Reliable responses from large language models (LLMs) require adherence to user instructions and retrieved information. While alignment techniques help LLMs align with human intentions and values, improving context-faithfulness through alignment remains underexplored. To address this, we propose Context-DPO, the first alignment method specifically designed to enhance LLMs’ context-faithfulness. We introduce ConFiQA, a benchmark that simulates Retrieval-Augmented Generation (RAG) scenarios with knowledge conflicts to evaluate context-faithfulness. By leveraging faithful and stubborn responses to questions with provided context from ConFiQA, our Context-DPO aligns LLMs through direct preference optimization. Extensive experiments demonstrate that our Context-DPO significantly improves context-faithfulness, achieving 35% to 280% improvements on popular open-source models. Further analysis demonstrates that Context-DPO preserves LLMs’ generative capabilities while providing interpretable insights into context utilization.
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
Interactive Fiction games (IF games) are where players interact through natural language commands. While recent advances in Artificial Intelligence agents have reignited interest in IF games as a domain for studying decision-making, existing approaches prioritize task-specific performance metrics over human-like comprehension of narrative context and gameplay logic. This work presents a cognitively inspired framework that guides Large Language Models (LLMs) to learn and play IF games systematically. Our proposed **L**earning to **P**lay **L**ike **H**umans (LPLH) framework integrates three key components: (1) structured map building to capture spatial and narrative relationships, (2) action learning to identify context-appropriate commands, and (3) feedback-driven experience analysis to refine decision-making over time. By aligning LLMs-based agents’ behavior with narrative intent and commonsense constraints, LPLH moves beyond purely exploratory strategies to deliver more interpretable, human-like performance. Crucially, this approach draws on cognitive science principles to more closely simulate how human players read, interpret, and respond within narrative worlds. As a result, LPLH reframes the IF games challenge as a learning problem for LLMs-based agents, offering a new path toward robust, context-aware gameplay in complex text-based environments.
Recent studies on reasoning in language models (LMs) have sparked a debate on whether they can learn systematic inferential principles or merely exploit superficial patterns in the training data. To understand and uncover the mechanisms adopted for formal reasoning in LMs, this paper presents a mechanistic interpretation of syllogistic inference. Specifically, we present a methodology for circuit discovery aimed at interpreting content-independent and formal reasoning mechanisms. Through two distinct intervention methods, we uncover a sufficient and necessary circuit involving middle-term suppression that elucidates how LMs transfer information to derive valid conclusions from premises. Furthermore, we investigate how belief biases manifest in syllogistic inference, finding evidence of partial contamination from additional attention heads responsible for encoding commonsense and contextualized knowledge. Finally, we explore the generalization of the discovered mechanisms across various syllogistic schemes, model sizes and architectures. The identified circuit is sufficient and necessary for syllogistic schemes on which the models achieve high accuracy (\geq 60%), with compatible activation patterns across models of different families. Overall, our findings suggest that LMs learn transferable content-independent reasoning mechanisms, but that, at the same time, such mechanisms do not involve generalizable and abstract logical primitives, being susceptible to contamination by the same world knowledge acquired during pre-training.
Explainable Depression Detection in Clinical Interviews with Personalized Retrieval-Augmented Generation
PDF ↗Depression is a widespread mental health disorder, and clinical interviews are the gold standard for assessment. However, their reliance on scarce professionals highlights the need for automated detection. Current systems mainly employ black-box neural networks, which lack interpretability, which is crucial in mental health contexts. Some attempts to improve interpretability use post-hoc LLM generation but suffer from hallucination. To address these limitations, we propose RED, a Retrieval-augmented generation framework for Explainable depression Detection. RED retrieves evidence from clinical interview transcripts, providing explanations for predictions. Traditional query-based retrieval systems use a one-size-fits-all approach, which may not be optimal for depression detection, as user backgrounds and situations vary. We introduce a personalized query generation module that combines standard queries with user-specific background inferred by LLMs, tailoring retrieval to individual contexts. Additionally, to enhance LLM performance in social intelligence, we augment LLMs by retrieving relevant knowledge from a social intelligence datastore using an event-centric retriever. Experimental results on the real-world benchmark demonstrate RED’s effectiveness compared to neural networks and LLM-based baselines.
Large vision-language models (LVLMs) have shown great promise in medical applications, particularly in visual question answering (MedVQA) and diagnosis from medical images. However, existing datasets and models often fail to consider critical aspects of medical diagnostics, such as the integration of historical records and the analysis of disease progression over time. In this paper, we introduce MMXU (Multimodal and MultiX-ray Understanding), a novel dataset for MedVQA that focuses on identifying changes in specific regions between two patient visits. Unlike previous datasets that primarily address single-image questions, MMXU enables multi-image questions, incorporating both current and historical patient data. We demonstrate the limitations of current LVLMs in identifying disease progression on MMXU-test, even those that perform well on traditional benchmarks. To address this, we propose a MedRecord-Augmented Generation (MAG) approach, incorporating both global and regional historical records.Our experiments show that integrating historical records significantly enhances diagnostic accuracy by at least 20%, bridging the gap between current LVLMs and human expert performance. Additionally, we fine-tune models with MAG on MMXU-dev, which demonstrates notable improvements. We hope this work could illuminate the avenue of advancing the use of LVLMs in medical diagnostics by emphasizing the importance of historical context in interpreting medical images.Our dataset is released at github.
Large language models (LLMs) have revolutionized the field of natural language processing, enabling remarkable progress in various tasks. Different from objective tasks such as commonsense reasoning and arithmetic question-answering, the performance of LLMs on subjective tasks is still limited, where the perspective on the specific problem plays crucial roles for better interpreting the context and giving proper response. For example, in certain scenarios, LLMs may perform better when answering from an expert role perspective, potentially eliciting their relevant domain knowledge. In contrast, in some scenarios, LLMs may provide more accurate responses when answering from a third-person standpoint, enabling a more comprehensive understanding of the problem and potentially mitigating inherent biases. In this paper, we propose Reasoning through Perspective Transition (RPT), a method based on in-context learning that enables LLMs to dynamically select among direct, role, and third-person perspectives for the best way to solve corresponding subjective problem. Through extensive experiments on totally 12 subjective tasks by using both closed-source and open-source LLMs including GPT-4, GPT-3.5, Llama-3, and Qwen-2, our method outperforms widely used single fixed perspective based methods such as chain-of-thought prompting and expert prompting, highlights the intricate ways that LLMs can adapt their perspectives to provide nuanced and contextually appropriate responses for different problems.
Multimodal learning is garnering significant attention for its capacity to represent diverse human perceptions (e.g., linguistic, acoustic, and visual signals), achieving more natural and intuitive interactions with technology.However, the frequent occurrence of incomplete data, either within a single modality (intra-modality) or across different modalities (inter-modality), presents substantial challenges in reliable semantic interpretation and model reasoning.Furthermore, there is currently no robust representation learning mechanism capable of managing both intra-modality and inter-modality real-data deficiencies.To address this challenge, we present \text{T}^2DR, a two-tier deficiency-resistant framework for incomplete multimodal learning, which comprises two main modules:(1) Intra-Modal Deficiency-Resistant module (IADR): To address fine-grained deficiencies, we introduce Intra-Attn to focus on the available data while avoiding excessive suppression of the missing regions.(2) Inter-Modal Deficiency-Resistant module (IEDR): To handle coarse-grained deficiencies, we propose the shared feature prediction (SFP) to leverage cross-modal shared features for preliminary data imputation. Subsequently, we apply Inter-Attn to allocate appropriate attention to each modality based on the results from the capability-aware scorer (CAS).Extensive experiments are performed on two well-known multimodal benchmarks, CMU-MOSI and CMU-MOSEI, across various missing scenarios for sentiment analysis. Experimental results show that \text{T}^2DR significantly outperforms the SOTA models. Code is available at https://github.com/LH019/T2DR.
Understanding robustness is essential for building reliable NLP systems. Unfortunately, in the context of machine translation, previous work mainly focused on documenting robustness failures or improving robustness. In contrast, we study robustness from a model representation perspective by looking at internal model representations of ungrammatical inputs and how they evolve through model layers. For this purpose, we perform Grammatical Error Detection (GED) probing and representational similarity analysis. Our findings indicate that the encoder first detects the grammatical error, then corrects it by moving its representation toward the correct form. To understand what contributes to this process, we turn to the attention mechanism where we identify what we term *Robustness Heads*. We find that *Robustness Heads* attend to interpretable linguistic units when responding to grammatical errors, and that when we fine-tune models for robustness, they tend to rely more on *Robustness Heads* for updating the ungrammatical word representation.
Large Language Models (LLMs) are often challenged by generating erroneous or hallucinated responses, especially in complex reasoning tasks. Leveraging Knowledge Graphs (KGs) as external knowledge sources has emerged as a viable solution. However, existing KG-enhanced methods, either retrieval-based or agent-based, encounter difficulties in accurately retrieving knowledge and efficiently traversing KGs at scale. In this paper, we propose a unified framework, FiDeLiS, designed to improve the factuality of LLM responses by anchoring answers to verifiable reasoning steps retrieved from KGs. To achieve this, we leverage step-wise beam search with a deductive scoring function, allowing the LLM to validate reasoning process step by step, and halt the search once the question is deducible. In addition, we propose a Path-RAG module to pre-select a smaller candidate set for each beam search step, reducing computational costs by narrowing the search space. Extensive experiments show that our method, as a training-free framework, not only improve the performance but also enhance the factuality and interpretability across different benchmarks.
Advances in large language models (LLMs) have spurred research into enhancing their reasoning capabilities, particularly in math-rich STEM (Science, Technology, Engineering, and Mathematics) documents.While LLMs can generate equations or solve math-related queries, their ability to fully understand and interpret abstract mathematical symbols in long, math-rich documents remains limited. In this paper, we introduce STEM-PoM, a comprehensive benchmark dataset designed to evaluate LLMs’ reasoning abilities on math symbols within contextual scientific text. The dataset, sourced from real-world ArXiv documents, contains over 2K math symbols classified as main attributes of variables, constants, operators, and unit descriptors, with additional sub-attributes including scalar/vector/matrix for variables and local/global/discipline-specific labels for both constants and operators. Our extensive experiments demonstrate that state-of-the-art LLMs achieve an average accuracy of 20-60% under in-context learning and 50-60% with fine-tuning, highlighting a substantial gap in their ability to classify mathematical symbols. By improving LLMs’ mathematical symbol classification, STEM-PoM further enhances models’ downstream mathematical reasoning capabilities. The code and data are available at https://github.com/jiaruzouu/STEM-PoM.
Large language models (LLMs) have made remarkable progress in various domains, yet they often suffer from repetitive text generation, a phenomenon we refer to as the ”Repeat Curse”. While previous studies have proposed decoding strategies to mitigate repetition, the underlying mechanism behind this issue remains insufficiently explored. In this work, we investigate the root causes of repetition in LLMs through the lens of mechanistic interpretability. Inspired by recent advances in Sparse Autoencoders (SAEs), which enable monosemantic feature extraction, we propose a novel approach—”Duplicatus Charm”—to induce and analyze the Repeat Curse. Our method systematically identifies “Repetition Features” -the key model activations responsible for generating repetitive outputs. First, we locate the layers most involved in repetition through logit analysis. Next, we extract and stimulate relevant features using SAE-based activation manipulation. To validate our approach, we construct a repetition dataset covering token and paragraph level repetitions and introduce an evaluation pipeline to quantify the influence of identified repetition features. Furthermore, by deactivating these features, we have effectively mitigated the Repeat Curse.
Why Not Act on What You Know? Unleashing Safety Potential of LLMs via Self-Aware Guard Enhancement
PDF ↗Large Language Models (LLMs) have shown impressive capabilities across various tasks but remain vulnerable to meticulously crafted jailbreak attacks. In this paper, we identify a critical safety gap: while LLMs are adept at detecting jailbreak prompts, they often produce unsafe responses when directly processing these inputs. Inspired by this insight, we propose SAGE(Self-Aware Guard Enhancement), a training-free defense strategy designed to align LLMs’ strong safety discrimination performance with their relatively weaker safety generation ability. SAGE consists of two core components: a Discriminative Analysis Module and a Discriminative Response Module, enhancing resilience against sophisticated jailbreak attempts through flexible safety discrimination instructions. Extensive experiments demonstrate SAGE’s effectiveness and robustness across various open-source and closed-source LLMs of different sizes and architectures, achieving an average 99% defense success rate against numerous complex and covert jailbreak methods while maintaining helpfulness on general benchmarks. We further conduct mechanistic interpretability analysis through hidden states and attention distributions, revealing the underlying mechanisms of this detection-generation discrepancy. Our work thus contributes to developing future LLMs with coherent safety awareness and generation behavior. Our code and datasets are publicly available at https://github.com/NJUNLP/SAGE.
Applying Large Language Models (LLM) to solve math problems is one of the hottest research topics at present. Traditional Chain-of-Thought-based methods typically generate the reasoning path in a chain structure, leading to unnecessary interference caused by non-zero self-attention among weakly related reasoning steps. Such a setting also differs from humans’ typical graph-structured reasoning habit (with an inter-step relationship graph in mind). To solve the problem, this paper proposes a novel decoding method for Transformer-based LLM, named Self-attention-based Graph-of-Thought (SaGoT). SaGoT constructs a thought graph simultaneously as an LLM inference (based on a newly defined inter-step self-attention indicator), and generates reasoning steps with a novel graph-structured self-attention mechanism. It is a significant contribution for SaGoT to enable an LLM’s graph-like reasoning ability by modifying its inner working operations, compared to SOTA prompting methods that are ex-post, rely on huge LLMs and redundant reasoning step generation to form a graph (inefficient & non-human-like). In addition, SaGoT is a training-free technique that can be seamlessly incorporated into pre-trained Transformer-based LLMs. Our experimental results have shown that SaGoT could significantly enhance mathematical reasoning accuracy without the reliance on huge computationally over-expensive LLMs. It also avoids SOTA methods’ performance degradation issues when the LLM is too small to comprehend complex prompts. Moreover, SaGoT integrates intrinsic interpretability into the LLM’s reasoning procedure, intuitively assisting humans in understanding how an LLM views the relationships among its reasoning steps, and why the LLM succeeds or fails.
Language has become a prominent modality in computer vision with the rise of LLMs. Despite supporting long context-lengths, their effectiveness in handling long-term information gradually declines with input length. This becomes critical, especially in applications such as long-form video understanding. In this paper, we introduce a Language Repository (LangRepo) for LLMs, that maintains concise and structured information as an interpretable (i.e., all-textual) representation. Our repository is updated iteratively based on multi-scale video chunks. We introduce write and read operations that focus on pruning redundancies in text, and extracting information at various temporal scales. The proposed framework is evaluated on zero-shot visual question-answering benchmarks, showing state-of-the-art performance at its scale. Our code is available at https://github.com/kkahatapitiya/LangRepo.
Is LLM an Overconfident Judge? Unveiling the Capabilities of LLMs in Detecting Offensive Language with Annotation Disagreement
PDF ↗Large Language Models (LLMs) have become essential for offensive language detection, yet their ability to handle annotation disagreement remains underexplored. Disagreement samples, which arise from subjective interpretations, pose a unique challenge due to their ambiguous nature. Understanding how LLMs process these cases, particularly their confidence levels, can offer insight into their alignment with human annotators. This study systematically evaluates the performance of multiple LLMs in detecting offensive language at varying levels of annotation agreement. We analyze binary classification accuracy, examine the relationship between model confidence and human disagreement, and explore how disagreement samples influence model decision-making during few-shot learning and instruction fine-tuning. Our findings reveal that LLMs struggle with low-agreement samples, often exhibiting overconfidence in these ambiguous cases. However, utilizing disagreement samples in training improves both detection accuracy and model alignment with human judgment. These insights provide a foundation for enhancing LLM-based offensive language detection in real-world moderation tasks.
From Phrases to Subgraphs: Fine-Grained Semantic Parsing for Knowledge Graph Question Answering
PDF ↗The recent emergence of large language models (LLMs) has brought new opportunities to knowledge graph question answering (KGQA), but also introduces challenges such as semantic misalignment and reasoning noise. Semantic parsing (SP), previously a mainstream approach for KGQA, enables precise graph pattern matching by mapping natural language queries to executable logical forms. However, it faces limitations in scalability and generalization, especially when dealing with complex, multi-hop reasoning tasks.In this work, we propose a Fine-Grained Semantic Parsing (FGSP) framework for KGQA. Our framework constructs a fine-grained mapping library via phrase-level segmentation of historical question-logical form pairs, and performs online retrieval and fusion of relevant subgraph fragments to answer complex queries. This fine-grained, compositional approach ensures tighter semantic alignment between questions and knowledge graph structures, enhancing both interpretability and adaptability to diverse query types. Experimental results on two KGQA benchmarks demonstrate the effectiveness of FGSP, with a notable 18.5% relative F1 performance improvement over the SOTA on the complex multi-hop CWQ dataset. Our code is available at https://github.com/NUSTM/From-Phrases-to-Subgraphs.
Understanding how feature representations evolve across layers in large language models (LLMs) is key to improving their interpretability and robustness. While recent studies have identified critical layers linked to specific functions or behaviors, these efforts typically rely on data-dependent analyses of fine-tuned models, limiting their use to post-hoc settings. In contrast, we introduce a data-oblivious approach to identify intrinsic critical layers in pre-fine-tuned LLMs by analyzing representation dynamics via Centered Kernel Alignment (CKA). We show that layers with significant shifts in representation space are also those most affected during fine-tuning—a pattern that holds consistently across tasks for a given model. Our spectral analysis further reveals that these shifts are driven by changes in the top principal components, which encode semantic transitions from rationales to conclusions.We further apply these findings to two practical scenarios: efficient domain adaptation, where fine-tuning critical layers leads to greater loss reduction compared to non-critical layers; and backdoor defense, where freezing them reduces attack success rates by up to 40%.
Video generation has many unique challenges beyond those of image generation. The temporal dimension introduces extensive possible variations across frames, over which consistency and continuity may be violated. In this work, we evaluate the emergence of new concepts and relation transitions as time progresses in a video, which we refer to as Temporal Compositionality. We propose TC-Bench, a benchmark of meticulously crafted text prompts, ground truth videos, and new evaluation metrics. The prompts articulate the initial and final states of scenes, effectively reducing ambiguities for frame development. In addition, by collecting corresponding ground-truth videos, the benchmark can be used for text-to-video and image-to-video generation. We develop new metrics to measure the completeness of component transitions, which demonstrate significantly higher correlations with human judgments than existing metrics. Our experiments reveal that contemporary video generators are still weak in prompt understanding and achieve less than 20% of the compositional changes, highlighting enormous improvement space. Our analysis indicates that current video generation models struggle to interpret descriptions of compositional changes and synthesize various components across different time steps.
Continued Pretraining and Interpretability-Based Evaluation for Low-Resource Languages: A Galician Case Study
PDF ↗Recent advances in Large Language Models (LLMs) have led to remarkable improvements in language understanding and text generation. However, challenges remain in enhancing their performance for underrepresented languages, ensuring continual learning without catastrophic forgetting, and developing robust evaluation methodologies. This work addresses these issues by investigating the impact of Continued Pretraining (CPT) on multilingual models and proposing a comprehensive evaluation framework for LLMs, focusing on the case of Galician language. Our first contribution explores CPT strategies for languages with limited representation in multilingual models. We analyze how CPT with Galician corpora improves text generation while assessing the trade-offs between linguistic enrichment and task-solving capabilities. Our findings show that CPT with small, high-quality corpora and diverse instructions enhances both task performance and linguistic quality. Our second contribution is a structured evaluation framework based on distinguishing task-based and language-based assessments, leveraging existing and newly developed benchmarks for Galician. Additionally, we contribute new Galician LLMs, datasets for evaluation and instructions, and an evaluation framework.
Document retrieval techniques are essential for developing large-scale information systems. The common approach involves using a bi-encoder to compute the semantic similarity between a query and documents. However, the scalar similarity often fail to reflect enough information, hindering the interpretation of retrieval results. In addition, this process primarily focuses on global semantics, overlooking the finer-grained semantic relationships between the query and the document’s content. In this paper, we introduce a novel method, \textbf{Ge}neration \textbf{A}ugmented \textbf{R}etrieval (\textbf{GeAR}), which not only improves the global document-query similarity through contrastive learning, but also integrates well-designed fusion and decoding modules. This enables GeAR to generate relevant context within the documents based on a given query, facilitating learning to retrieve local fine-grained information.Furthermore, when used as a retriever, GeAR does not incur any additional computational cost over bi-encoders. GeAR exhibits competitive retrieval performance across diverse scenarios and tasks. Moreover, qualitative analysis and the results generated by GeAR provide novel insights into the interpretation of retrieval results. The code, data, and models will be released at https://github.com/microsoft/LMOps.