论文检索

Haoyang Chen, Yi Liu, Jianzhi Shao, Tao Zhang, Chengfu Huo, Wei Hu

Thinking LLMs produce reasoning traces before answering. Prior activation steering work mainly targets on shaping these traces. It remains less understood how answer tokens actually read and integrate the reasoning to produce reliable outcomes. Focusing on quantitative reasoning, we analyze the answer-to-reasoning attention and observe a benign self-reading pattern aligned with correctness, characterized by a forward drift of the reading focus along the reasoning trace and a persistent concentration on key semantic anchors, whereas incorrect solutions exhibit diffuse and irregular attention pattern. We interpret this as internal certainty during answer decoding, where the model commits to a viable solution branch and integrates key evidence. Following this, we propose a training-free steering method driven by Self-Reading Quality (SRQ) scores combining geometric metrics for process control with semantic metrics for content monitoring. SRQ selects data to build steering vectors that guide inference toward benign self-reading and away from uncertain and disorganized reading. Experiments show that our method yields consistent accuracy gains.

Vallabhaneni Raj Kumar, Ashwin S, Supriya Manna, Niladri Sett, Cheedella V S N M S Hema Harshitha, Kurakula Harshitha, Basina Deepakraj, Anand Kumar Sharma, Tanuj Sarkar, Samanthapudi Shakeer 等

Sentiment analysis for low-resource languages remains challenging in an era where interpretability, human alignment, and fairness are increasingly non-negotiable aspects of modern machine learning systems. These challenges stem both from the scarcity of annotated data and from the resulting difficulty of conducting reliable, human-interpretable analyses that go beyond predictive accuracy. Telugu, one of the primary Dravidian languages with over 96 million speakers, is not an exception. In this work, we first introduce TeSent, a large-scale Telugu sentiment classification dataset annotated with sentiment labels and human-selected rationales from multiple native speakers. This resource enables the study of rationale-based supervision for aligning models with human reasoning in this low-resource setting. We fine-tune five transformer-based models with and without rationale supervision and evaluate them on classification performance, explanation quality, and social bias. To facilitate controlled fairness evaluation, we additionally construct TeEEC, an evaluation corpus for Telugu sentiment analysis. Our results show that incorporating human rationales consistently improves alignment and often leads to holistic gains in predictive performance. We further provide extensive analysis of multi-facade explanation quality and fairness, offering insights into the broader effects of alignment-oriented supervision in resource-scarce language contexts.

Min Jae Jung, YongTaek Lim, Chaeyun Kim, Junghwan Kim, Kihyun Kim, Minwoo Kim

While Large Language Models (LLMs) are widely used, they remain susceptible to jailbreak prompts that can elicit harmful or inappropriate responses. This paper introduces STAR-Teaming, a novel black-box framework for automated red teaming that effectively generates such prompts. STAR-Teaming integrates a Multi-Agent System (MAS) with a Strategy-Response Multiplex Network and employs network-driven optimization to sample effective attack strategies. This network-based approach recasts the intractable high-dimensional embedding space into a tractable structure, yielding two key advantages: it enhances the interpretability of the LLM’s strategic vulnerabilities, and it streamlines the search for effective strategies by organizing the search space into semantic communities, thereby preventing redundant exploration. Empirical results demonstrate that STAR-Teaming significantly surpasses existing methods, achieving a higher attack success rate (ASR) at a lower computational cost. Extensive experiments validate the effectiveness and explainability of the Multiplex Network. The code is available at https://github.com/selectstar-ai/STAR-Teaming-paper.

Micky C. Nnamdi, Benoit Louis Marteau, Yishan Zhong, J. Ben Tamo, May Dongmei Wang

Large Multimodal Models (LMMs) achieve state-of-the-art performance in high-stakes domains like healthcare, yet their reasoning remains opaque. Attention- and saliency-based methods often fail to faithfully represent the model’s decision process, particularly when integrating heterogeneous modalities. We introduce Tree-of-Evidence (ToE), an inference-time search algorithm that frames interpretability as a discrete optimization problem. Rather than relying on soft attention weights, ToE employs lightweight Evidence Bottlenecks that score coarse units of data (e.g., vital-sign windows, report chunks) and performs a beam search to identify the compact evidence set required to reproduce the model’s prediction. We evaluate ToE across six tasks spanning three datasets and two domains, including clinical prediction on MIMIC-IV, cross-center validation on eICU, and non-clinical fault detection on LEMMA-RCA. ToE retains over 98% of full-model AUROC with as few as five evidence units, achieves higher decision agreement and lower fidelity error than LIME, SHAP, saliency, and concept-bottleneck baselines under sparse budgets, and outperforms LLMs up to 70B parameters. ToE therefore provides a practical mechanism for auditing multimodal models.

Juncheng Hu, Jiming Yu, Rui Song, Kedi Lyu, Yingji Li, Zheli Liu

ocial bias in Multimodal Large Language Models (MLLMs) has become an increasingly important concern. Prompt-based approaches offer a lightweight solution for debiasing; however, existing methods rely heavily on handcrafted prompts that are brittle, highly context-sensitive, and difficult to generalize across tasks, bias types, and multimodal settings. In this work, we propose Historical Reflection-Guided Prompt Optimization (HRPO), an adaptive self-debiasing framework for black-box MLLMs that automatically optimizes task-specific debiasing prompts to suppress stereotypical outputs. To mitigate forgetting during prompt optimization, we introduce Historical Contrastive Self-Reflection (HCSR), which performs contrastive reflection over positive and negative optimization histories, enabling the model to retain effective prompts and avoid redundant exploration, thereby improving optimization efficiency. Experiments on three benchmarks involving eight open-source and two closed-source MLLMs, covering ten singular and two intersectional bias types, demonstrate that HRPO achieves strong debiasing performance while offering improved interpretability, generalization, and robustness. Code is available at: https://github.com/liyingji1996/HRPO.

Yongchan Kwon, Shang Zhu, Federico Bianchi, Kaitlyn Zhou, James Zou

The ability of large language models (LLMs) to follow user instructions is central to their reliability, safety, and usefulness. While prior studies assess instruction adherence in the model’s main responses, we argue that it is also critical for large reasoning models (LRMs) to follow user instructions throughout their reasoning process. Reasoning instruction following makes LRMs more controllable and transparent, while reducing risks of undesirable shortcuts, hallucinations, or reward hacking within reasoning traces. To evaluate this dimension, we introduce ReasonIF, a systematic benchmark for assessing reasoning instruction following. ReasonIF includes six categories of instruction prompts, spanning multilingual reasoning, and length control. Across many open-source LRMs including GPT-OSS, Qwen3, and DeepSeek-R1, we find substantial failures in reasoning instruction adherence: the highest instruction following score (IFS) remains below 0.25, meaning that fewer than 25% of reasoning traces comply with the given instructions. Notably, as task difficulty increases, reasoning instruction following degrades further. We also explore two strategies to enhance reasoning instruction fidelity: (1) multi-turn reasoning and (2) Reasoning Instruction Finetuning (RIF) using synthetic data. RIF improves the IFS of GPT-OSS-20B from 0.11 to 0.27, indicating measurable progress but leaving ample room for improvement. We hope this work draws attention to reasoning-level instruction adherence as an underexplored but critical aspect of model alignment, and helps pave the way toward more controllable, interpretable, and trustworthy reasoning models.

Ailiang Lin, Zhuoyun Li, Keyu Mao, Kotaro Funakoshi, Manabu Okumura

Large language models (LLMs) have been widely explored for embedding generation. While recent studies show that in-context learning (ICL) effectively enhances the representational capability of LLMs by prepending a few task-related demonstrations, it causes substantial token overhead due to the increased sequence length. In this work, we propose EPIC, a novel embedding-based in-context prompt training strategy that leverages ICL to generate high-quality embeddings while reducing computational burden during both training and inference. This approach replaces discrete text demonstrations with their corresponding continuous embeddings, which not only encourages the LLM to align semantically-related text pairs during contrastive learning, but also requires the model to interpret demonstration embeddings as part of the in-context prompt. Consequently, EPIC-trained models achieve excellent embedding performance both with or without in-context prompts at inference time. Comprehensive experiments demonstrate that our method establishes new state-of-the-art results on the MTEB benchmark, surpassing frontier models trained solely on publicly available retrieval data. Extensive ablation studies further validate the effectiveness and necessity of our mechanism.

Abid Ali, Diego Molla, Usman Naseem

Multimodal Large Language Models (MLLMs) have facilitated Multimodal Summarization with Multimodal Output (MSMO), wherein systems generate concise textual summaries accompanied by salient visuals from multimodal sources. However, current MSMO evaluation remains fragmented: text quality, image-text alignment, and visual diversity are typically assessed in isolation using unimodal metrics, making it difficult to capture whether the modalities jointly support a faithful and useful summary. To address this gap, we introduce MM-Eval, a unified evaluation framework that integrates assessments of textual quality, cross-modal alignment, and visual diversity. MM-Eval comprises three components: (1) text quality, measured using OpenFActScore for factual consistency and G-Eval for coherence, fluency, and relevance; (2) image-text relevance, evaluated via an MLLM-as-a-judge approach; and (3) image-set diversity, quantified using Truncated CLIP Entropy. We calibrate -Eval through a learned aggregation model trained on the mLLM-EVAL news benchmark, aligning component contributions with human preferences. Our analysis reveals a text-dominant hierarchy in this setting, where factual consistency acts as a critical determinant of perceived overall quality, while visual relevance and diversity provide complementary signals. MM-Eval improves over heuristic aggregation baselines and provides an interpretable, reference-weak framework for comparative evaluation of multimodal summaries.

Wangjie Gan, Miao Pan, Linbo Xi, Wenqi Zhang, Jintao Chen, Jianwei Yin, Xuhong Zhang

Large language models are typically post-trained using supervised fine-tuning (SFT) and reinforcement learning (RL), yet effectively unifying efficient knowledge injection with robust generalization remains challenging. In this work, we provide a training-dynamics analysis showing that SFT can be interpreted as a special case of policy gradient optimization with an extremely sparse implicit reward and unstable inverse-probability weighting, which together lead to single-path dependency, entropy collapse, and gradient explosion. Motivated by this diagnosis, we propose Group Fine-Tuning (GFT), a unified post-training framework that addresses these intrinsic limitations through two mechanisms: Group Advantage Learning, which constructs diverse response groups and derives normalized contrastive supervision to alleviate reward sparsity, and Dynamic Coefficient Rectification, which adaptively bounds inverse-probability weights to stabilize optimization while preserving efficient knowledge injection. Experiments demonstrate that GFT consistently surpasses SFT-based methods and yields policies that integrate more smoothly with subsequent RL training.Our code is publicly available athttps://github.com/ZJU-OmniAI/GFT.

Yusen Hou, Weicai Long, Haitao HU, Houcheng su, Junning Feng, Yanlin Zhang

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.

Ahmad Aljanaideh, Saeb Ganideh

The complexity of recent natural language classification models led to interest in developing methods for improving the performance of explainable models (e.g. Logistic Regression). Existing methods focus on clustering word embeddings to discover fine-grained contextual features that can be used to train a linear model. While those methods help reduce the gap in performance between black-box models and explainable models, they are based on discovering a large number of features, and this affects interpretability. In this work, we propose a model that leverages Large Language Models (LLMs) and clustering algorithms to discover a compact set of interpretable features. The proposed model first uses GPT-4o mini to extract rationales (i.e. phrases which explain an item’s label) from labeled text, and then clusters those rationales to obtain a compact, interpretable feature space. Across 3 Style Classification tasks, the resulting features achieve comparable performance to word-cluster baselines on most tasks, while reducing the number of features by 85–99%. These results highlight the potential of LLMs to improve the compactness of explainable AI models.

Wuya Chen, Yihao Yang, Yue Lin

Code edit suggestion, which encompasses modifying, refactoring, and maintaining existing code, represents the most frequent software development activity and has become a focal point for AI-powered tools. Traditional methods translate explicit natural language instructions into code edits, while pattern-based approaches learn from users’ historical editing patterns to provide style-consistent and more accurate suggestions. However, these pattern-based methods still face two critical challenges: (1) difficulty handling edits that demand deep contextual reasoning, and (2) lack of interpretability in editing decisions. To tackle this, we propose CoT-Edit, a reinforcement learning framework that guides LLMs to discover chain-of-thought (CoT) reasoning paths for code editing without requiring human-annotated CoT data. Specifically, we design multi-step reasoning framework that enable: (1) analysis-guided code editing, and (2) seamless switching between CoT and non-CoT inference modes. Building on this, we introduce Edit-Aware Reward Modeling (EARM), a fine-grained diff-based reward approach for effective learning. Furthermore, we discover a LoRA merging strategy that enhances model generalization. Evaluations on an industrial dataset show that our approach achieves 60.2% edit accuracy, outperforming all strong baselines. Online A/B tests further confirm its effectiveness in production. Code is available at https://github.com/202230483077yyh/CoT-Edit.

Jianming Chen, Yawen Wang, Junjie Wang, Xiaofei Xie, Shoubin Li, Qing Wang, Fanjiang Xu

Embodied agents in safety-critical applications such as Vision-Language Navigation (VLN) rely on multiple interdependent capabilities (e.g., perception, memory, planning, decision), making failures difficult to localize and attribute. Existing testing methods are largely system-level and provide limited insight into which capability deficiencies cause task failures. We propose a capability-oriented testing approach that enables failure detection and attribution by combining (1) adaptive test case generation via seed selection and mutation, (2) capability oracles for identifying capability-specific errors, and (3) a feedback mechanism that attributes failures to capabilities and guides further test generation. Experiments show that our method discovers more failure cases and more accurately pinpoints capability-level deficiencies than state-of-the-art baselines, providing more interpretable and actionable guidance for improving embodied agents.

Xiangtao Meng, Yingkai Dong, Ning Yu, Li Wang, Zheng Li, Shanqing Guo

Text-to-image (T2I) generative models have achieved remarkable visual fidelity, yet remain vulnerable to generating unsafe content. Existing safety defenses typically intervene internally within the generative model, but suffer from severe concept entanglement, leading to degradation of benign generation quality—a trade-off we term the Safety Tax. To overcome this limitation, we advocate a paradigm shift from destructive internal editing to external safety rectification. Following this principle, we propose SafePatch, a structurally isolated safety module that performs external, interpretable rectification without modifying the base model. The core backbone of SafePatch is architecturally instantiated as a trainable clone of the base model’s encoder, allowing it to inherit rich semantic priors and maintain representation consistency. To enable interpretable safety rectification, we construct a strictly aligned counterfactual safety dataset (ACS) for differential supervision training. Across nudity and multi-category bench- marks and recent adversarial prompt attacks, SafePatch achieves robust unsafe suppression (7% unsafe on I2P) while preserving image quality and semantic alignment.

Fuwen Luo, Zihao Wan, Ziyue Wang, Yaluo Liu, Pau Tong Lin Xu, Xuanjia Qiao, Xiaolong Wang, Peng Li, Yang Liu

Hieroglyphs, as logographic writing systems, encode rich semantic and cultural information within their internal structural composition. Yet, current advanced Large Language Models (LLMs) and Multimodal LLMs (MLLMs) usually remain structurally blind to this information. LLMs process characters as textual tokens, while MLLMs additionally view them as raw pixel grids. Both fall short to model the underlying logic of character strokes. Furthermore, existing structural analysis methods are often script-specific and labor-intensive. In this paper, we propose Hieroglyphic Stroke Analyzer (HieroSA), a novel and generalizable framework that enables MLLMs to automatically derive stroke-level structures from character bitmaps without handcrafted data. It transforms modern logographic and ancient hieroglyphs character images into explicit, interpretable line-segment representations in a normalized coordinate space, allowing for cross-lingual generalization. Extensive experiments demonstrate that HieroSA effectively captures character-internal structures and semantics, bypassing the need for language-specific priors. Experimental results highlight the potential of our work as a graphematics analysis tool for a deeper understanding of hieroglyphic scripts.

Youmi Ma, Naoaki Okazaki

Advances in mechanistic interpretability have identified special attention heads, known as retrieval heads, that are responsible for retrieving information from the context. However, the role of these retrieval heads in improving model performance remains unexplored. This work investigates whether retrieval heads can be leveraged to enhance the long-context capabilities of LLMs. Specifically, we propose RetMask, a method that generates training signals by contrasting normal model outputs with those from an ablated variant in which the retrieval heads are masked. This mechanism-based approach achieves substantial improvements: +2.28 points on HELMET at 128K for Llama-3.1, with +70% gains on generation with citation and +32% on passage re-ranking, while preserving performance on general tasks. Experiments across three model families demonstrate that RetMask consistently improves long-context performance, with gains correlating with the sparsity of the retrieval score distribution: models with sparser distributions, where retrieval capabilities are concentrated in a small set of heads, respond more strongly, while those with less sparse distributions show more modest gains. These results validate the functional role of retrieval heads and show that mechanistic insights can be transformed into performance enhancements.

Yang Liu, Yinghao Zhang, Lin Liu, Jiuyong Li, Debo Cheng, Zaiwen Feng

A prevalent approach to interpretable representation learning involves creating a mask that weights the significance of each input feature, followed by deriving a masked representation by applying this mask to the input representation. However, the identifiability of these learned masked representations is often uncertain, making the origin of these representations ambiguous or unreliable. Furthermore, the approaches to interpreting Transformer based on attention weights have been criticized for their faithfulness. To address these limitations, we propose a novel causal framework that directly learns identifiable and explainable representations from attention weights, rather than relying on importance masks. Our framework leverages identifiability theory and causal representation learning to extract explainable representations within a subspace of input representations, effectively transforming frozen representation learning methods into self-explaining systems. Experimental results on real-world datasets demonstrate that, compared to well-established state-of-the-art methods, our approach provides identifiable and more trustworthy explanations while guaranteeing faithfulness.

Donghoon Jung, Jiwoo Choi, Songeun Chae, Seohyon Jung

We introduce a constraint-selection-based experiment design for measuring narrative preferences of Large Language Models (LLMs). This design offers an interpretable lens on LLMs’ narrative selection behavior. We developed a library of 200 narratology-grounded constraints and prompted selections from six LLMs under three different instruction types: basic, quality-focused, and creativity-focused. Findings demonstrate that models consistently prioritize Style over narrative content elements like Event, Character, and Setting. Style preferences remain stable across models and instruction types, whereas content elements show cross-model divergence and instructional sensitivity. These results suggest that LLMs have latent narrative preferences, which should inform how the NLP community evaluates and deploys models in creative domains.

Jeonghyun Park, Ingeol Baek, Seunghyun Yoon, Haeun Jang, Aparna Garimella, Akriti Jain, Nedim Lipka, Hwanhee Lee

Real-world multi-hop QA is naturally linked with ambiguity, where a single query can trigger multiple reasoning paths that require independent resolution. Since ambiguity can occur at any stage, models must navigate layered uncertainty throughout the entire reasoning chain. Despite its prevalence in real-world user queries, previous benchmarks have primarily focused on single-hop ambiguity, leaving the complex interaction between multi-step inference and layered ambiguity underexplored. In this paper, we introduce MARCH, a benchmark for their intersection, with 2,209 multi-hop ambiguous questions curated via multi-LLM verification and validated by human annotation with strong agreement. Our experiments reveal that even state-of-the-art models struggle with MARCH, confirming that combining ambiguity resolution with multi-step reasoning is a significant challenge. To address this, we propose CLARION, a two-stage agentic framework that explicitly decouples ambiguity planning from evidence-driven reasoning, significantly outperforms existing approaches, and paves the way for robust reasoning systems.

Pingjun Hong, Beiduo Chen, Siyao Peng, Marie-Catherine de Marneffe, Benjamin Roth, Barbara Plank

Natural Language Inference (NLI) datasets often exhibit human label variation. To better understand these variations, explanation-based approaches analyze the underlying reasoning behind annotators’ decisions. One such approach is the LiTEx taxonomy, which categorizes free-text explanations in English into reasoning categories. However, previous work applying LiTEx has focused on within-label variation: cases where annotators agree on the NLI label but provide different explanations. This paper broadens the scope by examining how annotators may diverge not only in the reasoning category but also in the labeling. We use explanations as a lens to analyze variation in NLI annotations and to examine individual differences in reasoning. We apply LiTEx to two NLI datasets and align annotation variation from multiple aspects: NLI label agreement, explanation similarity, and taxonomy agreement, with an additional compounding factor of annotators’ selection bias. We observe instances where annotators disagree on the label but provide similar explanations, suggesting that surface-level disagreement may mask underlying agreement in interpretation. Moreover, our analysis reveals individual preferences in explanation strategies and label choices. These findings highlight that agreement in reasoning categories better reflects the semantic similarity of explanations than label agreement alone. Our findings underscore the richness of reasoning-based explanations and the need for caution in treating labels as ground truth.