Mechanistic interpretability of large language models (LLMs) aims to uncover the internal processes of information propagation and reasoning. Sparse autoencoders (SAEs) have demonstrated promise in this domain by extracting interpretable and monosemantic features. However, prior works primarily focus on feature extraction from a single layer, failing to effectively capture activations that span multiple layers. In this paper, we introduce Route Sparse Autoencoder (RouteSAE), a new framework that integrates a routing mechanism with a shared SAE to efficiently extract features from multiple layers. It dynamically assigns weights to activations from different layers, incurring minimal parameter overhead while achieving high interpretability and flexibility for targeted feature manipulation. We evaluate RouteSAE through extensive experiments on Llama-3.2-1B-Instruct. Specifically, under the same sparsity constraint of 64, RouteSAE extracts 22.5% more features than baseline SAEs while achieving a 22.3% higher interpretability score. These results underscore the potential of RouteSAE as a scalable and effective method for LLM interpretability, with applications in feature discovery and model intervention. Our codes are available at https://github.com/swei2001/RouteSAEs.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
FlightGPT: Towards Generalizable and Interpretable UAV Vision-and-Language Navigation with Vision-Language Models
PDF ↗Unmanned Aerial Vehicle (UAV) Vision-and-Language Navigation (VLN) is vital for applications such as disaster response, logistics delivery, and urban inspection. However, existing methods often struggle with insufficient multimodal fusion, weak generalization, and poor interpretability. To address these challenges, we propose FlightGPT, a novel UAV VLN framework built upon Vision-Language Models (VLMs) with powerful multimodal perception capabilities. We design a two-stage training pipeline: first, Supervised Fine-Tuning (SFT) using high-quality demonstrations to improve initialization and structured reasoning; then, Group Relative Policy Optimization (GRPO) algorithm, guided by a composite reward that considers goal accuracy, reasoning quality, and format compliance, to enhance generalization and adaptability. Furthermore, FlightGPT introduces a Chain-of-Thought (CoT)-based reasoning mechanism to improve decision interpretability. Extensive experiments on the city-scale dataset CityNav demonstrate that FlightGPT achieves state-of-the-art performance across all scenarios, with a 9.22% higher success rate than the strongest baseline in unseen environments. Our implementation is publicly available.
What Makes a Good Reasoning Chain? Uncovering Structural Patterns in Long Chain-of-Thought Reasoning
PDF ↗Recent advances in reasoning with large language models (LLMs) have popularized Long Chain-of-Thought (LCoT), a strategy that encourages deliberate and step-by-step reasoning before producing a final answer. While LCoTs have enabled expert-level performance in complex tasks, how the internal structures of their reasoning chains drive, or even predict, the correctness of final answers remains a critical yet underexplored question. In this work, we present LCoT2Tree, an automated framework that converts sequential LCoTs into hierarchical tree structures and thus enables deeper structural analysis of LLM reasoning. Using graph neural networks (GNNs), we reveal that structural patterns extracted by LCoT2Tree, including exploration, backtracking, and verification, serve as stronger predictors of final performance across a wide range of tasks and models. Leveraging an explainability technique, we further identify critical thought patterns such as over-branching that account for failures. Beyond diagnostic insights, the structural patterns by LCoT2Tree support practical applications, including improving Best-of-N decoding effectiveness. Overall, our results underscore the critical role of internal structures of reasoning chains, positioning LCoT2Tree as a powerful tool for diagnosing, interpreting, and improving reasoning in LLMs.
Counter-arguments (CAs) are a good means to improve the critical-thinking skills of learners, especially given that one has to thoroughly consider the logic of initial arguments (IA) when composing their CA. Although several tasks have been created for identifying the logical structure of CAs, no prior work has focused on capturing multiple interpretations of logical structures due to their complexity. In this work, we create CALSA+, a dataset consisting of 134 CAs annotated with 13 logical predicate questions. CALSA+ contains 1,742 instances annotated by 3 expert annotators (5,226 total annotations) with good agreement (Krippendorff \alpha=0.46). Using CALSA+, we train a model with Reinforcement Learning with Verifiable Rewards (RLVR) to identify multiple logical interpretations and show that models trained with RLVR can perform on par with much bigger proprietary models. Our work is the first to attempt to annotate all the interpretations of logical structure on top of CAs. We publicly release our dataset to facilitate research in CA logical structure identification.
Improving Low-Resource Sequence Labeling with Knowledge Fusion and Contextual Label Explanations
PDF ↗Sequence labeling remains a significant challenge in low-resource, domain-specific scenarios, particularly for character-dense languages. Existing methods primarily focus on enhancing model comprehension and improving data diversity to boost performance. However, these approaches still struggle with inadequate model applicability and semantic distribution biases in domain-specific contexts. To overcome these limitations, we propose a novel framework that combines an LLM-based knowledge enhancement workflow with a span-based Knowledge Fusion for Rich and Efficient Extraction (KnowFREE) model. Our workflow employs explanation prompts to generate precise contextual interpretations of target entities, effectively mitigating semantic biases and enriching the model’s contextual understanding. The KnowFREE model further integrates extension label features, enabling efficient nested entity extraction without relying on external knowledge during inference. Experiments on multiple domain-specific sequence labeling datasets demonstrate that our approach achieves state-of-the-art performance, effectively addressing the challenges posed by low-resource settings.
Multimodal misinformation, encompassing textual, visual, and cross-modal distortions, poses an increasing societal threat that is amplified by generative AI. Existing methods typically focus on a single type of distortion and struggle to generalize to unseen scenarios. In this work, we observe that different distortion types share common reasoning capabilities while also requiring task-specific skills. We hypothesize that joint training across distortion types facilitates knowledge sharing and enhances the model’s ability to generalize. To this end, we introduce TRUST-VL, a unified and explainable vision-language model for general multimodal misinformation detection. TRUST-VL incorporates a novel Question-Aware Visual Amplifier module, designed to extract task-specific visual features. To support training, we also construct TRUST-Instruct, a large-scale instruction dataset containing 198K samples featuring structured reasoning chains aligned with human fact-checking workflows. Extensive experiments on both in-domain and zero-shot benchmarks demonstrate that TRUST-VL achieves state-of-the-art performance, while also offering strong generalization and interpretability.
Graph Neural Networks (GNNs) have shown immense potential in improving the performance of large-scale models by effectively incorporating structured relational information. However, current approaches face two key challenges: (1) achieving robust semantic alignment between graph representations and large models, and (2) ensuring interpretability in the generated outputs. To address these challenges, we propose ExGLM (Explainable Graph Language Model), a novel training framework designed to seamlessly integrate graph and language modalities while enhancing transparency. Our framework introduces two core components: (1) a graph-language synergistic alignment module, which aligns graph structures with language model to ensure semantic consistency across modalities; and (2) a judge-and-improve paradigm, which allows the language model to iteratively evaluate, refine, and prioritize responses with higher interpretability, thereby improving both performance and transparency. Extensive experiments conducted on three benchmark datasets—ogbn-arxiv, Cora, and PubMed—demonstrate that ExGLM not only surpasses existing methods in efficiency but also generates outputs that are significantly more interpretable, effectively addressing the primary limitations of current approaches.
Text-based person search (TBPS) enables the retrieval of person images from large-scale databases using natural language descriptions, offering critical value in surveillance applications. However, a major challenge lies in the labor-intensive process of obtaining high-quality textual annotations, which limits scalability and practical deployment. To address this, we introduce two complementary modules: Multi-Turn Text Generation (MTG) and Multi-Turn Text Interaction (MTI). MTG generates rich pseudo-labels through simulated dialogues with MLLMs, producing fine-grained and diverse visual descriptions without manual supervision. MTI refines user queries at inference time through dynamic, dialogue-based reasoning, enabling the system to interpret and resolve vague, incomplete, or ambiguous descriptions—characteristics often seen in real-world search scenarios. Together, MTG and MTI form a unified and annotation-free framework that significantly improves retrieval accuracy, robustness, and usability. Extensive evaluations demonstrate that our method achieves competitive or superior results while eliminating the need for manual captions, paving the way for scalable and practical deployment of TBPS systems.
LLMs have shown strong performance on human-centric reasoning tasks. While previous evaluations have explored whether LLMs can infer intentions or detect deception, they often overlook the individualized reasoning styles that influence how people interpret and act in social contexts. Social deduction games (SDGs) provide a natural testbed for evaluating individualized reasoning styles, where different players may adopt diverse but contextually valid reasoning strategies under identical conditions. To address this, we introduce InMind, a cognitively grounded evaluation framework designed to assess whether LLMs can capture and apply personalized reasoning styles in SDGs. InMind enhances structured gameplay data with round-level strategy traces and post-game reflections, collected under both Observer and Participant modes. It supports four cognitively motivated tasks that jointly evaluate both static alignment and dynamic adaptation. As a case study, we apply InMind to the game Avalon, evaluating 11 state-of-the-art LLMs. General-purpose LLMs, even GPT-4o frequently rely on lexical cues, struggling to anchor reflections in temporal gameplay or adapt to evolving strategies. In contrast, reasoning-enhanced LLMs like DeepSeek-R1 exhibit early signs of style-sensitive reasoning. These findings reveal key limitations in current LLMs’ capacity for individualized, adaptive reasoning, and position InMind as a step toward cognitively aligned human–AI interaction.
Large Language Models (LLMs) based agent systems have made great strides in real-world applications beyond traditional NLP tasks. This paper proposes a new LLM-based Multi-Agent System (LLM-MAS) benchmark, Collab-Overcooked, built on the popular Overcooked-AI game with more applicable and challenging tasks in interactive environments. Collab-Overcooked extends existing benchmarks in two novel ways. First, it provides a multi-agent framework supporting diverse tasks and objectives and encourages collaboration through natural language communication. Second, it introduces a spectrum of process-oriented evaluation metrics to assess the fine-grained collaboration capabilities of different LLM agents, a dimension often overlooked in prior work. We conduct extensive experiments with 13 popular LLMs and show that, while the LLMs exhibit a strong ability in goal interpretation, there are significant shortcomings in active collaboration and continuous adaptation, which are critical for efficiently fulfilling complex tasks. Notably, we highlight the strengths and weaknesses of LLM-MAS and provide insights for improving and evaluating LLM-MAS on a unified and open-source benchmark. The environments, 30 open-ended tasks, and the evaluation package are publicly available at https://github.com/YusaeMeow/Collab-Overcooked.
Co-Evolving LLMs and Embedding Models via Density-Guided Preference Optimization for Text Clustering
PDF ↗Large language models (LLMs) have shown strong potential in enhancing text clustering when combined with traditional embedding models. However, existing methods predominantly treat LLMs as static pseudo-oracles, i.e., unidirectionally querying them for similarity assessment or data augmentation, while never seeking feedback from embedding models to improve them. In this work, we propose a training framework that enables bidirectional refinement between LLMs and embedding models. We first design task-aware prompts to guide the LLM in generating interpretations for the input texts. These interpretations are projected into the embedding space, in which interpretations that are preferred by the embedding model are selected based on their distribution densities. The selected interpretations are then used to fine-tune the LLM via preference optimization to prioritize the generation of helpful interpretations. Meanwhile, we enhance the embedding model via contrastive learning on the generated interpretations and perform clustering on the output embeddings, leading to iterative co-training between the LLM and the embedding model. Experiments on 14 benchmark datasets across 5 tasks demonstrate the effectiveness of our method.
Contrastive Language–Image Pre-training (CLIP) has recently demonstrated remarkable success in aligning vision and language. Aligning time series with text leverages the rich semantic cues of language to enhance interpretability and generalization, addressing a largely underexplored area of research. Although applying the CLIP training paradigm to time-series and language pairs is promising, it may result in label collapse due to the sparse semantic annotations and the absence of visual cues in time-series data. To address this, we introduce Time Series CLIP (TS-CLIP), a novel approach that tackles label collapse using a synonym bank mechanism. Synonym bank exploits word analogy phenomena to generate potential synonym embeddings as alignment targets. Specifically, the synonym bank facilitates aligning time series with a word distribution instead of a precise textual description. We conducted extensive zero-shot and few-shot experiments on 128 sub-datasets from the UCR archive. The results show that TS-CLIP achieves state-of-the-art (SOTA) performance in zero-shot settings on 51 datasets. Comprehensive ablation studies and visualization analyzes reveal that TS-CLIP effectively aligns time series with natural language. To the best of our knowledge, this is the first foundational model to achieve general time series and natural language alignment. TS-CLIP introduces a new paradigm for the semantic understanding of time series and opens the possibility of integrating the time series modality into multimodal large models.
Pruning the Paradox: How CLIP’s Most Informative Heads Enhance Performance While Amplifying Bias
PDF ↗CLIP is one of the most popular foundation models and is heavily used for many vision-language tasks, yet little is known about its inner workings. As CLIP is increasingly deployed in real-world applications, it is becoming even more critical to understand its limitations and embedded social biases to mitigate potentially harmful downstream consequences. However, the question of what internal mechanisms drive both the impressive capabilities as well as problematic shortcomings of CLIP has largely remained unanswered. To bridge this gap, we study the conceptual consistency of text descriptions for attention heads in CLIP-like models. Specifically, we propose Concept Consistency Score (CCS), a novel interpretability metric that measures how consistently individual attention heads in CLIP models align with specific concepts. Our soft-pruning experiments reveal that high CCS heads are critical for preserving model performance, as pruning them leads to a significantly larger performance drop than pruning random or low CCS heads. Notably, we find that high CCS heads capture essential concepts and play a key role in out-of-domain detection, concept-specific reasoning, and video-language understanding. Moreover, we prove that high CCS heads learn spurious correlations which amplify social biases. These results position CCS as a powerful interpretability metric exposing the paradox of performance and social biases in CLIP models.
Multi-hop question answering (QA) remains challenging, as solutions must reliably integrate and reconcile evidence from multiple sources without succumbing to error propagation. While large language models (LLMs) have achieved substantial improvements via chain-of-thought (CoT) prompting and retrieval-augmented generation, these methods typically adopt a forward-only workflow—early mistakes persist throughout inference, and contradictions discovered later cannot systematically trigger re-evaluation. To address this limitation, we present ReAgent, a reversible multi-agent reasoning framework. Specifically, ReAgent enables agents to backtrack to earlier valid states when conflicts arise, thereby isolating and rectifying flawed assumptions before they undermine subsequent reasoning. Our approach combines explicit local and global rollback protocols with modular role specialization, resulting in a flexible and error-tolerant pipeline. Empirical evaluation on three multi-hop QA benchmarks demonstrates consistent performance gains of approximately 6% over forward-only baselines, in addition to enhanced interpretability. These findings highlight the value of non-monotonic, backtracking-driven inference in complex QA scenarios and point to broader implications for multi-agent collaboration in knowledge-intensive tasks.
Results reported in large-scale multilingual evaluations are often fragmented and confounded by factors such as target languages, differences in experimental setups, and model choices. We propose a framework that disentangles these confounding variables and introduces three interpretable metrics—the performance realisation ratio, its coefficient of variation, and language potential—enabling a finer-grained and more insightful quantification of actual performance disparities across both (i) models and (ii) languages. Through a case study of 13 model variants on 11 multilingual datasets, we demonstrate that our framework provides a more reliable measurement of model performance and language disparities, particularly for low-resource languages, which have so far proven challenging to evaluate. Importantly, our results reveal that higher overall model performance does not necessarily imply greater fairness across languages.
Figure-to-Text (F2T) tasks aim to convert structured figure information into natural language text, serving as a bridge between visual perception and language understanding.However, existing evaluation methods remain limited: 1) Reference-based methods can only capture shallow semantic similarities and rely on costly labeled reference text; 2) Reference-free methods depend on multimodal large language models, which suffer from low efficiency and instruction sensitivity; 3) Existing methods provide only sample-level evaluations, lacking interpretability and alignment with expert-level multi-dimensional evaluation criteria.Accordingly, we propose F2TEval, a five-dimensional reference-free evaluation method aligned with expert criteria, covering faithfulness, completeness, conciseness, logicality, and analysis, to support fine-grained evaluation. We design a lightweight mixture-of-experts model that incorporates independent scoring heads and applies the Hilbert-Schmidt Independence Criterion to optimize the disentanglement of scoring representations across dimensions. Furthermore, we construct F2TBenchmark, a human-annotated benchmark dataset covering 21 chart types and 35 application domains, to support research on F2T evaluation. Experimental results demonstrate our model’s superior performance and efficiency, outperforming Gemini-2.0 and Claude-3.5 with only 0.9B parameters.
In recent years, Large Language Models (LLMs) have been widely applied to legal tasks. To enhance their understanding of legal texts and improve reasoning accuracy, a promising approach is to incorporate legal theories. One of the most widely adopted theories is the Four-Element Theory (FET), which defines the crime constitution through four elements: Subject, Object, Subjective Aspect, and Objective Aspect. While recent work has explored prompting LLMs to follow FET, our evaluation demonstrates that LLM-generated four-elements are often incomplete and less representative, limiting their effectiveness in legal reasoning.To address these issues, we present JUREX-4E, an expert-annotated four-element knowledge base covering 155 criminal charges. The annotations follow a progressive hierarchical framework grounded in legal source validity and incorporate diverse interpretive methods to ensure precision and authority. We evaluate JUREX-4E on the Similar Charge Disambiguation task and apply it to Legal Case Retrieval. Experimental results validate the high quality of JUREX-4E and its substantial impact on downstream legal tasks, underscoring its potential for advancing legal AI applications. The dataset and code are available at: https://github.com/THUlawtech/JUREX
Machine-generated text detection is critical for preventing misuse of large language models (LLMs). Although LLMs have recently excelled at mimicking human writing styles, they still suffer from factual hallucinations manifested as entity-relation inconsistencies with real-world knowledge. Current detection methods inadequately address the authenticity of the entity graph, which is a key discriminative feature for identifying machine-generated content. To bridge this gap, we propose a fact-aware model that assesses discrepancies between textual and factual entity graphs through graph comparison. In order to holistically analyze context information, our approach employs hierarchical feature extraction with gating units, enabling the adaptive fusion of multi-grained features from entity, sentence, and document levels. Experimental results on three public datasets demonstrate that our approach outperforms the state-of-the-art methods. Interpretability analysis shows that our model can capture the differences in entity graphs between machine-generated and human-written texts.
Memes have emerged as a popular form of multimodal online communication, where their interpretation heavily depends on the specific context in which they appear. Current approaches predominantly focus on isolated meme analysis, either for harmful content detection or standalone interpretation, overlooking a fundamental challenge: the same meme can express different intents depending on its conversational context. This oversight creates an evaluation gap: although humans intuitively recognize how context shapes meme interpretation, Large Vision Language Models (LVLMs) can hardly understand context-dependent meme intent. To address this critical limitation, we introduce MemeReaCon, a novel benchmark specifically designed to evaluate how LVLMs understand memes in their original context. We collected memes from five different Reddit communities, keeping each meme’s image, the post text, and user comments together. We carefully labeled how the text and meme work together, what the poster intended, how the meme is structured, and how the community responded. Our tests with leading LVLMs show a clear weakness: models either fail to interpret critical information in the contexts, or overly focus on visual details while overlooking communicative purpose. MemeReaCon thus serves both as a diagnostic tool exposing current limitations and as a challenging benchmark to drive development toward more sophisticated LVLMs of the context-aware understanding.
Humanizing Machines: Rethinking LLM Anthropomorphism Through a Multi-Level Framework of Design
PDF ↗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.