Fact verification on knowledge graphs (KGs) uses the structured representation of entities and relations as evidence for validating claims. Previous methods for KG-based fact verification predominantly use natural language inference (NLI) models to predict entailment between claims and KG triples, based on implicit reasoning. We propose Programmatic Graph Reasoning (PGR), a novel framework that integrates large language models (LLMs) for fact verification on KGs. PGR explicitly encodes the reasoning process as a graph reasoning program composed of predefined functions to verify claims step by step. These functions are executed sequentially for graph reasoning and final result prediction. By making the graph reasoning process explicit, PGR ensures more precise and transparent reasoning steps compared to implicit methods. Experimental results on the FactKG dataset demonstrate that PGR achieves state-of-the-art performance with 86.82% accuracy, outperforming all the baseline models. Further analysis confirms the interpretability and effectiveness of our method in handling complex graph reasoning.
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
SCE: Semantic Consistency Enhanced Reinforcement Learning for Multi-Hop Knowledge Graph Reasoning
PDF ↗Multi-hop reasoning with reinforcement learning has proven effective in discovering inference paths in incomplete knowledge graphs. However, a major challenge remains: spurious paths (incorrect reasoning paths that accidentally lead to correct answers) often arise due to reward mechanisms that prioritize final results over reasoning quality. While existing approaches attempt to mitigate this issue using external rules, they often neglect the internal semantic consistency between the target triple and the intermediate triples along the reasoning path. In this paper, we propose a novel framework, \textbf{S}emantic \textbf{C}onsistency \textbf{E}nhanced Reinforcement Learning (SCE), which incorporates semantic consistency into the reward function to guide multi-hop reasoning. Experimental results demonstrate that SCE outperforms strong baseline methods and facilitates the discovery of more interpretable reasoning paths.
X-Boundary: Establishing Exact Safety Boundary to Shield LLMs from Jailbreak Attacks without Compromising Usability
PDF ↗With the widespread application of large language models (LLMs) across various domains, techniques for enhancing their security have progressed rapidly. In this paper, we reveal that although existing defense methods can improve the robustness of LLMs against jailbreaks, they compromise usability, i.e., reducing general capabilities or causing the over-refusal problem. From the perspective of LLM mechanism interpretability, we discover that these methods fail to establish a boundary that exactly distinguishes safe and harmful feature representations. Therefore, boundary-safe representations close to harmful representations are inevitably disrupted, leading to a decline in usability. To address this issue, we propose X-Boundary to push harmful representations away from boundary-safe representations and obtain an exact distinction boundary. In this way, harmful representations can be precisely erased without disrupting safe ones. Experimental results show that X-Boundary achieves state-of-the-art defense performance against both single-turn and multi-turn jailbreak attacks, while reducing the over-refusal rate by about 20% and maintaining nearly complete general capability. Furthermore, we theoretically prove and empirically verify that X-Boundary can accelerate the convergence process during training.
Cross-lingual consistency should be considered to assess cross-lingual transferability, maintain the factuality of the model knowledge across languages, and preserve the parity of language model performance. We are thus interested in analyzing, evaluating, and interpreting cross-lingual consistency for factual knowledge.To facilitate our study, we examine multiple pretrained models and tuned models with code-mixed coreferential statements that convey identical knowledge across languages. Interpretability approaches are leveraged to analyze the behavior of a model in cross-lingual contexts, showing different levels of consistency in multilingual models, subject to language families, linguistic factors, scripts, and a bottleneck in cross-lingual consistency on a particular layer. Code-switching training and cross-lingual word alignment objectives show the most promising results, emphasizing the worthiness of cross-lingual alignment supervision and code-switching strategies for both multilingual performance and cross-lingual consistency enhancement. In addition, experimental results suggest promising result for calibrating consistency on test time via activation patching.
Pre-trained language models (PLMs) have achieved remarkable knowledge graph completion(KGC) success. However, most methods derive KGC results mainly from triple-level and text-described learning, which lack the capability to capture long-term relational and structural information. Moreover, the absence of a visible reasoning process leads to poor interpretability and credibility of the completions. In this paper, we propose a path-enhanced pre-trained language model-based knowledge graph completion method (PEKGC), which employs multi-view generation to infer missing facts in triple-level and path-level simultaneously to address lacking long-term relational information and interpretability issues. Furthermore, a neighbor selector module is proposed to filter neighbor triples to provide the adjacent structural information. Besides, we propose a fact-level re-evaluation and a heuristic fusion ranking strategy for candidate answers to fuse multi-view predictions. Extensive experiments on the benchmark datasets demonstrate that our model significantly improves the performance of the KGC task.
While Vision-Language Models (VLMs) have achieved competitive performance in various tasks, their comprehension of the underlying structure and semantics of a scene remains understudied. To investigate the understanding of VLMs, we study their capability regarding object properties and relations in a controlled and interpretable manner. To this scope, we introduce CIVET, a novel and extensible framework for systematiC evaluatIon Via controllEd sTimuli. CIVET addresses the lack of standardized systematic evaluation for assessing VLMs’ understanding, enabling researchers to test hypotheses with statistical rigor. With CIVET, we evaluate five state-of-the-art VLMs on exhaustive sets of stimuli, free from annotation noise, dataset-specific biases, and uncontrolled scene complexity. Our findings reveal that 1) current VLMs can accurately recognize only a limited set of basic object properties; 2) their performance heavily depends on the position of the object in the scene; 3) they struggle to understand basic relations among objects. Furthermore, a comparative evaluation with human annotators reveals that VLMs still fall short of achieving human-level accuracy.
FG-PRM: Fine-grained Hallucination Detection and Mitigation in Language Model Mathematical Reasoning
PDF ↗Hallucinations in large language models (LLMs) pose significant challenges in tasks requiring complex multi-step reasoning, such as mathematical problem-solving. Existing approaches primarily detect the presence of hallucinations but lack a nuanced understanding of their types and manifestations. In this paper, we first introduce a comprehensive taxonomy that categorizes the common hallucinations in mathematical reasoning tasks into six types. We then propose FG-PRM (Fine-Grained Process Reward Model), an augmented model designed to detect and mitigate hallucinations in a fine-grained, step-level manner. To address the limitations of manually labeling training data, we propose an automated method for generating fine-grained hallucination data using LLMs. Our FG-PRM demonstrates superior performance across two key tasks: 1) Fine-grained hallucination detection: classifying hallucination types for each reasoning step; and 2) Verification: ranking multiple LLM-generated outputs to select the most accurate solution. Our experiments show that FG-PRM excels in fine-grained hallucination detection and substantially boosts the performance of LLMs on GSM8K and MATH benchmarks. These results highlight the benefits of fine-grained supervision in enhancing the reliability and interpretability of LLM reasoning processes. Codes and datasets are available at: https://github.com/du-nlp-lab/FG-PRM.
ReLoop: “Seeing Twice and Thinking Backwards” via Closed-loop Training to Mitigate Hallucinations in Multimodal understanding
PDF ↗While Multimodal Large Language Models (MLLMs) have achieved remarkable progress in open-ended visual question answering, they remain vulnerable to hallucinations. These are outputs that contradict or misrepresent input semantics, posing a critical challenge to the reliability and factual consistency. Existing methods often rely on external verification or post-hoc correction, lacking an internal mechanism to validate outputs directly during training. To bridge this gap, we propose ReLoop, a unified closed-loop training framework that encourages multimodal consistency for cross-modal understanding in MLLMs. ReLoop adopts a ring-shaped structure that integrates three complementary consistency feedback mechanisms, obliging MLLMs to “seeing twice and thinking backwards”. Specifically, ReLoop employs the frozen Consistency Feedback Plugin (CFP), comprising semantic reconstruction, visual description, and an attention supervision module for attention alignment. These components collectively enforce semantic reversibility, visual consistency, and interpretable attention, enabling the model to correct its outputs during training. Extensive evaluations and analyses demonstrate the effectiveness of ReLoop in reducing hallucination rates across multiple benchmarks, establishing a robust method for hallucination mitigation in MLLMs. We will release our source code and data in the camera-ready version. The code is available at: https://github.com/ZiyanHuang11/Reloop-hallucinations.
DocAssistant: Integrating Key-region Reading and Step-wise Reasoning for Robust Document Visual Question Answering
PDF ↗Understanding the multimodal documents is essential for accurately extracting relevant evidence and using it for reasoning. Existing document understanding models struggle to focus on key information and tend to generate answers straightforwardly, ignoring evidence from source documents and lacking interpretability. In this work, we improve the visual encoder to focus on key information relevant to the question and address the shortcomings of existing document visual question-answering datasets to provide the model with the ability to answer questions step-wise, dubbed DocAssistant. Specifically, for the visual side, we propose an effective vision-language adaptation that fuses text into visual encoders without compromising the performance of the original model. For the language side, we use Multimodal Large Language Models (MLLMs) as data generators and checkers to produce high-quality step-wise question-and-answer pairs for document images. We then use the generated high-quality data to train our enhanced model, specifically designed to solve complex questions that require reasoning or multi-hop question answering. The experimental results demonstrate the effectiveness of the model.
As language models become integral to critical workflows, assessing their behavior remains a fundamental challenge – human evaluation is costly and noisy, while automated metrics provide only coarse, difficult-to-interpret signals. We introduce natural language unit tests, a paradigm that decomposes response quality into explicit, testable criteria, along with a unified scoring model, LMUnit, which combines multi-objective training across preferences, direct ratings, and natural language rationales. Through controlled human studies, we show this paradigm significantly improves inter-annotator agreement and enables more effective LLM development workflows. LMUnit achieves state-of-the-art performance on evaluation benchmarks including FLASK, BigGenBench, and RewardBench 2, while maintaining competitive results on the original RewardBench. These results validate both our proposed paradigm and scoring model, suggesting a promising path forward for language model evaluation and development. Our code has been released at github.com/ContextualAI/LMUnit with an MIT license.
Scientific claim verification against tables typically requires predicting whether a claim is supported or refuted given a table. However, we argue that predicting the final label alone is insufficient: it reveals little about the model’s reasoning and offers limited interpretability. To address this, we reframe table–text alignment as an explanation task, requiring models to identify the table cells essential for claim verification. We build a new dataset by extending the SciTab benchmark with human-annotated cell-level rationales. Annotators verify the claim label and highlight the minimal set of cells needed to support their decision. After the annotation process, we utilize the collected information and propose a taxonomy for handling ambiguous cases. Our experiments show that (i) incorporating table alignment information improves claim verification performance, and (ii) most LLMs, while often predicting correct labels, fail to recover human-aligned rationales, suggesting that their predictions do not stem from faithful reasoning.
Toward Inclusive Language Models: Sparsity-Driven Calibration for Systematic and Interpretable Mitigation of Social Biases in LLMs
PDF ↗Large Language Models (LLMs) such as GPT and LLaMA excel in natural language tasks, e.g., text generation and machine translation. However, inherent biases from training on vast Internet datasets potentially amplify harmful stereotypes—widely held, oversimplified, and often inaccurate generalizations about groups of people. Our contribution introduces a novel, systematic, and architecture-aware method to identify and mitigate stereotypical bias in decoder-only transformer models. This interpretable approach operates without gradient access or retraining from scratch. We first evaluate bias and then apply a bias localization mechanism that correlates internal activations with a newly defined Context Influence (CI) Score. Our method pinpoints specific attention heads that consistently align with biased shifts in model predictions. To mitigate this, we introduce a soft pruning strategy that scales attention head parameters based on their correlation strength, followed by lightweight fine-tuning to maintain fluent text generation. Experiments across five models demonstrate our approach reduces bias by up to 37% on BBQ, 32% on StereoSet, and 33% on CrowS-Pairs while simultaneously improving reasoning performance on MMLU by up to 10%.
Accurate analysis of pathological images is essential for automated tumor diagnosis but remains challenging due to high structural similarity and subtle morphological variations in tissue images. Current vision-language (VL) models often struggle to capture the complex reasoning required for interpreting structured pathological reports. To address these limitations, we propose PathoHR-Bench, a novel benchmark designed to evaluate VL models’ abilities in hierarchical semantic understanding and compositional reasoning within the pathology domain. Results of this benchmark reveal that existing VL models fail to effectively model intricate cross-modal relationships, hence limiting their applicability in clinical setting. To overcome this, we further introduce a pathology-specific VL training scheme that generates enhanced and perturbed samples for multimodal contrastive learning. Experimental evaluations demonstrate that our approach achieves state-of-the-art performance on PathoHR-Bench and six additional pathology datasets, highlighting its effectiveness in fine-grained pathology representation.
This paper proposes Complete Textual Concept Bottleneck Model (CT-CBM), a novel TCBM generator building concept labels in a fully unsupervised manner using a small language model, eliminating both the need for predefined human labeled concepts and LLM annotations. CT-CBM iteratively targets and adds important and identifiable concepts in the bottleneck layer to create a complete concept basis. CT-CBM achieves striking results against competitors in terms of concept basis completeness and concept detection accuracy, offering a promising solution to reliably enhance interpretability of NLP classifiers.
A Structured Framework for Evaluating and Enhancing Interpretive Capabilities of Multimodal LLMs in Culturally Situated Tasks
PDF ↗This study aims to test and evaluate the capabilities and characteristics of current mainstream Visual Language Models (VLMs) in generating critiques for traditional Chinese painting. To achieve this, we first developed a quantitative framework for Chinese painting critique. This framework was constructed by extracting multi-dimensional evaluative features covering evaluative stance, feature focus, and commentary quality from human expert critiques using a zero-shot classification model. Based on these features, several representative critic personas were defined and quantified. This framework was then employed to evaluate selected VLMs such as Llama, Qwen, or Gemini. The experimental design involved persona-guided prompting to assess the VLM’s ability to generate critiques from diverse perspectives. Our findings reveal the current performance levels, strengths, and areas for improvement of VLMs in the domain of art critique, offering insights into their potential and limitations in complex semantic understanding and content generation tasks.
Understanding the decision-making processes of large language models (LLMs) is essential for their trustworthy development and deployment, however, current interpretability methods often face challenges such as low resolution and high computational cost. To address these limitations, we propose the Multi-Layer Attention Consistency Score (MACS), a novel, lightweight, and easily deployable heuristic for estimating the importance of input tokens in decoder-based models. MACS measures contributions of input tokens based on the consistency of maximal attention. Empirical evaluations demonstrate that MACS achieves a favorable trade-off between interpretability quality and computational efficiency, showing faithfulness comparable to complex techniques with a 22% decrease in VRAM usage and 30% reduction in latency.
A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models
PDF ↗Large Language Models (LLMs) have transformed natural language processing, yet their internal mechanisms remain largely opaque. Recently, mechanistic interpretability has attracted significant attention from the research community as a means to understand the inner workings of LLMs. Among various mechanistic interpretability approaches, Sparse Autoencoders (SAEs) have emerged as a promising method due to their ability to disentangle the complex, superimposed features within LLMs into more interpretable components. This paper presents a comprehensive survey of SAEs for interpreting and understanding the internal workings of LLMs. Our major contributions include: (1) exploring the technical framework of SAEs, covering basic architecture, design improvements, and effective training strategies; (2) examining different approaches to explaining SAE features, categorized into input-based and output-based explanation methods; (3) discussing evaluation methods for assessing SAE performance, covering both structural and functional metrics; and (4) investigating real-world applications of SAEs in understanding and manipulating LLM behaviors.
FinHEAR: Human Expertise and Adaptive Risk-Aware Temporal Reasoning for Financial Decision-Making
PDF ↗Financial decision-making presents unique challenges for language models, requiring them to handle temporally evolving, risk-sensitive, and event-driven contexts. While large language models (LLMs) demonstrate strong general reasoning abilities, they often overlook key behavioral patterns underlying human financial behavior—such as expert reliance under information asymmetry, loss-averse risk adjustment, and temporal adaptation. We propose FinHEAR, a multi-agent framework for Human Expertise and Adaptive Risk-aware reasoning. FinHEAR coordinates multiple LLM-based agents to capture historical trends, interpret current events, and incorporate expert knowledge within a unified, event-aware pipeline. Grounded in behavioral economics, FinHEAR features mechanisms for expert-guided retrieval to reduce information asymmetry, dynamic position sizing to reflect loss aversion, and feedback-driven refinement to enhance temporal consistency. Experiments on a curated real-world financial dataset show that FinHEAR consistently outperforms strong baselines in both trend forecasting and decision-making.
We introduce SQLSpace, a human-interpretable, generalizable, compact representation for text-to-SQL examples derived with minimal human intervention. We demonstrate the utility of these representations in evaluation with three use cases: (i) closely comparing and contrasting the composition of popular NL2SQL benchmarks to identify unique dimensions of examples they evaluate, (ii) understanding model performance at a granular level beyond overall accuracy scores, and (iii) improving model performance through targeted query rewriting based on learned correctness estimation. We show that SQLSpace enables analysis that would be difficult with raw examples alone: it reveals compositional differences between benchmarks, exposes performance patterns obscured by accuracy alone, and supports modeling of query success.
Text-to-image synthesis has made remarkable progress, yet accurately interpreting complex and lengthy prompts remains challenging, often resulting in semantic inconsistencies and missing details. Existing solutions, such as fine-tuning, are model-specific and require training, while prior automatic prompt optimization (APO) approaches typically lack systematic error analysis and refinement strategies, resulting in limited reliability and effectiveness. Meanwhile, test-time scaling methods operate on fixed prompts and on noise or sample numbers, limiting their interpretability and adaptability. To solve these, we introduce a flexible and efficient test-time prompt optimization strategy that operates directly on the input text. We propose a plug-and-play multi-agent system called GenPilot, integrating error analysis, clustering-based adaptive exploration, fine-grained verification, and a memory module for iterative optimization. Our approach is model-agnostic, interpretable, and well-suited for handling long and complex prompts. Simultaneously, we summarize the common patterns of errors and the refinement strategy, offering more experience and encouraging further exploration. Experiments on DPG-bench and Geneval with improvements of up to 16.9% and 5.7% demonstrate the strong capability of our methods in enhancing the text and image consistency and structural coherence of generated images, revealing the effectiveness of our test-time prompt optimization strategy. The code is available at https://github.com/27yw/GenPilot.