论文检索

Shuyao Xiao, Shengling Wang, Ke Chao

Large Language Models (LLMs) are widely used in high-stakes applications, making their interpretability increasingly important. Existing interpretability methods are typically categorized into internal and external perspectives, which are often studied in isolation and tend to overlook two key aspects: causality and temporal dynamics. Explanations are often limited to surface correlations or static dependencies, failing to capture how influences evolve during autoregressive generation. To address these limitations, we propose a causal and dynamic interpretability framework for LLM generation. We first characterize the backdoor-adjusted causal effects of both the generated prefix and the prompt on the current token using the Structural Causal Model. Next, we introduce two metrics to quantify contextual causal influence and question–answer causal influence. Overall, our work provides a unified causal view of internal consistency and external alignment in LLM generation dynamics.

Zikang Liu, Peilan Xu

Generative answer engines expose content through selective citation rather than ranked retrieval, fundamentally altering how visibility is determined. This shift calls for new optimization methods beyond traditional search engine optimization. Existing generative engine optimization (GEO) approaches primarily rely on token-level text rewriting, offering limited interpretability and weak control over the trade-off between citation visibility and content quality. We propose FeatGEO, a feature-level, multi-objective optimization framework that abstracts webpages into interpretable structural, content, and linguistic properties. Instead of directly editing text, FeatGEO optimizes over this feature space and uses a language model to realize feature configurations into natural language, decoupling high-level optimization from surface-level generation. Experiments on GEO-Bench across three generative engines demonstrate that FeatGEO consistently improves citation visibility while maintaining or improving content quality, substantially outperforming token-level baselines. Further analyses show that citation behavior is more strongly influenced by document-level content properties than by isolated lexical edits, and that the learned feature configurations generalize across language models of different scales.

Shichen Li, Zhouyang Wang, Zhongqing Wang, Peifeng Li

Large Language Models (LLMs) can perform sentiment analysis via natural language instructions, yet their predictions are highly sensitive to prompt phrasing. Prior work has shown that sentiment is encoded linearly in LLM representations, but the model’s ability to utilize this information remains surprisingly fragile to prompt variations. We leverage Sparse Autoencoders (SAEs) and circuit-level analysis to uncover causal mechanisms underlying sentiment prediction. We identify a sentiment analysis circuit and find that prompt sensitivity may stem from task activation failure. The model encodes the sentiment feature consistently, but different prompts trigger varying degrees of circuit activation. Based on this insight, we propose a simple inference-time intervention method that amplifies circuit features to compensate for insufficient activation. Experiments across diverse datasets, templates, and languages show consistent improvements, offering an interpretable and training-free alternative to manual prompt engineering.

Ai Jian, Jingqing Ruan, Xing Ma, Dailin Li, Weipeng Zhang, Ke Zeng, Xunliang Cai

Reward models (RMs) are central to reinforcement learning from human feedback (RLHF), providing the critical supervision signals that align large language models (LLMs) with human preferences.Generative reward models (GRMs) provide greater interpretability than traditional scalar RMs, but they come with a critical trade-off: pairwise methods are hindered by a training-inference mismatch, while pointwise methods require expensive absolute annotations.To bridge this gap, we propose the Preference-aware Task-adaptive Reward Model (PaTaRM).Unlike prior approaches, PaTaRM enables robust pointwise training using readily available pairwise data via a novel Preference-Aware Reward (PAR) mechanism, eliminating the need for explicit rating labels. Furthermore, it incorporates a task-adaptive rubric system that dynamically generates instance-specific criteria for precise evaluation.Extensive experiments demonstrate that PaTaRM achieves an average relative improvement of 8.7% over the corresponding base models on RewardBench and RMBench across the Qwen3-8B and Qwen3-14B backbones.Crucially, when used as a reward model for downstream RLHF, it yields an average relative improvement of 13.6% over the corresponding base policies on IFEval and InfoBench, validating its effectiveness for policy alignment.Our code, data, and checkpoints are available at https://huggingface.co/AIJian/PaTaRM

Shuyang Hou, Yi Hu, Muhan Zhang

Recent advancements in large language models (LLMs) have significantly enhanced their reasoning capabilities. However, they continue to struggle with basic character-level tasks, such as counting letters in words—a problem rooted in their tokenization process. While existing benchmarks have highlighted this weakness through basic character operations, such failures are often dismissed due to lacking practical relevance. Yet, many real-world applications, such as navigating text-based maps or interpreting structured tables, rely heavily on precise sub-token understanding. In this regard, we introduce SubTokenTest, a comprehensive benchmark that assesses sub-token understanding through **practical, utility-driven** tasks. Our benchmark includes ten tasks across four domains and isolates tokenization-related failures by decoupling performance from complex reasoning. We provide a comprehensive evaluation of nine advanced LLMs. Additionally, we investigate the impact of test-time scaling on sub-token reasoning and explore how character-level information is encoded within the hidden states.

Qiyuan Chen, Hongsen Huang, Jiahe Chen, Qian Shao, Jintai Chen, Hongxia Xu, Renjie Hua, Ren Chuan, Jian Wu

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.

Yanyi Su, Hongshuai Wang, Zhifeng Gao, Jun Cheng

Olfaction lies at the intersection of chemical structure, neural encoding, and linguistic perception, yet existing representation methods fail to fully capture this pathway. Current approaches typically model only isolated segments of the olfactory pathway, overlooking the complete chain from molecule to receptors to linguistic descriptions. Such fragmentation yields learned embeddings that lack both biological grounding and semantic interpretability. We propose NOSE (Neural Olfactory-Semantic Embedding), a representation learning framework that aligns three modalities along the olfactory pathway: molecular structure, receptor sequence, and natural language description. Rather than simply fusing these signals, we decouple their contributions via orthogonal constraints, preserving the unique encoded information of each modality. To address the sparsity of olfactory language, we introduce a weak positive sample strategy to calibrate semantic similarity, preventing erroneous repulsion of similar odors in the feature space. Extensive experiments demonstrate that NOSE achieves state-of-the-art (SOTA) performance and excellent zero-shot generalization, confirming the strong alignment between its representation space and human olfactory intuition. Code and data are available at https://github.com/Xianyusyy/NOSE

Yi Hu, Jiaqi Gu, Ruxin Wang, Zijun Yao, Hao Peng, Xiaobao Wu, Jianhui Chen, Muhan Zhang, Liangming Pan

Reinforcement learning (RL) has catalyzed the emergence of Large Reasoning Models (LRMs) that have pushed reasoning capabilities to new heights. While their performance has garnered significant excitement, exploring the internal mechanisms driving these behaviors has become an equally critical research frontier. This paper provides a comprehensive survey of the mechanistic understanding of LRMs, organizing recent findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors. By synthesizing these insights, we aim to bridge the gap between black-box performance and mechanistic transparency. Finally, we discuss under-explored challenges to outline a roadmap for future mechanistic studies, including the need for applied interpretability, improved methodologies, and a unified theoretical framework.

Xv Wang, Wang Zhenyu, Guanyu Zheng, Rui Zhang

While Large Language Models (LLMs) have significantly advanced the fluency of Emotional Support Conversation (ESC) systems, current research predominantly focuses on engineering increasingly complex architectures—from intricate reasoning chains to multi-agent collaborations. While these advancements (e.g., CoT) offer semantic traces of reasoning, they remain mechanistically opaque, obscuring the fundamental causal mechanisms between dialogue features and effective empathic strategies, leading to poor interpretability and susceptibility to distribution shifts in offline learning. To address these limitations, we propose a novel framework Causal-ESC. Departing from conventional paradigms that directly utilize raw dialogue history as input, our approach introduces Doubly Robust (DR) learning to explicitly model the causal effect of utterance features on strategy selection, effectively mitigating the biases and counterfactual unobservability inherent in offline datasets. We further integrate an LLM-based stylized rewriting mechanism to translate these rigorously learned causal strategies into natural, context-consistent responses. Comprehensive experiments, supported by statistical verification (e.g., Outcome R^2) and human-like evaluation, demonstrate that our framework not only significantly outperforms state-of-the-art baselines in empathy and helpfulness but also provides a theoretically grounded, interpretable solution to the mechanistic interpretability dilemma in affective computing.

Rachneet Kaur, Nishan Srishankar, Zhen Zeng, Sumitra Ganesh

Recent multimodal LLMs have shown promise in chart-based visual question answering, but their performance declines sharply on unannotated charts—those requiring precise visual interpretation rather than relying on textual shortcuts. To address this, we introduce ChartAgent, a novel agentic framework that explicitly performs visual reasoning directly within the chart’s spatial domain. Unlike textual chain-of-thought reasoning, ChartAgent iteratively decomposes queries into visual subtasks and actively manipulates and interacts with chart images through specialized actions such as drawing annotations, cropping regions (e.g., segmenting pie slices, isolating bars), and localizing axes, using a library of chart-specific vision tools to fulfill each subtask. This iterative reasoning process closely mirrors human cognitive strategies for chart comprehension. ChartAgent achieves state-of-the-art accuracy on the ChartBench and ChartX benchmarks, surpassing prior methods by up to 16.07% absolute gain overall and 17.31% on unannotated, numerically intensive queries. Furthermore, our analyses show that ChartAgent is (a) effective across diverse chart types, (b) achieves the highest scores across varying visual and reasoning complexity levels, and (c) serves as a plug-and-play framework that boosts performance across diverse underlying LLMs. Our work is among the first to demonstrate visually grounded reasoning for chart understanding using tool-augmented multimodal agents.

Yicheng Liu, Shumin Shi, Youchao Zhou, Xingchen Zhang

Large language models (LLMs) perform well on Standard Chinese but struggle with low-resource Chinese dialects due to substantial phonological divergence. We investigate whether incorporating Middle Chinese, the common historical ancestor of most of the modern Chinese dialects, can improve dialectal pronunciation modeling in a linguistically interpretable manner. We focus on two specific task variants: (1) conditional sound change rule induction (a variant of Sound Law Induction, SLI), where models infer executable phonological transformation rules from Middle Chinese to modern dialects, and (2) sentence-level dialectal pronunciation transcription (a variant of Grapheme-to-Phoneme, G2P), requiring dialect-specific International Phonetic Alphabet (IPA) generation. We construct a multi-source dataset covering Middle Chinese and 12 modern Chinese dialects, including character-level correspondences, rule exemplars, and sentence-level IPA transcription. We adopt a parameter-efficient training framework combining LoRA-based supervised fine-tuning and reinforcement learning via Group Relative Policy Optimization (GRPO) for the first task. Across both tasks and a wide range of dialects and evaluation metrics, our approach achieves overall improvements over strong baselines, including DeepSeek-V3.2 and ChatGPT-5.2, while revealing variation across dialects. These results demonstrate the value of leveraging historical linguistic knowledge for modeling low-resource Chinese dialects.

Xinhe Wang, Jin Huang, Xingjian Zhang, Tianhao Wang, Jiaqi W. Ma

Reasoning benchmarks such as the Abstraction and Reasoning Corpus (ARC) and ARC-AGI are widely used to assess progress in artificial intelligence and are often interpreted as probes of core, so-called “fluid” reasoning abilities. Despite their apparent simplicity for humans, these tasks remain challenging for frontier vision-language models (VLMs), a gap commonly attributed to deficiencies in machine reasoning. We challenge this interpretation and hypothesize that the gap arises primarily from limitations in visual perception rather than from shortcomings in inductive reasoning.To verify this hypothesis, we introduce a two-stage experimental pipeline that explicitly separates perception and reasoning. In the perception stage, each image is independently converted into a natural-language description, while in the reasoning stage a model induces and applies rules using these descriptions. This design prevents leakage of cross-image inductive signals and isolates reasoning from perception bottlenecks. Across three ARC-style datasets, Mini-ARC, ACRE, and Bongard-LOGO, we show that the perception capability is the dominant factor underlying the observed performance gap by comparing the two-stage pipeline with against standard end-to-end one-stage evaluation. Manual inspection of reasoning traces in the VLM outputs further reveals that approximately 80 percent of model failures stem from perception errors. Together, these results demonstrate that ARC-style benchmarks conflate perceptual and reasoning challenges and that observed performance gaps may overstate deficiencies in machine reasoning. Our findings underscore the need for evaluation protocols that disentangle perception from reasoning when assessing progress in machine intelligence.

Yun He, Wenzhe Li, Hejia Zhang, Songlin Li, Karishma Mandyam, Sopan Khosla, Yuanhao Xiong, Nanshu Wang, Xiaoliang Peng, Beibin Li 等

Recent progress in large language models (LLMs) has led to impressive performance on a range of tasks, yet advanced instruction following (IF)—especially for complex, multi-turn, and system-prompted instructions—remains a significant challenge. Rigorous evaluation and effective training for such capabilities are hindered by the lack of high-quality, human-annotated benchmarks and reliable, interpretable reward signals. In this work, we introduce AdvancedIF, a comprehensive benchmark featuring over 1,600 prompts and expert-curated rubrics that assess LLMs’ ability to follow complex, multi-turn, and system-level instructions. We also open-source the evaluation script of AdvancedIF. We further propose RIFL (Rubric-based Instruction-Following Learning), a novel post-training pipeline that leverages rubric generation, a finetuned rubric verifier, and reward shaping to enable effective reinforcement learning for instruction following. Extensive experiments demonstrate that RIFL substantially improves the instruction-following abilities of LLMs, achieving a 6.7% absolute gain on AdvancedIF and strong results on public benchmarks. Our ablation studies confirm the effectiveness of each component in RIFL. This work establishes rubrics as a powerful tool for both training and evaluating advanced IF in LLMs, paving the way for more capable and reliable AI systems.

Wenxi Li, Jingyu Peng

The necessity of explicit linguistic representations has been increasingly questioned in the era of large language models (LLMs). In this work, we revisit this issue using Universal Dependencies (UD) as a case study, examining whether and in what ways this cross-lingual syntactic framework can still benefit contemporary LLMs. We focus on a cross-lingual adversarial paraphrase identification task that is designed to foreground the role of syntactic structure in semantic interpretation across languages. Within this setting, we systematically evaluate three strategies for integrating UD into LLMs: UD-Prompt, UD-Tuning, and UD-Attention. Our experiments show that, although the magnitude of gains depends on how UD-based structural priors interact with model behavior and cross-lingual variation, UD-augmented models consistently outperform their syntax-agnostic counterparts. Across strategies, we observe average accuracy improvements of 2.67%, 8.24%, and 2.53%, respectively. These findings demonstrate that linguistic knowledge remains informative for LLMs, offering practical value in cross-lingual settings where structural alignment is challenging.

Aaron Mueller, Andrew Lee, Shruti Joshi, Ekdeep Singh Lubana, Dhanya Sridhar, Patrik Reizinger

A goal of interpretability is to recover disentangled representations of latent concepts (features) from the activations of neural networks. The quality of features is typically evaluated in isolation, and under implicit independence assumptions that may not hold in practice. Thus, it is unclear to what extent common featurization methods such as sparse autoencoders (SAEs) and probes disentangle one concept from another. We propose a multi-concept evaluation setting using concepts such as sentiment, domain, voice, and tense. We evaluate how well featurizers produce disentangled representations of each concept, observing that features are typically sensitive to only one concept, but also that concepts are distributed across many features. Then, we steer these features, measuring whether each concept is independently manipulable, and whether features interact. Even in idealized settings, steering a feature often affects many concepts, despite a near absence of interaction effects. These results suggest that correlational metrics are insufficient to establish steering selectivity, and that demonstrating that two features operate in separate spaces is insufficient to claim that they will be selective for one concept. These results underscore the importance of multi-concept evaluations in interpretability research.

Shira Wein, Anna Serbina, Jiyuan Ji, Nathan Wolf, Jason DeGraaff, Prajakta Kini, Maria Leonor Pacheco

Translationese refers to the statistical patterns that distinguish translated texts from original texts, which are often subtle and imperceptible to human readers. When translated texts appear in either training or testing data, these patterns can negatively affect model performance or warp model evaluation. We approach the task of discerning whether a text was originally written in English or translated into English by fine-tuning contemporary foundation models at distinct item lengths and achieve state-of-the-art performance (94% Macro F1). Given that these linguistic cues are subtle and often imperceptible to humans, we analyze the features which enable our model’s high performance. Employing a suite of interpretability-based techniques, we find that: (1) our high accuracy is enabled by a collection of linguistic features, a number of which correspond with linguistic theories of translationese, and (2) pretrained neural models are adept at picking up these features without any fine-tuning.

Zizhen Wang, Bo Feng, Zhengfeng Lai, Shiyu Li, Yang Lu, Meng Cao, Ping Huang, Xiaoming Simon Wang

Evaluating video captioning remains a critical challenge for Visual Large Language Models (VLLMs). Existing metrics primarily rely on matching generated text against ground-truth references. This paradigm suffers from the “one-to-many” nature of video description, where high-quality captions are often penalized for lexical mismatches or valid shifts in visual focus. Furthermore, such assessments are typically one-dimensional, failing to provide a fine-grained analysis of caption quality. To address this, we redefine caption quality through the lens of information fidelity: A caption must maximize the coverage of salient visual information while ensuring strict factuality. We introduce CapQuiz, a novel reference-free benchmark that assesses captions based on their utility in answering human-verified, fine-grained, multiple-choice questions derived from the video. CapQuiz features a hierarchical taxonomy of 10 question types (spanning Descriptive and Inferential categories) across 24 diverse video domains. Extensive experiments demonstrate that CapQuiz correlates significantly better with human judgments than existing metrics and offers interpretable insights into model performance. We will release the benchmark to facilitate reproducible research.

Mingfei Lu, Yi Zhang, Mengjia Wu, Yue Feng

Legal consultation question answering (Legal CQA) presents unique challenges compared to traditional legal QA tasks, including the scarcity of high-quality training data, complex task composition, and strong contextual dependencies. To address these, we construct JurisCQAD, a large-scale dataset of over 43,000 real-world Chinese legal queries annotated with expert-validated positive and negative responses, and design a structured task decomposition that converts each query into a legal element graph integrating entities, events, intents, and legal issues. We further propose JurisMA, a modular multi-agent framework supporting dynamic routing, statutory grounding, and stylistic optimization. Combined with the element graph, the framework enables strong context-aware reasoning, effectively capturing dependencies across legal facts, norms, and procedural logic. Trained on JurisCQAD and evaluated on a refined LawBench, our system significantly outperforms both general-purpose and legal-domain LLMs across multiple lexical and semantic metrics, demonstrating the benefits of interpretable decomposition and modular collaboration in Legal CQA.

Yusuke Yamauchi, Akiko Aizawa

Psychological research has long utilized circumplex models to structure emotions, placing similar emotions adjacently and opposing ones diagonally. Although frequently used to interpret deep learning representations, these models are rarely directly incorporated into the representation learning of language models, leaving their geometric validity unexplored. This paper proposes a method to induce circular emotion representations within language model embeddings via contrastive learning on a hypersphere. We show that while this circular alignment offers superior interpretability and robustness against dimensionality reduction, it underperforms compared to conventional designs in high-dimensional settings and fine-grained classification. Our findings elucidate the trade-offs involved in applying psychological circumplex models to deep learning architectures.

Itay Yona, Amir Sarid, Michael Karasik, Yossi Gandelsman

We introduce Doublespeak, a simple in-context representation hijacking attack against language models. The attack works by systematically replacing a harmful keyword (e.g., bomb) with a benign token (e.g., carrot) across multiple in-context examples, provided as a prefix to a harmful request. We demonstrate that this substitution leads to the internal representation of the benign token converging toward that of the harmful one, effectively embedding the harmful semantics under a euphemism. As a result, superficially innocuous prompts (e.g., "How to build a carrot?") are internally interpreted as disallowed instructions ("How to build a bomb?"), thereby bypassing the model’s safety alignment. We use interpretability tools to show this semantic shift occurs progressively across layers. Doublespeak is optimization-free, broadly transferable across model families, and achieves strong success rates on closed-source systems, reaching 74% on Llama-3.3-70B-Instruct with a single-sentence context override. Our findings highlight a new attack surface in LM latent space, indicating that current alignment strategies are insufficient and should instead operate at the representation level.