Live-stream E-commerce faces significant challenges from morphs, deliberate linguistic variants used to evade real-time voice filters and amplify product claims illegally. While critical for regulatory enforcement, Live Auditory Morph Resolution (LiveAMR) research is hindered by limited datasets: prior work relied on narrow, redundant health domain corpora, restricting model robustness. To bridge this gap, we introduce two datasets: (1) HealthAMR, a refined health-domain corpus via deduplication and re-annotation. (2) GeneralAMR, a general domain benchmark with 28K annotated sentences from 77 channels across 7 E-commerce categories. Further, we propose JointMRE, a multi-task framework that jointly resolves morphs and generates structured explanations, transferring grammatical insights from large language models to enhance generalization. Predictions are refined by our Conflict-aware Dual-output Refinement Framework (CDRF), which detects inconsistencies between corrections and explanations. Experiments show CDRF significantly improves morph resolution accuracy and interpretability. Our datasets and code are available [<https://anonymous.4open.science/r/Morph-Resolution-Datasets-and-Methods-611E>].
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
Temporal Token Matters: Investigating and Interpreting the Consistency of Temporal Ordering in Large Language Models
PDF ↗Despite the remarkable performance across numerous tasks, Large Language Models (LLMs) still exhibit notable deficiencies in temporal reasoning, even in simple event ordering tasks. For instance, a slight alteration in the temporal phrasing of the question (e.g., changing "Is event A before B?” to "Is event A after B?") can lead LLMs to hallucinate and produce inconsistent answers, reflecting a lack of robust temporal reasoning. Although many prior studies have focused on benchmarking and improving the temporal reasoning ability of LLMs, little is known about the intrinsic mechanisms within LLMs when performing temporal reasoning. In this work, we investigate the mechanistic interpretability of temporal ordering within event temporal reasoning through a structured "Identify-Interpret-Verify” pipeline. We first employ path patching to identify a sparse subset of attention heads that are causally responsible for reasoning outcomes. Detailed pattern analysis reveals that these key heads specialize in attending to either temporal keywords (semantic cues) or structural delimiters (syntactic cues). Furthermore, we rigorously validate the observed mechanism through comprehensive intervention-based experiments, ranging from head ablation to targeted attention modulation. We demonstrate that dynamically modulating the attention of these specific heads can robustly enhance model performance, which serves as strong empirical evidence that our identified mechanism faithfully captures the internal logic of temporal ordering in LLMs.
In the era of Large Language Models (LLMs), the Mixture of Experts (MoE) architecture has emerged as an effective approach for training extremely large models with improved computational efficiency. This success builds upon extensive prior research aimed at enhancing expert specialization in MoE-based LLMs. However, the nature of such specializations and how they can be systematically interpreted remain open research challenges. In this work, we investigate this gap by posing a fundamental question: *Do domain-specific experts exist in MoE-based LLMs?* To answer the question, we evaluate ten advanced MoE-based LLMs ranging from 3.8B to 120B parameters and provide empirical evidence for the existence of domain-specific experts. Building on this finding, we propose **Domain Steering Mixture of Experts (DSMoE)**, a training-free framework that introduces zero additional inference cost and outperforms both well-trained MoE-based LLMs and strong baselines, including Supervised Fine-Tuning (SFT). Experiments on four advanced open-source MoE-based LLMs across both target and non-target domains demonstrate that our method achieves strong performance and robust generalization without increasing inference cost or requiring additional retraining.
Treating random masking as a performance plug-in for large language models (LLMs) offers three advantages: low coupling to the task, the model, and training resources. However, the critical drawback is that its gains are highly stochastic. Motivated by this, we propose play-it-by-ear masking performance plug-in (PibE-MPP), which enables LLMs to adaptively select masking target combinations for each task, retaining these advantages and mitigating the drawback. Specifically, we pose two core questions—what are the masking targets and what is the masking strategy under 7 constraints obtained from these advantages and a drawback. For the first question, we select all attention heads in the last layer as masking targets by constructing a first-order Markov process with alternating hidden state and information fusion. The feasibility of this target is validated by random masking experiments. For the second question, we first construct a small yet interpretable candidate set by proposing a three-axis mapping and a mean-based criterion for fusion features of masking targets. We then propose an axis-variance minimization to select a compact masking-target combination, reducing sensitivity to outlier targets. Experiments on 6 LLMs (Qwen and LLaMA) and 24 datasets demonstrate PibE-MPP’s effectiveness and generality, gain stability, and domain performance, and verify the necessity of its final module, providing empirical evidence of its transferability across tasks and models. The code is available at https://github.com/wtctcop/PibE-MPP.
Large Language Models (LLMs) exhibit strong general reasoning but struggle in molecular science due to the lack of explicit priors required for precise chemical reasoning. Current solutions inject priors into parameters, but this static coupling hinders rapid knowledge updates and often compromises the model’s general capabilities. To address this, we introduce REAP, a training-free framework that equips LLMs with an external knowledge base, enabling them to reason over retrieved chemical priors dynamically. REAP implements a structured reasoning pipeline that autonomously selects relevant priors from our constructed atom-level knowledge base, retrieves analogue exemplars, and synthesizes these information to guide the LLM’s decision-making. This architecture ensures interpretability and adaptability while preserving the LLM’s intrinsic general intelligence. Experiments show that REAP outperforms current reasoning methods and rivals state-of-the-art training-based models, demonstrating the effectiveness of our framework.
Role-playing models (RPMs) are widely used in real-world applications but underperform when deployed in the wild. This degradation can be attributed to distribution shifts, including user, character, and dialogue compositional shifts. Existing methods like LLM-as-a-judge fall short in providing a fine-grained diagnosis of how these shifts affect RPM generalization, and thus there lack formal frameworks to characterize RPM generalization behaviors. To bridge these gaps, we introduce an information-theoretic metric, named reasoning-based effective mutual information difference (R-EMID), to measure RPM performance degradation in an interpretable way. We also derive an upper bound on R-EMID to predict the worst-case generalization performance of RPMs and theoretically reveal how various shifts contribute to the RPM performance degradation. Moreover, we propose a co-evolving reinforcement learning framework to adaptively model the connection among user, character, and dialogue context and thus enhance the estimation of dialogue response generation probability, which is critical for calculating R-EMID. Finally, we evaluate the generalization performance of various RPMs using R-EMID, finding that user shift poses the highest risk among all shifts and reinforcement learning is the most effective approach for enhancing RPM generalization. Code and data are available at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/RPM-Generalization.
Rethinking Research on Stereotypes: An Analysis through Social Psychological and Computational Perspectives
PDF ↗Stereotypes are social constructs shaping human perception and behavior that can produce harmful outcomes under specific conditions. Recent work shows that large language models (LLMs) may inherit and amplify such social harms. However, most existing research often focuses only upon stereotypical biases and overlooks stereotypes and the rich social psychological literature on them, resulting in resource wastage and slowed progress in stereotype research.We argue that meaningful progress in mitigating stereotypes in LLMs requires tighter integration between social psychology and computational research. To address this gap, we review core social psychological theories and frameworks and analyze their computational operationalization, highlighting substantial open opportunities.We also analyze computational progress across media narratives, body imaging, and multilingual, multicultural, and multimodal contexts, identifying key gaps and limitations in each domain.We also present a unified analysis of challenges in stereotype research.We further discuss implications for responsible AI, highlighting stereotypes as a major source of downstream harms, and briefly examine the limitations of current mitigation approaches along with potential improvements via explainability and interpretability. We frame stereotypes in AI as socio-technical phenomena and urge further research in responsible AI, informed by the perspectives and future directions presented in this paper.
Large Language Models (LLMs) face a fundamental challenge with delayed disambiguation: **How is the meaning of an ambiguous word updated when clarifying context arrives only after it has been processed?** While LLMs possess the latent capacity to resolve such ambiguities—as revealed when a full, non-causal context is provided—their unidirectional architecture prevents immediate updates. We investigate the underlying computational mechanism and show this semantic re-evaluation is deferred to subsequent tokens in a process we term "Deferred Semantic Drift (DSD)". Through targeted analysis of attentional pathways, we find that later tokens actively retrieve context-dependent "informational packets" from the ambiguous word’s value vector to steer the final interpretation. We demonstrate this mechanism in metaphor comprehension and provide causal validation by steering model outputs towards literal or metaphorical meanings via targeted activation interventions. This research uncovers a key computational strategy for meaning construction, offering crucial insights for understanding and guiding the behavior of LLMs.
Large language models (LLMs) seem to offer an easy path to interpretability: just ask them to explain their answers. Yet the features driving an answer often differ from those emphasized in its explanation, meaning post-hoc rationales can misrepresent what actually shaped the model’s output. We quantify this gap by comparing the feature-importance distributions of answers and their explanations. Prior analyses reveal such discrepancies, but large-scale study has been limited by the high computational cost of attribution methods. To address this, we introduce the Post-hoc Self-Consistency Bank (PSCB), a large-scale benchmark linking model decisions with diverse explanations and attribution vectors across datasets, methods, and model families. Using PSCB, we find that Spearman rank correlation provides a more reliable signal of alignment than cosine similarity. Building on this insight, we apply Direct Preference Optimization (DPO) to attribution-based preference data, improving alignment without degrading task accuracy, and show that standard supervised fine-tuning on the same data fails to achieve comparable gains. These improvements generalize robustly across domains, paving the way toward scalable and faithful alignment between LLM decisions and their natural language explanations.
Visual text grounding provides interpretable evidence for document question answering. Due to the complex layouts and mixed visual-text contents in text-rich images, effective visual text grounding requires strong visual and spatial reasoning to localize multiple referenced regions. Existing multimodal large language model (MLLM) approaches often struggle to align query tokens with visual–text patches, heavily relying on lengthy OCR inputs. To tackle this problem, we propose Doc-AGround, an OCR-free approach that leverages the MLLM’s inherent multi-head attention for multi-patch grounding. Doc-AGround extracts a patch-wise attention map as the grounding prediction. Concurrently, it introduces an effective multi-head weighting mechanism to amplify the attention heads’ intrinsic role in connecting vision and text. Empirical results of Doc-AGround show state-of-the-art performance on challenging document grounding benchmarks, demonstrating the effectiveness of the proposed attention-based grounding design.
Large language models (LLMs) encode problem difficulty as an internal signal that can be linearly decoded from their residuals. Given their multilingual capabilities, we investigate whether this meta-cognitive signal is language-agnostic and how it is organized across the model’s layers by training linear probes on the AMC subset of the Easy2Hard benchmark, translated into 21 languages. We found that difficulty-related signals emerge at two distinct stages of the model internals, corresponding to shallow (early-layers) and deep (later-layers) internal representations, that exhibit functionally different behaviors. Probes trained on deep representations achieve high accuracy when evaluated on the same language but exhibit weaker cross-lingual transfer. In contrast, probes trained on shallow representations generalize better across languages, despite achieving lower within-language performance. This closely aligns with existing findings in LLM interpretability, showing that models tend to operate in an abstract conceptual space before producing language-specific outputs. Our results suggest that this two-stage organizational principle extends beyond simple semantic processing to meta-cognitive properties such as problem difficulty, highlighting an internal control signal that is not tied to surface meaning.
This paper introduces Late Code Chunking (LC^2), a chunking strategy designed to improve the semantic understanding of code segments for Large Language Models (LLMs). Repository-level code completion requires predicting the completion of unfinished code by leveraging cross-file context spread across a repository. However, when retrieved fragments have missing semantics—the loss of structural or behavioral information during chunking—LLMs struggle to interpret the target code. To address this, LC^2 refines retrieved chunks by constructing a dual context: a "Code Retrieval Context" optimized for similarity-based search, and a "Code Comprehension Context" that serves as a late enrichment step through context expansion and augmentation. This dual-context design reduces information loss due to chunking and enhances the ability of LLMs to utilize retrieved code. Additionally, we introduce an Asymmetric Query-Chunk Sizing strategy to further optimize retrieval quality by minimizing query noise. Our experiments demonstrate that LC^2 provides robust performance gains, achieving a statistically significant 19.7% improvement in Exact Match accuracy on the CrossCodeEval benchmark compared to the best existing chunking method.
Commonsense reasoning often involves evaluating multiple plausible interpretations rather than selecting a single atomic answer, yet most benchmarks rely on single-label evaluation, obscuring whether statements are jointly plausible, mutually exclusive, or jointly implausible. We introduce LOGICAL-COMMONSENSEQA, a benchmark that reframes commonsense reasoning as logical composition over pairs of atomic statements using plausibility-level operators (AND, OR and NEITHER/NOR). Evaluating instruction-tuned, reasoning-specialized, and fine-tuned models under zero-shot, few-shot, and chain-of-thought prompting, we find that while models perform reasonably on conjunctive and moderately on disjunctive reasoning, performance degrades sharply on negation-based questions. LOGICAL-COMMONSENSEQA exposes fundamental reasoning limitations and provides a controlled framework for advancing compositional commonsense reasoning.
Large language models (LLMs) are fluent but often brittle when interpretation depends on external information (e.g., events or participant roles), as next-token prediction does not explicitly encode situation-level semantic constraints. FrameNet provides a structured account of semantics through its inventory of frames, roles, and relations. We present a scalable framework that injects frame-semantic knowledge into LLMs via LoRA, moving from fact-oriented prompting to principle-oriented supervision over the full FrameNet inventory. The supervision encodes semantic constraints through semantic types, sense-aware definitions, frame relations, and role-annotated examples. To test whether this knowledge generalizes beyond surface cues, we use Natural Language Inference (NLI) as a diagnostic task for event-level reasoning. Experiments on CONFER and SNLI show consistent gains over Meta-Llama-3.1-8B-Instruct in zero-shot and few-shot settings, especially for entailment and contradiction. Complementary semantic role labeling analyses further indicate improved sensitivity to frame, role, and span structure.
When an AI assistant remembers that Sarah is a single mother working two jobs, does it interpret her stress differently than if she were a wealthy executive? As personalized AI systems increasingly incorporate long-term user memory, understanding how this memory shapes emotional reasoning is critical. We investigate how user memory affects emotional intelligence in large language models (LLMs) by evaluating 15 models on human validated emotional intelligence tests. We find that identical scenarios paired with different user profiles produce systematically divergent emotional interpretations. Across validated user-independent emotional scenarios and diverse user profiles, systematic biases emerged in several high-performing LLMs where advantaged profiles received more accurate emotional interpretations. Moreover, LLMs demonstrate significant disparities across demographic factors in emotion understanding and supportive recommendations tasks, indicating that personalization mechanisms can embed social hierarchies into models’ emotional reasoning. These results highlight a key challenge for memory-enhanced AI: systems designed for personalization may inadvertently reinforce social inequalities.
Events exhibit rich semantic relations that are essential for understanding the unfolding of real-world processes. Although large language models (LLMs) have achieved strong performance on event relation extraction, how event relations are internally represented and utilized remains unclear. In this paper, we present UERLens, an interpretability framework for understanding event relations in LLMs. Specifically, we first construct UERBench, a counterfactual dataset for event relation analysis that covers causal, temporal, and sub-event relations. Based on counterfactual pairs, we identify relation-sensitive internal features by comparing model activations. We then examine the functional role of these features through model manipulation, including model intervention and model training. Experimental results show that event relations are encoded through structured and layer-specific internal features. Disabling relation-sensitive features leads to performance drops of over 22%, while enhancing them yields improvements of up to 7%. Furthermore, leveraging these interpretable features to train a lightweight classifier significantly improves event relation extraction, achieving F1 gains of up to 24% for causal relations.
Retrieval-augmented generation (RAG) over long documents typically involves splitting the text into smaller chunks, which serve as the basic units for retrieval. However, due to dependencies across the original document, contextual information is often essential for accurately interpreting each chunk. To address this, prior work has explored encoding longer context windows to produce embeddings for longer chunks, yet their gains in retrieval and downstream tasks remain limited. This is because (1) longer chunks strain the capacity of embedding models due to the increased amount of information they must encode, and (2) many real-world applications still require returning localized evidence due to constraints on model or human bandwidth. To this end, we propose an alternative approach to this challenge by representing short chunks in a way that is conditioned on a broader context window to enhance retrieval performance – i.e., situating a chunk’s meaning within its context. We further show that existing embedding models are not well-equipped to encode such situated context effectively, and thus introduce a new training paradigm and develop the first situated embedding model. To evaluate our method, we curate a book-plot retrieval dataset specifically designed to assess situated retrieval capabilities. On this benchmark, our 1B-parameter model substantially outperforms state-of-the-art embedding models, including several with up to 7B parameters.
Is Chain-of-Thought Really Not Explainability? Chain-of-Thought Can Be Faithful without Hint Verbalization
PDF ↗Recent work, using the Biasing Features metric, labels a CoT as unfaithful if it omits a prompt-injected hint that affected the prediction. We argue this metric adopts a narrow notion of faithfulness and confuses unfaithfulness with incompleteness, the lossy compression needed to turn distributed transformer computation into a linear natural language narrative. On multi-hop reasoning tasks with instruct-tuned and reasoning models, many CoTs flagged as unfaithful by Biasing Features are judged faithful by other metrics, exceeding 50% in some models. With a new faithful@k metric, we show that larger inference-time budgets greatly increase hint verbalization (up to 90% in some settings), suggesting much apparent unfaithfulness is due to tight token limits. Using Causal Mediation Analysis, we further show that even non-verbalized hints can causally mediate prediction changes through the CoT. We therefore caution against relying solely on hint-based evaluations and advocate a broader interpretability toolkit, including causal mediation and corruption-based metrics. We do not claim all CoTs are faithful, only that the absence of hint words alone does not prove unfaithfulness.
Understanding how large language models (LLMs) store, retain, and remove knowledge is critical for interpretability, reliability, and privacy compliance. We reveal a key phenomenon: machine unlearning imprints distinct geometric signatures in the model’s input loss landscape (ILL), with unlearned examples forming flat, low-curvature plateaus that contrast sharply with the high-curvature basins of retained or unseen examples. Remarkably, these patterns emerge even when pointwise losses overlap, exposing residual memorization through input-output behavior alone. Building on this insight, we introduce **REMIND (Residual Memorization in Neighborhood Dynamics)**, a framework that diagnoses memorization states (retained, forgotten, holdout) by probing local ILL curvature over semantically coherent neighborhoods. REMIND operates using only loss queries and a novel embedding-proximity perturbation method to generate controlled, interpretable variants. In evaluations, REMIND achieves 82% multi-class ROC-AUC, outperforming baselines like ROUGE-L and MIN-K%++, with roughly 2× higher AUC at 1% FPR, and remains robust on paraphrased inputs. This neighborhood-level geometric analysis provides a practical, interpretable lens on LLM knowledge retention and unlearning, detecting subtle residual signals missed by pointwise or aggregated metrics.
Reinforcement Learning from Human Feedback (RLHF) has emerged as a crucial technique for aligning large language models (LLMs) with human preferences. However, existing RLHF methods face key challenges, including poor sample efficiency, high computational overhead, and slow convergence. Recent studies highlight the importance of data selection in RL, but how to effectively select the most beneficial experiences for RL training remains an open problem. Existing data selection methods for RL rely on heuristic metrics, failing to establish an interpretable connection between data and optimization objectives. To address this problem, we propose InfOES (Influence-based Online Experience Selection), a novel data selection method for RLHF that dynamically estimates the influence of individual training samples on policy optimization. By incorporating data attribution into the policy gradient, InfOES can identify and filter out detrimental samples on the fly, ensuring effective convergence toward alignment objectives. Our approach is compatible with various RL algorithms (e.g., PPO, GRPO, REINFORCE++). Extensive experiments demonstrate that InfOES significantly enhances training effectiveness, achieving superior alignment performance with fewer optimization steps.