupation classification method known as FewSOC that achieves accuracy significantly higher than the original occupation labels in the resume dataset. Analysis of 228,710 career trajectories reveals that intra-firm occupation change has been found to facilitate upward mobility most strongly, followed by inter-firm occupation change and inter-firm lateral move. Women and Black college graduates experience significantly lower returns from job changes than men and White peers. Multilevel sensitivity analyses confirm that these disparities are robust to cluster-level heterogeneity and reveal additional intersectional patterns.
Authors: Chenyang Shao, Dehao Huang, Yu Li, Keyu Zhao, Weiquan Lin, Yining Zhang, Qingbin Zeng, Zhiyu Chen, Tianxing Li, Yifei Huang, Taozhong Wu, Xinyang Liu, Ruotong Zhao, Mengsheng Zhao, Jiaoyang Li, Xuhua Zhang, Yue Wang, Yuanyi Zhen, Fengli Xu, Yong Li, Tie-Yan Liu
Abstract: With the rapid development of Large Language Models (LLMs), AI agents have demonstrated increasing proficiency in scientific tasks, ranging from hypothesis generation and experimental design to manuscript writing. Such agent systems are commonly referred to as "AI Scientists." However, existing AI Scientists predominantly formulate scientific discovery as a standalone search or optimization problem, overlooking the fact that scientific research is inherently a social and collaborative endeavor. Real-world science relies on a complex scientific infrastructure composed of collaborative mechanisms, contribution attribution, peer review, and structured scientific knowledge networks. Due to the lack of modeling for these critical dimensions, current systems struggle to establish a genuine research ecosystem or interact deeply with the human scientific community. To bridge this gap, we introduce OmniScientist, a framework that explicitly encodes the underlying mechanisms of human research into the AI scientific workflow. OmniScientist not only achieves end-to-end automation across data foundation, literature review, research ideation, experiment automation, scientific writing, and peer review, but also provides comprehensive infrastructural support by simulating the human scientific system, comprising: (1) a structured knowledge system built upon citation networks and conceptual correlations; (2) a collaborative research protocol (OSP), which enables seamless multi-agent collaboration and human researcher participation; and (3) an open evaluation platform (ScienceArena) based on blind pairwise user voting and Elo rankings. This infrastructure empowers agents to not only comprehend and leverage human knowledge systems but also to collaborate and co-evolve, fostering a sustainable and scalable innovation ecosystem.
Authors: Yixin Wan, Lei Ke, Wenhao Yu, Kai-Wei Chang, Dong Yu
Abstract: We introduce MotionEdit, a novel dataset for motion-centric image editing-the task of modifying subject actions and interactions while preserving identity, structure, and physical plausibility. Unlike existing image editing datasets that focus on static appearance changes or contain only sparse, low-quality motion edits, MotionEdit provides high-fidelity image pairs depicting realistic motion transformations extracted and verified from continuous videos. This new task is not only scientifically challenging but also practically significant, powering downstream applications such as frame-controlled video synthesis and animation. To evaluate model performance on the novel task, we introduce MotionEdit-Bench, a benchmark that challenges models on motion-centric edits and measures model performance with generative, discriminative, and preference-based metrics. Benchmark results reveal that motion editing remains highly challenging for existing state-of-the-art diffusion-based editing models. To address this gap, we propose MotionNFT (Motion-guided Negative-aware Fine Tuning), a post-training framework that computes motion alignment rewards based on how well the motion flow between input and model-edited images matches the ground-truth motion, guiding models toward accurate motion transformations. Extensive experiments on FLUX.1 Kontext and Qwen-Image-Edit show that MotionNFT consistently improves editing quality and motion fidelity of both base models on the motion editing task without sacrificing general editing ability, demonstrating its effectiveness. Our code is at https://github.com/elainew728/motion-edit/.
Authors: Devanshu Sahoo, Manish Prasad, Vasudev Majhi, Jahnvi Singh, Vinay Chamola, Yash Sinha, Murari Mandal, Dhruv Kumar
Abstract: The landscape of scientific peer review is rapidly evolving with the integration of Large Language Models (LLMs). This shift is driven by two parallel trends: the widespread individual adoption of LLMs by reviewers to manage workload (the "Lazy Reviewer" hypothesis) and the formal institutional deployment of AI-powered assessment systems by conferences like AAAI and Stanford's Agents4Science. This study investigates the robustness of these "LLM-as-a-Judge" systems (both illicit and sanctioned) to adversarial PDF manipulation. Unlike general jailbreaks, we focus on a distinct incentive: flipping "Reject" decisions to "Accept," for which we develop a novel evaluation metric which we term as WAVS (Weighted Adversarial Vulnerability Score). We curated a dataset of 200 scientific papers and adapted 15 domain-specific attack strategies to this task, evaluating them across 13 Language Models, including GPT-5, Claude Haiku, and DeepSeek. Our results demonstrate that obfuscation strategies like "Maximum Mark Magyk" successfully manipulate scores, achieving alarming decision flip rates even in large-scale models. We will release our complete dataset and injection framework to facilitate more research on this topic.