Large language models (LLMs) are designed for discrete tokens, yet they operate in a continuous embedding space. Recent context compression methods exploit this property by encoding text into dense vectors for frozen LLM decoding. However, a key question remains unanswered: how does a frozen LLM interpret continuous vectors that encode complex semantics? We investigate this through controlled reconstruction experiments. Our analysis reveals a critical geometric property: compression encoders learn to produce vectors with L2 norms two orders of magnitude higher than standard embeddings. We show that this high-norm signal is causally necessary for the frozen LLM to decode compressed information. Based on this finding, we propose a landmark-based compression framework for long contexts. Our encoder uses bidirectional attention over landmark tokens. This design captures global dependencies and avoids semantic fragmentation from segment-based methods. Experiments on text reconstruction and four QA benchmarks validate our approach. At 4x and 16x compression ratios, our method outperforms prior soft compression baselines.
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
Video misinformation detection is often approached as a binary veracity classification problem, overlooking the complex reasoning required to explain how and why content misleads. Existing benchmarks fail to capture the diversity of manipulation strategies, such as AI-generated edits and out-of-context manipulation, and do not evaluate whether models can provide process-level justifications for their judgments. We address these limitations with MisVideoQA, a multi-turn benchmark designed to assess comprehensive understanding and reasoning in video misinformation analysis. MisVideoQA covers 12 fine-grained deception categories and evaluates models along six dimensions, progressing from perceptual attribution to intent and persuasion analysis. Recognizing that standard MLLMs struggle to sustain such structured, evidence-based deduction, we propose MisAgent, a Delphi-inspired multi-agent framework in which specialized agents collaboratively integrate multimodal cues with external evidence. Experimental results show that state-of-the-art multimodal large language models perform poorly on MisVideoQA, while MisAgent consistently improves reasoning accuracy and explanation quality. Together, our benchmark and framework establish a unified foundation for reliable, interpretable, and evidence-grounded video misinformation analysis.
R2IF: Aligning Reasoning with Decisions via Composite Rewards for Interpretable LLM Function Calling
PDF ↗Function calling empowers large language models (LLMs) to interface with external tools, yet existing RL-based approaches suffer from misalignment between reasoning processes and tool-call decisions. We propose R2IF, a reasoning-aware RL framework for interpretable function calling, adopting a composite reward integrating format/correctness constraints, Chain-of-Thought Effectiveness Reward (CER), and Specification-Modification-Value (SMV) reward, optimized via GRPO. Experiments on BFCL/ACEBench show R2IF outperforms baselines by up to 34.62% (Llama3.2-3B on BFCL) with positive Average CoT Effectiveness (0.05 for Llama3.2-3B), enhancing both function-calling accuracy and interpretability for reliable tool-augmented LLM deployment.
Memory-Augmented Generation (MAG) extends large language models with external memory to support long-context reasoning, but existing approaches largely rely on semantic similarity over monolithic memory stores, entangling temporal, causal, and entity information. This design limits interpretability and alignment between query intent and retrieved evidence, leading to suboptimal reasoning accuracy. In this paper, we propose MAGMA, a multi-graph agentic memory architecture that represents each memory item across orthogonal semantic, temporal, causal, and entity graphs. MAGMA formulates retrieval as policy-guided traversal over these relational views, enabling query-adaptive selection and structured context construction. By decoupling memory representation from retrieval logic, MAGMA provides transparent reasoning paths and fine-grained control over retrieval. Experiments on LoCoMo and LongMemEval demonstrate that MAGMA consistently outperforms state-of-the-art agentic memory systems in long-horizon reasoning task.
ARCHITECT: Uncertainty-Aware Dynamic Tool Learning via Causal Intervention for Open-World Agents
PDF ↗Dynamic tool generation empowers Large Language Model (LLM) agents to synthesize tools on demand, yet a critical challenge remains: 32.4% of generated tools fail on first invocation. We present Causal Tool Diagnosis (CTD), a principled framework that moves beyond black-box reliability prediction to interpretable failure attribution. CTD constructs a Structural Causal Model (SCM) capturing how specification quality, code characteristics, and execution environment jointly determine tool outcomes. Uniquely leveraging code’s intervenability, we conduct controlled sandbox experiments to estimate causal effects—an advantage unavailable in pure text generation. CTD jointly predicts confidence (Spearman rank correlation coefficient \rho=0.90) and root cause attribution (78% accuracy), with attributions directly guiding targeted repairs (+9.6% success rate over error-type classification). Our ARCHITECT framework, integrating CTD throughout the tool lifecycle, achieves state-of-the-art on four benchmarks including StableToolBench (+3.8%), MINT (+4.6%), T-Eval (+3.7%), and SWE-bench Lite (+2.4%), with consistent improvements across all settings.
Post-cutoff performance decay of LLMs has been widely interpreted as a temporal signal for benchmark contamination, where public information released before the training cutoff may have been included into training corpora and inflated model performance by memorization. We critically examine this view and demonstrate that this temporal signal is highly sensitive to how benchmark questions are constructed, even if the underlying source material remains invariant. Specifically, we show that LLM-transformed questions can produce remarkably different temporal patterns compared to fill-in-the-blank (cloze) questions directly retrieved from the very same documents. We validate this effect on prior benchmarks that report clear post-cutoff decay (LiveCodeBench), and show that a simple LLM-driven transformation of the same problems can effectively remove the temporal pattern. We further provide a mechanistic understanding of this phenomenon using influence function analysis. Overall, our results suggest that post-cutoff performance decay is a sensitive contamination signal, motivating more robust contamination probes for reliable LLM evaluation.
Warning: This paper may contain content that could be disturbing or offensive. Content moderation in online platforms faces persistent challenges due to the evolving complexity of user-generated content and the limitations of traditional rule-based and machine learning approaches. While recent advances in large language models (LLMs) have enabled more sophisticated moderation via direct prompting or fine-tuning, these approaches often exhibit limited generalization, interpretability, and adaptability to unseen or ambiguous cases.In this work, we propose a novel moderation framework that leverages analogical examples to enhance rule induction and decision reliability. Our approach integrates end-to-end optimization of analogical retrieval, rule generation, and moderation classification, enabling the dynamic adaptation of moderation rules to diverse content scenarios. Through comprehensive experiments, we demonstrate that our method significantly outperforms both rule-injected fine-tuning baselines and multi-stage static RAG pipelines in terms of moderation accuracy and rule quality. Further evaluations—including human assessments and external model generalization tests confirm the superiority of rules generated by our framework in terms of clarity, interpretability, and applicability. These findings highlight the potential of analogical example-driven methods for advancing robust, explainable, and generalizable content moderation in real-world applications.
Multi-View Attention Multiple-Instance Learning Enhanced by LLM Reasoning for Cognitive Distortion Detection
PDF ↗Cognitive distortions have been closely linked to mental health disorders, yet their automatic detection remains challenging due to contextual ambiguity, co-occurrence, and semantic overlap. We propose a novel framework that combines Large Language Models (LLMs) with a Multiple-Instance Learning (MIL) architecture to enhance interpretability and expression-level reasoning. Each utterance is decomposed into Emotion, Logic, and Behavior (ELB) components, which are processed by LLMs to infer multiple distortion instances, each with a predicted type, expression, and model-assigned salience score. These instances are integrated via a Multi-View Gated Attention mechanism for final classification. Experiments on Korean (KoACD) and English (Therapist QA) datasets demonstrate that incorporating ELB and LLM-inferred salience scores improves classification performance, especially for distortions with high interpretive ambiguity. Our results suggest a psychologically grounded and generalizable approach for fine-grained reasoning in mental health NLP. The dataset and implementation details are publicly accessible.
Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillation
PDF ↗Recent advances in reasoning-oriented Large Language Models (LLMs) have been driven by the introduction of Chain-of-Thought (CoT) traces, where models generate intermediate reasoning traces before producing an answer. These traces, as in DeepSeek R1, are not only used to guide model inference but also serve as supervision signals for Knowledge Distillation (KD) to improve smaller models. A prevailing but under-examined implicit assumption is that these CoT traces when emitted at inference time are both semantically correct and interpretable for the end-users. While there are reasons to believe that these intermediate tokens help improve solution accuracy, in this work, we question their validity (semantic correctness) and interpretability to the end user. To isolate the effect of trace semantics, we design experiments in the Question Answering (QA) domain using a rule-based problem decomposition method. This enables us to create Supervised Fine-Tuning (SFT) datasets for LLMs where - each QA problem is paired with either verifiably correct or incorrect CoT traces, while always providing the correct final solution. Trace correctness at inference time is then evaluated by checking the accuracy of every sub-step in decomposed reasoning chains. To assess end-user interpretability, we finetune LLMs with three additional types of CoT traces: R1 traces, R1 trace summaries, and post-hoc explanations of R1 traces. We further conduct a human-subject study with 100 participants asking them to rate the interpretability of each trace type on a standardized Likert scale. Our experiments reveal two key findings - (1) CoT trace correctness is not reliably correlated with the model’s generation of correct final answers: correct traces led to correct solutions only for 28% test-set problems while incorrect traces don’t necessarily degrade solution accuracy. (2) In end-user interpretability studies, fine-tuning on verbose R1 traces produced the best model performance but these traces were rated as least interpretable by users, scoring on average 3.39 for interpretability and 4.59 for cognitive load metrics on a 5-point Likert scale. In contrast, the decomposed traces that are judged significantly more interpretable don’t lead to comparable solution accuracy. Together, these findings challenge the assumption in question suggesting that researchers and practitioners should decouple model supervision objectives from end-user-facing trace design.
ReflectRM: Boosting Generative Reward Models via Self-Reflection within a Unified Judgment Framework
PDF ↗Reward Models (RMs) are critical components in the Reinforcement Learning from Human Feedback (RLHF) pipeline, directly determining the alignment quality of Large Language Models (LLMs). Recently, Generative Reward Models (GRMs) have emerged as a superior paradigm, offering higher interpretability and stronger generalization than traditional scalar RMs. However, existing methods for GRMs focus primarily on outcome-level supervision, neglecting analytical process quality, which constrains their potential. To address this, we propose ReflectRM, a novel GRM that leverages self-reflection to assess analytical quality and enhance preference modeling. ReflectRM is trained under a unified generative framework for joint modeling of response preference and analysis preference. During inference, we use its self-reflection capability to identify the most reliable analysis, from which the final preference prediction is derived. Experiments across four benchmarks show that ReflectRM consistently improves performance, achieving an average accuracy gain of +3.7 on Qwen3-4B. Further experiments confirm that response preference and analysis preference are mutually reinforcing. Notably, ReflectRM substantially mitigates positional bias, yielding +10.2 improvement compared with leading GRMs and establishing itself as a more stable evaluator. Our code is available at https://github.com/yuliangCarmelo/ReflectRM.
Automated interlinear gloss prediction with neural networks is a promising approach to accelerate language documentation efforts. However, while state-of-the-art models like GlossLM (Ginn et al., 2024) achieve high scores on glossing benchmarks, user studies with linguists have found critical barriers to the usefulness of such models in real-world scenarios (Rice et al., 2025). In particular, existing models typically generate morpheme-level glosses but assign them to whole words without predicting the actual morpheme boundaries, making the predictions less interpretable and thus untrustworthy to human annotators.We conduct the first study on neural models that jointly predict interlinear glosses and the corresponding morphological segmentation from raw text. We run experiments to determine the optimal way to train models that balance segmentation and glossing accuracy, as well as the alignment between the two tasks. We extend the training corpus of GlossLM and pretrain PolyGloss, a family of seq2seq multilingual models for joint segmentation and glossing that outperforms GlossLM on glossing and beats various open-source LLMs on segmentation, glossing, and alignment. In addition, we demonstrate that PolyGloss can be quickly adapted to a new dataset via low-rank adaptation.
Large Language Models enhanced with Retrieval Augmented Generation show strong potential in knowledge intensive tasks. However, they often encounter knowledge conflicts, where retrieved information contradicts the model’s internal knowledge or exhibits internal inconsistencies. Existing methods treat this as a simplistic binary choice, forcing models to blindly trust external contexts or rigidly rely on memory, resulting in unreliable predictions that swing between sycophancy and stubbornness. We argue that a more principled approach is to embrace contradictions as opportunities for deeper reasoning. To this end, we introduce Debate-of-Thoughts (DoT), a framework that transforms conflict resolution into an active deliberation process. DoT guides a single model through three phases: 1) hypothesis generation, which forms competing perspectives; 2) internal debate, where the model acts as both a proponent and a critic to stress test each view; and 3) adjudication, where a judge module evaluates arguments based on evidence and logical consistency. We implement DoT via two complementary strategies: inference time prompt chaining and supervised fine tuning. Experiments across multiple conflict benchmarks show that DoT consistently outperforms state-of-the-art methods, while generating transparent debate transcripts that explain its decisions. By improving both accuracy and interpretability under knowledge conflicts, DoT establishes a more reliable paradigm for retrieval augmented generation systems.
Large language model (LLM)-based agents are increasingly deployed in e-commerce shopping. To perform thorough, user-tailored product searches, agents should interpret personal preferences, engage in multi-turn dialogues, and ultimately retrieve and discriminate among highly similar products. However, existing research has yet to provide a unified simulation environment that consistently captures all of these aspects, and always focuses solely on evaluation benchmarks without training support. In this paper, we introduce ShopSimulator, a large-scale and challenging Chinese shopping environment. Leveraging ShopSimulator, we evaluate LLMs across diverse scenarios, finding that even the best-performing models achieve less than 40% full-success rate. Error analysis reveals that agents struggle with deep search and product selection in long trajectories, fail to balance the use of personalization cues, and to effectively engage with users. Further training exploration provides practical guidance for overcoming these weaknesses, with the combination of supervised fine-tuning (SFT) and reinforcement learning (RL) yielding significant performance improvements.
Conversational recommender systems (CRSs) operate under incremental preference revelation, requiring recommendation decisions under uncertainty. While recent LLM-based approaches achieve strong performance on proxy metrics such as Recall@K and BLEU, they often fail to deliver high-quality, user-aligned recommendations in practice, as they optimize intermediate objectives like retrieval accuracy or fluent generation rather than recommendation quality itself. We propose HARPO (Hierarchical Agentic Reasoning with Preference Optimization), an agentic framework that reframes conversational recommendation as a structured decision-making process optimized for multi-dimensional recommendation quality. HARPO integrates (i) hierarchical preference learning that decomposes recommendation quality into interpretable dimensions (relevance, diversity, satisfaction, and engagement) with context-dependent weighting; (ii) deliberative tree-search reasoning guided by a learned value network evaluating candidate paths on predicted quality; and (iii) domain-agnostic reasoning abstractions through Virtual Tool Operations and multi-agent refinement. We evaluate HARPO on ReDial, INSPIRED, and MUSE, demonstrating consistent improvements over strong baselines on recommendation-centric metrics while maintaining competitive response quality.
Specializing Large Models for Oracle Bone Script Interpretation via Component-Grounded Multimodal Knowledge Augmentation
PDF ↗Deciphering ancient Chinese Oracle Bone Script (OBS) is a challenging task that offers insights into the beliefs, systems, and culture of the ancient era. Existing approaches treat decipherment as a closed-set image recognition problem, which fails to bridge the “interpretation gap”: while individual characters are often unique and rare, they are composed of a limited set of recurring, pictographic components that carry transferable semantic meanings. To leverage this structural logic, we propose an agent-driven Vision-Language Model (VLM) framework that integrates a VLM for precise visual grounding with an LLM-based agent to automate a reasoning chain of component identification, graph-based knowledge retrieval, and relationship inference for linguistically accurate interpretation. To support this, we also introduce OB-Radix, an expert-annotated dataset providing structural and semantic data absent from prior corpora, comprising 1,022 character images (934 unique characters) and 1,853 fine-grained component images across 478 distinct components with verified explanations. By evaluating our system across three benchmarks of different tasks, we demonstrate that our framework yields more detailed and precise decipherments compared to baseline methods.
Towards Intrinsic Interpretability of Large Language Models: A Survey of Design Principles and Architectures
PDF ↗While Large Language Models (LLMs) have achieved strong performance across many NLP tasks, their opaque internal mechanisms hinder trustworthiness and safe deployment. Existing surveys in explainable AI largely focus on post-hoc explanation methods that interpret trained models through external approximations. In contrast, intrinsic interpretability, which builds transparency directly into model architectures and computations, has recently emerged as a promising alternative. This paper presents a systematic review of the recent advances in intrinsic interpretability for LLMs, categorizing existing approaches into five design paradigms: functional transparency, concept alignment, representational decomposability, explicit modularization, and latent sparsity induction. We further discuss open challenges and outline future research directions in this emerging field. The paper list is available at: Survey-Intrinsic-Interpretability-of-LLMs
Large language models (LLMs) often struggle to use tools reliably in domain-specific settings, where APIs may be idiosyncratic, under-documented, or tailored to private workflows. This highlights the need for effective adaptation to task-specific tools. We propose RIMRULE, a neuro-symbolic approach for LLM adaptation based on dynamic rule injection. Compact, interpretable rules are distilled from failure traces and injected into the prompt during inference to improve task performance. These rules are proposed by the LLM itself and consolidated using a Minimum Description Length (MDL) objective that favors generality and conciseness. Each rule is stored in both natural language and a structured symbolic form, supporting efficient retrieval at inference time. Experiments on tool-use benchmarks show that this approach improves accuracy on both seen and unseen tools without modifying LLM weights. It outperforms prompting-based adaptation methods and complements finetuning, offering an interpretable layer of inference-time generalization. Moreover, rules learned from one LLM can be reused to improve others, including long reasoning LLMs, highlighting the portability of symbolic knowledge across architectures.
Human conversation relies heavily on *conversational implicature*, in which speakers convey meanings that are suggested rather than explicitly stated. Although recent large language models (LLMs) exhibit strong conversational fluency, they remain unreliable when interpretation depends on reasoning that integrates social and contextual cues, a process rarely articulated in text. We introduce **DRinQ**, a benchmark for evaluating pragmatic reasoning about conversational implicature in question utterances, designed to isolate pragmatic variation while holding each question’s surface form fixed. To support scalable evaluation, we propose a semi-automated pipeline that produces question-context-interpretation instances with systematic variation. Across evaluations, we find a consistent generation-inference asymmetry: while state-of-the-art models can generate plausible pragmatic scenarios when guided, they often fail to recover the intended implication at inference time. For smaller models, structured prompting improves alignment with human judgments. A comparative writing study further reveals complementary strengths: human authors tend to produce safer, predictable contexts, whereas models generate varied scenarios with interpretations that sometimes exceed contextual support. These findings highlight persistent challenges in modeling conversational implicature and motivate more context-sensitive evaluation frameworks.
Confident, Calibrated, or Complicit: Safety Alignment and Ideological Bias in LLM Hate Speech Detection
PDF ↗We investigate the efficacy of Large Language Models (LLMs) in detecting implicit and explicit hate speech, examining how models with minimal safety alignment (uncensored) compare with more heavily aligned (censored) counterparts in a deployed-model setting when deployed using political personas. While uncensored models are often framed as offering a less constrained perspective, our results reveal a trade-off: censored models outperform their uncensored counterparts in both accuracy and robustness, achieving 69.0% versus 64.1% strict accuracy. However, this higher performance is also associated with greater resistance to persona-based influence, while uncensored models are more malleable to ideological framing. Furthermore, we identify critical failures across all models in understanding nuanced language such as irony. We also find alarming fairness disparities in performance across different targeted groups and systemic overconfidence that renders self-reported certainty unreliable. These findings challenge the notion of LLMs as objective arbiters and highlight the need for more sophisticated auditing frameworks that account for fairness, calibration, and ideological consistency. Taken together, these results point to censorship-as-deployed rather than safety alignment in isolation as the more appropriate frame for interpreting model differences.
Information-seeking is a core capability for AI agents, requiring them to gather and reason over tool-generated information across long trajectories. However, such multi-step information-seeking tasks remain challenging for agents backed by language models. While process reward models (PRMs) can guide agents by ranking candidate steps at test-time, existing PRMs, designed for short reasoning with binary judgment, cannot capture richer dimensions of information-seeking steps, such as tool interactions and reasoning over tool outputs, nor handle the rapidly growing context in long-horizon tasks. To address these limitations, we introduce PRInTS, a generative PRM trained with dual capabilities: (1) dense scoring based on the PRM’s reasoning across multiple step quality dimensions (e.g., interpretation of tool outputs, tool call informativeness) and (2) trajectory summarization that compresses the growing context while preserving essential information for step evaluation. Extensive evaluations across FRAMES, GAIA (levels 1-3), and WebWalkerQA (easy-hard) benchmarks on multiple models, with ablations, reveal that best-of-n sampling with PRInTS enhances information-seeking in open-source models as well as specialized agents, matching or surpassing frontier models with a much smaller backbone agent and outperforming other strong reward modeling baselines.