论文检索

Xiaofeng Zhang, Yihao Quan, Chen Shen, Chaochen Gu, Xiaosong Yuan, Shaotian Yan, Jiawei Cao, Hao Cheng, Kaijie Wu, Jieping Ye

Multimodal large language models (MLLMs) demonstrate excellent abilities for understanding visual information, while the hallucination remains. Albeit image tokens constitute the majority of the MLLMs input, the relation between image tokens and hallucinations is still unexplored. In this paper, we analyze the attention score distribution of image tokens across layers and attention heads in models, revealing an intriguing but common phenomenon: most hallucinations are closely linked to the attention sink patterns of image tokens attention matrix, where shallow layers exhibit dense sinks and deep layers exhibit the sparse. We further explore the attention heads of different layers, finding: heads with high-density attention sink of the image part act positively in mitigating hallucinations. Inspired by these findings, we propose a training-free approach called Enhancing Vision Attention Sinks (EVAS) to facilitate the convergence of the image token attention sink within shallow layers. Specifically, EVAS identifies the attention heads that emerge as the densest visual sink in shallow layers and extracts its attention matrix, which is then broadcast to other heads of the same layer, thereby strengthing the layer’s focus on the image itself. Extensive empirical results of various MLLMs illustrate the superior performance of the proposed EVAS, demonstrating its effectiveness and generality.

Yuhang He, Yash Jain, Xubo Liu, Andrew Markham, Vibhav Vineet

Existing text-to-audio (TTA) generation methods have neither systematically explored audio event relation modeling, nor proposed any new framework to enhance this capability. In this work, we systematically study audio event relation modeling in TTA generation models. We first establish a benchmark for this task by: (1) proposing a comprehensive relation corpus covering all potential relations in real-world scenarios; (2) introducing a new audio event corpus encompassing commonly heard audios; and (3) proposing new evaluation metrics to assess audio event relation modeling from various perspectives. Furthermore, we propose a gated prompt tuning strategy that improves existing TTA models’ relation modeling capability with negligible extra parameters. Specifically, we introduce learnable relation and event prompt that append to the text prompt before feeding to existing TTA models.

Qibin Li, Zhen Xu, Shengyuan Bai, Nianmin Yao, Kaili Sun, Bowen Wu, Ying Li, Baoxun Wang

Large Language Models (LLMs) have demonstrated significant advancements in various fields, notably in Role-Playing Conversational Agents (RPCAs). However, when confronted with role-specific professional inquiries, LLMs-based RPCAs tend to underperform due to their excessive emphasis on the conversational abilities of characters rather than effectively invoking and integrating relevant expert knowledge. This often results in inaccurate responses. We refer to this phenomenon as the “Knowledge Misalignment” which underscores the limitations of RPCAs in integrating expert knowledge. To mitigate this issue, we have introduced an Anchoring-Guidance Fine-Tuning (AnGFT) Framework into the RPCAs’ training process. This involves initially linking the Anchoring-Based System Prompt (ASP) with the LLM’s relevant expert domains through diverse prompt construction strategies and supervised fine-tuning (SFT). Following the role-play enriched SFT, the integration of ASP enables LLMs to better associate with relevant expert knowledge, thus enhancing their response capabilities in role-specific expert domains. Moreover, we have developed four comprehensive metrics—helpfulness, thoroughness, credibility, and feasibility—to evaluate the proficiency of RPCAs in responding to professional questions. Our method was tested across four professional fields, and the experimental outcomes suggest that the proposed AnGFT Framework substantially improves the RPCAs’ performance in handling role-specific professional queries, while preserving their robust role-playing abilities.

Ziyang Ma, Qingyue Yuan, Zhenglin Wang, Deyu Zhou

Previous research has primarily focused on the cognitive error detection capabilities of Large Language Models (LLMs), often prompting them to analyze mistakes in reasoning chains. However, few studies have examined the meta-cognitive abilities of LLMs (e.g., their self-awareness of step errors), which are crucial for their reliability. While studies on LLM self-evaluation present some measures, such as perplexity, which can reflect the answer correctness and be viewed as the lens of meta-cognition, they lack step-level analysis and adaptation. This paper studies the evaluation of LLM meta-cognition using the current lenses and how to improve these lenses. Specifically, we propose AutoMeco, an Automated Meta-cognition Evaluation framework for benchmarking the existing lenses. Furthermore, a training-free Markovian Intrinsic Reward Adjustment strategy, MIRA, is proposed to boost current meta-cognition lenses. Experimental results on three mathematical reasoning datasets and three LLMs show the reasonableness of AutoMeco by comparing it with Best-of-N verification. Moreover, the meta-cognition ability of LLMs can be better evaluated using MIRA.

Hao Nan Sheng, Zhi-Yong Wang, Hing Cheung So, Mingrui Yang

As large language models continue to grow in size, parameter-efficient fine-tuning (PEFT) has become increasingly crucial. While low-rank adaptation (LoRA) offers a solution through low-rank updates, its static rank allocation may yield suboptimal results. Adaptive low-rank adaptation (AdaLoRA) improves this with dynamic allocation but remains sensitive to initial and target rank configurations. We introduce AROMA, a framework that automatically constructs layer-specific updates by iteratively building up rank-one components with very few trainable parameters that gradually diminish to zero. Unlike existing methods that employ rank reduction mechanisms, AROMA introduces a dual-loop architecture for rank growth. The inner loop extracts information from each rank-one subspace, while the outer loop determines the number of rank-one subspaces, i.e., the optimal rank. We reset optimizer states to maintain subspace independence. AROMA significantly reduces parameters compared to LoRA and AdaLoRA while achieving superior performance on natural language understanding and generation, commonsense reasoning, offering new insights into adaptive PEFT.

Keisuke Kamahori, Jungo Kasai, Noriyuki Kojima, Baris Kasikci

Modern automatic speech recognition (ASR) models, such as OpenAI’s Whisper, rely on deep encoder-decoder architectures, and their encoders are a critical bottleneck for efficient deployment due to high computational intensity. We introduce LiteASR, a low-rank compression scheme for ASR encoders that significantly reduces inference costs while maintaining transcription accuracy. Our approach leverages the strong low-rank properties observed in intermediate activations: by applying principal component analysis (PCA) with a small calibration dataset, we approximate linear transformations with a chain of low-rank matrix multiplications, and further optimize self-attention to work in reduced dimensionality. Evaluation results show that our method can compress Whisper large-v3’s encoder size by over 50%, matching Whisper medium’s size with better transcription accuracy, thereby establishing a new Pareto frontier of accuracy and efficiency. The code of LiteASR is available at https://github.com/efeslab/LiteASR.

Ting-Yun Chang, Muru Zhang, Jesse Thomason, Robin Jia

Low-bit weight-only quantization significantly reduces the memory footprint of large language models (LLMs), but disproportionately affects certain examples. We analyze diverse 3-4 bit methods on LLMs ranging from 7B-70B in size and find that the quantization errors of 50 pairs of methods are strongly correlated (avg. \rho = 0.82) on FineWeb examples. Moreover, the residual stream magnitudes of full-precision models are indicative of future quantization errors. We further establish a hypothesis that relates the residual stream magnitudes to error amplification and accumulation over layers. Using LLM localization techniques, early exiting, and activation patching, we show that examples with large errors rely on precise residual activations in the late layers, and that the outputs of MLP gates play a crucial role in maintaining the perplexity. Our work reveals why certain examples result in large quantization errors and which model components are most critical for performance preservation.

Zhaofeng Wu, Michihiro Yasunaga, Andrew Cohen, Yoon Kim, Asli Celikyilmaz, Marjan Ghazvininejad

Reward models have become a staple in modern NLP, serving as not only a scalable text evaluator, but also an indispensable component in many alignment recipes and inference-time algorithms. However, while recent reward models increase performance on standard benchmarks, this may partly be due to overfitting effects, which would confound an understanding of their true capability. In this work, we scrutinize the robustness of reward models and the extent of such overfitting. We build reWordBench, which systematically transforms reward model inputs in meaning- or ranking-preserving ways. We show that state-of-the-art reward models suffer from substantial performance degradation even with minor input transformations, sometimes dropping to significantly below-random accuracy, suggesting brittleness. To improve reward model robustness, we propose to explicitly train them to assign similar scores to paraphrases, and find that this approach also improves robustness to other distinct kinds of transformations. For example, our robust reward model reduces such degradation by roughly half for the Chat Hard subset in RewardBench. Furthermore, when used in alignment, our robust reward models demonstrate better utility and lead to higher-quality outputs, winning in up to 59% of instances against a standardly trained RM.

Tu Anh Dinh, Jan Niehues

Quality Estimation (QE) is estimating quality of the model output during inference when the ground truth is not available. Deriving output quality from the models’ output probability is the most trivial and low-effort way. However, we show that the output probability of text-generation models can appear underconfident. At each output step, there can be multiple correct options, making the probability distribution spread out more. Thus, lower probability does not necessarily mean lower output quality. Due to this observation, we propose a QE approach called BoostedProb, which boosts the model’s confidence in cases where there are multiple viable output options. With no increase in complexity, BoostedProb is notably better than raw model probability in different settings, achieving on average +0.194 improvement in Pearson correlation to ground-truth quality. It also comes close to or outperforms more costly approaches like supervised or ensemble-based QE in certain settings.

Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, Wenjie Li

Chain-of-Thought (CoT) has been proven effective in enhancing the reasoning capabilities of large language models (LLMs). Recent advancements, such as OpenAI’s o1 and DeepSeek-R1, suggest that scaling up the length of CoT sequences during inference could further boost LLM reasoning performance. However, due to the autoregressive nature of LLM decoding, longer CoT outputs lead to a linear increase in inference latency, adversely affecting user experience, particularly when the CoT exceeds 10,000 tokens. To address this limitation, we analyze the semantic importance of tokens within CoT outputs and reveal that their contributions to reasoning vary. Building on this insight, we propose TokenSkip, a simple yet effective approach that enables LLMs to selectively skip less important tokens, allowing for controllable CoT compression. Extensive experiments across various models and tasks demonstrate the effectiveness of TokenSkip in reducing CoT token usage while preserving strong reasoning performance. Notably, when applied to Qwen2.5-14B-Instruct, TokenSkip reduces reasoning tokens by 40% (from 313 to 181) on GSM8K, with less than a 0.4% performance drop.

Yunze Xiao, Lynnette Hui Xian Ng, Jiarui Liu, Mona T. Diab

Large Language Models (LLMs) increasingly exhibit anthropomorphism characteristics – human-like qualities portrayed across their outlook, language, behavior, and reasoning functions. Such characteristics enable more intuitive and engaging human-AI interactions. However, current research on anthropomorphism remains predominantly risk-focused, emphasizing over-trust and user deception while offering limited design guidance. We argue that anthropomorphism should instead be treated as a concept of design that can be intentionally tuned to support user goals. Drawing from multiple disciplines, we propose that the anthropomorphism of an LLM-based artifact should reflect the interaction between artifact designers and interpreters. This interaction is facilitated by cues embedded in the artifact by the designers and the (cognitive) responses of the interpreters to the cues. Cues are categorized into four dimensions: perceptive, linguistic, behavioral, and cognitive. By analyzing the manifestation and effectiveness of each cue, we provide a unified taxonomy with actionable levers for practitioners. Consequently, we advocate for function-oriented evaluations of anthropomorphic design.

Carlos Mullov, Alexander Waibel

Recent work shows strong transfer learning capability to unseen languages in sequence-to-sequence neural networks, under the assumption that we have high-quality word representations for the target language. We evaluate whether this direction is a viable path forward for translation from low-resource languages by investigating how much data is required to learn such high-quality word representations. We first show that learning word embeddings separately from a translation model can enable rapid adaptation to new languages with only a few hundred sentences of parallel data. To see whether the current bottleneck in transfer to low-resource languages lies mainly with learning the word representations, we then train word embeddings models on varying amounts of data, to then plug them into a machine translation model. We show that in this simulated low-resource setting with only 500 parallel sentences and 31,250 sentences of monolingual data we can exceed 15 BLEU on Flores on unseen languages. Finally, we investigate why on a real low-resource language the results are less favorable and find fault with the publicly available multilingual language modelling datasets.

Bohan Lyu, Siqiao Huang, Zichen Liang, Qian Sun, Jiaming Zhang

Neural surrogate models are powerful and efficient tools in data mining. Meanwhile, large language models (LLMs) have demonstrated remarkable capabilities in code-related tasks, such as generation and understanding. However, an equally important yet underexplored question is whether LLMs can serve as surrogate models for code execution prediction. To systematically investigate it, we introduce SURGE, a comprehensive benchmark with 1160 problems covering 8 key aspects: multi-language programming tasks, competition-level programming problems, repository-level code analysis, high-cost scientific computing, time-complexity-intensive algorithms, buggy code analysis, programs dependent on specific compilers or execution environments, and formal mathematical proof verification. Through extensive analysis of 21 open-source and proprietary LLMs, we examine scaling laws, data efficiency, and predictive accuracy. Our findings reveal important insights about the feasibility of LLMs as efficient surrogates for computational processes. The benchmark and evaluation framework are available at https://github.com/Imbernoulli/SURGE.

Boyu Guan, Chuang Han, Yining Zhang, Yupu Liang, Zhiyang Zhang, Yang Zhao, Chengqing Zong

Video-guided Machine Translation (VMT) aims to improve translation quality by integrating contextual information from paired short video clips. Mainstream VMT approaches typically incorporate multimodal information by uniformly sampling frames from the input videos. However, this paradigm frequently incurs significant computational overhead and introduces redundant multimodal content, which degrades both efficiency and translation quality. To tackle these challenges, we propose SHIFT (Selected Helpful Informative Frame for Translation). It is a lightweight, plug-and-play framework designed for VMT with Multimodal Large Language Models (MLLMs). SHIFT adaptively selects a single informative key frame when visual context is necessary; otherwise, it relies solely on textual input. This process is guided by a dedicated clustering module and a selector module. Experimental results demonstrate that SHIFT enhances the performance of MLLMs on the VMT task while simultaneously reducing computational cost, without sacrificing generalization ability. The code will be released upon acceptance.

Jianyou Wang, Weili Cao, Longtian Bao, Youze Zheng, Gil Pasternak, Kaicheng Wang, Xiaoyue Wang, Ramamohan Paturi, Leon Bergen

Systems that answer questions by reviewing the scientific literature are becoming increasingly feasible. To draw reliable conclusions, these systems should take into account the quality of available evidence from different studies, placing more weight on studies that use a valid methodology. We present a benchmark for measuring the methodological strength of biomedical papers, drawing on the risk-of-bias framework used for systematic reviews. Derived from over 500 biomedical studies, the three benchmark tasks encompass expert reviewers’ judgments of studies’ research methodologies, including the assessments of risk of bias within these studies. The benchmark contains a human-validated annotation pipeline for fine-grained alignment of reviewers’ judgments with research paper sentences. Our analyses show that large language models’ reasoning and retrieval capabilities impact their effectiveness with risk-of-bias assessment. The dataset is available at https://github.com/RoBBR-Benchmark/RoBBR.

Gagan Bhatia, Maxime Peyrard, Wei Zhao

Modern BPE tokenisers often split calendar dates into meaningless fragments, e.g., “20250312” \rightarrow “202”, “503”, “12”, inflating token counts and obscuring the inherent structure needed for robust temporal reasoning. In this work, we (1) introduce a simple yet interpretable metric, termed date fragmentation ratio, that measures how faithfully a tokeniser preserves multi-digit date components; (2) release DateAugBench, a suite of 6500 examples spanning three temporal reasoning tasks: context-based date resolution, format-invariance puzzles, and date arithmetic across historical, contemporary, and future time periods; and (3) through layer-wise probing and causal attention-hop analyses, uncover an emergent date-abstraction mechanism whereby large language models stitch together the fragments of month, day, and year components for temporal reasoning. Our experiments show that excessive fragmentation correlates with accuracy drops of up to 10 points on uncommon dates like historical and futuristic dates. Further, we find that the larger the model, the faster the emergent date abstraction heals date fragments. Lastly, we observe a reasoning path that LLMs follow to assemble date fragments, typically differing from human interpretation (year \rightarrow month \rightarrow day).

Jiarui Yao, Ruida Wang, Tong Zhang

Large Language Models (LLMs) have displayed astonishing abilities in various tasks, especially in text generation, classification, question answering, etc. However, the reasoning ability of LLMs still faces many debates, especially in math reasoning. The inherent ambiguity of Natural Language (NL) limits LLMs’ ability to perform verifiable reasoning, making the answers lack coherence and trustworthy support. To tackle the above challenges, we propose a novel framework named FANS: Formal ANswer Selection for LLM Natural Language Math Reasoning Using Lean4. It is a pioneering framework that utilizes Lean4 to enhance LLMs’ NL math reasoning ability. In particular, given an NL math question and LLM-generated answers, FANS first translates it into Lean4 theorem statements. Then it invokes another Lean4 prover LLM to produce proofs, and finally verifies the proofs by Lean4 compiler. Answers are selected based on the verifications. It enhances LLMs’ NL math ability in providing a computer-verifiable solution for its correct answer and proposes an alternative method for answer selection beyond the reward model based ones. Our experiments demonstrate the effectiveness of FANS with an improvement of nearly 2% across several math benchmarks, and even higher further based on reward models or in subfields such as algebra and number theory that Lean4 is better at. The code is available in https://github.com/MaxwellJryao/FANS.

Huihan Li, You Chen, Siyuan Wang, Yixin He, Ninareh Mehrabi, Rahul Gupta, Xiang Ren

Large Language Models (LLMs) perform well on reasoning benchmarks but often fail when inputs alter slightly, raising concerns about the extent to which their success relies on memorization. This issue is especially acute in Chain-of-Thought (CoT) reasoning, where spurious memorized patterns can trigger intermediate errors that cascade into incorrect final answers. We introduce STIM, a novel framework for Source-aware Token-level Identification of Memorization, which attributes each token in a reasoning chain to one of multiple memorization sources – local, mid-range, or long-range – based on their statistical co-occurrence with the token in the pretraining corpus. Our token-level analysis across tasks and distributional settings reveals that models rely more on memorization in complex or long-tail cases, and that local memorization is often the dominant driver of errors, leading to up to 67% of wrong tokens. We also show that memorization scores from STIM can be effective in predicting the wrong tokens in the wrong reasoning step. STIM offers a powerful tool for diagnosing and improving model reasoning and can generalize to other structured step-wise generation tasks.

Sania Waheed, Na Min An

Image Captioning for state-of-the-art VLMs has significantly improved over time; however, this comes at the cost of increased computational complexity, making them less accessible for resource-constrained applications such as mobile devices and assistive technologies. Alternatively, comparably smaller VLMs prioritize high-level scene descriptions, overlooking finer details that contribute to a richer understanding of an image. In this paper, we introduce a training-free framework that enhances caption diversity and informativeness by explicitly attending to distinct image regions using a comparably small VLM, BLIP, as the backbone. Our approach leverages structured segmentation to produce hierarchical representations that capture both global and localized semantics. Without requiring additional model training, we demonstrate that our method allows smaller VLMs to achieve performance comparable to larger models in terms of image-caption alignment, semantic integrity, and diversity. We evaluate our framework on MSCOCO, Flickr30k, and Nocaps test datasets, achieving a Div-2 score of 0.735, 0.750, and 0.748 for each dataset, respectively, while maintaining strong image-caption relevancy and semantic integrity with the human-annotated captions. Our code is available at https://github.com/xfactlab/HBoP.

Jiazheng Li, Yuxiang Zhou, Junru Lu, Gladys Tyen, Lin Gui, Cesare Aloisi, Yulan He

Although preference optimization methods have improved reasoning performance in Large Language Models (LLMs), they often lack transparency regarding why one reasoning outcome is preferred over another. This limitation is especially critical in Automated Student Answer Scoring (ASAS), where explainability is essential to justify assessment outcomes. Verbal reinforcement learning offers the potential to generate explicit reflection, but it tends to produce superficial critiques that can harm assessment performance. Existing LLMs also struggle to reliably detect subtle reasoning errors in ASAS tasks. Moreover, manually identifying intermediate reasoning errors is expensive and difficult to scale. To address these challenges, we introduce a **contrastive reflection synthesis pipeline** that generates precise verbal feedback by identifying discrepancies in structure reasoning graph paths. Leveraging these synthetic reflection data, we propose *DARS*, a Dual-model Reflective Scoring framework featuring a dedicated Critic model trained for effective reflection. *DARS* achieves strong performance and consistently outperforms existing ASAS baselines across all evaluation metrics. Extensive experiments further provide novel insights into the value of reflection data, framework design, and the scaling behavior of *DARS*. We release the DARS code at https://github.com/lijiazheng99/DARS.