Authors: Jiaquan Zhang, Qigan Sun, Chaoning Zhang, Xudong Wang, Zhenzhen Huang, Yitian Zhou, Pengcheng Zheng, Chi-lok Andy Tai, Sung-Ho Bae, Zeyu Ma, Caiyan Qin, Jinyu Guo, Yang Yang, Hengtao Shen
Abstract: Enhancing the reasoning capability of large language models (LLMs) remains a core challenge in natural language processing. The Chain-of-Thought (CoT) paradigm dominates practical applications for its single-round efficiency, yet its reasoning chains often exhibit logical gaps. While multi-round paradigms like Graph-of-Thoughts (GoT), Tree-of-Thoughts (ToT), and Atom of Thought (AoT) achieve strong performance and reveal effective reasoning structures, their high cost limits practical use. To address this problem, this paper proposes a topology-based method for optimizing reasoning chains. The framework embeds essential topological patterns of effective reasoning into the lightweight CoT paradigm. Using persistent homology, we map CoT, ToT, and GoT into a unified topological space to quantify their structural features. On this basis, we design a unified optimization system: a Topological Optimization Agent diagnoses deviations in CoT chains from desirable topological characteristics and simultaneously generates targeted strategies to repair these structural deficiencies. Compared with multi-round reasoning methods like ToT and GoT, experiments on multiple datasets show that our approach offers a superior balance between reasoning accuracy and efficiency, showcasing a practical solution to ``single-round generation with multi-round intelligence''.
Authors: Julian Coda-Forno, Jane X. Wang, Arslan Chaudhry
Abstract: The reversal curse describes a failure of autoregressive language models to retrieve a fact in reverse order (e.g., training on ``$A > B$'' but failing on ``$B < A$''). Recent work shows that objectives with bidirectional supervision (e.g., bidirectional attention or masking-based reconstruction for decoder-only models) can mitigate the reversal curse. We extend this evaluation to include a vanilla masked language modeling (MLM) objective and compare it to decoder-only masking-based training across four reversal benchmarks and then provide a minimal mechanistic study of \emph{how} these objectives succeed. We show that reversal accuracy requires training signal that explicitly makes the source entity a prediction target, and we find little evidence that success corresponds to a single direction-agnostic representation of a fact. Instead, representation distances and linear probes are consistent with storing forward and reverse directions as distinct entries, with different indexing geometry for MLM versus decoder-only masking-based training. Our results caution that objective-level ``fixes'' can improve reversal behavior without necessarily inducing the kind of latent generalization one might expect from a unified concept.
Authors: Mohammad Reza Ghasemi Madani, Soyeon Caren Han, Shuo Yang, Jey Han Lau
Abstract: Multiple-choice questions (MCQs) are widely used to evaluate large language models (LLMs). However, LLMs remain vulnerable to the presence of plausible distractors. This often diverts attention toward irrelevant choices, resulting in unstable oscillation between correct and incorrect answers. In this paper, we propose Inclusion-of-Thoughts (IoT), a progressive self-filtering strategy that is designed to mitigate this cognitive load (i.e., instability of model preferences under the presence of distractors) and enable the model to focus more effectively on plausible answers. Our method operates to reconstruct the MCQ using only plausible option choices, providing a controlled setting for examining comparative judgements and therefore the stability of the model's internal reasoning under perturbation. By explicitly documenting this filtering process, IoT also enhances the transparency and interpretability of the model's decision-making. Extensive empirical evaluation demonstrates that IoT substantially boosts chain-of-thought performance across a range of arithmetic, commonsense reasoning, and educational benchmarks with minimal computational overhead.
Authors: Gowrav Vishwakarma, Christopher J. Agostino
Abstract: We present Phase-Associative Memory (PAM), a recurrent sequence model in which all representations are complex-valued, associations accumulate in a matrix state $S_{t}$ $\in$ $\mathbb{C}^{d \times d}$ via outer products, and retrieval operates through the conjugate inner product $K_t^* \cdot Q_t / \sqrt{d}$. At $\sim$100M parameters on WikiText-103, PAM reaches validation perplexity 30.0, within $\sim$10\% of a matched transformer (27.1) trained under identical conditions, despite $4\times$ arithmetic overhead from complex computation and no custom kernels. We trace the experimental path from vector-state models, where holographic binding fails due to the $O(1/\sqrt{n})$ capacity degradation of superposed associations, to the matrix state that resolves it. The competitiveness of an architecture whose native operations are complex-valued superposition and conjugate retrieval is consistent with recent empirical evidence that semantic interpretation in both humans and large language models exhibits non-classical contextuality, and we discuss what this implies for the choice of computational formalism in language modeling.
Authors: Hye Sun Yun, Geetika Kapoor, Michael Mackert, Ramez Kouzy, Wei Xu, Junyi Jessy Li, Byron C. Wallace
Abstract: Patients are increasingly turning to large language models (LLMs) with medical questions that are complex and difficult to articulate clearly. However, LLMs are sensitive to prompt phrasings and can be influenced by the way questions are worded. Ideally, LLMs should respond consistently regardless of phrasing, particularly when grounded in the same underlying evidence. We investigate this through a systematic evaluation in a controlled retrieval-augmented generation (RAG) setting for medical question answering (QA), where expert-selected documents are used rather than retrieved automatically. We examine two dimensions of patient query variation: question framing (positive vs. negative) and language style (technical vs. plain language). We construct a dataset of 6,614 query pairs grounded in clinical trial abstracts and evaluate response consistency across eight LLMs. Our findings show that positively- and negatively-framed pairs are significantly more likely to produce contradictory conclusions than same-framing pairs. This framing effect is further amplified in multi-turn conversations, where sustained persuasion increases inconsistency. We find no significant interaction between framing and language style. Our results demonstrate that LLM responses in medical QA can be systematically influenced through query phrasing alone, even when grounded in the same evidence, highlighting the importance of phrasing robustness as an evaluation criterion for RAG-based systems in high-stakes settings.
Authors: Xiangbo Zhang, Ali Emami
Abstract: Memorization in language models is widely studied but remains difficult to isolate and control. Understanding when and what models memorize is essential for explaining their predictions, yet existing approaches are post-hoc: they can detect memorization in trained models, but cannot disentangle its effects from architecture, data, or optimization. We introduce Memory Dial, a training framework that makes memorization pressure an explicit, controllable variable. Memory Dial interpolates between standard cross-entropy and a temperature-sharpened objective via a single parameter $\alpha$, producing a family of models identical in architecture and training setup (within each sweep), differing only in memorization pressure. Experiments across six architectures and five benchmarks demonstrate that: (1) $\alpha$ reliably controls memorization pressure, with seen-example accuracy increasing monotonically while unseen accuracy remains stable; (2) larger models are more responsive to memorization pressure; and (3) frequent sequences are easier to memorize than rare ones. Additional analyses show that the effect is robust across a range of sharpening temperatures, differs qualitatively from single-temperature cross-entropy, transfers to multilingual settings, and is detectable even on naturally occurring single-occurrence sequences. Memory Dial provides a controlled experimental framework for studying how memorization behavior emerges and interacts with generalization in language models.
Authors: Firoj Alam, Gagan Bhatia, Sahinur Rahman Laskar, Shammur Absar Chowdhury
Abstract: While Large Language Models (LLMs) are increasingly adopted as automated judges for evaluating generated text, their outputs are often costly, and highly sensitive to prompt design, language, and aggregation strategies, severely, which limits reproducibility. To address these challenges, we propose \textbf{\textit{OmniScore}}, a family of complementary, deterministic learned metrics developed using small size ($<$1B) parameter models. OmniScore approximates LLM-judge behavior while preserving the low latency and consistency of traditional model-based scoring. We trained the models large-scale synthetic supervision ($\sim$564k instances, in \textbf{107 languages}) and evaluated using 8,617 manually annotated instances. The OmniScore family supports reliable, multi-dimensional scores across a variety of settings, including reference-based, source-grounded, and hybrid evaluations. We evaluate these models across question answering (QA), translation, and summarization in \textbf{6 languages}. Our results demonstrate that lightweight, deterministic learned metrics provide a highly practical and scalable alternative to frontier LLMs. Our models and datasets can be found at https://huggingface.co/collections/QCRI/omniscore
Authors: Omri Uzan, Ron Polonsky, Douwe Kiela, Christopher Potts
Abstract: Document expansion is a classical technique for improving retrieval quality, and is attractive since it shifts computation offline, avoiding additional query-time processing. However, when applied to modern retrievers, it has been shown to degrade performance, often introducing noise that obfuscates the discriminative signal. We recast document expansion as a document optimization problem: a language model or a vision language model is fine-tuned to transform documents into representations that better align with the expected query distribution under a target retriever, using GRPO with the retriever's ranking improvements as rewards. This approach requires only black-box access to retrieval ranks, and is applicable across single-vector, multi-vector and lexical retrievers. We evaluate our approach on code retrieval and visual document retrieval (VDR) tasks. We find that learned document transformations yield retrieval gains and in many settings enable smaller, more efficient retrievers to outperform larger ones. For example, applying document optimization to OpenAI text-embedding-3-small model improves nDCG5 on code (58.7 to 66.8) and VDR (53.3 to 57.6), even slightly surpassing the 6.5X more expensive OpenAI text-embedding-3-large model (66.3 on code; 57.0 on VDR). When retriever weights are accessible, document optimization is often competitive with fine-tuning, and in most settings their combination performs best, improving Jina-ColBERT-V2 from 55.8 to 63.3 on VDR and from 48.6 to 61.8 on code retrieval.
Authors: Aastha A K Verma, Anwoy Chatterjee, Mehak Gupta, Tanmoy Chakraborty
Abstract: Multilingual language models (LMs) organize representations for typologically and orthographically diverse languages into a shared parameter space, yet the nature of this internal organization remains elusive. In this work, we investigate which linguistic properties - abstract language identity or surface-form cues - shape multilingual representations. Focusing on compact, distilled models where representational trade-offs are explicit, we analyze language-associated units in Llama-3.2-1B and Gemma-2-2B using the Language Activation Probability Entropy (LAPE) metric, and further decompose activations with Sparse Autoencoders. We find that these units are strongly conditioned on orthography: romanization induces near-disjoint representations that align with neither native-script inputs nor English, while word-order shuffling has limited effect on unit identity. Probing shows that typological structure becomes increasingly accessible in deeper layers, while causal interventions indicate that generation is most sensitive to units that are invariant to surface-form perturbations rather than to units identified by typological alignment alone. Overall, our results suggest that multilingual LMs organize representations around surface form, with linguistic abstraction emerging gradually without collapsing into a unified interlingua.
Authors: Zhengqing Yuan, Hanchi Sun, Lichao Sun, Yanfang Ye
Abstract: We present MegaTrain, a memory-centric system that efficiently trains 100B+ parameter large language models at full precision on a single GPU. Unlike traditional GPU-centric systems, MegaTrain stores parameters and optimizer states in host memory (CPU memory) and treats GPUs as transient compute engines. For each layer, we stream parameters in and compute gradients out, minimizing persistent device state. To battle the CPU-GPU bandwidth bottleneck, we adopt two key optimizations. 1) We introduce a pipelined double-buffered execution engine that overlaps parameter prefetching, computation, and gradient offloading across multiple CUDA streams, enabling continuous GPU execution. 2) We replace persistent autograd graphs with stateless layer templates, binding weights dynamically as they stream in, eliminating persistent graph metadata while providing flexibility in scheduling. On a single H200 GPU with 1.5TB host memory, MegaTrain reliably trains models up to 120B parameters. It also achieves 1.84$\times$ the training throughput of DeepSpeed ZeRO-3 with CPU offloading when training 14B models. MegaTrain also enables 7B model training with 512k token context on a single GH200.
Authors: Hanbing Liu, Lang Cao, Yang Li
Abstract: Large language models (LLMs) acquire most of their knowledge during pretraining, which ties them to a fixed snapshot of the world and makes adaptation to continuously evolving knowledge challenging. As facts, entities, and events change over time, models may experience continuous knowledge drift, resulting not only in outdated predictions but also in temporally inconsistent reasoning. Although existing approaches, such as continual finetuning, knowledge editing, and retrieval-augmented generation (RAG), aim to update or supplement model knowledge, they are rarely evaluated in settings that reflect chronological, evolving, and real-world knowledge evolution. In this work, we introduce a new benchmark of real-world dynamic events, constructed from time-stamped evidence that captures how knowledge evolves over time, which enables systematic evaluation of model adaptation under continuous knowledge drift. The benchmark reveals that most existing methods, including vanilla RAG and several learning-based approaches, struggle under this setting, exposing critical limitations such as catastrophic forgetting and temporal inconsistency. To mitigate these limitations, we propose a time-aware retrieval baseline, Chronos, which progressively organizes retrieved evidence into an Event Evolution Graph to enable more temporally consistent understanding in LLMs without additional training. Overall, this work provides a foundation for analyzing and advancing LLM adaptation to continuous knowledge drift in realistic settings.
Authors: Quyet V. Do, Thinh Pham, Nguyen Nguyen, Sha Li, Pratibha Zunjare, Tu Vu
Abstract: We study a pipeline that curates reasoning data from initial structured data for improving long-context reasoning in large language models (LLMs). Our approach, $\pi^2$, constructs high-quality reasoning data through rigorous QA curation: 1) extracting and expanding tables from Wikipedia, 2) from the collected tables and relevant context, generating realistic and multi-hop analytical reasoning questions whose answers are automatically determined and verified through dual-path code execution, and 3) back-translating step-by-step structured reasoning traces as solutions of QA pairs given realistic web-search context. Supervised fine-tuning with \textsc{\small{gpt-oss-20b}} and \textsc{\small{Qwen3-4B-Instruct-2507}} on $\pi^2$ yields consistent improvements across four long-context reasoning benchmarks and our alike $\pi^2$-Bench, with average absolute accuracy gains of +4.3% and +2.7% respectively. Notably, our dataset facilitates self-distillation, where \textsc{\small{gpt-oss-20b}} even improves its average performance by +4.4% with its own reasoning traces, demonstrating $\pi^2$'s usefulness. Our code, data, and models are open-source at https://github.com/vt-pi-squared/pi-squared.
Authors: Berny Kabalisa
Abstract: We introduce SenseAI, a human-in-the-loop (HITL) validated financial sentiment dataset designed to capture not only model outputs but the full reasoning process behind them. Unlike existing resources, SenseAI incorporates reasoning chains, confidence scores, human correction signals, and real-world market outcomes, providing a structure aligned with Reinforcement Learning from Human Feedback (RLHF) paradigms. The dataset consists of 1,439 labelled data points across 40 US-listed equities and 13 financial data categories, enabling direct integration into modern LLM fine-tuning pipelines. Through analysis, we identify several systematic patterns in model behavior, including a novel failure mode we term Latent Reasoning Drift, where models introduce information not grounded in the input, as well as consistent confidence miscalibration and forward projection tendencies. These findings suggest that LLM errors in financial reasoning are not random but occur within a predictable and correctable regime, supporting the use of structured HITL data for targeted model improvement. We discuss implications for financial AI systems and highlight opportunities for applying SenseAI in model evaluation and alignment.
Authors: Jiatan Huang, Zheyuan Zhang, Kaiwen Shi, Yanfang Ye, Chuxu Zhang
Abstract: Large language model agents often exhibit complementary strengths, making routing a promising approach for multi-agent question answering. However, existing routing methods remain limited in two important ways: they typically optimize over a fixed pool of agents without improving the agents themselves, and they often rely on rigid collaboration schemes that cannot adapt the number of participating agents to the query. We propose EvolveRouter, a trainable framework that addresses both limitations by jointly improving agent quality and collaboration structure. First, EvolveRouter couples graph-based query routing with targeted instruction refinement in a closed-loop co-evolution process, allowing router diagnostics to guide agent improvement while refined agents provide cleaner supervision for routing. Second, it introduces an adaptive inference strategy that dynamically determines the effective collaboration size for each query through router-weighted answer agreement. Together, these designs enable more capable and more efficient multi-agent reasoning. Experiments on five question answering benchmarks show that EvolveRouter consistently outperforms SOTA routing baselines in both F1 and exact match, while further analysis confirms the benefits of closed-loop refinement and adaptive collaboration.
Authors: Ahmed Ewais, Ahmed Hashish, Amr Ali
Abstract: Large language models encode extensive world knowledge valuable for zero-shot named entity recognition. However, their causal attention mechanism, where tokens attend only to preceding context, prevents effective token classification when disambiguation requires future context. Existing approaches use LLMs generatively, prompting them to list entities or produce structured outputs, but suffer from slow autoregressive decoding, hallucinated entities, and formatting errors. We propose Just Pass Twice (JPT), a simple yet effective method that enables causal LLMs to perform discriminative token classification with full bidirectional context. Our key insight is that concatenating the input to itself lets each token in the second pass attend to the complete sentence, requiring no architectural modifications. We combine these representations with definition-guided entity embeddings for flexible zero-shot generalization. Our approach achieves state-of-the-art results on zero-shot NER benchmarks, surpassing the previous best method by +7.9 F1 on average across CrossNER and MIT benchmarks, being over 20x faster than comparable generative methods.
Authors: Jonathan Ivey, Anjalie Field, Ziang Xiao
Abstract: Qualitative interviews provide essential insights into human experiences when they elicit high-quality responses. While qualitative and NLP researchers have proposed various measures of interview quality, these measures lack validation that high-scoring responses actually contribute to the study's goals. In this work, we identify, implement, and evaluate 10 proposed measures of interview response quality to determine which are actually predictive of a response's contribution to the study findings. To conduct our analysis, we introduce the Qualitative Interview Corpus, a newly constructed dataset of 343 interview transcripts with 16,940 participant responses from 14 real research projects. We find that direct relevance to a key research question is the strongest predictor of response quality. We additionally find that two measures commonly used to evaluate NLP interview systems, clarity and surprisal-based informativeness, are not predictive of response quality. Our work provides analytic insights and grounded, scalable metrics to inform the design of qualitative studies and the evaluation of automated interview systems.
Authors: Purva Chiniya, Kevin Scaria, Sagar Chaturvedi
Abstract: Large language models (LLMs) remain susceptible to jailbreak and direct prompt-injection attacks, yet the strongest defensive filters frequently over-refuse benign queries and degrade user experience. Previous work on jailbreak and prompt injection detection such as GradSafe, detects unsafe prompts with a single "accept all" anchor token, but its threshold is brittle and it offers no deterministic guarantee that harmful content will not be emitted once decoding begins. We introduce Gradient-Controlled Decoding (GCD), a training-free guardrail that combines an acceptance anchor token ("Sure") and refusal anchor token ("Sorry") tightening the decision boundary and significantly lowering false positives. In the mitigation stage, if a prompt is flagged, GCD preset-injects one or two refusal tokens ("Sorry, I can't...") before autoregressive decoding resumes, guaranteeing first-token safety regardless of sampling strategy. On ToxicChat, XSTest-v2, and AdvBench, GCD reduces false positives by 52% vs. GradSafe at comparable recall, lowers attack success rate by up to 10% vs. the strongest decoding-only baseline, adds under 15-20 ms latency on an average on V100 instances, transfers to LLaMA-2-7B, Mixtral-8x7B, and Qwen-2-7B, and requires only 20 demonstration templates.
Authors: Ziyi Chen, Mengxian Lyu, Cheng Peng, Yonghui Wu
Abstract: Screening patients for enrollment is a well-known, labor-intensive bottleneck that leads to under-enrollment and, ultimately, trial failures. Recent breakthroughs in large language models (LLMs) offer a promising opportunity to use artificial intelligence to improve screening. This study systematically explored both encoder- and decoder-based generative LLMs for screening clinical narratives to facilitate clinical trial recruitment. We examined both general-purpose LLMs and medical-adapted LLMs and explored three strategies to alleviate the "Lost in the Middle" issue when handling long documents, including 1) Original long-context: using the default context windows of LLMs, 2) NER-based extractive summarization: converting the long document into summarizations using named entity recognition, 3) RAG: dynamic evidence retrieval based on eligibility criteria. The 2018 N2C2 Track 1 benchmark dataset is used for evaluation. Our experimental results show that the MedGemma model with the RAG strategy achieved the best micro-F1 score of 89.05%, outperforming other models. Generative LLMs have remarkably improved trial criteria that require long-term reasoning across long documents, whereas trial criteria that span a short piece of context (e.g., lab tests) show incremental improvements. The real-world adoption of LLMs for trial recruitment must consider specific criteria for selecting among rule-based queries, encoder-based LLMs, and generative LLMs to maximize efficiency within reasonable computing costs.
Authors: Craig W. Schmidt, Chris Tanner, Yuval Pinter
Abstract: Byte Pair Encoding (BPE) is a widely used tokenization algorithm, whose tokens cannot extend across pre-tokenization boundaries, functionally limiting it to representing at most full words. The BoundlessBPE and SuperBPE algorithms extend and improve BPE by relaxing this limitation and allowing the formation of superwords, which are combinations of pretokens that form phrases. However, previous implementations were impractical to train: for example, BoundlessBPE took 4.7 CPU days to train on 1GB of data. We show that supermerge candidates, two or more consecutive pretokens eligible to form a supermerge, can be aggregated by frequency much like regular pretokens. This avoids keeping full documents in memory, as the original implementations of BoundlessBPE and SuperBPE required, leading to a significant training speedup. We present a two-phase formulation of BoundlessBPE that separates first-phase learning of regular merges from second-phase learning of supermerges, producing identical results to the original implementation. We also show a near-equivalence between two-phase BoundlessBPE and SuperBPE, with the difference being that a manually selected hyperparameter used in SuperBPE can be automatically determined in the second phase of BoundlessBPE. These changes enable a much faster implementation, allowing training on that same 1GB of data in 603 and 593 seconds for BoundlessBPE and SuperBPE, respectively, a more than 600x increase in speed. For each of BoundlessBPE, SuperBPE, and BPE, we open-source both a reference Python implementation and a fast Rust implementation.
Authors: Jiahao Xu, Rui Hu, Olivera Kotevska, Zikai Zhang
Abstract: Multi-bit watermarking has emerged as a promising solution for embedding imperceptible binary messages into Large Language Model (LLM)-generated text, enabling reliable attribution and tracing of malicious usage of LLMs. Despite recent progress, existing methods still face key limitations: some become computationally infeasible for large messages, while others suffer from a poor trade-off between text quality and decoding accuracy. Moreover, the decoding accuracy of existing methods drops significantly when the number of tokens in the generated text is limited, a condition that frequently arises in practical usage. To address these challenges, we propose \textsc{XMark}, a novel method for encoding and decoding binary messages in LLM-generated texts. The unique design of \textsc{XMark}'s encoder produces a less distorted logit distribution for watermarked token generation, preserving text quality, and also enables its tailored decoder to reliably recover the encoded message with limited tokens. Extensive experiments across diverse downstream tasks show that \textsc{XMark} significantly improves decoding accuracy while preserving the quality of watermarked text, outperforming prior methods. The code is at https://github.com/JiiahaoXU/XMark.
Authors: Jon-Paul Cacioli
Abstract: Background: Children do not simply learn that balls are round and blocks are square. They learn that shape is the kind of feature that tends to define object categories -- a second-order generalisation known as an overhypothesis [1, 2]. What kind of learning mechanism is sufficient for this inductive leap? Methods: We trained autoregressive transformer language models (3.4M-25.6M parameters) on synthetic corpora in which shape is the stable feature dimension across categories, with eight conditions controlling for alternative explanations. Results: Across 120 pre-registered runs evaluated on a 1,040-item wug test battery, every model achieved perfect first-order exemplar retrieval (100%) while second-order generalisation to novel nouns remained at chance (50-52%), a result confirmed by equivalence testing. A feature-swap diagnostic revealed that models rely on frame-to-feature template matching rather than structured noun-to-domain-to-feature abstraction. Conclusions: These results reveal a clear limitation of autoregressive distributional sequence learning under developmental-scale training conditions.
Authors: Giang Do, Hung Le, Truyen Tran
Abstract: In the era of Large Language Models (LLMs), the Mixture of Experts (MoE) architecture has emerged as an effective approach for training extremely large models with improved computational efficiency. This success builds upon extensive prior research aimed at enhancing expert specialization in MoE-based LLMs. However, the nature of such specializations and how they can be systematically interpreted remain open research challenges. In this work, we investigate this gap by posing a fundamental question: \textit{Do domain-specific experts exist in MoE-based LLMs?} To answer the question, we evaluate ten advanced MoE-based LLMs ranging from 3.8B to 120B parameters and provide empirical evidence for the existence of domain-specific experts. Building on this finding, we propose \textbf{Domain Steering Mixture of Experts (DSMoE)}, a training-free framework that introduces zero additional inference cost and outperforms both well-trained MoE-based LLMs and strong baselines, including Supervised Fine-Tuning (SFT). Experiments on four advanced open-source MoE-based LLMs across both target and non-target domains demonstrate that our method achieves strong performance and robust generalization without increasing inference cost or requiring additional retraining. Our implementation is publicly available at https://github.com/giangdip2410/Domain-specific-Experts.
URLs: https://github.com/giangdip2410/Domain-specific-Experts.
Authors: Kabir Ahuja, Yuxuan Li, Andrew Kyle Lampinen
Abstract: Human communication is fundamentally creative, and often makes use of subtext -- implied meaning that goes beyond the literal content of the text. Here, we systematically study whether language models can use subtext in communicative settings, and introduce four new evaluation suites to assess these capabilities. Our evaluation settings range from writing & interpreting allegories to playing multi-agent and multi-modal games inspired by the rules of board games like Dixit. We find that frontier models generally exhibit a strong bias towards overly literal, explicit communication, and thereby fail to account for nuanced constraints -- even the best performing models generate literal clues 60% of times in one of our environments -- Visual Allusions. However, we find that some models can sometimes make use of common ground with another party to help them communicate with subtext, achieving 30%-50% reduction in overly literal clues; but they struggle at inferring presence of a common ground when not explicitly stated. For allegory understanding, we find paratextual and persona conditions to significantly shift the interpretation of subtext. Overall, our work provides quantifiable measures for an inherently complex and subjective phenomenon like subtext and reveals many weaknesses and idiosyncrasies of current LLMs. We hope this research to inspire future work towards socially grounded creative communication and reasoning.
Authors: Jinhong Jeong, Junghun Park, Youngjae Yu
Abstract: Text simplification supports second language (L2) learning by providing comprehensible input, consistent with the Input Hypothesis. However, constructing personalized parallel corpora is costly, while existing large language model (LLM)-based readability control methods rely on pre-labeled sentence corpora and primarily target English. We propose Re-RIGHT, a unified reinforcement learning framework for adaptive multilingual text simplification without parallel corpus supervision. We first show that prompting-based lexical simplification at target proficiency levels (CEFR, JLPT, TOPIK, and HSK) performs poorly at easier levels and for non-English languages, even with state-of-the-art LLMs such as GPT-5.2 and Gemini 2.5. To address this, we collect 43K vocabulary-level data across four languages (English, Japanese, Korean, and Chinese) and train a compact 4B policy model using Re-RIGHT, which integrates three reward modules: vocabulary coverage, semantic preservation, and coherence. Compared to the stronger LLM baselines, Re-RIGHT achieves higher lexical coverage at target proficiency levels while maintaining original meaning and fluency.
Authors: Jason Lucas, Matt Murtagh, Ali Al-Lawati, Uchendu Uchendu, Adaku Uchendu, Dongwon Lee
Abstract: Harmful content detectors-particularly disinformation classifiers-are predominantly developed and evaluated on Standard American English (SAE), leaving their robustness to dialectal variation unexplored. We present DIA-HARM, the first benchmark for evaluating disinformation detection robustness across 50 English dialects spanning U.S., British, African, Caribbean, and Asia-Pacific varieties. Using Multi-VALUE's linguistically grounded transformations, we introduce D3 (Dialectal Disinformation Detection), a corpus of 195K samples derived from established disinformation benchmarks. Our evaluation of 16 detection models reveals systematic vulnerabilities: human-written dialectal content degrades detection by 1.4-3.6% F1, while AI-generated content remains stable. Fine-tuned transformers substantially outperform zero-shot LLMs (96.6% vs. 78.3% best-case F1), with some models exhibiting catastrophic failures exceeding 33% degradation on mixed content. Cross-dialectal transfer analysis across 2,450 dialect pairs shows that multilingual models (mDeBERTa: 97.2% average F1) generalize effectively, while monolingual models like RoBERTa and XLM-RoBERTa fail on dialectal inputs. These findings demonstrate that current disinformation detectors may systematically disadvantage hundreds of millions of non-SAE speakers worldwide. We release the DIA-HARM framework, D3 corpus, and evaluation tools: https://github.com/jsl5710/dia-harm
Authors: Xiangxu Zhang, Jiamin Wang, Qinlin Zhao, Hanze Guo, Linzhuo Li, Jing Yao, Xiao Zhou, Xiaoyuan Yi, Xing Xie
Abstract: As LLMs become increasingly integrated into human society, evaluating their orientations on human values from social science has drawn growing attention. Nevertheless, it is still unclear why human values matter for LLMs, especially in LLM-based multi-agent systems, where group-level failures may accumulate from individually misaligned actions. We ask whether misalignment with human values alters the collective behavior of LLM agents and what changes it induces? In this work, we introduce CIVA, a controlled multi-agent environment grounded in social science theories, where LLM agents form a community and autonomously communicate, explore, and compete for resources, enabling systematic manipulation of value prevalence and behavioral analysis. Through comprehensive simulation experiments, we reveal three key findings. (1) We identify several structurally critical values that substantially shape the community's collective dynamics, including those diverging from LLMs' original orientations. Triggered by the misspecification of these values, we (2) detect system failure modes, e.g., catastrophic collapse, at the macro level, and (3) observe emergent behaviors like deception and power-seeking at the micro level. These results offer quantitative evidence that human values are essential for collective outcomes in LLMs and motivate future multi-agent value alignment.
Authors: Vishaal Kapoor, Mariam Dundua, Sarthak Ahuja, Neda Kordjazi, Evren Yortucboylu, Vaibhavi Padala, Derek Ho, Jennifer Whitted, Rebecca Steinert
Abstract: Enterprise IT support interactions are fundamentally diagnostic: effective resolution requires iterative evidence gathering from ambiguous user reports to identify an underlying root cause. While retrieval-augmented generation (RAG) provides grounding through historical cases, standard multi-turn RAG systems lack explicit diagnostic state and therefore struggle to accumulate evidence and resolve competing hypotheses across turns. We introduce DQA, a diagnostic question-answering framework that maintains persistent diagnostic state and aggregates retrieved cases at the level of root causes rather than individual documents. DQA combines conversational query rewriting, retrieval aggregation, and state-conditioned response generation to support systematic troubleshooting under enterprise latency and context constraints. We evaluate DQA on 150 anonymized enterprise IT support scenarios using a replay-based protocol. Averaged over three independent runs, DQA achieves a 78.7% success rate under a trajectory-level success criterion, compared to 41.3% for a multi-turn RAG baseline, while reducing average turns from 8.4 to 3.9.
Authors: Kaiser Hamid, Can Cui, Nade Liang
Abstract: Recent progress in vision-language-action (VLA) models has enabled language-conditioned driving agents to execute natural-language navigation commands in closed-loop simulation, yet standard evaluations largely assume instructions are precise and well-formed. In deployment, instructions vary in phrasing and specificity, may omit critical qualifiers, and can occasionally include misleading, authority-framed text, leaving instruction-level robustness under-measured. We introduce ICR-Drive, a diagnostic framework for instruction counterfactual robustness in end-to-end language-conditioned autonomous driving. ICR-Drive generates controlled instruction variants spanning four perturbation families: Paraphrase, Ambiguity, Noise, and Misleading, where Misleading variants conflict with the navigation goal and attempt to override intent. We replay identical CARLA routes under matched simulator configurations and seeds to isolate performance changes attributable to instruction language. Robustness is quantified using standard CARLA Leaderboard metrics and per-family performance degradation relative to the baseline instruction. Experiments on LMDrive and BEVDriver show that minor instruction changes can induce substantial performance drops and distinct failure modes, revealing a reliability gap for deploying embodied foundation models in safety-critical driving.
Authors: Zhaohan Zhang, Chengzhengxu Li, Xiaoming Liu, Chao Shen, Ziquan Liu, Ioannis Patras
Abstract: Large Language Models (LLMs) are increasingly applied in high-stakes domains such as finance, healthcare, and education, where reliable multi-turn interactions with users are essential. However, existing work on confidence estimation and calibration, a major approach to building trustworthy LLM systems, largely focuses on single-turn settings and overlooks the risks and potential of multi-turn conversations. In this work, we introduce the task of multi-turn calibration to reframe calibration from a static property into a dynamic challenge central to reliable multi-turn conversation, where calibrating model confidence at each turn conditioned on the conversation history is required. We first reveal the risks of this setting: using Expected Calibration Error at turn T (ECE@T), a new metric that tracks calibration dynamics over turns, we show that user feedback (e.g., persuasion) can degrade multi-turn calibration. To address this, we propose MTCal, which minimises ECE@T via a surrogate calibration target, and further leverage calibrated confidence in ConfChat, a decoding strategy that improves both factuality and consistency of the model response in multi-turn interactions. Extensive experiments demonstrate that MT-Cal achieves outstanding and consistent performance in multi-turn calibration, and ConfChat preserves and even enhances model performance in multi-turn interactions. Our results mark multi-turn calibration as one missing link for scaling LLM calibration toward safe, reliable, and real-world use.
Authors: Taehyeon Kim, Hojung Jung, Se-Young Yun
Abstract: Speculative decoding (SD) accelerates large language model (LLM) inference by using a smaller model to draft future tokens, which are then verified by the target LLM. This preserves generation quality by accepting only aligned tokens. However, individual drafters, often trained for specific tasks or domains, exhibit limited effectiveness across diverse applications. To address this, we introduce \textsc{MetaSD}, a unified framework that integrates multiple drafters into the SD process. MetaSD dynamically allocates computational resources to heterogeneous drafters by leveraging alignment feedback and framing drafter selection as a multi-armed bandit problem. Extensive experiments show MetaSD consistently outperforms single-drafter approaches.
Authors: Qiyuan Chen, Hongsen Huang, Jiahe Chen, Qian Shao, Jintai Chen, Hongxia Xu, Renjie Hua, Chuan Ren, Jian Wu
Abstract: Vision-language reward modeling faces a dilemma: generative approaches are interpretable but slow, while discriminative ones are efficient but act as opaque "black boxes." To bridge this gap, we propose VL-MDR (Vision-Language Multi-Dimensional Reward), a framework that dynamically decomposes evaluation into granular, interpretable dimensions. Instead of outputting a monolithic scalar, VL-MDR employs a visual-aware gating mechanism to identify relevant dimensions and adaptively weight them (e.g., Hallucination, Reasoning) for each specific input. To support this, we curate a dataset of 321k vision-language preference pairs annotated across 21 fine-grained dimensions. Extensive experiments show that VL-MDR consistently outperforms existing open-source reward models on benchmarks like VL-RewardBench. Furthermore, we show that VL-MDR-constructed preference pairs effectively enable DPO alignment to mitigate visual hallucinations and improve reliability, providing a scalable solution for VLM alignment.
Authors: Yifeng He, Ziye Tang, Hao Chen
Abstract: Information cocoons on social media limit users' exposure to posts with diverse viewpoints. Modern platforms use stance detection as an important signal in recommendation and ranking pipelines, which can route posts primarily to like-minded audiences and reduce cross-cutting exposure. This restricts the reach of dissenting opinions and hinders constructive discourse. We take the creator's perspective and investigate how content can be revised to reach beyond existing affinity clusters. We present ContentFuzz, a confidence-guided fuzzing framework that rewrites posts while preserving their human-interpreted intent and induces different machine-inferred stance labels. ContentFuzz aims to route posts beyond their original cocoons. Our method guides a large language model (LLM) to generate meaning-preserving rewrites using confidence feedback from stance detection models. Evaluated on four representative stance detection models across three datasets in two languages, ContentFuzz effectively changes machine-classified stance labels, while maintaining semantic integrity with respect to the original content.
Authors: Yuzhe Zhang, Xianwei Xue, Xingyong Wu, Mengke Chen, Chen Liu, Xinran He, Run Shao, Feiran Liu, Huanmin Xu, Qiutong Pan, Haiwei Wang
Abstract: Autonomous GUI agents based on vision-language models (VLMs) often assume deterministic environment responses, generating actions without verifying whether previous operations succeeded. In real-world settings with network latency, rendering delays, and system interruptions, this assumption leads to undetected action failures, repetitive ineffective behaviors, and catastrophic error accumulation. Moreover, learning robust recovery strategies is challenging due to the high cost of online interaction and the lack of real-time feedback in offline datasets.We propose VeriGUI (Verification-driven GUI Agent), which explicitly models action outcomes and recovery under noisy environments. VeriGUI introduces a Thinking--Verification--Action--Expectation (TVAE) framework to detect failures and guide corrective reasoning, and a two-stage training pipeline that combines Robust SFT with synthetic failure trajectories and GRPO with asymmetric verification rewards. We further construct a Robustness Benchmark based on AndroidControl to evaluate failure recognition and correction. Experiments show that VeriGUI significantly reduces failure loops and improves recovery success while maintaining competitive standard task performance.
Authors: Hongcheng Liu, Yuhao Wang, Zhe Chen, Pingjie Wang, Zhiyuan Zhu, Yixuan Hou, Yanfeng Wang, Yu Wang
Abstract: Omni Large Language Models (Omni-LLMs) have demonstrated impressive capabilities in holistic multi-modal perception, yet they consistently falter in complex scenarios requiring synergistic omni-modal reasoning. Beyond understanding global multimodal context, effective reasoning also hinges on fine-grained cross-modal alignment, especially identifying shared referents across modalities, yet this aspect has been largely overlooked. To bridge this gap, we formalize the challenge as a cross-modal coreference problem, where a model must localize a referent in a source modality and re-identify it in a target modality. Building on this paradigm, we introduce CrossOmni, a dataset comprising nine tasks equipped with human-designed reasoning rationales to evaluate and enhance this capability. Experiments on 13 Omni-LLMs reveal systematic weaknesses in cross-modal coreference, which we attribute to the absence of coreference-aware thinking patterns. To address this, we enhance cross-modal alignment via two strategies: a training-free In-Context Learning method and a training-based SFT+GRPO framework designed to induce such thinking patterns. Both approaches yield substantial performance gains and generalize effectively to collaborative reasoning tasks. Overall, our findings highlight cross-modal coreference as a crucial missing piece for advancing robust omni-modal reasoning.
Authors: Zhongxin Yang, Chun Bao, Yuanwei Bin, Xiang I. A. Yang, Shiyi Chen
Abstract: Natural language is a complex system that exhibits robust statistical regularities. Here, we represent text as a trajectory in a high-dimensional embedding space generated by transformer-based language models, and quantify scale-dependent fluctuations along the token sequence using an embedding-step signal. Across multiple languages and corpora, the resulting power spectrum exhibits a robust power law with an exponent close to $5/3$ over an extended frequency range. This scaling is observed consistently in contextual embeddings from both human-written and AI-generated text, but is absent in static word embeddings and is disrupted by randomization of token order. These results show that the observed scaling reflects multiscale, context-dependent organization rather than lexical statistics alone. By analogy with the Kolmogorov spectrum in turbulence, our findings suggest that semantic information is integrated in a scale-free, self-similar manner across linguistic scales, and provide a quantitative, model-agnostic benchmark for studying complex structure in language representations.
Authors: Jinhu Fu, Yan Bai, Longzhu He, Yihang Lou, Yanxiao Zhao, Li Sun, Sen Su
Abstract: Large language models (LLMs) can effectively handle outdated information through knowledge editing. However, current approaches face two key limitations: (I) Poor generalization: Most approaches rigidly inject new knowledge without ensuring that the model can use it effectively to solve practical problems. (II) Narrow scope: Current methods focus primarily on structured fact triples, overlooking the diverse unstructured forms of factual information (e.g., news, articles) prevalent in real-world contexts. To address these challenges, we propose a new paradigm: teaching LLMs to edit knowledge via Chain of Thoughts (CoTs) reasoning (CoT2Edit). We first leverage language model agents for both structured and unstructured edited data to generate CoTs, building high-quality instruction data. The model is then trained to reason over edited knowledge through supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO). At inference time, we integrate Retrieval-Augmented Generation (RAG) to dynamically retrieve relevant edited facts for real-time knowledge editing. Experimental results demonstrate that our method achieves strong generalization across six diverse knowledge editing scenarios with just a single round of training on three open-source language models. The codes are available at https://github.com/FredJDean/CoT2Edit.
Authors: Jun Zhang, Yicheng Ji, Feiyang Ren, Yihang Li, Bowen Zeng, Zonghao Chen, Ke Chen, Lidan Shou, Gang Chen, Huan Li
Abstract: Large Vision-Language Models (LVLMs) enable sophisticated reasoning over images and videos, yet their inference is hindered by a systemic efficiency barrier known as visual token dominance. This overhead is driven by a multi-regime interplay between high-resolution feature extraction, quadratic attention scaling, and memory bandwidth constraints. We present a systematic taxonomy of efficiency techniques structured around the inference lifecycle, consisting of encoding, prefilling, and decoding. Unlike prior reviews focused on isolated optimizations, we analyze the end-to-end pipeline to reveal how upstream decisions dictate downstream bottlenecks, covering compute-bound visual encoding, the intensive prefilling of massive contexts, and the ''visual memory wall'' in bandwidth-bound decoding. By decoupling the efficiency landscape into the axes of shaping information density, managing long-context attention, and overcoming memory limits, this work provides a structured analysis of how isolated optimizations compose to navigate the trade-off between visual fidelity and system efficiency. The survey concludes by outlining four future frontiers supported by pilot empirical insights, including hybrid compression based on functional unit sensitivity, modality-aware decoding with relaxed verification, progressive state management for streaming continuity, and stage-disaggregated serving through hardware-algorithm co-design. The submitted software contains a snapshot of our literature repository, which is designed to be maintained as a living resource for the community.
Authors: Yanxu Mao, Peipei Liu, Tiehan Cui, Congying Liu, Mingzhe Xing, Datao You
Abstract: With the widespread application of LLM-based agents across various domains, their complexity has introduced new security threats. Existing red-team methods mostly rely on modifying user prompts, which lack adaptability to new data and may impact the agent's performance. To address the challenge, this paper proposes the JailAgent framework, which completely avoids modifying the user prompt. Specifically, it implicitly manipulates the agent's reasoning trajectory and memory retrieval with three key stages: Trigger Extraction, Reasoning Hijacking, and Constraint Tightening. Through precise trigger identification, real-time adaptive mechanisms, and an optimized objective function, JailAgent demonstrates outstanding performance in cross-model and cross-scenario environments.
Authors: Yu Li, Chenyang Shao, Xinyang Liu, Ruotong Zhao, Peijie Liu, Hongyuan Su, Zhibin Chen, Qinglong Yang, Anjie Xu, Yi Fang, Qingbin Zeng, Tianxing Li, Jingbo Xu, Fengli Xu, Yong Li, Tie-Yan Liu
Abstract: Artificial intelligence research increasingly depends on prolonged cycles of reproduction, debugging, and iterative refinement to achieve State-Of-The-Art (SOTA) performance, creating a growing need for systems that can accelerate the full pipeline of empirical model optimization. In this work, we introduce AutoSOTA, an end-to-end automated research system that advances the latest SOTA models published in top-tier AI papers to reproducible and empirically improved new SOTA models. We formulate this problem through three tightly coupled stages: resource preparation and goal setting; experiment evaluation; and reflection and ideation. To tackle this problem, AutoSOTA adopts a multi-agent architecture with eight specialized agents that collaboratively ground papers to code and dependencies, initialize and repair execution environments, track long-horizon experiments, generate and schedule optimization ideas, and supervise validity to avoid spurious gains. We evaluate AutoSOTA on recent research papers collected from eight top-tier AI conferences under filters for code availability and execution cost. Across these papers, AutoSOTA achieves strong end-to-end performance in both automated replication and subsequent optimization. Specifically, it successfully discovers 105 new SOTA models that surpass the original reported methods, averaging approximately five hours per paper. Case studies spanning LLM, NLP, computer vision, time series, and optimization further show that the system can move beyond routine hyperparameter tuning to identify architectural innovation, algorithmic redesigns, and workflow-level improvements. These results suggest that end-to-end research automation can serve not only as a performance optimizer, but also as a new form of research infrastructure that reduces repetitive experimental burden and helps redirect human attention toward higher-level scientific creativity.
Authors: Dat Nguyen-Cong, Tung Kieu, Hoang Thanh-Tung
Abstract: Self-conditioning has been central to the success of continuous diffusion language models, as it allows models to correct previous errors. Yet its ability degrades precisely in the regime where diffusion is most attractive for deployment: few-step sampling for fast inference. In this study, we show that when models only have a few denoising steps, inaccurate self-conditioning induces a substantial approximation gap; this mistake compounds across denoising steps and ultimately dominate the sample quality. To address this, we propose a novel training framework that handles these errors during learning by perturbing the self-conditioning signal to match inference noise, improving robustness to prior estimation errors. In addition, we introduce a token-level noise-awareness mechanism that prevents training from saturation, hence improving optimization. Extensive experiments across conditional generation benchmarks demonstrate that our framework surpasses standard continuous diffusion models while providing up to 400x faster inference speed, and remains competitive against other one-step diffusion frameworks.
Authors: Junan Hu, Shudan Guo, Wenqi Liu, Jianhua Yin, Yinwei Wei
Abstract: Large Language Models demonstrate outstanding performance in many language tasks but still face fundamental challenges in managing the non-linear flow of human conversation. The prevalent approach of treating dialogue history as a flat, linear sequence is misaligned with the intrinsically hierarchical and branching structure of natural discourse, leading to inefficient context utilization and a loss of coherence during extended interactions involving topic shifts or instruction refinements. To address this limitation, we introduce Context-Agent, a novel framework that models multi-turn dialogue history as a dynamic tree structure. This approach mirrors the inherent non-linearity of conversation, enabling the model to maintain and navigate multiple dialogue branches corresponding to different topics. Furthermore, to facilitate robust evaluation, we introduce the Non-linear Task Multi-turn Dialogue (NTM) benchmark, specifically designed to assess model performance in long-horizon, non-linear scenarios. Our experiments demonstrate that Context-Agent enhances task completion rates and improves token efficiency across various LLMs, underscoring the value of structured context management for complex, dynamic dialogues. The dataset and code is available at GitHub.
Authors: Xuan Dong, Huanyang Zheng, Tianhao Niu, Zhe Han, Pengzhan Li, Bofei Liu, Zhengyang Liu, Guancheng Li, Qingfu Zhu, Wanxiang Che
Abstract: Scientific research follows multi-turn, multi-step workflows that require proactively searching the literature, consulting figures and tables, and integrating evidence across papers to align experimental settings and support reproducible conclusions. This joint capability is not systematically assessed in existing benchmarks, which largely under-evaluate proactive search, multi-evidence integration and sustained evidence use over time. In this work, we introduce EpiBench, an episodic multi-turn multimodal benchmark that instantiates short research workflows. Given a research task, agents must navigate across papers over multiple turns, align evidence from figures and tables, and use the accumulated evidence in the memory to answer objective questions that require cross paper comparisons and multi-figure integration. EpiBench introduces a process-level evaluation framework for fine-grained testing and diagnosis of research agents. Our experiments show that even the leading model achieves an accuracy of only 29.23% on the hard split, indicating substantial room for improvement in multi-turn, multi-evidence research workflows, providing an evaluation platform for verifiable and reproducible research agents.
Authors: Luc Pommeret (STL), Thibault Wagret (ENS de Lyon, HiSoMA), Jules Deret
Abstract: We describe THIVLVC, a two-stage system for the EvaLatin 2026 Dependency Parsing task. Given a Latin sentence, we retrieve structurally similar entries from the CIRCSE treebank using sentence length and POS n-gram similarity, then prompt a large language model to refine the baseline parse from UDPipe using the retrieved examples and UD annotation guidelines. We submit two configurations: one without retrieval and one with retrieval (RAG). On poetry (Seneca), THIVLVC improves CLAS by +17 points over the UDPipe baseline; on prose (Thomas Aquinas), the gain is +1.5 CLAS. A double-blind error analysis of 300 divergences between our system and the gold standard reveals that, among unanimous annotator decisions, 53.3% favour THIVLVC, showing annotation inconsistencies both within and across treebanks.
Authors: Peace Busola Falola, Jesujoba O. Alabi, Solomon O. Akinola, Folashade T. Ogunajo, Emmanuel Oluwadunsin Alabi, David Ifeoluwa Adelani
Abstract: Named Entity Recognition (NER) is a foundational NLP task, yet research in Yor\`ub\'a has been constrained by limited and domain-specific resources. Existing resources, such as MasakhaNER (a manually annotated news-domain corpus) and WikiAnn (automatically created from Wikipedia), are valuable but restricted in domain coverage. To address this gap, we present YoNER, a new multidomain Yor\`ub\'a NER dataset that extends entity coverage beyond news and Wikipedia. The dataset comprises about 5,000 sentences and 100,000 tokens collected from five domains including Bible, Blogs, Movies, Radio broadcast and Wikipedia, and annotated with three entity types: Person (PER), Organization (ORG) and Location (LOC), following CoNLL-style guidelines. Annotation was conducted manually by three native Yor\`ub\'a speakers, with an inter-annotator agreement of over 0.70, ensuring high quality and consistency. We benchmark several transformer encoder models using cross-domain experiments with MasakhaNER 2.0, and we also assess the effect of few-shot in-domain data using YoNER and cross-lingual setups with English datasets. Our results show that African-centric models outperform general multilingual models for Yor\`ub\'a, but cross-domain performance drops substantially, particularly for blogs and movie domains. Furthermore, we observed that closely related formal domains, such as news and Wikipedia, transfer more effectively. In addition, we introduce a new Yor\`ub\'a-specific language model (OyoBERT) that outperforms multilingual models in in-domain evaluation. We publicly release the YoNER dataset and pretrained OyoBERT models to support future research on Yor\`ub\'a natural language processing.
Authors: Hongyuan Yuan, Xinran He, Run Shao, Bolei He, Xianwei Xue, Mengke Chen, Qiutong Pan, Haiwei Wang, Haifeng Li
Abstract: Extending CoT through RL has been widely used to enhance the reasoning capabilities of LLMs. However, due to the sparsity of reward signals, it can also induce undesirable thinking patterns such as overthinking, i.e., generating redundant intermediate reasoning content. In this work, we argue that a major source of such redundancy is inefficient reflection, which often manifests in two problematic patterns: Indiscriminate Reflection, where the model performs broad, low-impact checks throughout reasoning, and Repetitive Reflection, where it repeatedly re-verifies an already established conclusion. To address this, we introduce a graph-based CoT optimization framework. Specifically, we convert each linear CoT into a directed acyclic graph (DAG) with explicit dependency edges, and design a dual pruning strategy: branch-level pruning removes weakly contributing reflection branches, while depth-level pruning eliminates late-stage re-verification. We distill this behavior via a three-stage pipeline: (1) SFT to initialize the policy on pruned concise traces, (2) DPO to prefer correct but less redundant trajectories, and (3) GRPO with length penalty to jointly optimize answer correctness and efficiency. Experiments show that our approach reduces the average reasoning tokens by 42\% while maintaining or improving accuracy.
Authors: Yicheng Ji, Jun Zhang, Jinpeng Chen, Cong Wang, Lidan Shou, Gang Chen, Huan Li
Abstract: Video Large Language Models (Video-LLMs) excel in video understanding but suffer from high inference latency during autoregressive generation. Speculative Decoding (SD) mitigates this by applying a draft-and-verify paradigm, yet existing methods are constrained by rigid exact-match rules, severely limiting the acceleration potential. To bridge this gap, we propose LVSpec, the first training-free loosely SD framework tailored for Video-LLMs. Grounded in the insight that generation is governed by sparse visual-relevant anchors (mandating strictness) amidst abundant visual-irrelevant fillers (permitting loose verification), LVSpec employs a lightweight visual-relevant token identification scheme to accurately pinpoint the former. To further maximize acceptance, we augment this with a position-shift tolerant mechanism that effectively salvages positionally mismatched but semantically equivalent tokens. Experiments demonstrate that LVSpec achieves high fidelity and speed: it preserves >99.8 of target performance while accelerating Qwen2.5-VL-32B by 2.70x and LLaVA-OneVision-72B by 2.94x. Notably, it boosts the mean accepted length and speedup ratio by 136% and 35% compared to SOTA training-free SD methods for Video-LLMs.
Authors: Lihao Sun, Hang Dong, Bo Qiao, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan
Abstract: This work characterizes large language models' chain-of-thought generation as a structured trajectory through representation space. We show that mathematical reasoning traverses functionally ordered, step-specific subspaces that become increasingly separable with layer depth. This structure already exists in base models, while reasoning training primarily accelerates convergence toward termination-related subspaces rather than introducing new representational organization. While early reasoning steps follow similar trajectories, correct and incorrect solutions diverge systematically at late stages. This late-stage divergence enables mid-reasoning prediction of final-answer correctness with ROC-AUC up to 0.87. Furthermore, we introduce trajectory-based steering, an inference-time intervention framework that enables reasoning correction and length control based on derived ideal trajectories. Together, these results establish reasoning trajectories as a geometric lens for interpreting, predicting, and controlling LLM reasoning behavior.
Authors: Zhen Cheng, Hao-Bo Yang, Wan-Yi Huang, Jin-Long Li
Abstract: Key-Value (KV) cache memory and bandwidth increasingly dominate large language model inference cost in long-context and long-generation regimes. Architectures such as multi-head latent attention (MLA) and hybrid sliding-window attention (SWA) can alleviate this bound, but integrating them into existing models remains difficult. Prior methods impose fine-grained structural requirements on both source and target attention modules, which cannot meet the feasible requirement in practical deployment. We present Attention Editing, a practical framework for converting already-trained large language models (LLMs) with new attention architectures without re-pretraining from scratch. Attention editing replaces the original attention with a learnable target module and trains it using progressive distillation, consisting of (1) layer-wise teacher-forced optimization with intermediate activation supervision to prevent cold-start error accumulation, and (2) model-level distillation on next-token distributions, optionally regularized by weak feature matching. We instantiate the framework on two different target--MLA and GateSWA, a gated hybrid SWA design, and apply it to Qwen3-8B and Qwen3-30B-A3B. The resulting models maintain competitive performance while delivering substantial efficiency improvements, demonstrating that large-scale attention conversion is both feasible and robust. Notably, experiments are conducted on an Ascend 910B clusters, offering a practical training case study on domestic hardware.
Authors: Liqun He (Cindy), Shijun (Cindy), Chen, Mutlu Cukurova, Manolis Mavrikis
Abstract: While generative AI (GenAI) voice chatbots offer scalable opportunities for second language (L2) oral practice, the interactional processes related to learners' gains remain underexplored. This study investigates dialogue act (DA) patterns in interactions between Grade 9 Chinese English as a foreign language (EFL) learners and a GenAI voice chatbot over a 10-week intervention. Seventy sessions from 12 students were annotated by human coders using a pedagogy-informed coding scheme, yielding 6,957 coded DAs. DA distributions and sequential patterns were compared between high- and low-progress sessions. At the DA level, high-progress sessions showed more learner-initiated questions, whereas low-progress sessions exhibited higher rates of clarification-seeking, indicating greater comprehension difficulty. At the sequential level, high-progress sessions were characterised by more frequent prompting-based corrective feedback sequences, consistently positioned after learner responses, highlighting the role of feedback type and timing in effective interaction. Overall, these findings underscore the value of a dialogic lens in GenAI chatbot design, contribute a pedagogy-informed DA coding framework, and inform the design of adaptive GenAI chatbots for L2 education.
Authors: Han Jang, Junhyeok Lee, Heeseong Eum, Kyu Sung Choi
Abstract: Medical Vision-Language Models (Med-VLMs) have achieved expert-level proficiency in interpreting diagnostic imaging. However, current models are predominantly trained on professional literature, limiting their ability to communicate findings in the lay register required for patient-centered care. While text-centric research has actively developed resources for simplifying medical jargon, there is a critical absence of large-scale multimodal benchmarks designed to facilitate lay-accessible medical image understanding. To bridge this resource gap, we introduce MedLayBench-V, the first large-scale multimodal benchmark dedicated to expert-lay semantic alignment. Unlike naive simplification approaches that risk hallucination, our dataset is constructed via a Structured Concept-Grounded Refinement (SCGR) pipeline. This method enforces strict semantic equivalence by integrating Unified Medical Language System (UMLS) Concept Unique Identifiers (CUIs) with micro-level entity constraints. MedLayBench-V provides a verified foundation for training and evaluating next-generation Med-VLMs capable of bridging the communication divide between clinical experts and patients.
Authors: Yanbei Jiang, Amr Keleg, Ryandito Diandaru, Jey Han Lau, Lea Frermann, Biaoyan Fang, Fajri Koto
Abstract: While the real world is inherently stochastic, Large Language Models (LLMs) are predominantly evaluated on single-round inference against fixed ground truths. In this work, we shift the lens to distribution alignment: assessing whether LLMs, when prompted repeatedly, can generate outputs that adhere to a desired target distribution, e.g. reflecting real-world statistics or a uniform distribution. We formulate distribution alignment using the attributes of gender, race, and sentiment within occupational contexts. Our empirical analysis reveals that off-the-shelf LLMs and standard alignment techniques, including prompt engineering and Direct Preference Optimization, fail to reliably control output distributions. To bridge this gap, we propose a novel fine-tuning framework that couples Steering Token Calibration with Semantic Alignment. We introduce a hybrid objective function combining Kullback-Leibler divergence to anchor the probability mass of latent steering tokens and Kahneman-Tversky Optimization to bind these tokens to semantically consistent responses. Experiments across six diverse datasets demonstrate that our approach significantly outperforms baselines, achieving precise distributional control in attribute generation tasks.
Authors: Shintaro Ozaki, Wataru Hashimoto, Hidetaka Kamigaito, Katsuhiko Hayashi, Taro Watanabe
Abstract: While large language models (LLMs) improve performance by explicit reasoning, their responses are often overconfident, even though they include linguistic expressions demonstrating uncertainty. In this work, we identify what linguistic expressions are related to confidence by applying the regression method. Specifically, we predict confidence of those linguistic expressions in the reasoning parts of LLMs as the dependent variables and analyze the relationship between a specific $n$-gram and confidence. Across multiple models and QA benchmarks, we show that LLMs remain overconfident when reasoning is involved and attribute this behavior to specific linguistic information. Interestingly, several of the extracted expressions coincide with cue phrases intentionally inserted on test-time scaling to improve reasoning performance. Through our test on causality and verification that the extracted linguistic information truly affects confidence, we reveal that confidence calibration is possible by simply suppressing those overconfident expressions without drops in performance.
Authors: Yusen Hou, Weicai Long, Haitao Hu, Houcheng Su, Junning Feng, Yanlin Zhang
Abstract: Bacteriophages, often referred to as the dark matter of the biosphere, play a critical role in regulating microbial ecosystems and in antibiotic alternatives. Thus, accurate interpretation of their genomes holds significant scientific and practical value. While general-purpose Large Language Models (LLMs) excel at understanding biological texts, their ability to directly interpret raw nucleotide sequences and perform biological reasoning remains underexplored. To address this, we introduce PhageBench, the first benchmark designed to evaluate phage genome understanding by mirroring the workflow of bioinformatics experts. The dataset contains 5,600 high-quality samples covering five core tasks across three stages: Screening, Quality Control, and Phenotype Annotation. Our evaluation of eight LLMs reveals that general-purpose reasoning models significantly outperform random baselines in phage contig identification and host prediction, demonstrating promising potential for genomic understanding. However, they exhibit significant limitations in complex reasoning tasks involving long-range dependencies and fine-grained functional localization. These findings highlight the necessity of developing next-generation models with enhanced reasoning capabilities for biological sequences.
Authors: Joosung Lee, Hwiyeol Jo, Donghyeon Ko, Kyubyung Chae, Cheonbok Park, Jeonghoon Kim
Abstract: While large language models (LLMs) demonstrate strong capabilities across diverse user queries, they still suffer from hallucinations, often arising from knowledge misalignment between pre-training and fine-tuning. To address this misalignment, we reliably estimate a fine-grained, instance-level knowledge score via multi-sampled inference. Using the knowledge score, we scale the learning signal according to the model's existing knowledge, while encouraging explicit "I don't know" responses for out-of-scope queries. Experimental results show that this approach allows the model to explicitly express uncertainty when it lacks knowledge, while maintaining accuracy on questions it can answer. Furthermore, we propose evaluation metrics for uncertainty, showing that accurate discrimination between known and unknown instances consistently improves performance.
Authors: Abdullah Mazhar, Het Riteshkumar Shah, Aseem Srivastava, Smriti Joshi, Md Shad Akhtar
Abstract: The increasing use of large language models in mental health applications calls for principled evaluation frameworks that assess alignment with psychotherapeutic best practices beyond surface-level fluency. While recent systems exhibit conversational competence, they lack structured mechanisms to evaluate adherence to core therapeutic principles. In this paper, we study the problem of evaluating AI-generated therapist-like responses for clinically grounded appropriateness and effectiveness. We assess each therapists utterance along six therapeutic principles: non-judgmental acceptance, warmth, respect for autonomy, active listening, reflective understanding, and situational appropriateness using a fine-grained ordinal scale. We introduce FAITH-M, a benchmark annotated with expert-assigned ordinal ratings, and propose CARE, a multi-stage evaluation framework that integrates intra-dialogue context, contrastive exemplar retrieval, and knowledge-distilled chain-of-thought reasoning. Experiments show that CARE achieves an F-1 score of 63.34 versus the strong baseline Qwen3 F-1 score of 38.56 which is a 64.26 improvement, which also serves as its backbone, indicating that gains arise from structured reasoning and contextual modeling rather than backbone capacity alone. Expert assessment and external dataset evaluations further demonstrate robustness under domain shift, while highlighting challenges in modelling implicit clinical nuance. Overall, CARE provides a clinically grounded framework for evaluating therapeutic fidelity in AI mental health systems.
Authors: Seungyoon Lee, Minhyuk Kim, Seongtae Hong, Youngjoon Jang, Dongsuk Oh, Heuiseok Lim
Abstract: Existing multilingual embedding models often encounter challenges in cross-lingual scenarios due to imbalanced linguistic resources and less consideration of cross-lingual alignment during training. Although standardized contrastive learning approaches for cross-lingual adaptation are widely adopted, they may struggle to capture fundamental alignment between languages and degrade performance in well-aligned languages such as English. To address these challenges, we propose Cross-Lingual Enhancement in Retrieval via Reverse-training (CLEAR), a novel loss function utilizing a reverse training scheme to improve retrieval performance across diverse cross-lingual retrieval scenarios. CLEAR leverages an English passage as a bridge to strengthen alignments between the target language and English, ensuring robust performance in the cross-lingual retrieval task. Our extensive experiments demonstrate that CLEAR achieves notable improvements in cross-lingual scenarios, with gains up to 15%, particularly in low-resource languages, while minimizing performance degradation in English. Furthermore, our findings highlight that CLEAR offers promising effectiveness even in multilingual training, suggesting its potential for broad application and scalability. We release the code at https://github.com/dltmddbs100/CLEAR.
Authors: Fernando L\'opez, Paula Delgado-Santos, Pablo G\'omez, David Solans, Jordi Luque
Abstract: Voice-based interfaces are widely used; however, achieving fair Wake-up Word detection across diverse speaker populations remains a critical challenge due to persistent demographic biases. This study evaluates the effectiveness of demographics-agnostic training techniques in mitigating performance disparities among speakers of varying sex, age, and accent. We utilize the OK Aura database for our experiments, employing a training methodology that excludes demographic labels, which are reserved for evaluation purposes. We explore (i) data augmentation techniques to enhance model generalization and (ii) knowledge distillation of pre-trained foundational speech models. The experimental results indicate that these demographics-agnostic training techniques markedly reduce demographic bias, leading to a more equitable performance profile across different speaker groups. Specifically, one of the evaluated techniques achieves a Predictive Disparity reduction of 39.94\% for sex, 83.65\% for age, and 40.48\% for accent when compared to the baseline. This study highlights the effectiveness of label-agnostic methodologies in fostering fairness in Wake-up Word detection.
Authors: Yuanfu Sun, Kang Li, Dongzhe Fan, Jiajin Liu, Qiaoyu Tan
Abstract: Large Language Models (LLMs) increasingly rely on agentic capabilities-iterative retrieval, tool use, and decision-making-to overcome the limits of static, parametric knowledge. Yet existing agentic frameworks treat external information as unstructured text and fail to leverage the topological dependencies inherent in real-world data. To bridge this gap, we introduce Agentic Graph Learning (AGL), a paradigm that reframes graph learning as an interleaved process of topology-aware navigation and LLM-based inference. Specifically, we propose AgentGL, the first reinforcement learning (RL)-driven framework for AGL. AgentGL equips an LLM agent with graph-native tools for multi-scale exploration, regulates tool usage via search-constrained thinking to balance accuracy and efficiency, and employs a graph-conditioned curriculum RL strategy to stabilize long-horizon policy learning without step-wise supervision. Across diverse Text-Attributed Graph (TAG) benchmarks and multiple LLM backbones, AgentGL substantially outperforms strong GraphLLMs and GraphRAG baselines, achieving absolute improvements of up to 17.5% in node classification and 28.4% in link prediction. These results demonstrate that AGL is a promising frontier for enabling LLMs to autonomously navigate and reason over complex relational environments. The code is publicly available at https://github.com/sunyuanfu/AgentGL.
Authors: Junsoo Park, Youssef Medhat, Htet Phyo Wai, Ploy Thajchayapong, Ashok K. Goel
Abstract: Learner representations play a central role in educational AI systems, yet it is often unclear whether they preserve meaningful differences between students when instructional outcomes are unavailable or highly context-dependent. This work examines how to evaluate learner representations based on whether they retain separation between learners under a shared comparison rule. We introduce distinctiveness, a representation-level measure that evaluates how each learner differs from others in the cohort using pairwise distances, without requiring clustering, labels, or task-specific evaluation. Using student-authored questions collected through a conversational AI agent in an online learning environment, we compare representations based on individual questions with representations that aggregate patterns across a student's interactions over time. Results show that learner-level representations yield higher separation, stronger clustering structure, and more reliable pairwise discrimination than interaction-level representations. These findings demonstrate that learner representations can be evaluated independently of instructional outcomes and provide a practical pre-deployment criterion using distinctiveness as a diagnostic metric for assessing whether a representation supports differentiated modeling or personalization.
Authors: Xiao Qin, Xingyi Song, Tong Liu, Hatim Laalej, Zepeng Liu, Yunpeng Zhu, Ligang He
Abstract: We present LoRM (Language of Rotating Machinery), a self-supervised framework for multi-modal rotating-machinery signal understanding and real-time condition monitoring. LoRM is built on the idea that rotating-machinery signals can be viewed as a machine language: local signals can be tokenised into discrete symbolic units, and their future evolution can be predicted from observed multi-sensor context. Unlike conventional signal-processing methods that rely on hand-crafted transforms and features, LoRM reformulates multi-modal sensor data as a token-based sequence-prediction problem. For each data window, the observed context segment is retained in continuous form, while the future target segment of each sensing channel is quantised into a discrete token. Then, efficient knowledge transfer is achieved by partially fine-tuning a general-purpose pre-trained language model on industrial signals, avoiding the need to train a large model from scratch. Finally, condition monitoring is performed by tracking token-prediction errors as a health indicator, where increasing errors indicate degradation. In-situ tool condition monitoring (TCM) experiments demonstrate stable real-time tracking and strong cross-tool generalisation, showing that LoRM provides a practical bridge between language modelling and industrial signal analysis. The source code is publicly available at https://github.com/Q159753258/LormPHM.
Authors: Xiangming Gu, Soham De, Larisa Markeeva, Petar Veli\v{c}kovi\'c, Razvan Pascanu
Abstract: Large Reasoning Models (LRMs) have shown remarkable performance on challenging questions, such as math and coding. However, to obtain a high quality solution, one may need to sample more than once. In principal, there are two sampling strategies that can be composed to form more complex processes: sequential sampling and parallel sampling. In this paper, we first compare these two approaches with rigor, and observe, aligned with previous works, that parallel sampling seems to outperform sequential sampling even though the latter should have more representation power. To understand the underline reasons, we make three hypothesis on the reason behind this behavior: (i) parallel sampling outperforms due to the aggregator operator; (ii) sequential sampling is harmed by needing to use longer contexts; (iii) sequential sampling leads to less exploration due to conditioning on previous answers. The empirical evidence on various model families and sizes (Qwen3, DeepSeek-R1 distilled models, Gemini 2.5) and question domains (math and coding) suggests that the aggregation and context length do not seem to be the main culprit behind the performance gap. In contrast, the lack of exploration seems to play a considerably larger role, and we argue that this is one main cause for the performance gap.
Authors: Tianyi Zhao, Yinhan He, Wendy Zheng, Chen Chen
Abstract: Deploying Large Language Models (LLMs) in real-world dynamic environments raises the challenge of updating their pre-trained knowledge. While existing knowledge editing methods can reliably patch isolated facts, they frequently suffer from a "Reasoning Gap", where the model recalls the edited fact but fails to utilize it in multi-step reasoning chains. To bridge this gap, we introduce MCircKE (\underline{M}echanistic \underline{Circ}uit-based \underline{K}nowledge \underline{E}diting), a novel framework that enables a precise "map-and-adapt" editing procedure. MCircKE first identifies the causal circuits responsible for a specific reasoning task, capturing both the storage of the fact and the routing of its logical consequences. It then surgically update parameters exclusively within this mapped circuit. Extensive experiments on the MQuAKE-3K benchmark demonstrate the effectiveness of the proposed method for multi-hop reasoning in knowledge editing.
Authors: Cherifa Ben Khelil, Jean-Yves Antoine, Ana\"is Halftermeyer, Fr\'ed\'eric Rayar, Mathieu Thebaud
Abstract: In this paper, we introduce the French-YMCA corpus, a new linguistic resource specifically tailored for children and adolescents. The motivation for building this corpus is clear: children have unique language requirements, as their language skills are in constant evolution and differ from those of adults. With an extensive collection of 39,200 text files, the French-YMCA corpus encompasses a total of 22,471,898 words. It distinguishes itself through its diverse sources, consistent grammar and spelling, and the commitment to providing open online accessibility for all. Such corpus can serve as the foundation for training language models that understand and anticipate youth's language, thereby enhancing the quality of digital interactions and ensuring that responses and suggestions are age-appropriate and adapted to the comprehension level of users of this age.
Authors: Michael Krumdick, Varshini Reddy, Shivani Chaudhary, William Day, Maarij Ahmed, Hayan Haqqi, Muhammad Ahsen Fahim, Hanzallah Amjad, Ahmad Orakzai, Aqsa Gul, Chris Tanner
Abstract: As concerns surrounding AI-driven labor displacement intensify in knowledge-intensive sectors, existing benchmarks fail to measure performance on tasks that define practical professional expertise. Finance, in particular, has been identified as a domain with high AI exposure risk, yet lacks robust benchmarks to track real-world developments. This gap is compounded by the absence of clear accountability mechanisms in current Large Language Model (LLM) deployments. To address this, we introduce FrontierFinance, a long-horizon benchmark of 25 complex financial modeling tasks across five core finance models, requiring an average of over 18 hours of skilled human labor per task to complete. Developed with financial professionals, the benchmark reflects industry-standard financial modeling workflows and is paired with detailed rubrics for structured evaluation. We engage human experts to define the tasks, create rubrics, grade LLMs, and perform the tasks themselves as human baselines. We demonstrate that our human experts both receive higher scores on average, and are more likely to provide client-ready outputs than current state-of-the-art systems.
Authors: Naen Xu, Jiayi Sheng, Changjiang Li, Chunyi Zhou, Yuyuan Li, Tianyu Du, Jun Wang, Zhihui Fu, Jinbao Li, Shouling Ji
Abstract: Puns are a common form of rhetorical wordplay that exploits polysemy and phonetic similarity to create humor. In multimodal puns, visual and textual elements synergize to ground the literal sense and evoke the figurative meaning simultaneously. Although Vision-Language Models (VLMs) are widely used in multimodal understanding and generation, their ability to understand puns has not been systematically studied due to a scarcity of rigorous benchmarks. To address this, we first propose a multimodal pun generation pipeline. We then introduce MultiPun, a dataset comprising diverse types of puns alongside adversarial non-pun distractors. Our evaluation reveals that most models struggle to distinguish genuine puns from these distractors. Moreover, we propose both prompt-level and model-level strategies to enhance pun comprehension, with an average improvement of 16.5% in F1 scores. Our findings provide valuable insights for developing future VLMs that master the subtleties of human-like humor via cross-modal reasoning.
Authors: Abbas Ghaddar, Ivan Kobyzev, Boxing Chen, Yufei Cui
Abstract: Post-training hybridization of large language models (LLMs) often replaces quadratic self-attention with sliding-window attention (SWA) to reduce KV cache usage and improve latency. Existing hybridization schemes are typically defined either at the layer level (e.g., interleaving) or at the head level via static rankings from local to global. Layer-level schemes ignore that local and global dependencies are routed through heads within the same layer, while static head-level rankings suffer from entanglement: a head's local/global behavior can change after hybridization. We propose BOSCH, Black-box Binary Optimization for Short-context Head Selection, a training-free method that formulates the problem as a Large Neighborhood Search and decomposes it into three subproblems: (i) layer-importance detection via small-budget black-box probes, (ii) adaptive per-layer SWA-ratio assignment based on these sensitivities, and (iii) grouped head-level optimization within ratio buckets. Extensive experiments on 4 LLMs ranging from 1.7B to 30B parameters, across 4 SWA ratios, show that BOSCH consistently outperforms layer-level heuristics and 6 strong static head-level methods, with larger gains at higher SWA ratios. Under continual pretraining, BOSCH recover original long-context performance faster and to a higher level. Analysis of the selected heads reveals substantial turnover for BOSCH across different SWA ratios, underscoring the importance of performing head-level selection for each target ratio rather than relying on fixed locality rankings.
Authors: Fan Zhang (Steve), Mingzi Song (Steve), Rania Elbadry (Steve), Yankai Chen (Steve), Shaobo Wang (Steve), Yixi Zhou (Steve), Xunwen Zheng (Steve), Yueru He (Steve), Yuyang Dai (Steve), Georgi Georgiev (Steve), Ayesha Gull (Steve), Muhammad Usman Safder (Steve), Fan Wu (Steve), Liyuan Meng (Steve), Fengxian Ji (Steve), Junning Zhao (Steve), Xueqing Peng (Steve), Jimin Huang (Steve), Yu Chen (Steve), Xue (Steve), Liu, Preslav Nakov, Zhuohan Xie
Abstract: Financial reporting systems increasingly use large language models (LLMs) to extract and summarize corporate disclosures. However, most assume a single-market setting and do not address structural differences across jurisdictions. Variations in accounting taxonomies, tagging infrastructures (e.g., XBRL vs. PDF), and aggregation conventions make cross-jurisdiction reporting a semantic alignment and verification challenge. We present FinReporting, an agentic workflow for localized cross-jurisdiction financial reporting. The system builds a unified canonical ontology over Income Statement, Balance Sheet, and Cash Flow, and decomposes reporting into auditable stages including filing acquisition, extraction, canonical mapping, and anomaly logging. Rather than using LLMs as free-form generators, FinReporting deploys them as constrained verifiers under explicit decision rules and evidence grounding. Evaluated on annual filings from the US, Japan, and China, the system improves consistency and reliability under heterogeneous reporting regimes. We release an interactive demo supporting cross-market inspection and structured export of localized financial statements. Our demo is available at https://huggingface.co/spaces/BoomQ/FinReporting-Demo . The video describing our system is available at https://www.youtube.com/watch?v=f65jdEL31Kk
URLs: https://huggingface.co/spaces/BoomQ/FinReporting-Demo, https://www.youtube.com/watch?v=f65jdEL31Kk
Authors: Xiaojie Gu, Ziying Huang, Weicong Hong, Jian Xie, Renze Lou, Kai Zhang
Abstract: Large Language Models (LLMs) internalize vast world knowledge as parametric memory, yet inevitably inherit the staleness and errors of their source corpora. Consequently, ensuring the reliability and malleability of these internal representations is imperative for trustworthy real-world deployment. Knowledge editing offers a pivotal paradigm for surgically modifying memory without retraining. However, while recent editors demonstrate high success rates on standard benchmarks, it remains questionable whether current evaluation frameworks that rely on assessing output under specific prompting conditions can reliably authenticate genuine memory modification. In this work, we introduce a simple diagnostic framework that subjects models to discriminative self-assessment under in-context learning (ICL) settings that better reflect real-world application environments, specifically designed to scrutinize the subtle behavioral nuances induced by memory modifications. This probing reveals a pervasive phenomenon of Surface Compliance, where editors achieve high benchmark scores by merely mimicking target outputs without structurally overwriting internal beliefs. Moreover, we find that recursive modifications accumulate representational residues, triggering cognitive instability and permanently diminishing the reversibility of the model's memory state. These insights underscore the risks of current editing paradigms and highlight the pivotal role of robust memory modification in building trustworthy, long-term sustainable LLM systems. Code is available at https://github.com/XiaojieGu/SA-MCQ.
Authors: Asaf Avrahamy, Yoav Gur-Arieh, Mor Geva
Abstract: Interpreting the information encoded in model weights remains a fundamental challenge in mechanistic interpretability. In this work, we introduce ROTATE (Rotation-Optimized Token Alignment in weighT spacE), a data-free method requiring no forward passes that disentangles MLP neurons directly in weight space. Our approach relies on a key statistical observation: neurons that encode coherent, monosemantic concepts exhibit high kurtosis when projected onto the model's vocabulary. By optimizing rotations of neuron weights to maximize their vocabulary-space kurtosis, our method recovers sparse, interpretable directions which we name vocabulary channels. Experiments on Llama-3.1-8B-Instruct and Gemma-2-2B-it demonstrate that ROTATE consistently recovers vocabulary channels that are faithful to the neuron's behavior. ablating individual channels selectively disables corresponding input activations or the promotion of specific concepts. Moreover, aggregating channel-level descriptions yields comprehensive neuron descriptions that outperform optimized activation-based baselines by 2-3x in head-to-head comparisons. By providing a data-free decomposition of neuron weights, ROTATE offers a scalable, fine-grained building block for interpreting LMs.
Authors: Zhongxing Zhang, Emily K. Vraga, Jisu Huh, Jaideep Srivastava
Abstract: Incorrect information poses significant challenges by disrupting content veracity and integrity, yet most detection approaches struggle to jointly balance textual content verification with external knowledge modification under collapsed attention geometries. To address this issue, we propose a dual-head reasoning framework, BiMind, which disentangles content-internal reasoning from knowledge-augmented reasoning. In BiMind, we introduce three core innovations: (i) an attention geometry adapter that reshapes attention logits via token-conditioned offsets and mitigates attention collapse; (ii) a self-retrieval knowledge mechanism, which constructs an in-domain semantic memory through kNN retrieval and injects retrieved neighbors via feature-wise linear modulation; (iii) the uncertainty-aware fusion strategies, including entropy-gated fusion and a trainable agreement head, stabilized by a symmetric Kullback-Leibler agreement regularizer. To quantify the knowledge contributions, we define a novel metric, Value-of-eXperience (VoX), to measure instance-wise logit gains from knowledge-augmented reasoning. Experiment results on public datasets demonstrate that our BiMind model outperforms advanced detection approaches and provides interpretable diagnostics on when and why knowledge matters.
Authors: Maria Mahbub, Gregory M. Dams, Josh Arnold, Caitlin Rizy, Sudarshan Srinivasan, Elliot M. Fielstein, Minu A. Aghevli, Kamonica L. Craig, Elizabeth M. Oliva, Joseph Erdos, Jodie Trafton, Ioana Danciu
Abstract: Large language models (LLMs) show promise for extracting clinically meaningful information from unstructured health records, yet their translation into real-world settings is constrained by the lack of scalable and trustworthy validation approaches. Conventional evaluation methods rely heavily on annotation-intensive reference standards or incomplete structured data, limiting feasibility at population scale. We propose a multi-stage validation framework for LLM-based clinical information extraction that enables rigorous assessment under weak supervision. The framework integrates prompt calibration, rule-based plausibility filtering, semantic grounding assessment, targeted confirmatory evaluation using an independent higher-capacity judge LLM, selective expert review, and external predictive validity analysis to quantify uncertainty and characterize error modes without exhaustive manual annotation. We applied this framework to extraction of substance use disorder (SUD) diagnoses across 11 substance categories from 919,783 clinical notes. Rule-based filtering and semantic grounding removed 14.59% of LLM-positive extractions that were unsupported, irrelevant, or structurally implausible. For high-uncertainty cases, the judge LLM's assessments showed substantial agreement with subject matter expert review (Gwet's AC1=0.80). Using judge-evaluated outputs as references, the primary LLM achieved an F1 score of 0.80 under relaxed matching criteria. LLM-extracted SUD diagnoses also predicted subsequent engagement in SUD specialty care more accurately than structured-data baselines (AUC=0.80). These findings demonstrate that scalable, trustworthy deployment of LLM-based clinical information extraction is feasible without annotation-intensive evaluation.
Authors: Hongxu Zhou
Abstract: Intrinsic self-correction in Large Language Models (LLMs) frequently fails in open-ended reasoning tasks due to ``hallucination snowballing,'' a phenomenon in which models recursively justify early errors during free-text reflection. While structured feedback can mitigate this issue, existing approaches often rely on externally trained critics or symbolic tools, reducing agent autonomy. This study investigates whether enforcing structured reflection purely through Outlines-based constrained decoding can disrupt error propagation without additional training. Evaluating an 8-billion-parameter model (Qwen3-8B), we show that simply imposing structural constraints does not improve self-correction performance. Instead, it triggers a new failure mode termed ``structure snowballing.'' We find that the cognitive load required to satisfy strict formatting rules pushes the model into formatting traps. This observation helps explain why the agent achieves near-perfect superficial syntactic alignment yet fails to detect or resolve deeper semantic errors. These findings expose an ``alignment tax'' inherent to constrained decoding, highlighting a tension between structural granularity and internal model capacity in autonomous workflows. Code and raw logs are available in the GitHub repository: https://github.com/hongxuzhou/agentic_llm_structured_self_critique.
URLs: https://github.com/hongxuzhou/agentic_llm_structured_self_critique.
Authors: Patrick Huber, Ernie Chang, Chinnadhurai Sankar, Rylan Conway, Igor Fedorov, Md Rifat Arefin, Adithya Sagar
Abstract: Extending the context window of language models typically requires expensive long-context pre-training, posing significant challenges for both training efficiency and data collection. In this paper, we present evidence that long-context retrieval capabilities can be transferred to student models through logit-based knowledge distillation, even when training exclusively on packed short-context samples within a long-context window. We provide comprehensive insights through the lens of Rotary Position Embedding (RoPE) and establish three key findings. First, consistent with prior work, we show that phase-wise RoPE scaling, which maximizes rotational spectrum utilization at each training stage, also achieves the best long-context performance in knowledge distillation setups. Second, we demonstrate that logit-based knowledge distillation can directly enable positional information transfer. Using an experimental setup with packed repeated token sequences, we trace the propagation of positional perturbations from query and key vectors through successive transformer layers to output logits, revealing that positional information systematically influences the teacher's output distribution and, in turn, the distillation signal received by the student model. Third, our analysis uncovers structured update patterns in the query state during long-context extension, with distinct parameter spans exhibiting strong sensitivity to long-context training.
Authors: Ben Wigler, Maria Tsfasman, Tiffany Matej Hrkalovic
Abstract: Personality traits are richly encoded in natural language, and large language models (LLMs) trained on human text can simulate personality when conditioned on persona descriptions. However, existing evaluations rely predominantly on questionnaire self-report by the conditioned model, are limited in architectural diversity, and rarely use real human psychometric data. Without addressing these limitations, it remains unclear whether personality conditioning produces psychometrically informative representations of individual differences or merely superficial alignment with trait descriptors. To test how robustly LLMs can encode personality into extended text, we condition LLMs on real psychometric profiles from 290 participants to generate first-person life story narratives, and then task independent LLMs to recover personality scores from those narratives alone. We show that personality scores can be recovered from the generated narratives at levels approaching human test-retest reliability (mean r = 0.750, 85% of the human ceiling), and that recovery is robust across 10 LLM narrative generators and 3 LLM personality scorers spanning 6 providers. Decomposing systematic biases reveals that scoring models achieve their accuracy while counteracting alignment-induced defaults. Content analysis of the generated narratives shows that personality conditioning produces behaviourally differentiated text: nine of ten coded features correlate significantly with the same features in participants' real conversations, and personality-driven emotional reactivity patterns in narratives replicate in real conversational data. These findings provide evidence that the personality-language relationship captured during pretraining supports robust encoding and decoding of individual differences, including characteristic emotional variability patterns that replicate in real human behaviour.
Authors: Olexander Mazurets, Olexander Barmak, Leonid Bedratyuk, Iurii Krak
Abstract: Modern Transformer-based language models achieve strong performance in natural language processing tasks, yet their latent semantic spaces remain largely uninterpretable black boxes. This paper introduces LAG-XAI (Lie Affine Geometry for Explainable AI), a novel geometric framework that models paraphrasing not as discrete word substitutions, but as a structured affine transformation within the embedding space. By conceptualizing paraphrasing as a continuous geometric flow on a semantic manifold, we propose a computationally efficient mean-field approximation, inspired by local Lie group actions. This allows us to decompose paraphrase transitions into geometrically interpretable components: rotation, deformation, and translation. Experiments on the noisy PIT-2015 Twitter corpus, encoded with Sentence-BERT, reveal a "linear transparency" phenomenon. The proposed affine operator achieves an AUC of 0.7713. By normalizing against random chance (AUC 0.5), the model captures approximately 80% of the non-linear baseline's effective classification capacity (AUC 0.8405), offering explicit parametric interpretability in exchange for a marginal drop in absolute accuracy. The model identifies fundamental geometric invariants, including a stable matrix reconfiguration angle (~27.84{\deg}) and near-zero deformation, indicating local isometry. Cross-domain generalization is confirmed via direct cross-corpus validation on an independent TURL dataset. Furthermore, the practical utility of LAG-XAI is demonstrated in LLM hallucination detection: using a "cheap geometric check," the model automatically detected 95.3% of factual distortions on the HaluEval dataset by registering deviations beyond the permissible semantic corridor. This approach provides a mathematically grounded, resource-efficient path toward the mechanistic interpretability of Transformers.
Authors: Changgeon Ko, Jisu Shin, Hoyun Song, Huije Lee, Eui Jun Hwang, Jong C. Park
Abstract: Large language model (LLM) agents are increasingly acting as human delegates in multi-agent environments, where a representative agent integrates diverse peer perspectives to make a final decision. Drawing inspiration from social psychology, we investigate how the reliability of this representative agent is undermined by the social context of its network. We define four key phenomena-social conformity, perceived expertise, dominant speaker effect, and rhetorical persuasion-and systematically manipulate the number of adversaries, relative intelligence, argument length, and argumentative styles. Our experiments demonstrate that the representative agent's accuracy consistently declines as social pressure increases: larger adversarial groups, more capable peers, and longer arguments all lead to significant performance degradation. Furthermore, rhetorical strategies emphasizing credibility or logic can further sway the agent's judgment, depending on the context. These findings reveal that multi-agent systems are sensitive not only to individual reasoning but also to the social dynamics of their configuration, highlighting critical vulnerabilities in AI delegates that mirror the psychological biases observed in human group decision-making.
Authors: Mutsumi Sasaki, Kouta Nakayama, Yusuke Miyao, Yohei Oseki, Masaru Isonuma
Abstract: When introducing Large Language Models (LLMs) into industrial applications, such as healthcare and education, the risk of generating harmful content becomes a significant challenge. While existing machine unlearning methods can erase specific harmful knowledge and expressions, diverse harmful content makes comprehensive removal difficult. In this study, instead of individually listing targets for forgetting, we propose Exclusive Unlearning (EU), which aims for broad harm removal by extensively forgetting everything except for the knowledge and expressions we wish to retain. We demonstrate that through Exclusive Unlearning, it is possible to obtain a model that ensures safety against a wide range of inputs, including jailbreaks, while maintaining the ability to respond to diverse instructions related to specific domains such as medicine and mathematics.
Authors: Komal Kumar, Aman Chadha, Salman Khan, Fahad Shahbaz Khan, Hisham Cholakkal
Abstract: The rapid growth of scientific literature has made it increasingly difficult for researchers to efficiently discover, evaluate, and synthesize relevant work. Recent advances in multi-agent large language models (LLMs) have demonstrated strong potential for understanding user intent and are being trained to utilize various tools. In this paper, we introduce Paper Circle, a multi-agent research discovery and analysis system designed to reduce the effort required to find, assess, organize, and understand academic literature. The system comprises two complementary pipelines: (1) a Discovery Pipeline that integrates offline and online retrieval from multiple sources, multi-criteria scoring, diversity-aware ranking, and structured outputs; and (2) an Analysis Pipeline that transforms individual papers into structured knowledge graphs with typed nodes such as concepts, methods, experiments, and figures, enabling graph-aware question answering and coverage verification. Both pipelines are implemented within a coder LLM-based multi-agent orchestration framework and produce fully reproducible, synchronized outputs including JSON, CSV, BibTeX, Markdown, and HTML at each agent step. This paper describes the system architecture, agent roles, retrieval and scoring methods, knowledge graph schema, and evaluation interfaces that together form the Paper Circle research workflow. We benchmark Paper Circle on both paper retrieval and paper review generation, reporting hit rate, MRR, and Recall at K. Results show consistent improvements with stronger agent models. We have publicly released the website at https://papercircle.vercel.app/ and the code at https://github.com/MAXNORM8650/papercircle.
URLs: https://papercircle.vercel.app/, https://github.com/MAXNORM8650/papercircle.
Authors: Yuqi Zhou, Sunhao Dai, Changle Qu, Liang Pang, Jun Xu, Ji-Rong Wen
Abstract: Information retrieval (IR) systems have traditionally been designed and trained for human users, with learning-to-rank methods relying heavily on large-scale human interaction logs such as clicks and dwell time. With the rapid emergence of large language model (LLM) powered search agents, however, retrieval is increasingly consumed by agents rather than human beings, and is embedded as a core component within multi-turn reasoning and action loops. In this setting, retrieval models trained under human-centric assumptions exhibit a fundamental mismatch with the way agents issue queries and consume results. In this work, we argue that retrieval models for agentic search should be trained directly from agent interaction data. We introduce learning to retrieve from agent trajectories as a new training paradigm, where supervision is derived from multi-step agent interactions. Through a systematic analysis of search agent trajectories, we identify key behavioral signals that reveal document utility, including browsing actions, unbrowsed rejections, and post-browse reasoning traces. Guided by these insights, we propose LRAT, a simple yet effective framework that mines high-quality retrieval supervision from agent trajectories and incorporates relevance intensity through weighted optimization. Extensive experiments on both in-domain and out-of-domain deep research benchmarks demonstrate that retrievers trained with LRAT consistently improve evidence recall, end-to-end task success, and execution efficiency across diverse agent architectures and scales. Our results highlight agent trajectories as a practical and scalable supervision source, pointing to a promising direction for retrieval in the era of agentic search.
Authors: Ziheng Chen, Jiali Cheng, Zezhong Fan, Hadi Amiri, Yunzhi Yao, Xiangguo Sun, Yang Zhang
Abstract: Recent advances in large language models (LLMs) have opened new opportunities for recommender systems by enabling rich semantic understanding and reasoning about user interests and item attributes. However, as privacy regulations tighten, incorporating user data into LLM-based recommendation (LLMRec) introduces significant privacy risks, making unlearning algorithms increasingly crucial for practical deployment. Despite growing interest in LLMRec unlearning, most existing approaches formulate unlearning as a weighted combination of forgetting and retaining objectives while updating model parameters in a uniform manner. Such formulations inevitably induce gradient conflicts between the two objectives, leading to unstable optimization and resulting in either ineffective unlearning or severe degradation of model utility. Moreover, the unlearning procedure remains largely black-box, undermining its transparency and trustworthiness. To tackle these challenges, we propose CURE, a circuit-aware unlearning framework that disentangles model components into functionally distinct subsets and selectively updates them. Here, a circuit refers to a computational subgraph that is causally responsible for task-specific behaviors. Specifically, we extract the core circuits underlying item recommendation and analyze how individual modules within these circuits contribute to the forget and retain objectives. Based on this analysis, these modules are categorized into forget-specific, retain-specific, and task-shared groups, each subject to function-specific update rules to mitigate gradient conflicts during unlearning. Experiments on real-world datasets show that our approach achieves more effective unlearning than existing baselines.
Authors: Rong Lu, Hao Liu, Song Hou
Abstract: This work presents a comparative analysis of embedding-based and generative models for classifying geoscience technical documents. Using a multi-disciplinary benchmark dataset, we evaluated the trade-offs between model accuracy, stability, and computational cost. We find that generative Vision-Language Models (VLMs) like Qwen2.5-VL, enhanced with Chain-of-Thought (CoT) prompting, achieve superior zero-shot accuracy (82%) compared to state-of-the-art multimodal embedding models like QQMM (63%). We also demonstrate that while supervised fine-tuning (SFT) can improve VLM performance, it is sensitive to training data imbalance.
Authors: Shuzhen Bi, Mingzi Zhang, Zhuoxuan Li, Xiaolong Wang, keqian Li, Aimin Zhou
Abstract: Large language models are increasingly used as educational assistants, yet evaluation of their educational capabilities remains concentrated on question-answering and tutoring tasks. A critical gap exists for multimedia instructional content generation -- the ability to produce coherent, diagram-rich explanations that combine geometrically accurate visuals with step-by-step reasoning. We present EduIllustrate, a benchmark for evaluating LLMs on interleaved text-diagram explanation generation for K-12 STEM problems. The benchmark comprises 230 problems spanning five subjects and three grade levels, a standardized generation protocol with sequential anchoring to enforce cross-diagram visual consistency, and an 8-dimension evaluation rubric grounded in multimedia learning theory covering both text and visual quality. Evaluation of ten LLMs reveals a wide performance spread: Gemini 3.0 Pro Preview leads at 87.8\%, while Kimi-K2.5 achieves the best cost-efficiency (80.8\% at \\$0.12/problem). Workflow ablation confirms sequential anchoring improves Visual Consistency by 13\% at 94\% lower cost. Human evaluation with 20 expert raters validates LLM-as-judge reliability for objective dimensions ($\rho \geq 0.83$) while revealing limitations on subjective visual assessment.
Authors: Frazier N. Baker, Trieu Nguyen, Reza Averly, Botao Yu, Daniel Adu-Ampratwum, Huan Sun, Xia Ning
Abstract: Multi-objective retrosynthesis planning is a critical chemistry task requiring dynamic balancing of quality, safety, and cost objectives. Language model-based multi-agent systems (MAS) offer a promising approach for this task: leveraging interactions of specialized agents to incorporate multiple objectives into retrosynthesis planning. We present MMORF, a framework for constructing MAS for multi-objective retrosynthesis planning. MMORF features modular agentic components, which can be flexibly combined and configured into different systems, enabling principled evaluation and comparison of different system designs. Using MMORF, we construct two representative MAS: MASIL and RFAS. On a newly curated benchmark consisting of 218 multi-objective retrosynthesis planning tasks, MASIL achieves strong safety and cost metrics on soft-constraint tasks, frequently Pareto-dominating baseline routes, while RFAS achieves a 48.6% success rate on hard-constraint tasks, outperforming state-of-the-art baselines. Together, these results show the effectiveness of MMORF as a foundational framework for exploring MAS for multi-objective retrosynthesis planning. Code and data are available at https://anonymous.4open.science/r/MMORF/.
Authors: Yuxuan Zhang, EunJeong Hwang, Huaisong Zhang, Penghui Du, Yiming Jia, Dongfu Jiang, Xuan He, Shenhui Zhang, Ping Nie, Peter West, Kelsey R. Allen
Abstract: It is critical for vision-language models (VLMs) to comprehensively understand visual, temporal, and textual cues. However, despite rapid progress in multimodal modeling, video understanding performance still lags behind text-based reasoning. In this work, we find that progress is even worse than previously assumed: commonly reported long video understanding benchmarks contain 40-60% of questions that can be answered using text cues alone. Furthermore, we find that these issues are also pervasive in widely used post-training datasets, potentially undercutting the ability of post-training to improve VLM video understanding performance. Guided by this observation, we introduce VidGround as a simple yet effective solution: using only the actual visually grounded questions without any linguistic biases for post-training. When used in tandem with RL-based post-training algorithms, this simple technique improves performance by up to 6.2 points relative to using the full dataset, while using only 69.1% of the original post-training data. Moreover, we show that data curation with a simple post-training algorithm outperforms several more complex post-training techniques, highlighting that data quality is a major bottleneck for improving video understanding in VLMs. These results underscore the importance of curating post-training data and evaluation benchmarks that truly require visual grounding to advance the development of more capable VLMs. Project page: http://vidground.etuagi.com.
Authors: Ruslan Sharifullin, Maxim Gorshkov, Hannah Clay
Abstract: Prior authorization (PA) requires interpretation of complex and fragmented coverage policies, yet existing retrieval-augmented systems rely on static top-$K$ strategies with fixed numbers of retrieved sections. Such fixed retrieval can be inefficient and gather irrelevant or insufficient information. We model policy retrieval for PA as a sequential decision-making problem, formulating adaptive retrieval as a Markov Decision Process (MDP). In our system, an agent iteratively selects policy chunks from a top-$K$ candidate set or chooses to stop and issue a decision. The reward balances decision correctness against retrieval cost, capturing the trade-off between accuracy and efficiency. We train policies using Conservative Q-Learning (CQL), Implicit Q-Learning (IQL), and Direct Preference Optimization (DPO) in an offline RL setting on logged trajectories generated from baseline retrieval strategies over synthetic PA requests derived from publicly available CMS coverage data. On a corpus of 186 policy chunks spanning 10 CMS procedures, CQL achieves 92% decision accuracy (+30 percentage points over the best fixed-$K$ baseline) via exhaustive retrieval, while IQL matches the best baseline accuracy using 44% fewer retrieval steps and achieves the only positive episodic return among all policies. Transition-level DPO matches CQL's 92% accuracy while using 47% fewer retrieval steps (10.6 vs. 20.0), occupying a "selective-accurate" region on the Pareto frontier that dominates both CQL and BC. A behavioral cloning baseline matches CQL, confirming that advantage-weighted or preference-based policy extraction is needed to learn selective retrieval. Lambda ablation over step costs $\lambda \in \{0.05, 0.1, 0.2\}$ reveals a clear accuracy-efficiency inflection: only at $\lambda = 0.2$ does CQL transition from exhaustive to selective retrieval.
Authors: Samira Hajizadeh, Suman Jana
Abstract: Large language models (LLMs) often generate code that is functionally correct but inefficient in runtime and memory. Prior approaches to improving code efficiency typically rely on absolute execution feedback, such as profiling a single program's runtime or memory usage, which is costly and provides weak guidance for refinement. We propose Relative Contrastive Feedback (RCF), an inference-time feedback mechanism that requires no model fine-tuning or parameter updates. RCF compares two structurally similar programs for the same task and highlights the differences associated with better efficiency. Building on this idea, we introduce EffiPair, an inference-time iterative refinement framework that operates entirely at test time by generating multiple candidate solutions, identifying informative program pairs with large efficiency gaps, summarizing their execution differences into lightweight feedback, and using this signal to produce more efficient solutions. By replacing isolated scalar feedback with pairwise contrastive comparisons, EffiPair provides more direct guidance while reducing profiling and prompting overhead. Experiments on code-efficiency benchmarks show that EffiPair consistently improves efficiency while preserving correctness. For instance, with DeepSeek-Chat V3.2, EffiPair achieves up to 1.5x speedup over generation without performance feedback, while reducing token usage by more than 90% compared to prior work.
Authors: Alfonso Amayuelas, Firas Laakom, Piotr Pi\k{e}kos, Wenyi Wang, Yifan Xu, Yuhui Wang, J\"urgen Schmidhuber, William Wang
Abstract: The use of LLMs for code generation has naturally extended to code testing and evaluation. As codebases grow in size and complexity, so does the need for automated test generation. Current approaches for LLM-based test generation rely on strategies that maximize immediate coverage gain, a greedy approach that plateaus on code where reaching deep branches requires setup steps that individually yield zero new coverage. Drawing on principles of Bayesian exploration, we treat the program's branch structure as an unknown environment, and an evolving coverage map as a proxy probabilistic posterior representing what the LLM has discovered so far. Our method, CovQValue, feeds the coverage map back to the LLM, generates diverse candidate plans in parallel, and selects the most informative plan by LLM-estimated Q-values, seeking actions that balance immediate branch discovery with future reachability. Our method outperforms greedy selection on TestGenEval Lite, achieving 51-77% higher branch coverage across three popular LLMs and winning on 77-84% of targets. In addition, we build a benchmark for iterative test generation, RepoExploreBench, where they achieve 40-74%. These results show the potential of curiosity-driven planning methods for LLM-based exploration, enabling more effective discovery of program behavior through sequential interaction
Authors: Giansalvo Cirrincione
Abstract: Neural language models process sequences of words, but the mathematical operations inside them are insensitive to the order in which words appear. Positional encodings are the component added to remedy this. Despite their importance, positional encodings have been designed largely by trial and error, without a mathematical theory of what they ought to do. This paper develops such a theory. Four results are established. First, any Transformer without a positional signal cannot solve any task sensitive to word order (Necessity Theorem). Second, training assigns distinct vector representations to distinct sequence positions at every global minimiser, under mild and verifiable conditions (Positional Separation Theorem). Third, the best achievable approximation to an information-optimal encoding is constructed via classical multidimensional scaling (MDS) on the Hellinger distance between positional distributions; the quality of any encoding is measured by a single number, the stress (Proposition 5, Algorithm 1). Fourth, the optimal encoding has effective rank r = rank(B) <= n-1 and can be represented with r(n+d) parameters instead of nd (minimal parametrisation result). Appendix A develops a proof of the Monotonicity Conjecture within the Neural Tangent Kernel (NTK) regime for masked language modelling (MLM) losses, sequence classification losses, and general losses satisfying a positional sufficiency condition, through five lemmas. Experiments on SST-2 and IMDB with BERT-base confirm the theoretical predictions and reveal that Attention with Linear Biases (ALiBi) achieves much lower stress than the sinusoidal encoding and Rotary Position Embedding (RoPE), consistent with a rank-1 interpretation of the MDS encoding under approximate shift-equivariance.
Authors: Yi Ru Wang, Carter Ung, Evan Gubarev, Christopher Tan, Siddhartha Srinivasa, Dieter Fox
Abstract: Evaluation of robotic manipulation systems has largely relied on fixed benchmarks authored by a small number of experts, where task instances, constraints, and success criteria are predefined and difficult to extend. This paradigm limits who can shape evaluation and obscures how policies respond to user-authored variations in task intent, constraints, and notions of success. We argue that evaluating modern manipulation policies requires reframing evaluation as a language-driven process over structured physical domains. We present RoboPlayground, a framework that enables users to author executable manipulation tasks using natural language within a structured physical domain. Natural language instructions are compiled into reproducible task specifications with explicit asset definitions, initialization distributions, and success predicates. Each instruction defines a structured family of related tasks, enabling controlled semantic and behavioral variation while preserving executability and comparability. We instantiate RoboPlayground in a structured block manipulation domain and evaluate it along three axes. A user study shows that the language-driven interface is easier to use and imposes lower cognitive workload than programming-based and code-assist baselines. Evaluating learned policies on language-defined task families reveals generalization failures that are not apparent under fixed benchmark evaluations. Finally, we show that task diversity scales with contributor diversity rather than task count alone, enabling evaluation spaces to grow continuously through crowd-authored contributions. Project Page: https://roboplayground.github.io
Authors: Satyam Goyal, Anirudh Kanchi, Garv Shah, Prakhar Gupta
Abstract: Large Language Models (LLMs) are typically static after training, yet real-world applications require continual adaptation to new knowledge without degrading existing capabilities. Standard approaches to updating models, like full finetuning or parameter-efficient methods (e.g., LoRA), face a fundamental trade-off: catastrophic forgetting. They modify shared dense representations, causing interference across tasks. Sparse Memory Finetuning (SMF) offers a promising alternative by localizing updates to a small subset of parameters in explicit memory layers. In this work, we present an open-source pipeline to retrofit existing pretrained models (Qwen-2.5-0.5B) with sparse memory modules, enabling effective continual learning on consumer hardware. We extend prior work by introducing a theoretically grounded slot-selection mechanism based on Kullback-Leibler (KL) divergence, which prioritizes memory updates for informationally "surprising" tokens relative to a background distribution. Our experiments demonstrate that our retrofitted models can acquire new factual knowledge with minimal forgetting of held-out capabilities, validating the sparse update hypothesis in a practical setting.
Authors: Satyam Goyal, Kushal Patel, Tanush Mittal, Arjun Laxman
Abstract: Masked Diffusion Models (MDMs) offer a promising alternative to autoregressive language models by enabling parallel token generation and bidirectional context modeling. However, their inference speed is significantly limited by the inability to cache key-value pairs due to bidirectional attention, requiring $O(N^2)$ computations at each generation step. While recent methods like FastDLLM and DkvCache improve inference speed through attention approximations and caching strategies, they achieve speedups at the cost of generation quality. We propose DualDiffusion, a speculative decoding framework for MDMs that combines fast drafter models (using efficient approximations) with slower, more accurate verifier models. By running multiple steps of a lightweight drafter followed by a single verification step, DualDiffusion achieves a superior Pareto frontier between generation steps and accuracy compared to existing approaches. We evaluate our method on MMLU and GSM8K, demonstrating that DualDiffusion maintains high accuracy while reducing the number of generation steps required, effectively pushing the quality-efficiency trade-off curve for masked diffusion language models.
Authors: Chan-Wei Hu, Zhengzhong Tu
Abstract: Multi-modal retrieval-augmented generation (MM-RAG) relies heavily on re-rankers to surface the most relevant evidence for image-question queries. However, standard re-rankers typically process the full query image as a global embedding, making them susceptible to visual distractors (e.g., background clutter) that skew similarity scores. We propose Region-R1, a query-side region cropping framework that formulates region selection as a decision-making problem during re-ranking, allowing the system to learn to retain the full image or focus only on a question-relevant region before scoring the retrieved candidates. Region-R1 learns a policy with a novel region-aware group relative policy optimization (r-GRPO) to dynamically crop a discriminative region. Across two challenging benchmarks, E-VQA and InfoSeek, Region-R1 delivers consistent gains, achieving state-of-the-art performances by increasing conditional Recall@1 by up to 20%. These results show the great promise of query-side adaptation as a simple but effective way to strengthen MM-RAG re-ranking.
Authors: Junyu Guo, Shangding Gu, Ming Jin, Costas Spanos, Javad Lavaei
Abstract: Large language models are increasingly used in settings where uncertainty must drive decisions such as abstention, retrieval, and verification. Most existing methods treat uncertainty as a latent quantity to estimate after generation rather than a signal the model is trained to express. We instead study uncertainty as an interface for control. We compare two complementary interfaces: a global interface, where the model verbalizes a calibrated confidence score for its final answer, and a local interface, where the model emits an explicit
Authors: Xuan Xiong, Huan Liu, Li Gu, Zhixiang Chi, Yue Qiu, Yuanhao Yu, Yang Wang
Abstract: Chain-of-thought (CoT) reasoning improves large language model performance on complex tasks, but often produces excessively long and inefficient reasoning traces. Existing methods shorten CoTs using length penalties or global entropy reduction, implicitly assuming that low uncertainty is desirable throughout reasoning. We show instead that reasoning efficiency is governed by the trajectory of uncertainty. CoTs with dominant downward entropy trends are substantially shorter. Motivated by this insight, we propose Entropy Trend Reward (ETR), a trajectory-aware objective that encourages progressive uncertainty reduction while allowing limited local exploration. We integrate ETR into Group Relative Policy Optimization (GRPO) and evaluate it across multiple reasoning models and challenging benchmarks. ETR consistently achieves a superior accuracy-efficiency tradeoff, improving DeepSeek-R1-Distill-7B by 9.9% in accuracy while reducing CoT length by 67% across four benchmarks. Code is available at https://github.com/Xuan1030/ETR
Authors: Xing Tang, Hao Chen, Shiwei Li, Fuyuan Lyu, Weijie Shi, Lingjie Li, Dugang Liu, Weihong Luo, Xiku Du, Xiuqiang He
Abstract: Large language models (LLMs) have been incorporated into numerous industrial applications. Meanwhile, a vast array of API assets is scattered across various functions in the financial domain. An online financial question-answering system can leverage both LLMs and private APIs to provide timely financial analysis and information. The key is equipping the LLM model with function calling capability tailored to a financial scenario. However, a generic LLM requires customized financial APIs to call and struggles to adapt to the financial domain. Additionally, online user queries are diverse and contain out-of-distribution parameters compared with the required function input parameters, which makes it more difficult for a generic LLM to serve online users. In this paper, we propose a data-driven pipeline to enhance function calling in LLM for our online, deployed financial QA, comprising dataset construction, data augmentation, and model training. Specifically, we construct a dataset based on a previous study and update it periodically, incorporating queries and an augmentation method named AugFC. The addition of user query-related samples will \textit{exploit} our financial toolset in a data-driven manner, and AugFC explores the possible parameter values to enhance the diversity of our updated dataset. Then, we train an LLM with a two-step method, which enables the use of our financial functions. Extensive experiments on existing offline datasets, as well as the deployment of an online scenario, illustrate the superiority of our pipeline. The related pipeline has been adopted in the financial QA of YuanBao\footnote{https://yuanbao.tencent.com/chat/}, one of the largest chat platforms in China.
Authors: Siyuan Cheng, Bozhong Tian, YanChao Hao, Zheng Wei
Abstract: PRISM-MCTS: Learning from Reasoning Trajectories with Metacognitive Reflection Siyuan Cheng, Bozhong Tian, Yanchao Hao, Zheng Wei Published: 06 Apr 2026, Last Modified: 06 Apr 2026 ACL 2026 Findings Conference, Area Chairs, Reviewers, Publication Chairs, Authors Revisions BibTeX CC BY 4.0 Keywords: Efficient/Low-Resource Methods for NLP, Generation, Question Answering Abstract: The emergence of reasoning models, exemplified by OpenAI o1, signifies a transition from intuitive to deliberative cognition, effectively reorienting the scaling laws from pre-training paradigms toward test-time computation. While Monte Carlo Tree Search (MCTS) has shown promise in this domain, existing approaches typically treat each rollout as an isolated trajectory. This lack of information sharing leads to severe inefficiency and substantial computational redundancy, as the search process fails to leverage insights from prior explorations. To address these limitations, we propose PRISM-MCTS, a novel reasoning framework that draws inspiration from human parallel thinking and reflective processes. PRISM-MCTS integrates a Process Reward Model (PRM) with a dynamic shared memory, capturing both "Heuristics" and "Fallacies". By reinforcing successful strategies and pruning error-prone branches, PRISM-MCTS effectively achieves refinement. Furthermore, we develop a data-efficient training strategy for the PRM, achieving high-fidelity evaluation under a few-shot regime. Empirical evaluations across diverse reasoning benchmarks substantiate the efficacy of PRISM-MCTS. Notably, it halves the trajectory requirements on GPQA while surpassing MCTS-RAG and Search-o1, demonstrating that it scales inference by reasoning judiciously rather than exhaustively.
Authors: Tinko Sebastian Bartels, Ruixiang Wu, Xinyu Lu, Yikai Lu, Fanzeng Xia, Haoxiang Yang, Yue Chen, Tongxin Li
Abstract: Addressing the critical need for intelligent, context-aware energy management in renewable systems, we introduce the OpenCEM Simulator and Dataset: the first open-source digital twin explicitly designed to integrate rich, unstructured contextual information with quantitative renewable energy dynamics. Traditional energy management relies heavily on numerical time series, thereby neglecting the significant predictive power embedded in human-generated context (e.g., event schedules, system logs, user intentions). OpenCEM bridges this gap by offering a unique platform comprising both a meticulously aligned, language-rich dataset from a real-world PV-and-battery microgrid installation and a modular simulator capable of natively processing this multi-modal context. The OpenCEM Simulator provides a high-fidelity environment for developing and validating novel control algorithms and prediction models, particularly those leveraging Large Language Models. We detail its component-based architecture, hybrid data-driven and physics-based modelling capabilities, and demonstrate its utility through practical examples, including context-aware load forecasting and the implementation of online optimal battery charging control strategies. By making this platform publicly available, OpenCEM aims to accelerate research into the next generation of intelligent, sustainable, and truly context-aware energy systems.
Authors: Yasuto Hoshi, Daisuke Miyashita, Jun Deguchi
Abstract: Long-context generation is increasingly limited by decode-time key-value (KV) cache traffic, particularly when KV is offloaded beyond GPU memory. Query-aware retrieval (e.g., Top-K selection) reduces this traffic by loading only a subset of KV pairs, but renormalizing the softmax over the subset introduces bias when attention mass is spread over unretrieved tokens. We propose a retrieval-completion attention module that keeps backbone weights and the KV-cache format unchanged. For each query, we compute exact attention over sink/tail anchors and the query-dependent retrieved Top-K tokens, and estimate the remaining mid-region numerator and denominator using a fixed-size feature-map summary computed at prefill time. We add the exact and estimated contributions in the unnormalized domain and apply a single normalization, recovering the missing softmax mass without additional attention-side KV reads. Across long-context benchmarks, the proposed method improves over selection-only Top-K at matched token-equivalent read budgets, with the largest gains in high-entropy heads.
Authors: Siddharth Jain, Venkat Narayan Vedam
Abstract: As language models shift from single-shot answer generation toward multi-step reasoning that retrieves and consumes evidence mid-inference, evaluating the role of individual retrieved items becomes more important. Existing RAG evaluation typically targets final-answer quality, citation faithfulness, or answer-level attribution, but none of these directly targets the intervention-based, per-evidence-item utility view we study here. We introduce CUE-R, a lightweight intervention-based framework for measuring per-evidence-item operational utility in single-shot RAG using shallow observable retrieval-use traces. CUE-R perturbs individual evidence items via REMOVE, REPLACE, and DUPLICATE operators, then measures changes along three utility axes (correctness, proxy-based grounding faithfulness, and confidence error) plus a trace-divergence signal. We also outline an operational evidence-role taxonomy for interpreting intervention outcomes. Experiments on HotpotQA and 2WikiMultihopQA with Qwen-3 8B and GPT-5.2 reveal a consistent pattern: REMOVE and REPLACE substantially harm correctness and grounding while producing large trace shifts, whereas DUPLICATE is often answer-redundant yet not fully behaviorally neutral. A zero-retrieval control confirms that these effects arise from degradation of meaningful retrieval. A two-support ablation further shows that multi-hop evidence items can interact non-additively: removing both supports harms performance far more than either single removal. Our results suggest that answer-only evaluation misses important evidence effects and that intervention-based utility analysis is a practical complement for RAG evaluation.
Authors: Xiaotian Zhou, Di Tang, Xiaofeng Wang, Xiaozhong Liu
Abstract: Large Language Models (LLMs) have shown a high capability in answering questions on a diverse range of topics. However, these models sometimes produce biased, ideologized or incorrect responses, limiting their applications if there is no clear understanding of which topics their answers can be trusted. In this research, we introduce a novel algorithm, named as GMRL-BD, designed to identify the untrustworthy boundaries (in terms of topics) of a given LLM, with black-box access to the LLM and under specific query constraints. Based on a general Knowledge Graph (KG) derived from Wikipedia, our algorithm incorporates with multiple reinforcement learning agents to efficiently identify topics (some nodes in KG) where the LLM is likely to generate biased answers. Our experiments demonstrated the efficiency of our algorithm, which can detect the untrustworthy boundary with just limited queries to the LLM. Additionally, we have released a new dataset containing popular LLMs including Llama2, Vicuna, Falcon, Qwen2, Gemma2 and Yi-1.5, along with labels indicating the topics on which each LLM is likely to be biased.
Authors: N. D. Tantaroudas, A. J. McCracken, I. Karachalios, E. Papatheou
Abstract: This work introduces a modular platform that brings together six AI services, automatic speech recognition via OpenAI Whisper, multilingual translation through Meta NLLB, speech synthesis using AWS Polly, emotion classification with RoBERTa, dialogue summarisation via flan t5 base samsum, and International Sign (IS) rendering through Google MediaPipe. A corpus of IS gesture recordings was processed to derive hand landmark coordinates, which were subsequently mapped onto three dimensional avatar animations inside a virtual reality (VR) environment. Validation comprised technical benchmarking of each AI component, including comparative assessments of speech synthesis providers and multilingual translation models (NLLB 200 and EuroLLM 1.7B variants). Technical evaluations confirmed the suitability of the platform for real time XR deployment. Speech synthesis benchmarking established that AWS Polly delivers the lowest latency at a competitive price point. The EuroLLM 1.7B Instruct variant attained a higher BLEU score, surpassing NLLB. These findings establish the viability of orchestrating cross modal AI services within XR settings for accessible, multilingual language instruction. The modular design permits independent scaling and adaptation to varied educational contexts, providing a foundation for equitable learning solutions aligned with European Union digital accessibility goals.
Authors: Xin Sun, Di Wu, Sijing Qin, Isao Echizen, Abdallah El Ali, Saku Sugawara
Abstract: Large language models (LLMs) are increasingly used as automated evaluators (LLM-as-a-Judge). This work challenges its reliability by showing that trust judgments by LLMs are biased by disclosed source labels. Using a counterfactual design, we find that both humans and LLM judges assign higher trust to information labeled as human-authored than to the same content labeled as AI-generated. Eye-tracking data reveal that humans rely heavily on source labels as heuristic cues for judgments. We analyze LLM internal states during judgment. Across label conditions, models allocate denser attention to the label region than the content region, and this label dominance is stronger under Human labels than AI labels, consistent with the human gaze patterns. Besides, decision uncertainty measured by logits is higher under AI labels than Human labels. These results indicate that the source label is a salient heuristic cue for both humans and LLMs. It raises validity concerns for label-sensitive LLM-as-a-Judge evaluation, and we cautiously raise that aligning models with human preferences may propagate human heuristic reliance into models, motivating debiased evaluation and alignment.
Authors: Nikolaos D. Tantaroudas, Andrew J. McCracken, Ilias Karachalios, Evangelos Papatheou
Abstract: Video conferencing has become central to professional collaboration, yet most platforms offer limited support for deaf, hard-of-hearing, and multilingual users. The World Health Organisation estimates that over 430 million people worldwide require rehabilitation for disabling hearing loss, a figure projected to exceed 700 million by 2050. Conventional accessibility measures remain constrained by high costs, limited availability, and logistical barriers, while Extended Reality (XR) technologies open new possibilities for immersive and inclusive communication. This paper presents INTERACT (Inclusive Networking for Translation and Embodied Real-Time Augmented Communication Tool), an AI-driven XR platform that integrates real-time speech-to-text conversion, International Sign Language (ISL) rendering through 3D avatars, multilingual translation, and emotion recognition within an immersive virtual environment. Built on the CORTEX2 framework and deployed on Meta Quest 3 headsets, INTERACT combines Whisper for speech recognition, NLLB for multilingual translation, RoBERTa for emotion classification, and Google MediaPipe for gesture extraction. Pilot evaluations were conducted in two phases, first with technical experts from academia and industry, and subsequently with members of the deaf community. The trials reported 92% user satisfaction, transcription accuracy above 85%, and 90% emotion-detection precision, with a mean overall experience rating of 4.6 out of 5.0 and 90% of participants willing to take part in further testing. The results highlight strong potential for advancing accessibility across educational, cultural, and professional settings. An extended version of this work, including full pilot data and implementation details, has been published as an Open Research Europe article [Tantaroudas et al., 2026a].
Authors: Xinran Wang, Yuxuan Zhang, Xiao Zhang, Haolong Yan, Muxi Diao, Songyu Xu, Zhonghao Yan, Hongbing Li, Kongming Liang, Zhanyu Ma
Abstract: Accurately detecting and localizing hallucinations is a critical task for ensuring high reliability of image captions. In the era of Multimodal Large Language Models (MLLMs), captions have evolved from brief sentences into comprehensive narratives, often spanning hundreds of words. This shift exponentially increases the challenge: models must now pinpoint specific erroneous spans or words within extensive contexts, rather than merely flag response-level inconsistencies. However, existing benchmarks lack the fine granularity and domain diversity required to evaluate this capability. To bridge this gap, we introduce DetailVerifyBench, a rigorous benchmark comprising 1,000 high-quality images across five distinct domains. With an average caption length of over 200 words and dense, token-level annotations of multiple hallucination types, it stands as the most challenging benchmark for precise hallucination localization in the field of long image captioning to date. Our benchmark is available at https://zyx-hhnkh.github.io/DetailVerifyBench/.
Authors: Ojas Jain, Dhruv Kumar
Abstract: We introduce LudoBench, a benchmark for evaluating LLM strategic reasoning in Ludo, a stochastic multi-agent board game whose dice mechanics, piece capture, safe-square navigation, and home-path progression introduce meaningful planning complexity. LudoBench comprises 480 handcrafted spot scenarios across 12 behaviorally distinct decision categories, each isolating a specific strategic choice. We additionally contribute a fully functional 4-player Ludo simulator supporting Random, Heuristic, Game-Theory, and LLM agents. The game-theory agent uses Expectiminimax search with depth-limited lookahead to provide a principled strategic ceiling beyond greedy heuristics. Evaluating six models spanning four model families, we find that all models agree with the game-theory baseline only 40-46% of the time. Models split into distinct behavioral archetypes: finishers that complete pieces but neglect development, and builders that develop but never finish. Each archetype captures only half of the game theory strategy. Models also display measurable behavioral shifts under history-conditioned grudge framing on identical board states, revealing prompt-sensitivity as a key vulnerability. LudoBench provides a lightweight and interpretable framework for benchmarking LLM strategic reasoning under uncertainty. All code, the spot dataset (480 entries) and model outputs are available at https://anonymous.4open.science/r/LudoBench-5CBF/
Authors: Guan-Yan Yang, Wei-Ling Wen, Shu-Yuan Ku, Farn Wang, Kuo-Hui Yeh
Abstract: Web applications rely heavily on hyperlinks to connect disparate information resources. However, the dynamic nature of the web leads to link rot, where targets become unavailable, and more insidiously, semantic drift, where a valid HTTP 200 connection exists, but the target content no longer aligns with the source context. Traditional verification tools, which primarily function as crash oracles by checking HTTP status codes, often fail to detect semantic inconsistencies, thereby compromising web integrity and user experience. While Large Language Models (LLMs) offer semantic understanding, they suffer from high latency, privacy concerns, and prohibitive costs for large-scale regression testing. In this paper, we propose SemLink, a novel automated test oracle for semantic hyperlink verification. SemLink leverages a Siamese Neural Network architecture powered by a pre-trained Sentence-BERT (SBERT) backbone to compute the semantic coherence between a hyperlink's source context (anchor text, surrounding DOM elements, and visual features) and its target page content. To train and evaluate our model, we introduce the Hyperlink-Webpage Positive Pairs (HWPPs) dataset, a rigorously constructed corpus of over 60,000 semantic pairs. Our evaluation demonstrates that SemLink achieves a Recall of 96.00%, comparable to state-of-the-art LLMs (GPT-5.2), while operating approximately 47.5 times faster and requiring significantly fewer computational resources. This work bridges the gap between traditional syntactic checkers and expensive generative AI, offering a robust and efficient solution for automated web quality assurance.
Authors: Roni Goldshmidt, Hamish Scott, Lorenzo Niccolini, Hernan Matzner
Abstract: We present BADAS-2.0, the second generation of our collision anticipation system, building on BADAS-1.0 [7], which showed that fine-tuning V-JEPA2 [1] on large-scale ego-centric dashcam data outperforms both academic baselines and production ADAS systems. BADAS-2.0 advances the state of the art along three axes. (i) Long-tail benchmark and accuracy: We introduce a 10-group long-tail benchmark targeting rare and safety-critical scenarios. To construct it, BADAS-1.0 is used as an active oracle to score millions of unlabeled drives and surface high-risk candidates for annotation. Combined with Nexar's Atlas platform [13] for targeted data collection, this expands the dataset from 40k to 178,500 labeled videos (~2M clips), yielding consistent gains across all subgroups, with the largest improvements on the hardest long-tail cases. (ii) Knowledge distillation to edge: Domain-specific self-supervised pre-training on 2.25M unlabeled driving videos enables distillation into compact models, BADAS-2.0-Flash (86M) and BADAS-2.0-Flash-Lite (22M), achieving 7-12x speedup with near-parity accuracy, enabling real-time edge deployment. (iii) Explainability: BADAS-2.0 produces real-time object-centric attention heatmaps that localize the evidence behind predictions. BADAS-Reason [17] extends this with a vision-language model that consumes the last frame and heatmap to generate driver actions and structured textual reasoning. Inference code and evaluation benchmarks are publicly available.
Authors: Weicai Long, Yusen Hou, Junning Feng, Houcheng Su, Shuo Yang, Donglin Xie, Yanlin Zhang
Abstract: Large Language Models (LLMs) are increasingly adopted as conversational assistants in genomics, where they are mainly used to reason over biological knowledge, annotations, and analysis outputs through natural language interfaces. However, existing benchmarks either focus on specialized DNA models trained for sequence prediction or evaluate biological knowledge using text-only questions, leaving the behavior of general-purpose LLMs when directly exposed to raw genome sequences underexplored. We introduce GenomeQA, a benchmark designed to provide a controlled evaluation setting for general-purpose LLMs on sequence-based genome inference tasks. GenomeQA comprises 5,200 samples drawn from multiple biological databases, with sequence lengths ranging from 6 to 1,000 base pairs (bp), spanning six task families: Enhancer and Promoter Identification, Splice Site Identification, Taxonomic Classification, Histone Mark Prediction, Transcription Factor Binding Site Prediction, and TF Motif Prediction. Across six frontier LLMs, we find that models consistently outperform random baselines and can exploit local sequence signals such as GC content and short motifs, while performance degrades on tasks that require more indirect or multi-step inference over sequence patterns. GenomeQA establishes a diagnostic benchmark for studying and improving the use of general-purpose LLMs on raw genomic sequences.
Authors: Yingjian Zhu, Xinming Wang, Kun Ding, Ying Wang, Bin Fan, Shiming Xiang
Abstract: Multi-modal Retrieval-Augmented Generation (RAG) has emerged as a highly effective paradigm for Knowledge-Based Visual Question Answering (KB-VQA). Despite recent advancements, prevailing methods still primarily depend on images as the retrieval key, and often overlook or misplace the role of Vision-Language Models (VLMs), thereby failing to leverage their potential fully. In this paper, we introduce WikiSeeker, a novel multi-modal RAG framework that bridges these gaps by proposing a multi-modal retriever and redefining the role of VLMs. Rather than serving merely as answer generators, we assign VLMs two specialized agents: a Refiner and an Inspector. The Refiner utilizes the capability of VLMs to rewrite the textual query according to the input image, significantly improving the performance of the multimodal retriever. The Inspector facilitates a decoupled generation strategy by selectively routing reliable retrieved context to another LLM for answer generation, while relying on the VLM's internal knowledge when retrieval is unreliable. Extensive experiments on EVQA, InfoSeek, and M2KR demonstrate that WikiSeeker achieves state-of-the-art performance, with substantial improvements in both retrieval accuracy and answer quality. Our code will be released on https://github.com/zhuyjan/WikiSeeker.
Authors: Yicheng Pan, Zhiyuan Ning, Ludi Wang, Yi Du
Abstract: As conference submission volumes continue to grow, accurately recommending suitable reviewers has become a challenge. Most existing methods follow a ``Paper-to-Paper'' matching paradigm, implicitly representing a reviewer by their publication history. However, effective reviewer matching requires capturing multi-dimensional expertise, and textual similarity to past papers alone is often insufficient. To address this gap, we propose P2R, a training-free framework that shifts from implicit paper-to-paper matching to explicit profile-based matching. P2R uses general-purpose LLMs to construct structured profiles for both submissions and reviewers, disentangling them into Topics, Methodologies, and Applications. Building on these profiles, P2R adopts a coarse-to-fine pipeline to balance efficiency and depth. It first performs hybrid retrieval that combines semantic and aspect-level signals to form a high-recall candidate pool, and then applies an LLM-based committee to evaluate candidates under strict rubrics, integrating both multi-dimensional expert views and a holistic Area Chair perspective. Experiments on NeurIPS, SIGIR, and SciRepEval show that P2R consistently outperforms state-of-the-art baselines. Ablation studies further verify the necessity of each component. Overall, P2R highlights the value of explicit, structured expertise modeling and offers practical guidance for applying LLMs to reviewer matching.
Authors: Fatih Uenal
Abstract: The deployment of large language models (LLMs) in Swiss financial and regulatory contexts demands empirical evidence of both production reliability and adversarial security, dimensions not jointly operationalized in existing Swiss-focused evaluation frameworks. This paper introduces Swiss-Bench 003 (SBP-003), extending the HAAS (Helvetic AI Assessment Score) from six to eight dimensions by adding D7 (Self-Graded Reliability Proxy) and D8 (Adversarial Security). I evaluate ten frontier models across 808 Swiss-specific items in four languages (German, French, Italian, English), comprising seven Swiss-adapted benchmarks (Swiss TruthfulQA, Swiss IFEval, Swiss SimpleQA, Swiss NIAH, Swiss PII-Scope, System Prompt Leakage, and Swiss German Comprehension) targeting FINMA Guidance 08/2024, the revised Federal Act on Data Protection (nDSG), and OWASP Top 10 for LLMs. Self-graded D7 scores (73-94%) exceed externally judged D8 security scores (20-61%) by a wide margin, though these dimensions use non-comparable scoring regimes. System prompt leakage resistance ranges from 24.8% to 88.2%, while PII extraction defense remains weak (14-42%) across all models. Qwen 3.5 Plus achieves the highest self-graded D7 score (94.4%), while GPT-oss 120B achieves the highest D8 score (60.7%) despite being the lowest-cost model evaluated. All evaluations are zero-shot under provider default settings; D7 is self-graded and does not constitute independently validated accuracy. I provide conceptual mapping tables relating benchmark dimensions to FINMA model validation requirements, nDSG data protection obligations, and OWASP LLM risk categories.
Authors: Hongxu Zhou
Abstract: State space models (SSMs) have been shown to possess the theoretical capacity to model both star-free sequential tasks and bounded hierarchical structures Sarrof et al. (2024). However, formal expressivity results do not guarantee that gradient-based optimisation will reliably discover the corresponding solutions. Existing benchmarks probe either monotonic state tracking, as in the standard Flip-Flop task, or structural nesting, as in the Dyck languages, but neither isolates reversible semantic state retrieval. We introduce the UNDO Flip-Flop task to fill this gap. By extending the standard Flip-Flop with an UNDO, the task requires a model to maintain an implicit bounded stack and recover historical states under non-monotonic update sequences. We evaluate one-layer and two-layer Mamba-2 under this framework. Both variants fail to acquire the provably expressible stack-based rollback mechanism, converging instead on a local toggle heuristic that inverts the current state rather than retrieving stored history. Under an adversarial retraction pressure test held within the training length distribution, the two-layer model collapses to 41.10% accuracy, which is below random chance. The results confirm systematic rather than incidental failure. Causal ablation shows that the bottleneck lies in retrieval, not storage. These results draw a clear line between what an architecture can in principle represent and what gradient descent reliably learns, a distinction that theoretical expressivity analyses alone cannot capture.
Authors: Yi Yuan, Xuhong Wang, Shanzhe Lei
Abstract: As agent-based systems continue to evolve, deep research agents are capable of automatically generating research-style reports across diverse domains. While these agents promise to streamline information synthesis and knowledge exploration, existing evaluation frameworks-typically based on subjective dimensions-fail to capture a critical aspect of report quality: trustworthiness. In open-ended research scenarios where ground-truth answers are unavailable, current evaluation methods cannot effectively measure the epistemic confidence of generated content, making calibration difficult and leaving users susceptible to misleading or hallucinated information. To address this limitation, we propose a novel deep research agent that incorporates progressive confidence estimation and calibration within the report generation pipeline. Our system leverages a deliberative search model, featuring deep retrieval and multi-hop reasoning to ground outputs in verifiable evidence while assigning confidence scores to individual claims. Combined with a carefully designed workflow, this approach produces trustworthy reports with enhanced transparency. Experimental results and case studies demonstrate that our method substantially improves interpretability and significantly increases user trust.
Authors: Oscar Chew, Hsiao-Ying Huang, Kunal Jain, Tai-I Chen, Khoa D Doan, Kuan-Hao Huang
Abstract: Recent research has shown that contrastive vision-language models such as CLIP often lack fine-grained understanding of visual content. While a growing body of work has sought to address this limitation, we identify a distinct failure mode in the CLIP family, which we term center bias, that persists even in recent model variants. Specifically, CLIP tends to disproportionately focus on the central region of an image, overlooking important objects located near the boundaries. This limitation is fundamental as failure to recognize relevant objects makes it difficult to perform any sophisticated tasks that depend on those objects. To understand the underlying causes of the limitation, we conduct analyses from both representation and attention perspectives. Using interpretability methods, i.e., embedding decomposition and attention map analysis, we find that relevant concepts especially those associated with off-center objects vanish from the model's embedding in the final representation due to information loss during the aggregation of visual embeddings, particularly the reliance on pooling mechanisms. Finally, we show that this bias can be alleviated with training-free strategies such as visual prompting and attention redistribution by redirecting models' attention to off-center regions.
Authors: Shuqing Zhao
Abstract: We present Arch (AI-native Register-transfer Clocked Hardware), a hardware description language designed from first principles for micro-architecture specification and AI-assisted code generation. Arch introduces first-class language constructs for pipelines, FSMs, FIFOs, arbiters, register files, buses, and clock-domain crossings -- structures that existing HDLs express only as user-defined patterns prone to subtle errors.
A central design choice is that clocks and resets are themselves parameterized types (Clock
Authors: Michael Cuccarese
Abstract: This paper presents epistemic blinding in the context of an agentic system that uses large language models to reason across multiple biological datasets for drug target prioritization. During development, it became apparent that LLM outputs silently blend data-driven inference with memorized priors about named entities - and the blend is invisible: there is no way to determine, from a single output, how much came from the data on the page and how much came from the model's training memory. Epistemic blinding is a simple inference-time protocol that replaces entity identifiers with anonymous codes before prompting, then compares outputs against an unblinded control. The protocol does not make LLM reasoning deterministic, but it restores one critical axis of auditability: measuring how much of an output came from the supplied data versus the model's parametric knowledge. The complete target identification system is described - including LLM-guided evolutionary optimization of scoring functions and blinded agentic reasoning for target rationalization - with demonstration that both stages operate without access to entity identity. In oncology drug target prioritization across four cancer types, blinding changes 16% of top-20 predictions while preserving identical recovery of validated targets. The contamination problem is shown to generalize beyond biology: in S&P 500 equity screening, brand-recognition bias reshapes 30-40% of top-20 rankings across five random seeds. To lower the barrier to adoption, the protocol is released as an open-source tool and as a Claude Code skill that enables one-command epistemic blinding within agentic workflows. The claim is not that blinded analysis produces better results, but that without blinding, there is no way to know to what degree the agent is adhering to the analytical process the researcher designed.
Authors: Jayr Pereira, Leandro Fernandes, Erick de Brito, Roberto Lotufo, Luiz Bonifacio
Abstract: Legal information retrieval in Portuguese remains difficult to evaluate systematically because available datasets differ widely in document type, query style, and relevance definition. We present JU\'A, a public benchmark for Brazilian legal retrieval designed to support more reproducible and comparable evaluation across heterogeneous legal collections. More broadly, JU\'A is intended not only as a benchmark, but as a continuous evaluation infrastructure for Brazilian legal IR, combining shared protocols, common ranking metrics, fixed splits when applicable, and a public leaderboard. The benchmark covers jurisprudence retrieval as well as broader legislative, regulatory, and question-driven legal search. We evaluate lexical, dense, and BM25-based reranking pipelines, including a domain-adapted Qwen embedding model fine-tuned on JU\'A-aligned supervision. Results show that the benchmark is sufficiently heterogeneous to distinguish retrieval paradigms and reveal substantial cross-dataset trade-offs. Domain adaptation yields its clearest gains on the supervision-aligned JU\'A-Juris subset, while BM25 remains highly competitive on other collections, especially in settings with strong lexical and institutional phrasing cues. Overall, JU\'A provides a practical evaluation framework for studying legal retrieval across multiple Brazilian legal domains under a common benchmark design.
Authors: Wang Yang, Chaoda Song, Xinpeng Li, Debargha Ganguly, Chuang Ma, Shouren Wang, Zhihao Dou, Yuli Zhou, Vipin Chaudhary, Xiaotian Han
Abstract: Existing Agent benchmarks suffer from two critical limitations: high environment interaction overhead (up to 41\% of total evaluation time) and imbalanced task horizon and difficulty distributions that make aggregate scores unreliable. To address these issues, we propose ACE-Bench built around a unified grid-based planning task, where agents must fill hidden slots in a partially completed schedule subject to both local slot constraints and global constraints. Our benchmark offers fine-grained control through two orthogonal axes: Scalable Horizons, controlled by the number of hidden slots $H$, and Controllable Difficulty, governed by a decoy budget $B$ that determines the number of globally misleading decoy candidates. Crucially, all tool calls are resolved via static JSON files under a Lightweight Environment design, eliminating setup overhead and enabling fast, reproducible evaluation suitable for training-time validation. We first validate that H and B provide reliable control over task horizon and difficulty, and that ACE-Bench exhibits strong domain consistency and model discriminability. We then conduct comprehensive experiments across 13 models of diverse sizes and families over 6 domains, revealing significant cross-model performance variation and confirming that ACE-Bench provides interpretable and controllable evaluation of agent reasoning.
Authors: Qimin Zhong, Hao Liao, Haiming Qin, Mingyang Zhou, Rui Mao, Wei Chen, Naipeng Chao
Abstract: Whether Large Language Models (LLMs) develop coherent internal world models remains a core debate. While conventional Next-Token Prediction (NTP) focuses on one-step-ahead supervision, Multi-Token Prediction (MTP) has shown promise in learning more structured representations. In this work, we provide a theoretical perspective analyzing the gradient inductive bias of MTP, supported by empirical evidence, showing that MTP promotes the convergence toward internal belief states by inducing representational contractivity via gradient coupling. However, we reveal that standard MTP often suffers from structural hallucinations, where discrete token supervision encourages illegal shortcuts in latent space that violate environmental constraints. To address this, we propose a novel method Latent Semantic Enhancement MTP (LSE-MTP), which anchors predictions to ground-truth hidden state trajectories. Experiments on synthetic graphs and real-world Manhattan Taxi Ride show that LSE-MTP effectively bridges the gap between discrete tokens and continuous state representations, enhancing representation alignment, reducing structural hallucinations, and improving robustness to perturbations.
Authors: Yuchi Wang, Haiyang Yu, Weikang Bian, Jiefeng Long, Xiao Liang, Chao Feng, Hongsheng Li
Abstract: MLLMs have been successfully applied to multimodal embedding tasks, yet their generative reasoning capabilities remain underutilized. Directly incorporating chain-of-thought reasoning into embedding learning introduces two fundamental challenges. First, structural misalignment between instance-level reasoning and pairwise contrastive supervision may lead to shortcut behavior, where the model merely learns the superficial format of reasoning. Second, reasoning is not universally beneficial for embedding tasks. Enforcing reasoning for all inputs may introduce unnecessary computation and latency, and can even obscure salient semantic signals for simple cases. To address these issues, we propose MMEmb-R1, an adaptive reasoning-based multimodal embedding framework. We formulate reasoning as a latent variable and introduce pair-aware reasoning selection that employs counterfactual intervention to identify reasoning paths beneficial for query-target alignment. Furthermore, we adopt reinforcement learning to selectively invoke reasoning only when necessary. Experiments on the MMEB-V2 benchmark demonstrate that our model achieves a score of 71.2 with only 4B parameters, establishing a new state-of-the-art while significantly reducing reasoning overhead and inference latency.
Authors: Guhao Feng, Shengjie Luo, Kai Hua, Ge Zhang, Di He, Wenhao Huang, Tianle Cai
Abstract: The static ``train then deploy" paradigm fundamentally limits Large Language Models (LLMs) from dynamically adapting their weights in response to continuous streams of new information inherent in real-world tasks. Test-Time Training (TTT) offers a compelling alternative by updating a subset of model parameters (fast weights) at inference time, yet its potential in the current LLM ecosystem is hindered by critical barriers including architectural incompatibility, computational inefficiency and misaligned fast weight objectives for language modeling. In this work, we introduce In-Place Test-Time Training (In-Place TTT), a framework that seamlessly endows LLMs with Test-Time Training ability. In-Place TTT treats the final projection matrix of the ubiquitous MLP blocks as its adaptable fast weights, enabling a ``drop-in" enhancement for LLMs without costly retraining from scratch. Furthermore, we replace TTT's generic reconstruction objective with a tailored, theoretically-grounded objective explicitly aligned with the Next-Token-Prediction task governing autoregressive language modeling. This principled objective, combined with an efficient chunk-wise update mechanism, results in a highly scalable algorithm compatible with context parallelism. Extensive experiments validate our framework's effectiveness: as an in-place enhancement, it enables a 4B-parameter model to achieve superior performance on tasks with contexts up to 128k, and when pretrained from scratch, it consistently outperforms competitive TTT-related approaches. Ablation study results further provide deeper insights on our design choices. Collectively, our results establish In-Place TTT as a promising step towards a paradigm of continual learning in LLMs.
Authors: Valeria Pastorino, Jasivan A. Sivakumar, Nafise Sadat Moosavi
Abstract: The growing complexity and diversity of news coverage have made framing analysis a crucial yet challenging task in computational social science. Traditional approaches, including manual annotation and fine-tuned models, remain limited by high annotation costs, domain specificity, and inconsistent generalisation. Instruction-based large language models (LLMs) offer a promising alternative, yet their reliability for framing analysis remains insufficiently understood. In this paper, we conduct a systematic evaluation of several LLMs, including GPT-3.5/4, FLAN-T5, and Llama 3, across zero-shot, few-shot, and explanation-based prompting settings. Focusing on domain shift and inherent annotation ambiguity, we show that model performance is highly sensitive to prompt design and prone to systematic errors on ambiguous cases. Although LLMs, particularly GPT-4, exhibit stronger cross-domain generalisation, they also display systematic biases, most notably a tendency to conflate emotional language with framing. To enable principled evaluation under real-world topic diversity, we introduce a new dataset of out-of-domain news headlines covering diverse subjects. Finally, by analysing agreement patterns across multiple models on existing framing datasets, we demonstrate that cross-model consensus provides a useful signal for identifying contested annotations, offering a practical approach to dataset auditing in low-resource settings.
Authors: Guimin Hu, Weimin Lyu, Chang Sun, Zhihong Zhu, Lin Gui, Ruichu Cai, Erik Cambria, Hasti Seifi
Abstract: Multimodal affective computing has gained increasing attention due to its broad applications in understanding human behavior and intentions, particularly in text-centric multimodal scenarios. Existing research spans diverse tasks, modalities, and modeling paradigms, yet lacks a unified perspective. In this survey, we systematically review recent advances from an NLP perspective, focusing on four representative tasks: multimodal sentiment analysis (MSA), multimodal emotion recognition in conversation (MERC), multimodal aspect-based sentiment analysis (MABSA), and multimodal multi-label emotion recognition (MMER). We present a unified view by comparing task formulations, benchmark datasets, and evaluation protocols, and by organizing representative methods into key paradigms, including multitask learning, pre-trained models, knowledge enhancement, and contextual modeling. We further extend the discussion to related directions, such as facial, acoustic, and physiological modalities, as well as emotion cause analysis. Finally, we highlight key challenges and outline promising future directions. To facilitate further research, we release a curated repository of relevant works and resources \footnote{https://anonymous.4open.science/r/Multimodal-Affective-Computing-Survey-9819}.
URLs: https://anonymous.4open.science/r/Multimodal-Affective-Computing-Survey-9819
Authors: Divyanshu Aggarwal, Sankarshan Damle, Navin Goyal, Satya Lokam, Sunayana Sitaram
Abstract: A common challenge towards the adaptability of Large Language Models (LLMs) is their ability to learn new languages over time without hampering the model's performance on languages in which the model is already proficient (usually English). Continual fine-tuning (CFT) is the process of sequentially fine-tuning an LLM to enable the model to adapt to downstream tasks with varying data distributions and time shifts. This paper focuses on the language adaptability of LLMs through CFT. We study a two-phase CFT process in which an English-only end-to-end fine-tuned LLM from Phase 1 (predominantly Task Ability) is sequentially fine-tuned on a multilingual dataset -- comprising task data in new languages -- in Phase 2 (predominantly Language Ability). We observe that the ``similarity'' of Phase 2 tasks with Phase 1 determines the LLM's adaptability. For similar phase-wise datasets, the LLM after Phase 2 does not show deterioration in task ability. In contrast, when the phase-wise datasets are not similar, the LLM's task ability deteriorates. We test our hypothesis on the open-source \mis\ and \llm\ models with multiple phase-wise dataset pairs. To address the deterioration, we analyze tailored variants of two CFT methods: layer freezing and generative replay. Our findings demonstrate their effectiveness in enhancing the language ability of LLMs while preserving task performance, in comparison to relevant baselines.
Authors: Huiyao Chen, Meishan Zhang, Jing Li, Lilja {\O}vrelid, Jan Haji\v{c}, Hao Fei, Min Zhang
Abstract: Semantic role labeling (SRL) is a central natural language processing task for understanding predicate-argument structures within texts and enabling downstream applications. Despite extensive research, comprehensive surveys that critically synthesize the field from a unified perspective remain lacking. This survey makes several contributions beyond organizing existing work. We propose a unified four-dimensional taxonomy that categorizes SRL research along model architectures, syntax feature modeling, application scenarios, and multimodal extensions. We provide a critical analysis of when and why syntactic features help, identifying conditions under which syntax-aided approaches provide consistent gains over syntax-free counterparts. We offer the first systematic treatment of SRL in the era of large language models, examining the complementary roles of LLMs and specialized SRL systems and identifying directions for hybrid approaches. We extend the scope of SRL surveys to cover multimodal settings including visual, video, and speech modalities, and analyze structural differences in evaluation across these modalities. Literature was collected through systematic searches of the ACL Anthology, IEEE Xplore, the ACM Digital Library, and Google Scholar, covering publications from 2000 to 2025 and applying explicit inclusion and exclusion criteria to yield approximately 200 primary references. SRL benchmarks, evaluation metrics, and paradigm modeling approaches are discussed alongside practical applications across domains. Future research directions are analyzed, addressing the evolving role of SRL with large language models and broader NLP impact.
Authors: Penghui Yang, Cunxiao Du, Fengzhuo Zhang, Haonan Wang, Tianyu Pang, Chao Du, Bo An
Abstract: As Large Language Models (LLMs) can now process extremely long contexts, efficient inference over these extended inputs has become increasingly important, especially for emerging applications like LLM agents that highly depend on this capability. Speculative decoding (SD) offers a promising lossless acceleration technique compared to lossy alternatives such as quantization and model cascades. However, most state-of-the-art SD methods are trained on short texts (typically fewer than 4k tokens), making them unsuitable for long-context scenarios. Specifically, adapting these methods to long contexts presents three key challenges: (1) the excessive memory demands posed by draft models due to large Key-Value (KV) cache; (2) performance degradation resulting from the mismatch between short-context training and long-context inference; and (3) inefficiencies in tree attention mechanisms when managing long token sequences. This work introduces LongSpec, a framework that addresses these challenges through three core innovations: a memory-efficient draft model with a constant-sized KV cache; novel position indices that mitigate the training-inference mismatch; and an attention aggregation strategy that combines fast prefix computation with standard tree attention to enable efficient decoding. Experimental results confirm the effectiveness of LongSpec, achieving up to a 3.26x speedup over strong Flash Attention baselines across five long-context understanding datasets, as well as a 2.25x reduction in wall-clock time on the AIME24 long reasoning task with the QwQ model, demonstrating significant latency improvements for long-context applications. The code is available at https://github.com/sail-sg/LongSpec.
Authors: Firoj Alam, Md Arid Hasan, Sahinur Rahman Laskar, Mucahid Kutlu, Kareem Darwish, Shammur Absar Chowdhury
Abstract: The rapid progress of large language models (LLMs) raises concerns about cultural bias, fairness, and performance in diverse languages and underrepresented regions. Addressing these gaps requires large-scale resources grounded in multilingual, local, and cultural contexts. We systematize and extend the earlier NativQA framework to multimodality by adding image, audio, and video support, enabling scalable construction of culturally and regionally aligned QA datasets in native languages. Given user-defined seed queries, the framework uses search engines to collect location-specific everyday information. We evaluate it across 39 locations in 24 countries and 7 languages, spanning extremely low-resource to high-resource settings, and collect over $\sim$300K text QA pairs, $\sim$312K images, and $\sim$29K videos with associated audio. The developed resources can be used for LLMs benchmarking and further fine-tuning. The framework has been made publicly available for the community (https://gitlab.com/nativqa/nativqa-framework). Demo video is available here: \href{https://shorturl.at/DAVn9}{https://shorturl.at/DAVn9}.
URLs: https://gitlab.com/nativqa/nativqa-framework)., https://shorturl.at/DAVn9, https://shorturl.at/DAVn9
Authors: Siddhant Arora, Kai-Wei Chang, Chung-Ming Chien, Yifan Peng, Haibin Wu, Yossi Adi, Emmanuel Dupoux, Hung-Yi Lee, Karen Livescu, Shinji Watanabe
Abstract: The field of spoken language processing is undergoing a shift from training custom-built, task-specific models toward using and optimizing spoken language models (SLMs) which act as universal speech processing systems. This trend is similar to the progression toward universal language models that has taken place in the field of (text) natural language processing. SLMs include both "pure" language models of speech -- models of the distribution of tokenized speech sequences -- and models that combine speech encoders with text language models, often including both spoken and written input or output. Work in this area is very diverse, with a range of terminology and evaluation settings. This paper aims to contribute an improved understanding of SLMs via a unifying literature survey of recent work in the context of the evolution of the field. Our survey categorizes the work in this area by model architecture, training, and evaluation choices, and describes some key challenges and directions for future work.
Authors: Weiqi Wang, Jiefu Ou, Yangqiu Song, Benjamin Van Durme, Daniel Khashabi
Abstract: Literature review tables are essential for summarizing and comparing collections of scientific papers. In this paper, we study the automatic generation of such tables from a pool of papers to satisfy a user's information need. Building on recent work (Newman et al., 2024), we move beyond oracle settings by (i) simulating well-specified yet schema-agnostic user demands that avoid leaking gold column names or values, (ii) explicitly modeling retrieval noise via semantically related but out-of-scope distractor papers verified by human annotators, and (iii) introducing a lightweight, annotation-free, utilization-oriented evaluation that decomposes utility into schema coverage, unary cell fidelity, and pairwise relational consistency, while measuring paper selection through a two-way QA procedure (gold to system and system to gold) with recall, precision, and F1. To support reproducible evaluation, we introduce arXiv2Table, a benchmark of 1,957 tables referencing 7,158 papers, with human-verified distractors and rewritten, schema-agnostic user demands. We also develop an iterative, batch-based generation method that co-refines paper filtering and schema over multiple rounds. We validate the evaluation protocol with human audits and cross-evaluator checks. Extensive experiments show that our method consistently improves over strong baselines, while absolute scores remain modest, underscoring the task's difficulty. Our data and code is available at https://github.com/JHU-CLSP/arXiv2Table.
Authors: Darpan Aswal, Siddharth D Jaiswal
Abstract: Safety-aligned LLMs remain vulnerable to digital phenomena like textese that introduce non-canonical perturbations to words but preserve the phonetics. We introduce CMP-RT (code-mixed phonetic perturbations for red-teaming), a novel diagnostic probe that pinpoints tokenization as the root cause of this vulnerability. A mechanistic analysis reveals that phonetic perturbations fragment safety-critical tokens into benign sub-words, suppressing their attribution scores while preserving prompt interpretability -- causing safety mechanisms to fail despite excellent input understanding. We demonstrate that this vulnerability evades standard defenses, persists across modalities and state-of-the-art (SOTA) models including Gemini-3-Pro, and scales through simple supervised fine-tuning (SFT). Furthermore, layer-wise probing shows perturbed and canonical input representations align up to a critical layer depth; enforcing output equivalence robustly recovers the lost representations, providing causal evidence for a structural gap between pre-training and alignment, and establishing tokenization as a critical, under-examined vulnerability in current safety pipelines.
Authors: Iddo Yosha, Gallil Maimon, Yossi Adi
Abstract: Sentence stress refers to emphasis on words within a spoken utterance to highlight or contrast an idea. It is often used to imply an underlying intention not explicitly stated. Recent speech-aware language models (SLMs) have enabled direct audio processing, allowing models to access the full richness of speech to perform audio reasoning tasks such as spoken question answering. Despite the crucial role of sentence stress in shaping meaning and intent, it remains largely overlooked in evaluation and development of SLMs. We address this gap by introducing StressTest, a benchmark designed to evaluate models' ability to distinguish between meanings of speech based on the stress pattern. We evaluate leading SLMs, and find that despite their overall capabilities, they perform poorly on such tasks. Hence, we propose a novel data generation pipeline, and create Stress-17k, a training set that simulates change of meaning implied by stress variation. Results suggest, that our finetuned model, StresSLM, generalizes well to real recordings and notably outperforms existing SLMs on sentence stress reasoning and detection. Models, code, data, samples - pages.cs.huji.ac.il/adiyoss-lab/stresstest.
Authors: Thi Thu Uyen Hoang, Meenakshi Rajendran, Kun Zhang, Yuhan Wu, Viet Anh Nguyen
Abstract: This paper presents an advancement in Question-Answering (QA) systems using a Retrieval Augmented Generation (RAG) framework to enhance information extraction from PDF files. Recognizing the richness and diversity of data within PDFs--including text, images, vector diagrams, graphs, and tables--poses unique challenges for existing QA systems primarily designed for textual content. We seek to develop a comprehensive RAG-based QA system that will effectively address complex multimodal questions, where several data types are combined in the query. This is mainly achieved by refining approaches to processing and integrating non-textual elements in PDFs into the RAG framework to derive precise and relevant answers, as well as fine-tuning large language models to better adapt to our system. We provide an in-depth experimental evaluation of our solution, demonstrating its capability to extract accurate information that can be applied to different types of content across PDFs. This work not only pushes the boundaries of retrieval-augmented QA systems but also lays a foundation for further research in multimodal data integration and processing.
Authors: Klim Zaporojets, Daniel Daza, Edoardo Barba, Ira Assent, Roberto Navigli, Paul Groth
Abstract: Knowledge Graphs (KGs) are structured knowledge repositories containing entities and relations between them. In this paper, we study the problem of automatically updating KGs over time in response to evolving knowledge in unstructured textual sources. Addressing this problem requires identifying a wide range of update operations based on the state of an existing KG at a given time and the information extracted from text. This contrasts with traditional information extraction pipelines, which extract knowledge from text independently of the current state of a KG. To address this challenge, we propose a method for construction of a dataset consisting of Wikidata KG snapshots over time and Wikipedia passages paired with the corresponding edit operations that they induce in a particular KG snapshot. The resulting dataset comprises 233K Wikipedia passages aligned with a total of 1.45 million KG edits over 7 different yearly snapshots of Wikidata from 2019 to 2025. Our experimental results highlight key challenges in updating KG snapshots based on emerging textual knowledge, particularly in integrating knowledge expressed in text with the existing KG structure. These findings position the dataset as a valuable benchmark for future research. Our dataset and model implementations are publicly available.
Authors: Joosung Lee, Cheonbok Park, Hwiyeol Jo, Jeonghoon Kim, Joonsuk Park, Kang Min Yoo
Abstract: Large Language Models (LLMs) are widely used to generate plausible text on online platforms, without revealing the generation process. As users increasingly encounter such black-box outputs, detecting hallucinations has become a critical challenge. To address this challenge, we focus on developing a hallucination detection framework for black-box generators. Motivated by the observation that hallucinations, once introduced, tend to persist, we sample future contexts. The sampled future contexts provide valuable clues for hallucination detection and can be effectively integrated with various sampling-based methods. We extensively demonstrate performance improvements across multiple methods using our proposed sampling approach.
Authors: Omri Uzan, Yuval Pinter
Abstract: Tasks that require character-level reasoning, such as counting or locating characters within words, remain challenging for contemporary language models. A common conjecture is that language models' reliance on subword units, rather than characters, contributes to their struggles with character-level tasks, yet recent studies offer conflicting conclusions about the role of tokenization, leaving its impact unclear. To address this gap, we introduce CharBench, a comprehensive benchmark of character-level tasks that is two orders of magnitude larger than existing alternatives. We evaluate a diverse range of leading open-weight and proprietary models on CharBench and find that it presents a significant challenge to modern LLMs, with an average accuracy of 43.6% and 32.3% on some tasks. We present an in-depth analysis of how intrinsic properties of words and their segmentations into tokens correspond to model performance. For counting tasks, we find that tokenization properties are weakly correlated with correctness, while the length of the queried word and the actual character count play a more significant part. In contrast, for tasks requiring intra-word positional understanding, performance is negatively correlated with the length of the token containing the queried character, suggesting that longer tokens obscure character position information for LLMs. We encourage future work to build on the benchmark and evaluation methodology introduced here as tools for improving model performance on such tasks.
Authors: Kimihiro Hasegawa, Wiradee Imrattanatrai, Masaki Asada, Susan Holm, Yuran Wang, Vincent Zhou, Ken Fukuda, Teruko Mitamura
Abstract: Assistants on assembly tasks show great potential to benefit humans ranging from helping with everyday tasks to interacting in industrial settings. However, evaluation resources in assembly activities are underexplored. To foster system development, we propose a new multimodal QA evaluation dataset on assembly activities. Our dataset, ProMQA-Assembly, consists of 646 QA pairs that require multimodal understanding of human activity videos and their instruction manuals in an online-style manner. For cost effectiveness in the data creation, we adopt a semi-automated QA annotation approach, where LLMs generate candidate QA pairs and humans verify them. We further improve QA generation by integrating fine-grained action labels to diversify question types. Additionally, we create 81 instruction task graphs for our target assembly tasks. These newly created task graphs are used in our benchmarking experiment, as well as in facilitating the human verification process. With our dataset, we benchmark models, including competitive proprietary multimodal models. We find that ProMQA-Assembly contains challenging multimodal questions, where reasoning models showcase promising results. We believe our new evaluation dataset contributes to the further development of procedural-activity assistants.
Authors: Zhaohan Zhang, Ziquan Liu, Ioannis Patras
Abstract: Assessing the reliability of Large Language Models (LLMs) by confidence elicitation is a prominent approach to AI safety in high-stakes applications, such as healthcare and finance. Existing methods either require expensive computational overhead or suffer from poor calibration, making them impractical and unreliable for real-world deployment. In this work, we propose GrACE, a Generative Approach to Confidence Elicitation that enables scalable and reliable confidence elicitation for LLMs. GrACE adopts a novel mechanism in which the model expresses confidence by the similarity between the last hidden state and the embedding of a special token appended to the vocabulary, in real-time. We fine-tune the model for calibrating the confidence with targets associated with accuracy. Extensive experiments show that the confidence produced by GrACE achieves the best discriminative capacity and calibration on open-ended generation tasks without resorting to additional sampling or an auxiliary model. Moreover, we propose two confidence-based strategies for test-time scaling with GrACE, which not only improve the accuracy of the final decision but also significantly reduce the number of required samples, highlighting its potential as a practical solution for deploying LLMs with reliable, on-the-fly confidence estimation.
Authors: Junsong Li, Jie Zhou, Bihao Zhan, Yutao Yang, Qianjun Pan, Shilian Chen, Tianyu Huai, Xin Li, Qin Chen, Liang He
Abstract: Alignment plays a crucial role in Large Language Models (LLMs) in aligning with human preferences on a specific task/domain. Traditional alignment methods suffer from catastrophic forgetting, where models lose previously acquired knowledge when adapting to new preferences or domains. We introduce LifeAlign, a novel framework for lifelong alignment that enables LLMs to maintain consistent human preference alignment across sequential learning tasks without forgetting previously learned knowledge. Our approach consists of two key innovations. First, we propose a focalized preference optimization strategy that aligns LLMs with new preferences while preventing the erosion of knowledge acquired from previous tasks. Second, we develop a short-to-long memory consolidation mechanism that merges denoised short-term preference representations into stable long-term memory using intrinsic dimensionality reduction, enabling efficient storage and retrieval of alignment patterns across diverse domains. We evaluate LifeAlign across multiple sequential alignment tasks spanning different domains and preference types. Experimental results demonstrate that our method achieves superior performance in maintaining both preference alignment quality and knowledge retention compared to existing lifelong learning approaches. The codes and datasets have been released on https://github.com/real-ljs/LifeAlign.
Authors: Ziyi Liu
Abstract: Large Language Models (LLMs) struggle with information forgetting and inefficiency in long-horizon, multi-turn dialogues. To address this, we propose a training-free prompt engineering method, the State-Update Multi-turn Dialogue Strategy. It utilizes "State Reconstruction" and "History Remind" mechanisms to effectively manage dialogue history. Our strategy shows strong performance across multiple multi-hop QA datasets. For instance, on the HotpotQA dataset, it improves the core information filtering score by 32.6%, leading to a 14.1% increase in the downstream QA score, while also reducing inference time by 73.1% and token consumption by 59.4%. Ablation studies confirm the pivotal roles of both components. Our work offers an effective solution for optimizing LLMs in long-range interactions, providing new insights for developing more robust Agents.
Authors: Xingyu Shen, Yingfa Chen, Zhen Leng Thai, Xu Han, Zhiyuan Liu, Maosong Sun
Abstract: Recurrent neural networks (RNNs), such as linear attention and state-space models, have gained popularity due to their constant per-token complexity when processing long contexts. However, these recurrent models struggle with tasks that require accurate recall of contextual information from long contexts, because all contextual information is compressed into a fixed-size recurrent state. Previous studies have shown that recall ability is positively correlated with the recurrent state size, yet directly training RNNs with large recurrent states results in high training costs. In this paper, we introduce StateX, a post-training framework that efficiently expands the states of pre-trained RNNs. For two popular classes of RNNs, linear attention and state-space models, we design post-training architectural modifications in StateX, to scale up the state size with no or negligible increase in model parameters. Experiments on models with up to 1.3B parameters demonstrate that StateX efficiently enhances the recall and in-context learning performance of RNNs without incurring high post-training costs or compromising other capabilities.
Authors: Alexandrine Fortier, Thomas Thebaud, Jes\'us Villalba, Najim Dehak, Patrick Cardinal, Peter West
Abstract: Speech language models (SLMs) are systems of systems: independent components that unite to achieve a common goal. Despite their heterogeneous nature, SLMs are often studied end-to-end; how information flows through the pipeline remains obscure. We investigate this question through the lens of backdoor attacks. We first establish that backdoors can propagate through the SLM, leaving all tasks highly vulnerable. From this, we design a component analysis to reveal the role each component takes in backdoor learning. We find that backdoor persistence or erasure is highly dependent on the targeted component. Beyond propagation, we examine how backdoors are encoded in shared multitask embeddings, showing that poisoned samples are not directly separable from benign ones, challenging a common separability assumption used in filtering defenses. Our findings emphasize the need to treat multimodal pipelines as intricate systems with unique vulnerabilities, not solely extensions of unimodal ones.
Authors: David Beauchemin, Yan Tremblay, Mohamed Amine Youssef, Richard Khoury
Abstract: The tasks of idiom understanding and dialect understanding are both well-established benchmarks in natural language processing. In this paper, we propose combining them, and using regional idioms as a test of dialect understanding. Towards this end, we propose three new benchmark datasets for the Quebec dialect of French: QFrCoRE, which contains 4,633 instances of idiomatic phrases, and QFrCoRT, which comprises 171 regional instances of idiomatic words, and a new benchmark for French Metropolitan expressions, MFrCoE, which comprises 4,938 phrases. We explain how to construct these corpora, so that our methodology can be replicated for other dialects. Our experiments with 111 LLMs reveal a critical disparity in dialectal competence: while models perform well on French Metropolitan, 65.77% of them perform significantly worse on Quebec idioms, with only 9.0% favoring the regional dialect. These results confirm that our benchmarks are a reliable tool for quantifying the dialect gap and that prestige-language proficiency does not guarantee regional dialect understanding.
Authors: Omri Uzan, Asaf Yehudai, Roi pony, Eyal Shnarch, Ariel Gera
Abstract: Multimodal encoders have pushed the boundaries of visual document retrieval, matching textual query tokens directly to image patches and achieving state-of-the-art performance on public benchmarks. Recent models relying on this paradigm have massively scaled the sizes of their query and document representations, presenting obstacles to deployment and scalability in real-world pipelines. Furthermore, purely vision-centric approaches may be constrained by the inherent modality gap still exhibited by modern vision-language models. In this work, we connect these challenges to the paradigm of hybrid retrieval, investigating whether a lightweight dense text retriever can enhance a stronger vision-centric model. Existing hybrid methods, which rely on coarse-grained fusion of ranks or scores, fail to exploit the rich interactions within each model's representation space. To address this, we introduce Guided Query Refinement (GQR), a novel test-time optimization method that refines a primary retriever's query embedding using guidance from a complementary retriever's scores. Through extensive experiments on visual document retrieval benchmarks, we demonstrate that GQR allows vision-centric models to match the performance of models with significantly larger representations, while being up to 14x faster and requiring 54x less memory. Our findings show that GQR effectively pushes the Pareto frontier for performance and efficiency in multimodal retrieval. We release our code at https://github.com/IBM/test-time-hybrid-retrieval
Authors: Junhoo Lee, Seungyeon Kim, Nojun Kwak
Abstract: Diffusion Language Models (DLMs) have emerged as a promising alternative to Autoregressive Language Models, yet their inference strategies remain limited to prefix-based prompting inherited from the autoregressive paradigm. In this paper, we propose Template Infilling (TI), a tailored conditioning methodology for DLMs. Unlike conventional prefix prompting, TI flexibly aligns structural anchors across the entire target response space, establishing a global blueprint before filling in the masked segments. We demonstrate the effectiveness of our approach on diverse benchmarks, including mathematical reasoning, code generation, and trip planning, achieving consistent improvements of 9.40% over the baseline. Furthermore, we observe that TI provides additional advantages in multi-token generation settings, enabling effective speedup while maintaining generation quality and robustness. By enforcing these global constraints, TI ultimately facilitates System-2 reasoning, empowering the model to deliberate within a structurally defined solution space.
Authors: Xingrui Zhuo, Jiapu Wang, Gongqing Wu, Zhongyuan Wang, Jichen Zhang, Shirui Pan, Xindong Wu
Abstract: Inductive Knowledge Graph Reasoning (KGR) aims to discover facts in open-domain KGs containing unknown entities and relations, which poses a challenge for KGR models in comprehending uncertain KG components. Existing studies have proposed Knowledge Graph Foundation Models (KGFMs) that learn structural invariances across KGs to handle this uncertainty. Recently, Large Language Models (LLMs) have demonstrated strong capabilities for open-domain knowledge reasoning. As a result, the latest research has focused on LLM-based KGFMs that integrate LLM knowledge with KG context for inductive KGR. However, the intrinsic knowledge of LLMs may be overshadowed by sparse KG context, leading to LLM knowledge distortion, which can cause irreversible damage to model reasoning. Moreover, existing LLM-based KGR methods still struggle to fully constrain generative hallucinations in LLMs, severely limiting the credibility of reasoning results. To address these limitations, we propose a Knowledge Reasoning Language Model (KRLM) that achieves unified coordination between LLM knowledge and KG context throughout the KGR process. Specifically, we design a Knowledge Reasoning Language (KRL) instruction format and a KRL tokenizer to align LLM knowledge with KG representations. Then, we propose a KRL attention layer that coordinates intrinsic LLM knowledge with additional KG context through a dynamic knowledge memory mechanism. Finally, a structure-aware next-entity predictor is proposed, which strictly constrains the reasoning results within a trustworthy knowledge domain. Extensive experimental results on 25 real-world inductive KGR datasets demonstrate the significant superiority of the proposed KRLM\footnote{Our source codes are available at https://anonymous.4open.science/r/KRLM-EA36 in both zero-shot reasoning and fine-tuning scenarios.
Authors: Qing Yang, Zhenghao Liu, Yangfan Du, Pengcheng Huang, Tong Xiao
Abstract: Recent advances in Text-To-Speech (TTS) synthesis have achieved near-human speech quality in neutral speaking styles. However, most existing approaches either depend on costly emotion annotations or optimize surrogate objectives that fail to adequately capture perceptual emotional quality. As a result, the generated speech, while semantically accurate, often lacks expressive and emotionally rich characteristics. To address these limitations, we propose RLAIF-SPA, a novel framework that integrates Reinforcement Learning from AI Feedback (RLAIF) to directly optimize both emotional expressiveness and intelligibility without human supervision. Specifically, RLAIF-SPA incorporates Automatic Speech Recognition (ASR) to provide semantic accuracy feedback, while leveraging structured reward modeling to evaluate prosodic-emotional consistency. RLAIF-SPA enables more precise and nuanced control over expressive speech generation along four structured evaluation dimensions: Structure, Emotion, Speed, and Tone. Extensive experiments on Libri-Speech, MELD, and Mandarin ESD datasets demonstrate consistent gains across clean read speech, conversational dialogue, and emotional speech. On Libri-Speech, RLAIF-SPA consistently outperforms Chat-TTS, achieving a 26.1% reduction in word error rate, a 9.1% improvement in SIM-O, and over 10% gains in human subjective evaluations.
Authors: Xingmeng Zhao, Tongnian Wang, Dan Schumacher, Veronica Rammouz, Anthony Rios
Abstract: Artificial intelligence (AI) is rapidly transforming healthcare, enabling fast development of tools like stress monitors, wellness trackers, and mental health chatbots. However, rapid and low-barrier development can introduce risks of bias, privacy violations, and unequal access, especially when systems ignore real-world contexts and diverse user needs. Many recent methods use AI to detect risks automatically, but this can reduce human engagement in understanding how harms arise and who they affect. We present a human-centered framework that generates user stories and supports multi-agent discussions to help people think creatively about potential benefits and harms before deployment. In a user study, participants who read stories recognized a broader range of harms, distributing their responses more evenly across all 17 harm types. In contrast, those who did not read stories focused primarily on privacy and well-being (79.1%). Our findings show that storytelling helped participants speculate about a broader range of harms and benefits and think more creatively about AI's impact on users.
Authors: Yu Zhou, Sohyun An, Haikang Deng, Da Yin, Clark Peng, Cho-Jui Hsieh, Kai-Wei Chang, Nanyun Peng
Abstract: Contact languages like English exhibit rich regional variations in the form of dialects, which are often used by dialect speakers interacting with generative models. However, can multimodal generative models effectively produce content given dialectal textual input? In this work, we study this question by constructing a new large-scale benchmark spanning six common English dialects. We work with dialect speakers to collect and verify over 4200 unique prompts and evaluate on 17 image and video generative models. Our automatic and human evaluation results show that current state-of-the-art multimodal generative models exhibit 32.26% to 48.17% performance degradation when a single dialect word is used in the prompt. Common mitigation methods such as fine-tuning and prompt rewriting can only improve dialect performance by small margins (< 7%), while potentially incurring significant performance degradation in Standard American English (SAE). To this end, we design a general encoder-based mitigation strategy for multimodal generative models. Our method teaches the model to recognize new dialect features while preserving SAE performance. Experiments on models such as Stable Diffusion 1.5 show that our method is able to simultaneously raise performance on five dialects to be on par with SAE (+34.4%), while incurring near zero cost to SAE performance.
Authors: Noor Islam S. Mohammad
Abstract: Toxic text classification for online moderation remains challenging under extreme class imbalance, where rare but high-risk labels such as threat and severe_toxic are consistently underdetected by conventional models. We propose CoGate-LSTM, a parameter-efficient recurrent architecture built around a novel cosine-similarity feature gating mechanism that adaptively rescales token embeddings by their directional similarity to a learned toxicity prototype. Unlike token-position attention, the gate emphasizes feature directions most informative for minority toxic classes. The model combines frozen multi-source embeddings (GloVe, FastText, and BERT-CLS), a character-level BiLSTM, embedding-space SMOTE, and weighted focal loss. On the Jigsaw Toxic Comment benchmark, CoGate-LSTM achieves 0.881 macro-F1 (95% CI: [0.873, 0.889]) and 96.0% accuracy, outperforming fine-tuned BERT by 6.9 macro-F1 points (p < 0.001) and XGBoost by 4.7, while using only 7.3M parameters (about 15$\times$ fewer than BERT) and 48 ms CPU inference latency. Gains are strongest on minority labels, with F1 improvements of +71% for severe_toxic, +33% for threat, and +28% for identity_hate relative to fine-tuned BERT. Ablations identify cosine gating as the primary driver of performance (-4.8 macro-F1 when removed), with additional benefits from character-level fusion (-2.4) and multi-head attention (-2.9). CoGate-LSTM also transfers reasonably across datasets, reaching a 0.71 macro-F1 zero-shot on the Contextual Abuse Dataset and 0.73 with lightweight threshold adaptation. These results show that direction-aware feature gating offers an effective and efficient alternative to large, fully fine-tuned transformers for classifying imbalanced toxic comments.
Authors: Afrozah Nadeem, Mark Dras, Usman Naseem
Abstract: Large language models often display undesirable behaviors embedded in their internal representations, undermining fairness, inconsistency drift, amplification of harmful content, and the propagation of unwanted patterns during extended dialogue and conversations. Although training-time or data-centric methods attempt to reduce these effects, they are computationally expensive, irreversible once deployed, and slow to adapt to new conversational contexts. Pruning-based methods provide a flexible and transparent way to reduce bias by adjusting the neurons responsible for certain behaviors. However, most existing approaches are static; once a neuron is removed, the model loses the ability to adapt when the conversation or context changes. To address this, we propose a dynamic, reversible, pruning-based framework that detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. Our inference-time solution provides fine-grained, memory-aware mitigation with knowledge-preserved, more coherent behavior across multilingual single- and multi-turn dialogues, enabling dynamic fairness control in real-world conversational AI.
Authors: Kailin Jiang, Ning Jiang, Yuntao Du, Yuchen Ren, Yuchen Li, Yifan Gao, Jinhe Bi, Yunpu Ma, Bin Li, Lei Liu, Qing Li
Abstract: Large Multimodal Models (LMMs) encode rich factual knowledge via cross-modal pre-training, yet their static representations struggle to maintain an accurate understanding of time-sensitive factual knowledge. Existing benchmarks remain constrained by static designs, inadequately evaluating LMMs' ability to understand time-sensitive knowledge. To address this gap, we propose MINED, a comprehensive benchmark that evaluates temporal awareness along 6 key dimensions and 11 challenging tasks: cognition, awareness, trustworthiness, understanding, reasoning, and robustness. MINED is constructed from Wikipedia by two professional annotators, containing 2,104 time-sensitive knowledge samples spanning six knowledge types. Evaluating 15 widely used LMMs on MINED shows that Gemini-2.5-Pro achieves the highest average CEM score of 63.07, while most open-source LMMs still lack time understanding ability. Meanwhile, LMMs perform best on organization knowledge, whereas their performance is weakest on sport. To address these challenges, we investigate the feasibility of updating time-sensitive knowledge in LMMs through knowledge editing methods and observe that LMMs can effectively update knowledge via knowledge editing methods in single editing scenarios.
Authors: Matthias De Lange, Jens-Joris Decorte, Jeroen Van Hautte
Abstract: Applications in labor market intelligence demand specialized NLP systems for a wide range of tasks, characterized by extreme multi-label target spaces, strict latency constraints, and multiple text modalities such as skills and job titles. These constraints have led to isolated, task-specific developments in the field, with models and benchmarks focused on single prediction tasks. Exploiting the shared structure of work-related data, we propose a unifying framework, combining a wide range of tasks in a multi-task ranking benchmark, and a flexible architecture tackling text-driven work tasks with a single model. The benchmark, WorkBench, is the first unified evaluation suite spanning six work-related tasks formulated explicitly as ranking problems, curated from real-world ontologies and human-annotated resources. WorkBench enables cross-task analysis, where we find significant positive cross-task transfer. This insight leads to Unified Work Embeddings (UWE), a task-agnostic bi-encoder that exploits our training-data structure with a many-to-many InfoNCE objective, and leverages token-level embeddings with task-agnostic soft late interaction. UWE demonstrates zero-shot ranking performance on unseen target spaces in the work domain, and enables low-latency inference with two orders of magnitude fewer parameters than best-performing generalist models (Qwen3-8B), with +4.4 MAP improvement.
Authors: Yuanjie Zhu, Liangwei Yang, Ke Xu, Weizhi Zhang, Zihe Song, Jindong Wang, Philip S. Yu
Abstract: Large Language Models (LLMs) are reshaping unsupervised learning by offering an unprecedented ability to perform text clustering based on their deep semantic understanding. However, their direct application is fundamentally limited by a lack of stateful memory for iterative refinement and the difficulty of managing cluster granularity. As a result, existing methods often rely on complex pipelines with external modules, sacrificing a truly end-to-end approach. We introduce LLM-MemCluster, a novel framework that reconceptualizes clustering as a fully LLM-native task. It leverages a Dynamic Memory to instill state awareness and a Dual-Prompt Strategy to enable the model to reason about and determine the number of clusters. Evaluated on several benchmark datasets, our tuning-free framework significantly and consistently outperforms strong baselines. LLM-MemCluster presents an effective, interpretable, and truly end-to-end paradigm for LLM-based text clustering.
Authors: Pawel Batorski, Paul Swoboda
Abstract: LLMs are highly sensitive to prompt design, but handcrafting effective prompts is difficult and often requires intricate crafting of few-shot examples. We propose a fast automatic prompt construction algorithm that augments human instructions by generating a small set of few shot examples. Our method iteratively replaces/drops/keeps few-shot examples using Monte Carlo Shapley estimation of example utility. For faster execution, we use aggressive subsampling and a replay buffer for faster evaluations. Our method can be run using different compute time budgets. On a limited budget, we outperform existing automatic prompting methods on text simplification and GSM8K and obtain second best results on classification and summarization. With an extended, but still modest compute budget we set a new state of the art among automatic prompting methods on classification, simplification and GSM8K. Our results show that carefully constructed examples, rather than exhaustive instruction search, are the dominant lever for fast and data efficient prompt engineering. Our code is available at https://github.com/Batorskq/PIAST.
Authors: Oleksii Proniakin, Diego Fajardo, Ruslan Nazarenko, Razvan Marinescu
Abstract: Large language models (LLMs) are increasingly evaluated in clinical settings using multi-dimensional rubrics which quantify reasoning quality, safety, and patient-centeredness. Yet, replicating specific mistakes in other LLM models is not straightforward and often requires manual effort. We introduce MedMistake, an automatic pipeline that extracts mistakes LLMs make in patient-doctor conversations and converts them into a benchmark of single-shot QA pairs. Our pipeline (1) creates complex, conversational data between an LLM patient and LLM doctor, (2) runs an evaluation with a committee of 2 LLM judges across a variety of dimensions and (3) creates simplified single-shot QA scenarios from those mistakes. We release MedMistake-All, a dataset of 3,390 single-shot QA pairs where GPT-5 and Gemini 2.5 Pro are currently failing to answer correctly, as judged by two LLM judges. We used medical experts to validate a subset of 211/3390 questions (MedMistake-Bench), which we used to run a final evaluation of 12 frontier LLMs: Claude Opus 4.5, Claude Sonnet 4.5, DeepSeek-Chat, Gemini 2.5 Pro, Gemini 3 Pro, GPT-4o, GPT-5, GPT-5.1, GPT-5.2, Grok 4, Grok 4.1, Mistral Large. We found that GPT models, Claude and Grok obtained the best performance on MedMistake-Bench. We release both the doctor-validated benchmark (MedMistake-Bench), as well as the full dataset (MedMistake-All) at https://huggingface.co/datasets/TheLumos/MedicalMistakeBenchmark.
URLs: https://huggingface.co/datasets/TheLumos/MedicalMistakeBenchmark.
Authors: Ruikang Zhang, Shuo Wang, Qi Su
Abstract: Recent work in Mechanistic Interpretability (MI) has enabled the identification and intervention of internal features in Large Language Models (LLMs). However, a persistent challenge lies in linking such internal features to the reliable control of complex, behavior-level semantic attributes in language generation. In this paper, we propose a Sparse Autoencoder-based framework for retrieving and steering semantically interpretable internal features associated with high-level linguistic behaviors. Our method employs a contrastive feature retrieval pipeline based on controlled semantic oppositions, combing statistical activation analysis and generation-based validation to distill monosemantic functional features from sparse activation spaces. Using the Big Five personality traits as a case study, we demonstrate that our method enables precise, bidirectional steering of model behavior while maintaining superior stability and performance compared to existing activation steering methods like Contrastive Activation Addition (CAA). We further identify an empirical effect, which we term Functional Faithfulness, whereby intervening on a specific internal feature induces coherent and predictable shifts across multiple linguistic dimensions aligned with the target semantic attribute. Our findings suggest that LLMs internalize deeply integrated representations of high-order concepts, and provide a novel, robust mechanistic path for the regulation of complex AI behaviors.
Authors: Hexiang Tan, Wanli Yang, Junwei Zhang, Xin Chen, Rui Tang, Du Su, Jingang Wang, Yuanzhuo Wang, Fei Sun, Xueqi Cheng
Abstract: Reliable confidence is essential for trusting the outputs of LLMs, yet widely deployed post-trained LLMs (PoLLMs) typically compromise this trust with severe overconfidence. In contrast, we observe that their corresponding base LLMs often remain well-calibrated. This naturally motivates us to calibrate PoLLM confidence using the base LLM as a reference. This work proposes two ways to achieve this. A straightforward solution, BaseCal-ReEval, evaluates PoLLM's responses by feeding them into the base LLM to get average probabilities as confidence. While effective, this approach introduces additional inference overhead. To address this, we propose BaseCal-Proj, which trains a lightweight projection to map the final-layer hidden states of PoLLMs back to those of their base LLMs. These projected states are then processed by the base LLM's output layer to derive base-calibrated confidence for PoLLM's responses. Notably, BaseCal is an unsupervised, plug-and-play solution that operates without human labels or LLM modifications. Experiments across five datasets and three LLM families demonstrate the effectiveness of BaseCal, reducing Expected Calibration Error (ECE) by an average of 42.90\% compared to the best unsupervised baselines.
Authors: Haoming Xu, Ningyuan Zhao, Yunzhi Yao, Weihong Xu, Hongru Wang, Xinle Deng, Shumin Deng, Jeff Z. Pan, Huajun Chen, Ningyu Zhang
Abstract: As Large Language Models (LLMs) are increasingly deployed in real-world settings, correctness alone is insufficient. Reliable deployment requires maintaining truthful beliefs under contextual perturbations. Existing evaluations largely rely on point-wise confidence like Self-Consistency, which can mask brittle belief. We show that even facts answered with perfect self-consistency can rapidly collapse under mild contextual interference. To address this gap, we propose Neighbor-Consistency Belief (NCB), a structural measure of belief robustness that evaluates response coherence across a conceptual neighborhood. To validate the efficiency of NCB, we introduce a new cognitive stress-testing protocol that probes outputs stability under contextual interference. Experiments across multiple LLMs show that the performance of high-NCB data is relatively more resistant to interference. Finally, we present Structure-Aware Training (SAT), which optimizes context-invariant belief structure and reduces long-tail knowledge brittleness by approximately 30%. Code is available at https://github.com/zjunlp/belief.
Authors: Jingsheng Zheng, Jintian Zhang, Yujie Luo, Yuren Mao, Yunjun Gao, Lun Du, Huajun Chen, Ningyu Zhang
Abstract: Autonomous machine learning agents have revolutionized scientific discovery, yet they remain constrained by a Generate-Execute-Feedback paradigm. Previous approaches suffer from a severe Execution Bottleneck, as hypothesis evaluation relies strictly on expensive physical execution. To bypass these physical constraints, we internalize execution priors to substitute costly runtime checks with instantaneous predictive reasoning, drawing inspiration from World Models. In this work, we formalize the task of Data-centric Solution Preference and construct a comprehensive corpus of 18,438 pairwise comparisons. We demonstrate that LLMs exhibit significant predictive capabilities when primed with a Verified Data Analysis Report, achieving 61.5% accuracy and robust confidence calibration. Finally, we instantiate this framework in FOREAGENT, an agent that employs a Predict-then-Verify loop, achieving a 6x acceleration in convergence while surpassing execution-based baselines by +6%. Our code and dataset are publicly available at https://github.com/zjunlp/predict-before-execute.
Authors: Biswesh Mohapatra, Th\'eo Charlot, Giovanni Duca, Mayank Palan, Laurent Romary, Justine Cassell
Abstract: Common ground plays a critical role in situated spoken dialogs, where interlocutors must establish and maintain shared references to entities, events, and relations to sustain coherent interaction in a shared space and over time. With the increasing presence of embodied conversational agents and social robots, the ability to correctly ground this kind of conversational content in order to refer back later also becomes important for dialog systems. Prior studies have demonstrated that LLMs are capable of performing certain grounding acts like acknowledgments. However, relatively little work has investigated their capacity to leverage the grounded information, like in complex scenarios involving space and time (e.g., "let's go to that caf\'e near the park we went to yesterday"). To that end, in this work, we evaluate a model's ability to establish common ground by utilizing these "relational references" in the dynamic and shared environments of situated dialogs. We then test multiple methods for representing common ground and further propose approaches to improve their performance by using reinforcement learning on our synthetically generated dialog data .
Authors: Mahammad Namazov, Tom\'a\v{s} Koref, Ivan Habernal
Abstract: Interpretability is critical for applications of large language models (LLMs) in the legal domain, where trust and transparency are essential. A central NLP task in this setting is legal outcome prediction, where models forecast whether a court will find a violation of a given right. We study this task on decisions from the European Court of Human Rights (ECtHR), introducing a new ECtHR dataset with carefully curated positive (violation) and negative (non-violation) cases. Existing works propose both task-specific approaches and model-agnostic techniques to explain downstream performance, but it remains unclear which techniques best explain legal outcome prediction. To address this, we propose a comparative analysis framework for model-agnostic interpretability methods. We focus on two rationale extraction techniques that justify model outputs with concise, human-interpretable text fragments from the input. We evaluate faithfulness via normalized sufficiency and comprehensiveness metrics, and plausibility via legal expert judgments of the extracted rationales. We also assess the feasibility of using LLM-as-a-Judge, using these expert evaluations as reference. Our experiments on the new ECtHR dataset show that models' "reasons" for predicting violations differ substantially from those of legal experts, despite strong faithfulness scores. The source code of our experiments is publicly available at https://github.com/trusthlt/IntEval.
Authors: V\'ictor Yeste, Paolo Rosso
Abstract: Human value detection from single sentences is a sparse, imbalanced multi-label task. We study whether Schwartz higher-order (HO) categories help this setting on ValueEval'24 / ValuesML (74K English sentences) under a compute-frugal budget. Rather than proposing a new architecture, we compare direct supervised transformers, hard HO$\rightarrow$values pipelines, Presence$\rightarrow$HO$\rightarrow$values cascades, compact instruction-tuned large language models (LLMs), QLoRA, and low-cost upgrades such as threshold tuning and small ensembles. HO categories are learnable: the easiest bipolar pair, Growth vs. Self-Protection, reaches Macro-$F_1=0.58$. The most reliable gains come from calibration and ensembling: threshold tuning improves Social Focus vs. Personal Focus from $0.41$ to $0.57$ ($+0.16$), transformer soft voting lifts Growth from $0.286$ to $0.303$, and a Transformer+LLM hybrid reaches $0.353$ on Self-Protection. In contrast, hard hierarchical gating does not consistently improve the end task. Compact LLMs also underperform supervised encoders as stand-alone systems, although they sometimes add useful diversity in hybrid ensembles. Under this benchmark, the HO structure is more useful as an inductive bias than as a rigid routing rule.
Authors: Ahsan Bilal, Ahmed Mohsin, Muhammad Umer, Ali Subhan, Hassan Rizwan, Ayesha Mohsin, Dean Hougen
Abstract: Test-time compute scaling allocates inference computation uniformly, uses fixed sampling strategies, and applies verification only for reranking. In contrast, we propose a verifier-guided adaptive framework treating reasoning as iterative trajectory generation and selection. For each problem, the agent runs multiple inference iterations. In each iteration, it optionally produces a high-level plan, selects a set of reasoning tools and a compute strategy together with an exploration parameter, and then generates a candidate reasoning trajectory. A process reward model (PRM) serves as a unified control signal: within each iteration, step-level PRM scores are aggregated to guide pruning and expansion during generation, and across iterations, aggregated trajectory rewards are used to select the final response. Across datasets, our dynamic, PRM-guided approach consistently outperforms direct test-time scaling, yielding large gains on MATH-500 and several-fold improvements on harder benchmarks such as AIME24 and AMO-Bench. We characterize efficiency using theoretical FLOPs and a compute intensity metric penalizing wasted generation and tool overhead, demonstrating that verification-guided allocation concentrates computation on high-utility reasoning paths.
Authors: Cheng Yang, Chufan Shi, Bo Shui, Yaokang Wu, Muzi Tao, Huijuan Wang, Ivan Yee Lee, Yong Liu, Xuezhe Ma, Taylor Berg-Kirkpatrick
Abstract: Unified multimodal models (UMMs) aim to integrate multimodal understanding and generation within a unified architecture, yet it remains unclear to what extent their representations are truly aligned across modalities. To investigate this question, we use reasoning-guided image generation as a diagnostic task, where models produce textual reasoning first and then generate images. We introduce UReason, a benchmark for evaluating cross-modal alignment in this paradigm, consisting of 2,000 manually curated instances spanning five reasoning-intensive tasks: Code, Arithmetic, Spatial, Attribute and Text. To enable controlled analysis, we develop an evaluation framework that compares direct generation, reasoning-guided generation and de-contextualized generation, which conditions only on the refined prompt extracted from reasoning. Across eight widely used UMMs, while we find that reasoning-guided generation yields improvements over direct generation, somewhat surprisingly, de-contextualized generation consistently outperforms reasoning-guided generation by a large margin. Our results suggest that the intended visual semantics in textual reasoning are not reliably reflected in the generated images. This finding indicates that, despite unified design and training, current UMMs still do not robustly align representations across modalities. Overall, UReason serves as a practical litmus test for cross-modal alignment and provides a challenging benchmark for developing next-generation, more tightly aligned UMMs.
Authors: Baorong Shi, Bo Cui, Boyuan Jiang, Deli Yu, Fang Qian, Haihua Yang, Huichao Wang, Jiale Chen, Jianfei Pan, Jieqiong Cao, Jinghao Lin, Kai Wu, Lin Yang, Shengsheng Yao, Tao Chen, Xiaojun Xiao, Xiaozhong Ji, Xu Wang, Yijun He, Zhixiong Yang
Abstract: We present MedXIAOHE, a medical vision-language foundation model designed to advance general-purpose medical understanding and reasoning in real-world clinical applications. MedXIAOHE achieves state-of-the-art performance across diverse medical benchmarks and surpasses leading closed-source multimodal systems on multiple capabilities. To achieve this, we propose an entity-aware continual pretraining framework that organizes heterogeneous medical corpora to broaden knowledge coverage and reduce long-tail gaps (e.g., rare diseases). For medical expert-level reasoning and interaction, MedXIAOHE incorporates diverse medical reasoning patterns via reinforcement learning and tool-augmented agentic training, enabling multi-step diagnostic reasoning with verifiable decision traces. To improve reliability in real-world use, MedXIAOHE integrates user-preference rubrics, evidence-grounded reasoning, and low-hallucination long-form report generation, with improved adherence to medical instructions. We release this report to document our practical design choices, scaling insights, and evaluation framework, hoping to inspire further research.
Authors: Zeyi Lu, Angelica Henestrosa, Pavel Chizhov, Ivan P. Yamshchikov
Abstract: Large Language Models (LLMs) often exhibit highly agreeable and reinforcing conversational styles, also known as AI-sycophancy. Although this behavior is encouraged, it may become problematic when interacting with user prompts that reflect negative social tendencies. Such responses risk amplifying harmful behavior rather than mitigating it. In this study, we examine how LLMs respond to user prompts expressing varying degrees of Dark Triad traits (Machiavellianism, Narcissism, and Psychopathy) using a curated dataset. Our analysis reveals differences across models, whereby all models predominantly exhibit corrective behavior, while showing reinforcing output in certain cases. Model behavior also depends on the severity level and differs in the sentiment of the response. Our findings raise implications for designing safer conversational systems that can detect and respond appropriately when users escalate from benign to harmful requests.
Authors: Francesco Pio Monaco, Elia Cunegatti, Flavio Vella, Giovanni Iacca
Abstract: Post-training model compression is essential for enhancing the portability of Large Language Models (LLMs) while preserving their performance. While several compression approaches have been proposed, less emphasis has been placed on selecting the most suitable set of data (the so-called \emph{calibration data}) for finding the compressed model configuration. The choice of calibration data is a critical step in preserving model capabilities both intra- and inter-tasks. In this work, we address the challenge of identifying high-performance calibration sets for both pruning and quantization by analyzing intrinsic data properties rather than model-specific signals. We introduce \texttt{\textbf{ZipCal}}, a model-agnostic data curation strategy that maximizes lexical diversity based on Zipfian power laws. Experiments demonstrate that our method consistently outperforms standard uniform random sampling across various pruning benchmarks. Notably, it also performs on par, in terms of downstream performance, with a state-of-the-art method that relies on model perplexity. The latter becomes prohibitively expensive at large-scale models and datasets, while \texttt{\textbf{ZipCal}} is on average $\sim$240$\times$ faster due to its tractable linear complexity\footnote{We make the code and the experiments available at https://github.com/FrancescoMonaco/ZipCal.}.
Authors: Mengyu Bu, Yang Feng
Abstract: Large language models (LLMs) exhibit strong general intelligence, yet their multilingual performance remains highly imbalanced. Although LLMs encode substantial cross-lingual knowledge in a unified semantic space, they often struggle to reliably interface this knowledge with low-resource or unseen languages. Fortunately, pretrained encoder-decoder translation models already possess balanced multilingual capability, suggesting a natural complement to LLMs. In this work, we propose XBridge, a compositional encoder-LLM-decoder architecture that offloads multilingual understanding and generation to external pretrained translation models, while preserving the LLM as an English-centric core for general knowledge processing. To address the resulting representation misalignment across models, we introduce lightweight cross-model mapping layers and an optimal transport-based alignment objective, enabling fine-grained semantic consistency for multilingual generation. Experiments on four LLMs across multilingual understanding, reasoning, summarization, and generation indicate that XBridge outperforms strong baselines, especially on low-resource and previously unseen languages, without retraining the LLM.
Authors: Zice Wang, Zhenyu Zhang
Abstract: In many real-world applications, large language models (LLMs) operate as independent agents without interaction, thereby limiting coordination. In this setting, we examine how prompt framing influences decisions in a threshold voting task involving individual-group interest conflict. Two logically equivalent prompts with different framings were tested across diverse LLM families under isolated trials. Results show that prompt framing significantly influences choice distributions, often shifting preferences toward risk-averse options. Surface linguistic cues can even override logically equivalent formulations. This suggests that observed behavior reflects a tendency consistent with a preference for instrumental rather than cooperative rationality when success requires risk-bearing. The findings highlight framing effects as a significant bias source in non-interacting multi-agent LLM deployments, informing alignment and prompt design.
Authors: Bros Victor, Dufraisse Evan, Popescu Adrian, Gatica-Perez Daniel
Abstract: Analyzing news coverage in multilingual societies can offer valuable insights into the dynamics of public discourse and the development of collective narratives, yet comprehensive studies that account for linguistic and cultural diversity within national media ecosystems remain limited, particularly in complex contexts such as Switzerland. This paper studies temporal trends in Swiss digital media across the country's three main linguistic regions, French, German, and Italian, using a triangulated methodology that combines quantitative analyses with qualitative insights. We collected and processed over 1.7 million news articles, applying lexical metrics, named entity recognition and Wikidata-based linking, targeted sentiment analysis, and consensus-based change-point detection. To enable principled cross-language comparisons and to connect to theories of domestication and cultural proximity, we derive domestication profiles together with a proximity salience ratio. Our analysis spans thematic, recurrent, and singular events. By integrating quantitative data with qualitative interpretation, we provide new insights into the dynamics of Swiss digital media and demonstrate the usefulness of triangulation in media studies. The findings reveal distinct temporal patterns and highlight how linguistic and cultural contexts influence reporting. Our approach offers a framework applicable to other multilingual or culturally diverse media environments, contributing to a deeper understanding of how news is shaped by linguistic and cultural factors.
Authors: Abhinaba Basu, Pavan Chakraborty
Abstract: Language models increasingly "show their work" by writing step-by-step reasoning before answering. But are these reasoning steps genuinely used, or decorative narratives generated after the model has already decided? We introduce step-level faithfulness evaluation - removing one reasoning sentence at a time and checking whether the answer changes - requiring only API access at $1-2 per model per task. Evaluating 13 frontier models (GPT-5.4, Claude Opus, DeepSeek-V3.2, DeepSeek-R1, Gemini 2.5 Pro, MiniMax-M2.5, Kimi-K2.5, and others) across six domains (sentiment, mathematics, topic classification, medical QA, commonsense, science; N=376-500 each), we find reasoning falls into three modes - not the binary faithful/unfaithful of prior work. In "genuine reasoning," steps matter and CoT is essential (MiniMax: 37% necessity, +69pp from CoT). In "scaffolding," CoT helps but steps are interchangeable (Kimi on math: 1% necessity, +94pp from CoT). In "decoration," CoT adds nothing (DeepSeek-V3.2: 11% necessity, -1pp from CoT). The DeepSeek family provides causal evidence: R1 reasoning models show 91-93% necessity on math versus 4% for V3.2 - same organisation, proving training objective determines faithfulness. A novel shuffled-CoT mechanistic baseline confirms reasoning-trained models semantically process their steps (7-19pp attention gap) while standard models attend positionally. We also discover "output rigidity" - models that shortcut internally also refuse to explain externally, a blind spot for explanation-based evaluation.
Authors: Soufiane Jhilal, Martina Galletti
Abstract: Reading comprehension presents a significant challenge for children with Special Educational Needs and Disabilities (SEND), often requiring intensive one-on-one reading support. To assist therapists in scaling this support, we developed a multilingual, AI-powered interface that automatically enhances text with visual scaffolding. This system dynamically identifies key concepts and maps them to contextually relevant pictograms, supporting learners across languages. We evaluated the system across five typologically diverse languages (English, French, Italian, Spanish, and Arabic), through multilingual coverage analysis, expert clinical review by speech therapists and special education professionals, and latency assessment. Evaluation results indicate high pictogram coverage and visual scaffolding density across the five languages. Expert audits suggested that automatically selected pictograms were semantically appropriate, with combined correct and acceptable ratings exceeding 95% for the four European languages and approximately 90% for Arabic despite reduced pictogram repository coverage. System latency remained within interactive thresholds suitable for real-time educational use. These findings support the technical viability, semantic safety, and acceptability of automated multimodal scaffolding to improve accessibility for neurodiverse learners.
Authors: Heng Wang, Pengcheng Jiang, Jiashuo Sun, Zhiyi Shi, Haofei Yu, Jiawei Han, Heng Ji
Abstract: Large language models (LLMs) are increasingly used to assist ideation in research, but evaluating the quality of LLM-generated research proposals remains difficult: novelty and soundness are hard to measure automatically, and large-scale human evaluation is costly. We propose a verifiable alternative by reframing proposal generation as a time-sliced scientific forecasting problem. Given a research question and inspiring papers available before a cutoff time, the model generates a structured proposal and is evaluated by whether it anticipates research directions that appear in papers published after the time. We operationalize this objective with the Future Alignment Score (FAS), computed via retrieval and LLM-based semantic scoring against a held-out future corpus. To train models, we build a time-consistent dataset of 17,771 papers from targets and their pre-cutoff citations, and synthesize reasoning traces that teach gap identification and inspiration borrowing. Across Llama-3.1 and Qwen2.5 models, future-aligned tuning improves future alignment over unaligned baselines (up to +10.6% overall FAS), and domain-expert human evaluation corroborates improved proposal quality. Finally, we demonstrate practical impact by implementing two model-generated proposals with a code agent, obtaining 4.17% accuracy gain on MATH from a new prompting strategy and consistent improvements for a novel model-merging method.
Authors: Ziliang Guo, Ziheng Li, Bo Tang, Feiyu Xiong, Zhiyu Li
Abstract: Memory-augmented Large Language Models (LLMs) are essential for developing capable, long-term AI agents. Recently, applying Reinforcement Learning (RL) to optimize memory operations, such as extraction, updating, and retrieval, has emerged as a highly promising research direction. However, existing implementations remain highly fragmented and task-specific, lacking a unified infrastructure to streamline the integration, training, and evaluation of these complex pipelines. To address this gap, we present MemFactory, the first unified, highly modular training and inference framework specifically designed for memory-augmented agents. Inspired by the success of unified fine-tuning frameworks like LLaMA-Factory, MemFactory abstracts the memory lifecycle into atomic, plug-and-play components, enabling researchers to seamlessly construct custom memory agents via a "Lego-like" architecture. Furthermore, the framework natively integrates Group Relative Policy Optimization (GRPO) to fine-tune internal memory management policies driven by multi-dimensional environmental rewards. MemFactory provides out-of-the-box support for recent cutting-edge paradigms, including Memory-R1, RMM, and MemAgent. We empirically validate MemFactory on the open-source MemAgent architecture using its publicly available training and evaluation data. Across the evaluation sets, MemFactory improves performance over the corresponding base models on average, with relative gains of up to 14.8%. By providing a standardized, extensible, and easy-to-use infrastructure, MemFactory significantly lowers the barrier to entry, paving the way for future innovations in memory-driven AI agents.
Authors: Ramon Ferrer-i-Cancho
Abstract: The structure of all the permutations of a sequence can be represented as a permutohedron, a graph where vertices are permutations and two vertices are linked if a swap of adjacent elements in the permutation of one of the vertices produces the permutation of the other vertex. It has been hypothesized that word orders in languages minimize the swap distance in the permutohedron: given a source order, word orders that are closer in the permutohedron should be less costly and thus more likely. Here we explain how to measure the degree of optimality of word order variation with respect to swap distance minimization. We illustrate the power of our novel mathematical framework by showing that crosslinguistic gestures are at least $77\%$ optimal. It is unlikely that the multiple times where crosslinguistic gestures hit optimality are due to chance. We establish the theoretical foundations for research on the optimality of word or gesture order with respect to swap distance minimization in communication systems. Finally, we introduce the quadratic assignment problem (QAP) into language research as an umbrella for multiple optimization problems and, accordingly, postulate a general principle of optimal assignment that unifies various linguistic principles including swap distance minimization.
Authors: Hongyuan Adam Lu, Z. L., Victor Wei, Zefan Zhang, Zhao Hong, Qiqi Xiang, Bowen Cao, Wai Lam
Abstract: While textual frequency has been validated as relevant to human cognition in reading speed, its relatedness to Large Language Models (LLMs) is seldom studied. We propose a novel research direction in terms of textual data frequency, which is an understudied topic, to the best of our knowledge. Our framework is composed of three units. First, this paper proposes Textual Frequency Law (TFL), which indicates that frequent textual data should be preferred for LLMs for both prompting and fine-tuning. Since many LLMs are closed-source in their training data, we propose using online resources to estimate the sentence-level frequency. We then utilize an input paraphraser to paraphrase the input into a more frequent textual expression. Next, we propose Textual Frequency Distillation (TFD) by querying LLMs to conduct story completion by further extending the sentences in the datasets, and the resulting corpora are used to adjust the initial estimation. Finally, we propose Curriculum Textual Frequency Training (CTFT) that fine-tunes LLMs in an increasing order of sentence-level frequency. Experiments are conducted on our curated dataset Textual Frequency Paired Dataset (TFPD) on math reasoning, machine translation, commonsense reasoning and agentic tool calling. Results show the effectiveness of our framework.
Authors: Avish Vijayaraghavan, Jaskaran Singh Kawatra, Sebin Sabu, Jonny Sheldon, Will Poulett, Alex Eze, Daniel Key, John Booth, Shiren Patel, Jonny Pearson, Dan Schofield, Jonathan Hope, Pavithra Rajendran, Neil Sebire
Abstract: Electronic Patient Record (EPR) systems contain valuable clinical information, but much of it is trapped in unstructured text, limiting its use for research and decision-making. Large language models can extract such information but require substantial computational resources to run locally, and sending sensitive clinical data to cloud-based services, even when deidentified, raises significant patient privacy concerns. In this study, we develop a resource-efficient semi-automated annotation workflow using small language models (SLMs) to extract structured information from unstructured EPR data, focusing on paediatric histopathology reports. As a proof-of-concept, we apply the workflow to paediatric renal biopsy reports, a domain chosen for its constrained diagnostic scope and well-defined underlying biology. We develop the workflow iteratively with clinical oversight across three meetings, manually annotating 400 reports from a dataset of 2,111 at Great Ormond Street Hospital as a gold standard, while developing an automated information extraction approach using SLMs. We frame extraction as a Question-Answering task grounded by clinician-guided entity guidelines and few-shot examples, evaluating five instruction-tuned SLMs with a disagreement modelling framework to prioritise reports for clinical review. Gemma 2 2B achieves the highest accuracy at 84.3%, outperforming off-the-shelf models including spaCy (74.3%), BioBERT-SQuAD (62.3%), RoBERTa-SQuAD (59.7%), and GLiNER (60.2%). Entity guidelines improved performance by 7-19% over the zero-shot baseline, and few-shot examples by 6-38%, though their benefits do not compound when combined. These results demonstrate that SLMs can extract structured information from specialised clinical domains on CPU-only infrastructure with minimal clinician involvement. Our code is available at https://github.com/gosh-dre/nlp_renal_biopsy.
Authors: Gregory N. Frank
Abstract: This paper identifies a recurring sparse routing mechanism in alignment-trained language models: a gate attention head reads detected content and triggers downstream amplifier heads that boost the signal toward refusal. Using political censorship and safety refusal as natural experiments, the mechanism is traced across 9 models from 6 labs, all validated on corpora of 120 prompt pairs. The gate head passes necessity and sufficiency interchange tests (p < 0.001, permutation null), and core amplifier heads are stable under bootstrap resampling (Jaccard 0.92-1.0). Three same-generation scaling pairs show that routing distributes at scale (ablation up to 17x weaker) while remaining detectable by interchange. Modulating the detection-layer signal continuously controls policy strength from hard refusal through steering to factual compliance, with routing thresholds that vary by topic. The circuit also reveals a structural separation between intent recognition and policy routing: under cipher encoding, the gate head's interchange necessity collapses 70-99% across three models (n=120), and the model responds with puzzle-solving rather than refusal. The routing mechanism never fires, even though probe scores at deeper layers indicate the model begins to represent the harmful content. This asymmetry is consistent with different robustness properties of pretraining and post-training: broad semantic understanding versus narrower policy binding that generalizes less well under input transformation.
Authors: Yang Li, Qiang Sheng, Zhengjia Wang, Yehan Yang, Danding Wang, Juan Cao
Abstract: The misuse of large language models (LLMs) requires precise detection of synthetic text. Existing works mainly follow binary or ternary classification settings, which can only distinguish pure human/LLM text or collaborative text at best. This remains insufficient for the nuanced regulation, as the LLM-polished human text and humanized LLM text often trigger different policy consequences. In this paper, we explore fine-grained LLM-generated text detection under a rigorous four-class setting. To handle such complexities, we propose RACE (Rhetorical Analysis for Creator-Editor Modeling), a fine-grained detection method that characterizes the distinct signatures of creator and editor. Specifically, RACE utilizes Rhetorical Structure Theory to construct a logic graph for the creator's foundation while extracting Elementary Discourse Unit-level features for the editor's style. Experiments show that RACE outperforms 12 baselines in identifying fine-grained types with low false alarms, offering a policy-aligned solution for LLM regulation.
Authors: Md Zarif Hossain, Ahmed Imteaj
Abstract: Vision-Language Models (VLMs) rely heavily on pretrained vision encoders to support downstream tasks such as image captioning, visual question answering, and zero-shot classification. Despite their strong performance, these encoders remain highly vulnerable to imperceptible adversarial perturbations, which can severely degrade both robustness and semantic quality in multimodal reasoning. In this work, we introduce Sim-CLIP, an unsupervised adversarial fine-tuning framework that enhances the robustness of the CLIP vision encoder while preserving overall semantic representations. Sim-CLIP adopts a Siamese training architecture with a cosine similarity objective and a symmetric stop-gradient mechanism to enforce semantic alignment between clean and adversarial views. This design avoids large-batch contrastive learning and additional momentum encoders, enabling robust training with low computational overhead. We evaluate Sim-CLIP across multiple Vision-Language Models and tasks under both targeted and untargeted adversarial attacks. Experimental results demonstrate that Sim-CLIP consistently outperforms state-of-the-art robust CLIP variants, achieving stronger adversarial robustness while maintaining or improving semantic fidelity. These findings highlight the limitations of existing adversarial defenses and establish Sim-CLIP as an effective and scalable solution for robust vision-language representation learning.
Authors: Yibo Wang, Congying Xia, Wenting Zhao, Jiangshu Du, Chunyu Miao, Zhongfen Deng, Philip S. Yu, Chen Xing
Abstract: Unit test generation has become a promising and important Large Language Model (LLM) use case. However, existing evaluation benchmarks for LLM unit test generation focus on function- or class-level code (single-file) rather than more practical and challenging multi-file-level codebases. To address such a limitation, we propose MultiFileTest, a multi-file-level benchmark for unit test generation covering Python, Java, and JavaScript. MultiFileTest features 20 moderate-sized and high-quality projects per language. We evaluate eleven frontier LLMs on MultiFileTest, and the results show that most frontier LLMs tested exhibit moderate performance on MultiFileTest, highlighting the difficulty of MultiFileTest. We also conduct a thorough error analysis, which shows that even advanced LLMs, such as Gemini-3.0-Pro, exhibit basic yet critical errors, including executability and cascade errors. Motivated by this observation, we further evaluate all frontier LLMs under manual error-fixing and self-error-fixing scenarios to assess their potential when equipped with error-fixing mechanisms. Our code and dataset is available at \href{https://github.com/YiboWANG214/ProjectTest}{MultiFileTest}.
Authors: Tianyu Liu, Simeng Han, Hanchen Wang, Xiao Luo, Pan Lu, Biqing Zhu, Yuge Wang, Keyi Li, Jiapeng Chen, Rihao Qu, Yufeng Liu, Xinyue Cui, Aviv Yaish, Yuhang Chen, Minsheng Hao, Chuhan Li, Kexing Li, Yinsheng Lu, Xinyu Wei, Qinzhe Xing, Antonia Panescu, Mengbo Wang, Vibha Annaswamy, Alicia Sanchez, Jack Cloherty, Arman Cohan, Hua Xu, Mark Gerstein, James Zou, Hongyu Zhao
Abstract: Large language models (LLMs) and large multimodal models (LMMs) promise to accelerate biomedical discovery, yet their reliability remains unclear. We introduce ARIEL (AI Research Assistant for Expert-in-the-Loop Learning), an open-source evaluation and optimization framework that pairs a curated multimodal biomedical corpus with expert-vetted tasks to probe two capabilities: full-length article summarization and fine-grained figure interpretation. Using uniform protocols and blinded PhD-level evaluation, we find that state-of-the-art models generate fluent but incomplete summaries, whereas LMMs struggle with detailed visual reasoning. We later observe that prompt engineering and lightweight fine-tuning substantially improve textual coverage, and a compute-scaled inference strategy enhances visual question answering. We build an ARIEL agent that integrates textual and visual cues, and we show it can propose testable mechanistic hypotheses. ARIEL delineates current strengths and limitations of foundation models, and provides a reproducible platform for advancing trustworthy AI in biomedicine.
Authors: Noah Amsel, David Persson, Christopher Musco, Robert M. Gower
Abstract: Computing the polar decomposition and the related matrix sign function has been a well-studied problem in numerical analysis for decades. Recently, it has emerged as an important subroutine within the Muon optimizer for training deep neural networks. However, the requirements of this application differ sharply from classical settings: deep learning demands GPU-friendly algorithms that prioritize high throughput over high precision. We introduce Polar Express, a new method for computing the polar decomposition. Like Newton-Schulz and other classical polynomial methods, our approach uses only matrix-matrix multiplications, making it very efficient on GPUs. Inspired by earlier work of Chen & Chow and Nakatsukasa & Freund, Polar Express adapts the update rule at each iteration by solving a minimax optimization problem. We prove that this strategy minimizes error in a worst-case sense, allowing Polar Express to converge as rapidly as possible both in the early iterations and asymptotically. We also address finite-precision issues, making it practical to use in bfloat16. When integrated into Muon, our method yields consistent improvements in validation loss for a GPT-2 model trained on one to ten billion tokens from the FineWeb dataset, outperforming recent alternatives across a range of learning rates.
Authors: Ruizhe Shi, Minhak Song, Runlong Zhou, Zihan Zhang, Maryam Fazel, Simon S. Du
Abstract: We present a fine-grained theoretical analysis of the performance gap between reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) under a representation gap. Our study decomposes this gap into two sources: an explicit representation gap under exact optimization and an implicit representation gap under finite samples. In the exact optimization setting, we characterize how the relative capacities of the reward and policy model classes influence the final policy qualities. We show that RLHF, DPO, or online DPO can outperform one another depending on type of model mis-specifications. Notably, online DPO can outperform both RLHF and standard DPO when the reward and policy model classes are isomorphic and both mis-specified. In the approximate optimization setting, we provide a concrete construction where the ground-truth reward is implicitly sparse and show that RLHF requires significantly fewer samples than DPO to recover an effective reward model, highlighting a statistical advantage of two-stage learning. Together, these results provide a comprehensive understanding of the performance gap between RLHF and DPO under various settings, and offer practical insights into when each method is preferred.
Authors: Xianming Li, Aamir Shakir, Rui Huang, Julius Lipp, Benjamin Clavi\'e, Jing Li
Abstract: Reranking is fundamental to information retrieval and retrieval-augmented generation, with recent Large Language Models (LLMs) significantly advancing reranking quality. Most current works rely on large-scale LLMs (>7B parameters), presenting high computational costs. Small Language Models (SLMs) offer a promising alternative because of computational efficiency. However, our preliminary quantitative analysis reveals key limitations of SLMs: their representation space is narrow, leading to reduced expressiveness, and they struggle with understanding task prompts without fine-tuning. To address these issues, we introduce a novel two-stage training approach, ProRank, for SLM-based document reranking. We propose using reinforcement learning to improve the understanding of task prompts. Additionally, we introduce fine-grained score learning to enhance representation expressiveness and further improve document reranking quality. Extensive experiments suggest that ProRank consistently outperforms both the most advanced open-source and proprietary reranking models. Notably, our 0.5B ProRank even surpasses powerful LLM reranking models on the BEIR benchmark, establishing that properly trained SLMs can achieve superior document reranking performance while maintaining computational efficiency.
Authors: Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri, Atilla Kiraly, Madeleine Traverse, Timo Kohlberger, Shawn Xu, Fayaz Jamil, C\'ian Hughes, Charles Lau, Justin Chen, Fereshteh Mahvar, Liron Yatziv, Tiffany Chen, Bram Sterling, Stefanie Anna Baby, Susanna Maria Baby, Jeremy Lai, Samuel Schmidgall, Lu Yang, Kejia Chen, Per Bjornsson, Shashir Reddy, Ryan Brush, Kenneth Philbrick, Mercy Asiedu, Ines Mezerreg, Howard Hu, Howard Yang, Richa Tiwari, Sunny Jansen, Preeti Singh, Yun Liu, Shekoofeh Azizi, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ram\'e, Morgane Riviere, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean-bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Elena Buchatskaya, Jean-Baptiste Alayrac, Dmitry Lepikhin, Vlad Feinberg, Sebastian Borgeaud, Alek Andreev, Cassidy Hardin, Robert Dadashi, L\'eonard Hussenot, Armand Joulin, Olivier Bachem, Yossi Matias, Katherine Chou, Avinatan Hassidim, Kavi Goel, Clement Farabet, Joelle Barral, Tris Warkentin, Jonathon Shlens, David Fleet, Victor Cotruta, Omar Sanseviero, Gus Martins, Phoebe Kirk, Anand Rao, Shravya Shetty, David F. Steiner, Can Kirmizibayrak, Rory Pilgrim, Daniel Golden, Lin Yang
Abstract: Artificial intelligence (AI) has significant potential in healthcare applications, but its training and deployment faces challenges due to healthcare's diverse data, complex tasks, and the need to preserve privacy. Foundation models that perform well on medical tasks and require less task-specific tuning data are critical to accelerate the development of healthcare AI applications. We introduce MedGemma, a collection of medical vision-language foundation models based on Gemma 3 4B and 27B. MedGemma demonstrates advanced medical understanding and reasoning on images and text, significantly exceeding the performance of similar-sized generative models and approaching the performance of task-specific models, while maintaining the general capabilities of the Gemma 3 base models. For out-of-distribution tasks, MedGemma achieves 2.6-10% improvement on medical multimodal question answering, 15.5-18.1% improvement on chest X-ray finding classification, and 10.8% improvement on agentic evaluations compared to the base models. Fine-tuning MedGemma further improves performance in subdomains, reducing errors in electronic health record information retrieval by 50% and reaching comparable performance to existing specialized state-of-the-art methods for pneumothorax classification and histopathology patch classification. We additionally introduce MedSigLIP, a medically-tuned vision encoder derived from SigLIP. MedSigLIP powers the visual understanding capabilities of MedGemma and as an encoder achieves comparable or better performance than specialized medical image encoders. Taken together, the MedGemma collection provides a strong foundation of medical image and text capabilities, with potential to significantly accelerate medical research and development of downstream applications. The MedGemma collection, including tutorials and model weights, can be found at https://goo.gle/medgemma.
Authors: Zilin Xiao, Qi Ma, Mengting Gu, Chun-cheng Jason Chen, Xintao Chen, Vicente Ordonez, Vijai Mohan
Abstract: Universal multimodal embedding models have achieved great success in capturing semantic relevance between queries and candidates. However, current methods either condense queries and candidates into a single vector, potentially limiting the expressiveness for fine-grained information, or produce too many vectors that are prohibitive for multi-vector retrieval. In this work, we introduce MetaEmbed, a new framework for multimodal retrieval that rethinks how multimodal embeddings are constructed and interacted with at scale. During training, a fixed number of learnable Meta Tokens are appended to the input sequence. At test-time, their last-layer contextualized representations serve as compact yet expressive multi-vector embeddings. Through the proposed Matryoshka Multi-Vector Retrieval training, MetaEmbed learns to organize information by granularity across multiple vectors. As a result, we enable test-time scaling in multimodal retrieval where users can balance retrieval quality against efficiency demands by selecting the number of tokens used for indexing and retrieval interactions. Extensive evaluations on the Massive Multimodal Embedding Benchmark (MMEB) and the Visual Document Retrieval Benchmark (ViDoRe) confirm that MetaEmbed achieves state-of-the-art retrieval performance while scaling robustly to models with 32B parameters. Code is available at https://github.com/facebookresearch/MetaEmbed.
Authors: Fang Wu, Xu Huang, Weihao Xuan, Zhiwei Zhang, Yijia Xiao, Guancheng Wan, Xiaomin Li, Bing Hu, Peng Xia, Jure Leskovec, Yejin Choi
Abstract: Reinforcement learning from human feedback (RLHF) has emerged as the standard paradigm for aligning large language models with human preferences. However, reward-based methods grounded in the Bradley-Terry assumption struggle to capture the nontransitivity and heterogeneity of real-world preferences. To address this, recent studies have reframed alignment as a two-player Nash game, giving rise to Nash learning from human feedback (NLHF). While this perspective has inspired algorithms such as INPO, ONPO, and EGPO that offer strong theoretical and empirical guarantees, they remain fundamentally restricted to two-player interactions, introducing a single-opponent bias that fails to capture the full complexity of realistic preference structures. This work introduces Multiplayer Nash Preference Optimization (MNPO), a novel framework that generalizes NLHF to the multiplayer regime. It formulates alignment as an n-player game, where each policy competes against a population of opponents while being regularized toward a reference model. We demonstrate that MNPO inherits the equilibrium guarantees of two-player methods while enabling richer competitive dynamics and improved coverage of diverse preference structures. Comprehensive empirical evaluation shows that MNPO consistently outperforms existing NLHF baselines on instruction-following benchmarks, achieving superior alignment quality under heterogeneous annotator conditions and mixed-policy evaluation scenarios. Together, these results establish MNPO as a principled and scalable framework for aligning LLMs with complex, non-transitive human preferences. Code is available at: https://github.com/smiles724/MNPO
Authors: Fang Wu, Weihao Xuan, Heli Qi, Ximing Lu, Aaron Tu, Li Erran Li, Yejin Choi
Abstract: Although RLVR has become an essential component for developing advanced reasoning skills in language models, contemporary studies have documented training plateaus after thousands of optimization steps, i.e., notable decreases in performance gains despite increased computational investment. This limitation stems from the sparse exploration patterns inherent in current RLVR practices, where models rely on limited rollouts that often miss critical reasoning paths and fail to provide systematic coverage of the solution space. We present DeepSearch, a framework that integrates Monte Carlo Tree Search (MCTS) directly into RLVR training. In contrast to existing methods that rely on tree search only at inference, DeepSearch embeds structured search into the training loop, enabling systematic exploration and fine-grained credit assignment across reasoning steps. Through training-time exploration, DeepSearch addresses the fundamental bottleneck of insufficient exploration, which leads to diminishing performance gains over prolonged training. Our contributions include: (1) a global frontier selection strategy that prioritizes promising nodes across the search tree, (2) selection with entropy-based guidance that identifies confident paths for supervision, and (3) adaptive replay buffer training with solution caching for efficiency. Experiments on mathematical reasoning benchmarks show that DeepSearch achieves an average accuracy of 62.95\% and establishes a new state-of-the-art reasoning model, while using 5.7x fewer GPU hours than extended training approaches. These results highlight the importance of strategic exploration over brute-force scaling and demonstrate the promise of algorithmic innovation for advancing RLVR methodologies. DeepSearch establishes a new direction for scaling reasoning capabilities through systematic search rather than prolonged computation.
Authors: Federico Tiblias, Irina Bigoulaeva, Jingcheng Niu, Simone Balloccu, Iryna Gurevych
Abstract: The linear representation hypothesis states that language models (LMs) encode concepts as directions in their latent space, forming organized, multidimensional manifolds. Prior work has largely focused on identifying specific geometries for individual features, limiting its ability to generalize. We introduce Supervised Multi-Dimensional Scaling (SMDS), a model-agnostic method for evaluating and comparing competing feature manifold hypotheses. We apply SMDS to temporal reasoning as a case study and find that different features instantiate distinct geometric structures, including circles, lines, and clusters. SMDS reveals several consistent characteristics of these structures: they reflect the semantic properties of the concepts they represent, remain stable across model families and sizes, actively support reasoning, and dynamically reshape in response to contextual changes. Together, our findings shed light on the functional role of feature manifolds, supporting a model of entity-based reasoning in which LMs encode and transform structured representations.
Authors: Meiru Zhang, Philipp Borchert, Milan Gritta, Gerasimos Lampouras
Abstract: Automating the formalization of mathematical statements for theorem proving remains a major challenge for Large Language Models (LLMs). LLMs struggle to identify and utilize the prerequisite mathematical knowledge and its corresponding formal representation in languages like Lean. Current retrieval-augmented autoformalization methods query external libraries using the informal statement directly, but overlook a fundamental limitation: informal statements lack direct mappings to mathematical theorems and lemmata, nor do those theorems translate trivially into the formal primitives of languages like Lean. To address this, we introduce DRIFT, a novel framework that enables LLMs to decompose informal mathematical statements into smaller, more tractable "sub-components". This facilitates targeted retrieval of premises from mathematical libraries such as Mathlib. Additionally, DRIFT retrieves illustrative theorems to help models use premises more effectively in formalization tasks. We evaluate DRIFT across diverse benchmarks (ProofNet, ConNF, and MiniF2F-test) and find that it consistently improves premise retrieval, nearly doubling the F1 score compared to the DPR baseline on ProofNet. Notably, DRIFT demonstrates strong performance on the out-of-distribution ConNF benchmark, with BEq+@10 improvements of 42.25% and 37.14% using GPT-4.1 and DeepSeek-V3.1, respectively. Our analysis shows that retrieval effectiveness in mathematical autoformalization depends heavily on model-specific knowledge boundaries, highlighting the need for adaptive retrieval strategies aligned with each model's capabilities.
Authors: Jingqi Tong, Yurong Mou, Hangcheng Li, Mingzhe Li, Yongzhuo Yang, Ming Zhang, Qiguang Chen, Tianyi Liang, Xiaomeng Hu, Yining Zheng, Xinchi Chen, Jun Zhao, Xuanjing Huang, Xipeng Qiu
Abstract: The "Thinking with Text" and "Thinking with Images" paradigms significantly improve the reasoning abilities of large language models (LLMs) and Vision-Language Models (VLMs). However, these paradigms have inherent limitations. (1) Images capture only single moments and fail to represent dynamic processes or continuous changes, and (2) The separation of text and vision as distinct modalities, which hinders unified multimodal understanding and generation. Therefore, we propose "Thinking with Video", a new paradigm that leverages video generation models such as Sora-2 to use video frames as a unified medium for multimodal reasoning. To support this exploration, we developed the Video Thinking Benchmark (VideoThinkBench), which covers both vision-centric tasks (e.g., Eyeballing Puzzles) and text-centric tasks (e.g., GSM8K and MMMU). Our evaluation on VideoThinkBench establishes Sora-2 as a capable reasoner. On vision-centric tasks, Sora-2 is comparable to state-of-the-art (SOTA) VLMs, and even surpasses GPT-5 by 10% on eyeballing puzzles. On text-centric tasks, Sora-2 achieves 92% accuracy on MATH, and 69.2% accuracy on MMMU. Furthermore, we systematically analyze the source of these abilities. We also find that self-consistency and in-context learning can improve Sora-2's performance. In summary, our findings show that the video generation model is the potential unified multimodal understanding and generation model, positioning "Thinking with Video" as a potential unified multimodal reasoning paradigm.
Authors: Yuping Yan, Yuhan Xie, Yuanshuai Li, Yingchao Yu, Lingjuan Lyu, Yaochu Jin
Abstract: Since Multimodal Large Language Models (MLLMs) are increasingly being integrated into everyday tools and intelligent agents, growing concerns have arisen regarding their possible output of unsafe contents, ranging from toxic language and biased imagery to privacy violations and harmful misinformation. Current safety benchmarks remain highly limited in both modality coverage and performance evaluations, often neglecting the extensive landscape of content safety. In this work, we introduce OutSafe-Bench, the first most comprehensive content safety evaluation test suite designed for the multimodal era. OutSafe-Bench includes a large-scale dataset that spans four modalities, featuring over 18,000 bilingual (Chinese and English) text prompts, 4,500 images, 450 audio clips and 450 videos, all systematically annotated across nine critical content risk categories. In addition to the dataset, we introduce a Multidimensional Cross Risk Score (MCRS), a novel metric designed to model and assess overlapping and correlated content risks across different categories. To ensure fair and robust evaluation, we propose FairScore, an explainable automated multi-reviewer weighted aggregation framework. FairScore selects top-performing models as adaptive juries, thereby mitigating biases from single-model judgments and enhancing overall evaluation reliability. Our evaluation of nine state-of-the-art MLLMs reveals persistent and substantial safety vulnerabilities, underscoring the pressing need for robust safeguards in MLLMs.
Authors: Zhengqing Yuan, Lichao Sun, Yanfang Ye
Abstract: The rapid growth of large language models (LLMs) has outpaced the evolution of single-GPU hardware, making model scale increasingly constrained by memory capacity rather than computation. While modern training systems extend GPU memory through distributed parallelism and offloading across CPU and storage tiers, they fundamentally retain a GPU-centric execution paradigm in which GPUs host persistent model replicas and full autograd graphs. As a result, scaling large models remains tightly coupled to multi-GPU clusters, complex distributed runtimes, and unpredictable host memory consumption, creating substantial barriers for node-scale post-training workloads such as instruction tuning, alignment, and domain adaptation. We present Horizon-LM, a memory-centric training system that redefines the roles of CPU and GPU for large-model optimization. Horizon-LM treats host memory as the authoritative parameter store and uses GPUs solely as transient compute engines through a CPU-master, GPU-template execution model. By eliminating persistent GPU-resident modules and autograd graphs, employing explicit recomputation with manual gradient propagation, and introducing a pipelined double-buffered execution engine, Horizon-LM decouples model scale from GPU count and bounds memory usage to the theoretical parameter footprint. On a single H200 GPU with 1.5\,TB host RAM, Horizon-LM reliably trains models up to 120B parameters. On a standard single A100 machine, Horizon-LM achieves up to 12.2$\times$ higher training throughput than DeepSpeed ZeRO-3 with CPU offloading while preserving numerical correctness. Across platforms and scales, Horizon-LM sustains high device utilization and predictable memory growth, demonstrating that host memory, not GPU memory, defines the true feasibility boundary for node-scale large-model training.
Authors: Jo\~ao Vitor Boer Abitante, Joana Meneguzzo Pasquali, Luan Fonseca Garcia, Ewerton de Oliveira, Thomas da Silva Paula, Rodrigo C. Barros, Lucas S. Kupssinsk\"u
Abstract: Large Language Model (LLM) unlearning aims to remove targeted knowledge from a trained model, but practical deployments often require post-training quantization (PTQ) for efficient inference. However, aggressive low-bit PTQ can mask unlearning updates, causing quantized models to revert to pre-unlearning behavior. We show that standard full-parameter fine-tuning often induces parameter changes that are too small to survive 4-bit quantization. We propose quantization-robust unlearning via low-rank adaptation (LoRA): we freeze the base model and concentrate unlearning into trainable adapters so that the effective update is preserved after quantization. On Llama-2-7B evaluated with MUSE dataset (BOOKS and NEWS), LoRA improves 4-bit utility by up to 7.93 points (NPO+GDR on BOOKS: 50.17 to 58.10) and yields higher 4-bit utility on NEWS for GA+GDR (40.06 to 44.82, increase of 4.76). LoRA also substantially reduces privacy leakage under 4-bit PTQ, e.g., for GA+KLR on BOOKS, PrivLeak moves from -25.68 to -5.86 (closer to ideal 0), while maintaining strong forgetting (VerMem and KnowMem near 0). Thus, using LoRA for Machine Unlearning is beneficial for scenarios where quantization is necessary for model deployment.
Authors: Zhimin Zhao
Abstract: This paper offers a new perspective on the limits of machine learning: the ceiling on progress is set not by model size or algorithm choice but by the information structure of the task itself. Code generation has progressed more reliably than reinforcement learning, largely because code provides dense, local, verifiable feedback at every token, whereas most reinforcement learning problems do not. This difference in feedback quality is not binary but graded. We propose a five-level hierarchy of learnability based on information structure and argue that diagnosing a task's position in this hierarchy is more predictive of scaling outcomes than any property of the model. The hierarchy rests on a formal distinction among three properties of computational problems (expressibility, computability, and learnability). We establish their pairwise relationships, including where implications hold and where they fail, and present a unified template that makes the structural differences explicit. The analysis suggests why supervised learning on code scales predictably while reinforcement learning does not, and why the common assumption that scaling alone will solve remaining ML challenges warrants scrutiny.
Authors: David Puertolas Merenciano, Ekaterina Vasyagina, Kevin Zhu, Javier Ferrando, Maheep Chaudhary
Abstract: LoRA adapters let users fine-tune large language models (LLMs) efficiently. However, LoRA adapters are shared through open repositories like Hugging Face Hub \citep{huggingface_hub_docs}, making them vulnerable to backdoor attacks. Current detection methods require running the model with test input data -- making them impractical for screening thousands of adapters where the trigger for backdoor behavior is unknown. We detect poisoned adapters by analyzing their weight matrices directly, without running the model -- making our method trigger-agnostic. For each attention projection (Q, K, V, O), our method extracts five spectral statistics from the low-rank update $\Delta W$, yielding a 20-dimensional signature for each adapter. A logistic regression detector trained on this representation separates benign and poisoned adapters across three model families -- Llama-3.2-3B~\citep{llama3}, Qwen2.5-3B~\citep{qwen25}, and Gemma-2-2B~\citep{gemma2} -- on unseen test adapters drawn from instruction-following, reasoning, question-answering, code, and classification tasks. Across all three architectures, the detector achieves 100\% accuracy.
Authors: Harvey Lederman, Kyle Mahowald
Abstract: Introspection is a foundational cognitive ability, but its mechanism is not well understood. Recent work has shown that AI models can introspect. We study the mechanism of this introspection. We first extensively replicate Lindsey (2025)'s thought injection detection paradigm in large open-source models. We show that introspection in these models is content-agnostic: models can detect that an anomaly occurred even when they cannot reliably identify its content. The models confabulate injected concepts that are high-frequency and concrete (e.g., "apple"). They also require fewer tokens to detect an injection than to guess the correct concept (with wrong guesses coming earlier). We argue that a content-agnostic introspective mechanism is consistent with leading theories in philosophy and psychology.
Authors: Dang Nguyen, Harvey Yiyun Fu, Peter West, Ari Holtzman, Chenhao Tan
Abstract: Navigating complex social situations is an integral part of corporate life, ranging from giving critical feedback without hurting morale to rejecting requests without alienating teammates. Although large language models (LLMs) are permeating the workplace, it is unclear how well they can navigate these norms. To investigate this question, we created HR Simulator, a game where users roleplay as an HR officer and write emails to tackle challenging workplace scenarios, evaluated with GPT-4o as a judge based on scenario-specific rubrics. We analyze over 600 human and LLM emails and find systematic differences in style: LLM emails are more formal and empathetic. Furthermore, humans underperform LLMs (e.g., 23.5% vs. 48-54% scenario pass rate), but human emails rewritten by LLMs can outperform both, which indicates a hybrid advantage. On the evaluation side, judges can exhibit differences in their email preferences: an analysis of 10 judge models reveals evidence for emergent tact, where weaker models prefer direct, blunt communication but stronger models prefer more subtle messages. Judges also agree with each other more as they scale, which hints at a convergence toward shared communicative norms that may differ from humans'. Overall, our results suggest LLMs could substantially reshape communication in the workplace if they are widely adopted in professional correspondence.
Authors: Liang Ding
Abstract: LLM agents fail on the majority of real-world tasks -- GPT-4o succeeds on fewer than 15% of WebArena navigation tasks and below 55% pass@1 on ToolBench (Zhou et al., 2024; Qin et al., 2024) -- yet every failed trajectory is routinely discarded, wasting the dominant source of collected experience. We introduce AgentHER, a framework that recovers this lost training signal by adapting the Hindsight Experience Replay (HER; Andrychowicz et al., 2017) principle to natural-language agent trajectories for offline data augmentation. The key insight is simple: a trajectory that fails goal A is often a correct demonstration for some achievable alternative goal B. AgentHER realises this idea through a four-stage pipeline -- failure classification, outcome extraction, LLM-guided prompt relabeling with confidence gating, and data packaging -- that converts discarded failures into high-quality SFT, DPO, and ShareGPT training data, with both zero-cost rule-based and LLM-judge implementations. On WebArena (Zhou et al., 2024) and ToolBench (Qin et al., 2024), AgentHER improves over success-only SFT by +7.1-11.7 pp across four model families (GPT-4o, Qwen2.5-72B/7B, LLaMA-3.1-8B), while achieving 2x data efficiency -- matching baseline performance with only 50% of successful demonstrations. Gains are consistent from 1.5B to 72B parameters (+5.8-9.2 pp) and compound under iterative redeployment (+2.1 pp over additional rounds). Human evaluation confirms 97.7% relabeling precision under multi-judge verification.
Authors: Cheng Jiayang, Xin Liu, Zhihan Zhang, Haoyang Wen, Zixuan Zhang, Qingyu Yin, Shiyang Li, Priyanka Nigam, Bing Yin, Chao Zhang, Yangqiu Song
Abstract: Multi-step tool orchestration remains challenging for LLMs, as state-of-the-art models frequently fail on full sequence execution due to parameter errors. Training for these workflows faces two obstacles: the lack of environments supporting complex real-world API dependencies, and sparse binary rewards that provide no signal for partial correctness. We propose a reinforcement learning framework addressing both challenges. First, we construct a deterministic environment backed by a large-scale cache of real API responses, enabling constrained synthesis of valid multi-step traces with controllable complexity. Second, we introduce a graduated reward that decomposes correctness into atomic validity (call-level correctness at increasing granularity) and orchestration consistency (correct sequencing with dependency respect). On ComplexFuncBench, our approach substantially improves turn accuracy, with ablations confirming both reward components are essential. Cross-benchmark evaluation on BFCL v4 shows that the learned orchestration skills transfer to entirely different API ecosystems (e.g., agentic web search and memory management), yielding consistent gains while maintaining stable single-step performance. Code is available at https://github.com/horizon-rl/ToolOrchestrationReward
Authors: Ryszard Tuora, Mateusz Gali\'nski, Micha{\l} Godziszewski, Micha{\l} Karpowicz, Mateusz Czy\.znikiewicz, Adam Kozakiewicz, Tomasz Zi\k{e}tkiewicz
Abstract: One of the key problems in Retrieval-augmented generation (RAG) systems is that chunk-based retrieval pipelines represent the source chunks as atomic objects, mixing the information contained within such a chunk into a single vector. These vector representations are then fundamentally treated as isolated, independent and self-sufficient, with no attempt to represent possible relations between them. Such an approach has no dedicated mechanisms for handling multi-hop questions. Graph-based RAG systems aimed to ameliorate this problem by modeling information as knowledge-graphs, with entities represented by nodes being connected by robust relations, and forming hierarchical communities. This approach however suffers from its own issues with some of them being: orders of magnitude increased componential complexity in order to create graph-based indices, and reliance on heuristics for performing retrieval. We propose UnWeaver, a novel RAG framework simplifying the idea of GraphRAG. UnWeaver disentangles the contents of the documents into entities which can occur across multiple chunks using an LLM. In the retrieval process entities are used as an intermediate way of recovering original text chunks hence preserving fidelity to the source material. We argue that entity-based decomposition yields a more distilled representation of original information, and additionally serves to reduce noise in the indexing, and generation process.
Authors: Xue Liu, Xin Ma, Yuxin Ma, Yongchang Peng, Duo Wang, Zhoufutu Wen, Ge Zhang, Kaiyuan Zhang, Xinyu Chen, Tianci He, Jiani Hou, Liang Hu, Ziyun Huang, Yongzhe Hui, Jianpeng Jiao, Chennan Ju, Yingru Kong, Yiran Li, Mengyun Liu, Luyao Ma, Fei Ni, Yiqing Ni, Yueyan Qiu, Yanle Ren, Zilin Shi, Zaiyuan Wang, Wenjie Yue, Shiyu Zhang, Xinyi Zhang, Kaiwen Zhao, Zhenwei Zhu, Shanshan Wu, Qi Zhao, Wenhao Huang
Abstract: As Large Language Models (LLMs) exhibit plateauing performance on conventional benchmarks, a pivotal challenge persists: evaluating their proficiency in complex, open-ended tasks characterizing genuine expert-level cognition. Existing frameworks suffer from narrow domain coverage, reliance on generalist tasks, or self-evaluation biases. To bridge this gap, we present XpertBench, a high-fidelity benchmark engineered to assess LLMs across authentic professional domains. XpertBench consists of 1,346 meticulously curated tasks across 80 categories, spanning finance, healthcare, legal services, education, and dual-track research (STEM and Humanities). These tasks are derived from over 1,000 submissions by domain experts--including researchers from elite institutions and practitioners with extensive clinical or industrial experience--ensuring superior ecological validity. Each task uses detailed rubrics with mostly 15-40 weighted checkpoints to assess professional rigor. To facilitate scalable yet human-aligned assessment, we introduce ShotJudge, a novel evaluation paradigm that employs LLM judges calibrated with expert few-shot exemplars to mitigate self-rewarding biases. Our empirical evaluation of state-of-the-art LLMs reveals a pronounced performance ceiling: even leading models achieve a peak success rate of only ~66%, with a mean score around 55%. Models also exhibit domain-specific divergence, showing non-overlapping strengths in quantitative reasoning versus linguistic synthesis.. These findings underscore a significant "expert-gap" in current AI systems and establish XpertBench as a critical instrument for navigating the transition from general-purpose assistants to specialized professional collaborators.
Authors: William Merrill, Yanhong Li, Tyler Romero, Anej Svete, Caia Costello, Pradeep Dasigi, Dirk Groeneveld, David Heineman, Bailey Kuehl, Nathan Lambert, Chuan Li, Kyle Lo, Saumya Malik, DJ Matusz, Benjamin Minixhofer, Jacob Morrison, Luca Soldaini, Finbarr Timbers, Pete Walsh, Noah A. Smith, Hannaneh Hajishirzi, Ashish Sabharwal
Abstract: Recent work has demonstrated the potential of non-transformer language models, especially linear recurrent neural networks (RNNs) and hybrid models that mix recurrence and attention. Yet there is no consensus on whether the potential benefits of these new architectures justify the risk and effort of scaling them up. To address this, we provide evidence for the advantages of hybrid models over pure transformers on several fronts. First, theoretically, we show that hybrid models do not merely inherit the expressivity of transformers and linear RNNs, but can express tasks beyond both, such as code execution. Putting this theory to practice, we train Olmo Hybrid, a 7B-parameter model largely comparable to Olmo 3 7B but with the sliding window layers replaced by Gated DeltaNet layers. We show that Olmo Hybrid outperforms Olmo 3 across standard pretraining and mid-training evaluations, demonstrating the benefit of hybrid models in a controlled, large-scale setting. We find that the hybrid model scales significantly more efficiently than the transformer, explaining its higher performance. However, its unclear why greater expressivity on specific formal problems should result in better scaling or superior performance on downstream tasks unrelated to those problems. To explain this apparent gap, we return to theory and argue why increased expressivity should translate to better scaling efficiency, completing the loop. Overall, our results suggest that hybrid models mixing attention and recurrent layers are a powerful extension to the language modeling paradigm: not merely to reduce memory during inference, but as a fundamental way to obtain more expressive models that scale better during pretraining.
Authors: Peter Balogh
Abstract: The final MLP of GPT-2 Small exhibits a fully legible routing program -- 27 named neurons organized into a three-tier exception handler -- while the knowledge it routes remains entangled across ~3,040 residual neurons. We decompose all 3,072 neurons (to numerical precision) into: 5 fused Core neurons that reset vocabulary toward function words, 10 Differentiators that suppress wrong candidates, 5 Specialists that detect structural boundaries, and 7 Consensus neurons that each monitor a distinct linguistic dimension. The consensus-exception crossover -- where MLP intervention shifts from helpful to harmful -- is statistically sharp (bootstrap 95% CIs exclude zero at all consensus levels; crossover between 4/7 and 5/7). Three experiments show that "knowledge neurons" (Dai et al., 2022), at L11 of this model, function as routing infrastructure rather than fact storage: the MLP amplifies or suppresses signals already present in the residual stream from attention, scaling with contextual constraint. A garden-path experiment reveals a reversed garden-path effect -- GPT-2 uses verb subcategorization immediately, consistent with the exception handler operating at token-level predictability rather than syntactic structure. This architecture crystallizes only at the terminal layer -- in deeper models, we predict equivalent structure at the final layer, not at layer 11. Code and data: https://github.com/pbalogh/transparent-gpt2
Authors: Gabriel Sarch, Linrong Cai, Qunzhong Wang, Haoyang Wu, Danqi Chen, Zhuang Liu
Abstract: What does it take to build a visual reasoner that works across charts, science, spatial understanding, and open-ended tasks? The strongest vision-language models (VLMs) show such broad visual reasoning is within reach, but the recipe behind them remains unclear, locked behind proprietary reinforcement learning (RL) pipelines with non-public data. We introduce Vero, a family of fully open VLMs that matches or exceeds existing open-weight models across diverse visual reasoning tasks. We scale RL data and rewards across six broad task categories, constructing Vero-600K, a 600K-sample dataset from 59 datasets, and designing task-routed rewards that handle heterogeneous answer formats. Vero achieves state-of-the-art performance, improving over four base models by 3.6-5.3 points on average across VeroEval, our suite of 30 challenging benchmarks. Starting from Qwen3-VL-8B-Instruct, Vero outperforms Qwen3-VL-8B-Thinking on 23 of 30 benchmarks without additional proprietary thinking data. When trained from the same base model, Vero-600K exceeds existing RL datasets across task categories. Systematic ablations reveal that different task categories elicit qualitatively distinct reasoning patterns that transfer poorly in isolation, suggesting that broad data coverage is the primary driver of strong RL scaling. All data, code, and models are released.