Implicit hate speech detection is challenging due to its subtlety and reliance on contextual interpretation rather than explicit offensive words. Current approaches rely on contrastive learning, which are shown to be effective on distinguishing hate and non-hate sentences. Humans, however, detect implicit hate speech by first identifying specific targets within the text and subsequently interpreting how these target relate to their surrounding context. Motivated by this reasoning process, we propose AmpleHate, a novel approach designed to mirror human inference for implicit hate detection. AmpleHate identifies explicit target using a pretrained Named Entity Recognition model and capture implicit target information via [CLS] tokens. It computes attention-based relationships between explicit, implicit targets and sentence context and then, directly injects these relational vectors into the final sentence representation. This amplifies the critical signals of target-context relations for determining implicit hate. Experiments demonstrate that AmpleHate achieves state-of-the-art performance, outperforming contrastive learning baselines by an average of 82.14% and achieve faster convergence. Qualitative analyses further reveal that attention patterns produced by AmpleHate closely align with human judgement, underscoring its interpretability and robustness.
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
Understanding the internal origins of capabilities in large language models (LLMs) is crucial for interpretability and efficient adaptation. However, the emergence of specific capabilities remains poorly understood, as most existing approaches rely on external signals (e.g., performance shifts or gradient similarities) with limited structural grounding. To address these issues, this paper proposes a lightweight and highly interpretable approach that links LLM capabilities to internal components by identifying correspondences at the level of attention heads. Specifically, we first define five fundamental capabilities, namely Mathematical Reasoning, Reading Comprehension, Commonsense Reasoning, Scientific Reasoning, and Professional Expertise, and employ probing techniques to detect the attention heads most predictive of each, thereby establishing capability–head mappings. For targeted instruction tuning, complex tasks are decomposed into these fundamental capabilities, and training data are selected accordingly. Experiments on LLaMA3.1-8B and Qwen2.5-7B show over 70% discrimination accuracy in identifying capabilities. On MMLU and BBH, our method improves accuracy by 1 to 1.5 points over the gradient-based method LESS and by 5 to 6 points over other intermediate-state baselines.
Do All Autoregressive Transformers Remember Facts the Same Way? A Cross-Architecture Analysis of Recall Mechanisms
PDF ↗Understanding how Transformer-based language models store and retrieve factual associations is critical for improving interpretability and enabling targeted model editing. Prior work, primarily on GPT-style models, has identified MLP modules in early layers as key contributors to factual recall. However, it remains unclear whether these findings generalize across different autoregressive architectures. To address this, we conduct a comprehensive evaluation of factual recall across several models—including GPT, LLaMA, Qwen, and DeepSeek—analyzing where and how factual information is encoded and accessed. Consequently, we find that Qwen-based models behave differently from previous patterns: attention modules in the earliest layers contribute more to factual recall than MLP modules. Our findings suggest that even within the autoregressive Transformer family, architectural variations can lead to fundamentally different mechanisms of factual recall.
Seeing the Same Story Differently: Framing‐Divergent Event Coreference for Computational Framing Analysis
PDF ↗News articles often describe the same real-world event in strikingly different ways, shaping perception through framing rather than factual disagreement. However, traditional computational framing approaches often rely on coarse-grained topic classification, limiting their ability to capture subtle, event-level differences in how the same occurrences are presented across sources. We introduce Framing-divergent Event Coreference (FrECo), a novel task that identifies pairs of event mentions referring to the same underlying occurrence but differing in framing across documents to provide a event-centric lens for computational framing analysis. To support this task, we construct the high-agreement and diverse FrECo corpus. We evaluate the FrECo task on the corpus through supervised and preference-based tuning of large language models, providing strong baseline performance. To scale beyond the annotated data, we develop a bootstrapped mining pipeline that iteratively expands the training set with high-confidence FrECo pairs. Our approach enables scalable, interpretable analysis of how media frame the same events differently, offering a new lens for contrastive framing analysis at the event level.
LinguaLens: Towards Interpreting Linguistic Mechanisms of Large Language Models via Sparse Auto-Encoder
PDF ↗Large language models (LLMs) demonstrate exceptional performance on tasks requiring complex linguistic abilities, such as reference disambiguation and metaphor recognition/generation. Although LLMs possess impressive capabilities, their internal mechanisms for processing and representing linguistic knowledge remain largely opaque. Prior research on linguistic mechanisms is limited by coarse granularity, limited analysis scale, and narrow focus. In this study, we propose LinguaLens, a systematic and comprehensive framework for analyzing the linguistic mechanisms of large language models, based on Sparse Auto-Encoders (SAEs). We extract a broad set of Chinese and English linguistic features across four dimensions—morphology, syntax, semantics, and pragmatics. By employing counterfactual methods, we construct a large-scale counterfactual dataset of linguistic features for mechanism analysis. Our findings reveal intrinsic representations of linguistic knowledge in LLMs, uncover patterns of cross-layer and cross-lingual distribution, and demonstrate the potential to control model outputs. This work provides a systematic suite of resources and methods for studying linguistic mechanisms, offers strong evidence that LLMs possess genuine linguistic knowledge, and lays the foundation for more interpretable and controllable language modeling in future research.
IntentionFrame: A Semi-Structured, Multi-Aspect Framework for Fine-Grained Conversational Intention Understanding
PDF ↗Understanding user intentions in multi-turn dialogues is critical for conversational AI, yet existing approaches—relying on rigid slot-value structures or unstructured free-text—fail to fully capture conversational complexity. In this paper, we propose IntentionFrame, a semi-structured framework inspired by psychological and cognitive intention theories, which organizes conversational intents into four interrelated aspects: situation, emotion, action, and knowledge. This design not only retains interpretability but also provides LLMs with a rich context to accurately parse and respond to nuanced user inputs. To efficiently scale IntentionFrame annotations, we introduce a Weakly-supervised Reinforced Generation (WeRG) method that leverages a small set of high-quality human annotations in conjunction with abundant coarsely labeled data. By applying reinforcement learning to balance these diverse signals, WeRG aims to effectively generate reliable IntentionFrame annotations, which serve as essential grounding for downstream tasks—leading to substantial improvements in response generation and task completion. Our experiments, supported by both automatic metrics and human evaluations, show that integrating IntentionFrame with WeRG significantly improves LLMs’ conversational understanding and sets a new benchmark for intent analysis.
Puns are a form of humorous wordplay that exploits polysemy and phonetic similarity. While LLMs have shown promise in detecting puns, we show in this paper that their understanding often remains shallow, lacking the nuanced grasp typical of human interpretation. By systematically analyzing and reformulating existing pun benchmarks, we demonstrate how subtle changes in puns are sufficient to mislead LLMs. Our contributions include comprehensive and nuanced pun detection benchmarks, human evaluation across recent LLMs, and an analysis of the robustness challenges these models face in processing puns.
Frame First, Then Extract: A Frame-Semantic Reasoning Pipeline for Zero-Shot Relation Triplet Extraction
PDF ↗Large Language Models (LLMs) have shown impressive capabilities in language understanding and generation, leading to growing interest in zero-shot relation triplet extraction (ZeroRTE), a task that aims to extract triplets for unseen relations without annotated data. However, existing methods typically depend on costly fine-tuning and lack the structured semantic guidance required for accurate and interpretable extraction. To overcome these limitations, we propose FrameRTE, a novel ZeroRTE framework that adopts a “frame first, then extract” paradigm. Rather than extracting triplets directly, FrameRTE first constructs high-quality Relation Semantic Frames (RSFs) through a unified pipeline that integrates frame retrieval, synthesis, and enhancement. These RSFs serve as structured and interpretable knowledge scaffolds that guide frozen LLMs in the extraction process. Building upon these RSFs, we further introduce a human-inspired three-stage reasoning pipeline consisting of semantic frame evocation, frame-guided triplet extraction, and core frame elements validation to achieve semantically constrained extraction. Experiments demonstrate that FrameRTE achieves competitive zero-shot performance on multiple benchmarks. Moreover, the RSFs we construct serve as high-quality semantic resources that can enhance other extraction methods, showcasing the synergy between linguistic knowledge and foundation models.
Large language models (LLMs) store vast amounts of information, making them powerful yet raising privacy and safety concerns when selective knowledge removal is required. Existing unlearning strategies, ranging from gradient-based fine-tuning and model editing to sparse autoencoder (SAE) steering, either lack interpretability or fail to provide a robust defense against adversarial prompts. We propose **S**AE–Guided **S**ubspace **P**rojection **U**nlearning (**SSPU**), a novel framework that leverages SAE features to drive targeted updates in the model’s parameter space, enabling precise, interpretable, and robust unlearning. SSPU’s three-stage pipeline performs data-driven layer and feature selection, subspace construction via QR decomposition, and constrained optimization that controls activations into an “irrelevant” subspace while preserving retained knowledge. Overall, we use SAE features to construct a subspace that supervises unlearning, refining the loss and adding a regularization term to guide interpretable parameter updates. In experiments on the WMDP–Cyber forget set and three utility benchmarks (MMLU, TruthfulQA, GSM8K), SSPU reduces harmful knowledge accuracy by 3.22% compared to the strongest baseline. It also improves adversarial robustness, lowering malicious accuracy under jailbreak prompts compared to baselines. Our findings expose the limitations of prior unlearning methods and demonstrate how interpretable subspace-guided optimization can achieve robust, controllable model behavior.
Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval
PDF ↗Despite their strong performance, Dense Passage Retrieval (DPR) models suffer from a lackof interpretability. In this work, we propose a novel interpretability framework that leveragesSparse Autoencoders (SAEs) to decompose previously uninterpretable dense embeddings fromDPR models into distinct, interpretable latent concepts. We generate natural language descriptionsfor each latent concept, enabling human interpretations of both the dense embeddingsand the query-document similarity scores of DPR models. We further introduce Concept-Level Sparse Retrieval (CL-SR), a retrieval framework that directly utilizes the extractedlatent concepts as indexing units. CL-SR effectively combines the semantic expressiveness ofdense embeddings with the transparency and efficiency of sparse representations. We showthat CL-SR achieves high index-space and computational efficiency while maintaining robustperformance across vocabulary and semantic mismatches.
Real-world data combines structured (e.g., graph connections) and unstructured (e.g., text, visuals) formats, capturing explicit relationships (e.g., social links) and implicit semantic interdependencies (e.g., knowledge graphs). We propose GraphAgent, an automated agent pipeline addressing both explicit and implicit graph-enhanced semantic dependencies for predictive (e.g., node classification) and generative (e.g., text generation) tasks. GraphAgent integrates three components: (i) a Graph Generator Agent creating knowledge graphs for semantic dependencies; (ii) a Task Planning Agent interpreting user queries and formulating tasks via self-planning; and (iii) a Task Execution Agent automating task execution with tool matching. These agents combine language and graph language models to reveal complex relational and semantic patterns. Extensive experiments on diverse datasets validate GraphAgent’s effectiveness in graph-related predictive and text generative tasks. GraphAgent is open-sourced at: https://anonymous.4open.science/r/GraphAgent-Submit-6F52/.
A multitude of interconnected risk events—ranging from regulatory changes to geopolitical tensions—can trigger ripple effects across firms. Identifying inter-firm risk relations is thus crucial for applications like portfolio management and investment strategy. Traditionally, such assessments rely on expert judgment and manual analysis, which are, however, subjective, labor-intensive, and difficult to scale. To address this, we propose a systematic method for extracting inter-firm risk relations using Form 10-K filings—authoritative, standardized financial documents—as our data source. Leveraging recent advances in natural language processing, our approach captures implicit and abstract risk connections through unsupervised fine-tuning based on chronological and lexical patterns in the filings. This enables the development of a domain-specific financial encoder with a deeper contextual understanding and introduces a quantitative risk relation score for transparency, interpretable analysis. Extensive experiments demonstrate that our method outperforms strong baselines across multiple evaluation settings.
With the growing spread of misinformation online, understanding how true news evolves into fake news has become crucial for early detection and prevention. However, previous research has often assumed fake news inherently exists rather than exploring its gradual formation. To address this gap, we propose \textbf{FUSE} (\textbf{F}ake news evol\textbf{U}tion \textbf{S}imulation fram\textbf{E}work), a novel Large Language Model (LLM)-based simulation approach explicitly focusing on fake news evolution from real news. Our framework model a social network with four distinct types of LLM agents commonly observed in daily interactions: \textit{ spreaders} who propagate information, \textit{commentators} who provide interpretations, \textit{verifiers} who fact-check, and \textit{standers} who observe passively to simulate realistic daily interactions that progressively distort true news. To quantify these gradual distortions, we develop \textbf{FUSE-EVAL}, a comprehensive evaluation framework measuring truth deviation along multiple linguistic and semantic dimensions. Results show that FUSE effectively captures fake news evolution patterns and accurately reproduces known fake news, aligning closely with human evaluations. Experiments demonstrate that FUSE accurately reproduces known fake news evolution scenarios, aligns closely with human judgment, and highlights the importance of timely intervention at early stages. Our framework is extensible, enabling future research on broader scenarios of fake news:https://github.com/LiuYuHan31/FUSE
Large language models (LLMs) have achieved remarkable success across various domains, driving significant technological advancements and innovations. Despite the rapid growth in model scale and capability, systematic, data-driven research on how structural configurations affect performance remains scarce. To address this gap, we present a large-scale dataset encompassing diverse open-source LLM structures and their performance across multiple benchmarks. Leveraging this dataset, we conduct a systematic, data mining-driven analysis to validate and quantify the relationship between structural configurations and performance. Our study begins with a review of the historical development of LLMs and an exploration of potential future trends. We then analyze how various structural choices impact performance across benchmarks and further corroborate our findings using mechanistic interpretability techniques. By providing data-driven insights into LLM optimization, our work aims to guide the targeted development and application of future models.
Definition Generation for Word Meaning Modeling: Monolingual, Multilingual, and Cross-Lingual Perspectives
PDF ↗The task of Definition Generation has recently gained attention as an interpretable approach to modeling word meaning. Thus far, most research has been conducted in English, with limited work and resources for other languages. In this work, we expand Definition Generation beyond English to a suite of 22 languages and evaluate Llama-based models within a monolingual, multilingual, and cross-lingual setting. Our experiments show that monolingual fine-tuning consistently outperforms pretrained baselines, with the largest gains observed in languages with lower initial performance; and that multilingual fine-tuning does not consistently improve performance on the individual fine-tuning languages. Our cross-lingual evaluation reveals that models fine-tuned on a single language typically lose the ability to generate definitions in other languages, whereas multilingual models exhibit robust generalization even to languages unseen during fine-tuning.
Recent works have suggested that In-Context Learning (ICL) operates in dual modes, i.e. task retrieval (remember learned patterns from pre-training) and task learning (inference-time ”learning” from demonstrations). However, disentangling these the two modes remains a challenging goal. We introduce ICL CIPHERS, a class of task reformulations based on substitution ciphers borrowed from classic cryptography. In this approach, a subset of tokens in the in-context inputs are substituted with other (irrelevant) tokens, rendering English sentences less comprehensible to human eye. However, by design, there is a latent, fixed pattern to this substitution, making it reversible. This bijective (reversible) cipher ensures that the task remains a well-defined task in some abstract sense, despite the transformations. It is a curious question if LLMs can solve tasks reformulated by ICL CIPHERS with a BIJECTIVE mapping, which requires ”deciphering” the latent cipher. We show that LLMs are better at solving tasks reformulated by ICL CIPHERS with BIJECTIVE mappings than the NON-BIJECTIVE (irreversible) baseline, providing a novel approach to quantify ”learning” in ICL. While this gap is small, it is consistent across the board on four datasets and six models. Finally, our interpretability analysis shows evidence that LLMs can internally decode ciphered inputs.
Interpreting Noun-Noun Compounds remains a persistent challenge for Large Language Models (LLMs) because the semantic relation between the modifier and the head is rarely stated explicitly. Recent benchmarks frame Noun-Noun Compound Interpretation as a multiple-choice question. While this setting allows LLMs to produce more controlled results, it still faces two key limitations: vague relation descriptions as options and the inability to handle polysemous compounds. We introduce a dual-faceted textual enrichment framework that augments prompts. Description enrichment paraphrases relations into event‐oriented descriptions instantiated with the target compound to explicitly surface the hidden event connecting head and modifier. Conditioned context enrichment identifies polysemous compounds leveraging qualia-role binding and assigns each compound with condition cues for disambiguation. Our method yields consistently higher accuracy across three LLM families. These gains suggest that surfacing latent compositional structure and contextual constraint is a promising path toward deeper semantic understanding in language models.
Linguistic feature datasets such as URIEL+ are valuable for modelling cross-lingual relationships, but their high dimensionality and sparsity, especially for low-resource languages, limit the effectiveness of distance metrics. We propose a pipeline to optimize the URIEL+ typological feature space by combining feature selection and imputation, producing compact yet interpretable typological representations. We evaluate these feature subsets on linguistic distance alignment and downstream tasks, demonstrating that reduced-size representations of language typology can yield more informative distance metrics and improve performance in multilingual NLP applications.
Large language models (LLMs) are increasingly used as evaluators in natural language generation tasks, offering advantages in scalability and interpretability over traditional evaluation methods. However, existing LLM-based evaluations often suffer from biases and misalignment, particularly in domain-specific tasks, due to limited functional understanding and knowledge gaps. To address these challenges, we first investigate the relationship between an LLM-based evaluator’s familiarity with the target task and its evaluation performance. We then introduce the Co-Eval framework, which leverages a criteria planner model and optimized machine metrics to enhance the scalability and fairness of LLM-based evaluation. Experimental results on both general and domain-specific tasks demonstrate that Co-Eval reduces biases, achieving up to a 0.4903 reduction in self-preference bias, and improves alignment with human preferences, with gains of up to 0.324 in Spearman correlation.
Subtle Risks, Critical Failures: A Framework for Diagnosing Physical Safety of LLMs for Embodied Decision Making
PDF ↗Large Language Models (LLMs) are increasingly used for decision making in embodied agents, yet existing safety evaluations often rely on coarse success rates and domain-specific setups, making it difficult to diagnose why and where these models fail. This obscures our understanding of embodied safety and limits the selective deployment of LLMs in high-risk physical environments. We introduce SAFEL, the framework for systematically evaluating the physical safety of LLMs in embodied decision making. SAFEL assesses two key competencies: (1) rejecting unsafe commands via the Command Refusal Test, and (2) generating safe and executable plans via the Plan Safety Test. Critically, the latter is decomposed into functional modules, goal interpretation, transition modeling, action sequencing enabling fine-grained diagnosis of safety failures. To support this framework, we introduce EMBODYGUARD, a PDDL-grounded benchmark containing 942 LLM-generated scenarios covering both overtly malicious and contextually hazardous instructions. Evaluation across 13 state-of-the-art LLMs reveals that while models often reject clearly unsafe commands, they struggle to anticipate and mitigate subtle, situational risks. Our results highlight critical limitations in current LLMs and provide a foundation for more targeted, modular improvements in safe embodied reasoning.