Many vision-language models (VLMs) that prove very effective at a range of multimodal task, build on CLIP-based vision encoders, which are known to have various limitations. We investigate the hypothesis that the strong language backbone in VLMs compensates for possibly weak visual features by contextualizing or enriching them. Using three CLIP-based VLMs, we perform controlled self-attention ablations on a carefully designed probing task. Our findings show that despite known limitations, CLIP visual representations offer ready-to-read semantic information to the language decoder. However, in scenarios of reduced contextualization in the visual representations, the language decoder can largely compensate for the deficiency and recover performance. This suggests a dynamic division of labor in VLMs and motivates future architectures that offload more visual processing to the language decoder.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Aggregating multiple annotations into a single ground truth label may hide valuable insights into annotator disagreement, particularly in tasks where subjectivity plays a crucial role. In this work, we explore methods for identifying subjectivity in recognizing the human values that motivate arguments. We evaluate two main approaches: inferring subjectivity through value prediction vs. directly identifying subjectivity. Our experiments show that direct subjectivity identification significantly improves the model performance of flagging subjective arguments. Furthermore, combining contrastive loss with binary cross-entropy loss does not improve performance but reduces the dependency on per-label subjectivity. Our proposed methods can help identify arguments that individuals may interpret differently, fostering a more nuanced annotation process.
Multi-modal intent recognition (MIR) requires integrating non-verbal cues from real-world contexts to enhance human intention understanding, which has attracted substantial research attention in recent years. Despite promising advancements, a comprehensive survey summarizing recent advances and new frontiers remains absent. To this end, we present a thorough and unified review of MIR, covering different aspects including (1) Extensive survey: we take the first step to present a thorough survey of this research field covering textual, visual (image/video), and acoustic signals. (2) Unified taxonomy: we provide a unified framework including evaluation protocol and advanced methods to summarize the current progress in MIR. (3) Emerging frontiers: We discuss some future directions such as multi-task, multi-domain, and multi-lingual MIR, and give our thoughts respectively. (4) Abundant resources: we collect abundant open-source resources, including relevant papers, data corpora, and leaderboards. We hope this survey can shed light on future research in MIR.
Long-form egocentric video understanding provides rich contextual information and unique insights into long-term human behaviors, holding significant potential for applications in embodied intelligence, long-term activity analysis, and personalized assistive technologies. However, existing benchmark datasets primarily focus on single, short (e.g., minutes to tens of minutes) to moderately long videos, leaving a substantial gap in evaluating extensive, ultra-long egocentric video recordings. To address this, we introduce X-LeBench, a novel benchmark dataset meticulously designed to fill this gap by focusing on tasks requiring a comprehensive understanding of extremely long egocentric video recordings. Our X-LeBench develops a life-logging simulation pipeline that produces realistic, coherent daily plans aligned with real-world video data. This approach enables the flexible integration of synthetic daily plans with real-world footage from Ego4D—a massive-scale egocentric video dataset covers a wide range of daily life scenarios—resulting in 432 simulated video life logs spanning from 23 minutes to 16.4 hours. The evaluations of several baseline systems and multimodal large language models (MLLMs) reveal their poor performance across the board, highlighting the inherent challenges of long-form egocentric video understanding, such as temporal localization and reasoning, context aggregation, and memory retention, and underscoring the need for more advanced models.
Mitigating social bias in large language models (LLMs) has become an increasingly important research objective. However, existing debiasing methods often incur high human and computational costs, exhibit limited effectiveness, and struggle to scale to larger models and open-ended generation tasks. To address these limitations, this paper proposes BiasFilter, a model-agnostic, inference-time debiasing framework that integrates seamlessly with both open-source and API-based LLMs. Instead of relying on retraining with balanced data or modifying model parameters, BiasFilter enforces fairness by filtering generation outputs in real time. Specifically, it periodically evaluates intermediate outputs every few tokens, maintains an active set of candidate continuations, and incrementally completes generation by discarding low-reward segments based on a fairness reward signal. To support this process, we construct a fairness preference dataset and train an implicit reward model to assess token-level fairness in generated responses. Extensive experiments demonstrate that BiasFilter effectively mitigates social bias across a range of LLMs while preserving overall generation quality.
Dropping Experts, Recombining Neurons: Retraining-Free Pruning for Sparse Mixture-of-Experts LLMs
PDF ↗Sparse Mixture-of-Experts (SMoE) architectures are widely used in large language models (LLMs) due to their computational efficiency. However, though only a few experts are activated for each token, SMoE still requires loading all expert parameters, leading to high memory usage and challenges in deployment. Previous work has tried to reduce the overhead by pruning and merging experts, but primarily focused on expert-level operations, leaving neuron-level structure underexplored. We propose **DERN** (**D**ropping **E**xperts, **R**ecombining **N**eurons), a task-agnostic and retraining-free framework for expert pruning and reconstruction. We observe that experts are often misaligned and contain semantic conflicts at the neuron level, which poses challenges for direct merging. To solve this, DERN works in three steps: it first prunes redundant experts using router statistics; then it decomposes them into neuron-level expert segments, assigning each segment to its most compatible retained expert; and finally, it merges segments within each retained expert to build a compact representation. Experiments on Mixtral, Qwen, and DeepSeek SMoE models show that DERN improves performance by more than 5% on commonsense reasoning and MMLU benchmarks under 50% expert sparsity, without extra training. It also greatly reduces the number of experts and memory usage, making SMoE LLMs easier to deploy in practice.
Guess What I am Thinking: A Benchmark for Inner Thought Reasoning of Role-Playing Language Agents
PDF ↗Recent advances in Large Language Model (LLM)-based Role-Playing Language Agents (RPLAs) have attracted broad attention in various applications. While chain-of-thought reasoning has shown importance in many tasks for LLMs, the internal thinking processes of RPLAs remain unexplored. Understanding characters’ inner thoughts is crucial for developing advanced RPLAs. In this paper, we introduce ROLETHINK, a novel benchmark constructed from literature for evaluating character thought generation. We propose the task of inner thought reasoning, constructing 6,058 data entries from 76 books, which includes two sets: the gold set that compares generated thoughts with original character monologues, and the silver set that uses expert-synthesized character analyses as references. To address this challenge, we propose MIRROR, a chain-of-thought approach that generates character thoughts by retrieving memories, predicting character reactions, and synthesizing motivations. Through extensive experiments, we demonstrate the importance of inner thought reasoning for RPLAs, and MIRROR consistently outperforms existing methods.
As large language models (LLMs) are adapted to sensitive domains such as medicine, their fluency raises safety risks, particularly regarding provenance and accountability. Watermarking embeds detectable patterns to mitigate these risks, yet its reliability in medical contexts remains untested. Existing benchmarks focus on detection-quality tradeoffs and overlook factual risks. In medical text, watermarking often reweights low-entropy tokens, which are highly predictable and often carry critical medical terminology. Shifting these tokens can cause inaccuracy and hallucinations, risks that prior general-domain benchmarks fail to capture.We propose a medical-focused evaluation workflow that jointly assesses factual accuracy and coherence. Using GPT-Judger and further human validation, we introduce the Factuality-Weighted Score (FWS), a composite metric prioritizing factual accuracy beyond coherence to guide watermarking deployment in medical domains. Our evaluation shows current watermarking methods substantially compromise medical factuality, with entropy shifts degrading medical entity representation. These findings underscore the need for domain-aware watermarking approaches that preserve the integrity of medical content.
Q-PRM: Adaptive Query Rewriting for Retrieval-Augmented Generation via Step-level Process Supervision
PDF ↗Query rewriting plays a pivotal role in Retrieval-Augmented Generation (RAG) by refining real-world queries of varying complexity. Existing approaches typically rely on outcome-supervised training or heuristic rules to guide the rewriting process. However, these paradigms often struggle to handle queries with varying levels of complexity, posing over- and under-refinement problems. We identify the root cause of these issues as the absence of supervision signals for intermediate steps. To fully construct and utilize such signals, we propose Q-PRM, a novel query rewriting framework. Q-PRM reformulates the rewriting process as a Markov Decision Process (MDP) composed of atomic rewriting steps. In this way, Q-PRM can apply process-level supervision to each atomic step according to the query type, offering more targeted and effective guidance. Q-PRM comprises three key stages: (1) applying Monte Carlo Tree Search to generate step-level process supervision signals; (2) performing reinforced self-training for progressive process refinement; and (3) employing PRM-guided decoding during inference. Experiments on several open-domain QA benchmarks demonstrate that Q-PRM consistently outperforms baselines across different levels of query complexity.
Conversational derailment — when online discussions stray from their intended topics due to toxic or inappropriate remarks — is a common issue on online platforms. These derailments can have negative impacts on users and the online community. While previous work has focused on post hoc identification of toxic content, recent efforts emphasize proactive prediction of derailments before they occur, enabling early moderation. However, forecasting derailment is difficult due to the context-dependent emergence of toxicity and the need for timely alerts. We prompt pre-trained large language models (LLMs) to predict conversational derailment without task-specific fine-tuning. We compare a range of prompting strategies, including chain-of-thought reasoning (CoT) and few-shot exemplars, across small and large scale models, and evaluate their performance and inference-cost trade-offs on derailment benchmarks. Our experiments show that the best prompting configuration attains state-of-the-art performance, and forecasts derailments earlier than existing approaches. These results demonstrate that LLMs, even without fine-tuning, can serve as an effective tool for proactive conversational moderation.
Mixture-of-Experts (MoE) models have become increasingly powerful in multimodal learning by enabling modular specialization across modalities. However, their effectiveness remains unclear when additional modalities introduce more noise than complementary information. Existing approaches, such as the Partial Information Decomposition, struggle to scale beyond two modalities and lack the resolution needed for instance-level control. We propose **B**eyond **T**wo-modality **W**eighting (**BTW**), a bi-level, non-parametric weighting framework that combines instance-level Kullback-Leibler (KL) divergence and modality-level mutual information (MI) to dynamically adjust modality importance during training. Our method does not require additional parameters and can be applied to an arbitrary number of modalities. Specifically, BTW computes per-example KL weights by measuring the divergence between each unimodal and the current multimodal prediction, and modality-wide MI weights by estimating global alignment between unimodal and multimodal outputs. Extensive experiments on sentiment regression and clinical classification demonstrate that our method significantly improves regression performance and multiclass classification accuracy.
Neutral Is Not Unbiased: Evaluating Implicit and Intersectional Identity Bias in LLMs Through Structured Narrative Scenarios
PDF ↗Large Language Models often reproduce societal biases, yet most evaluations overlook how such biases evolve across nuanced contexts or intersecting identities. We introduce a scenario-based evaluation framework built on 100 narrative tasks, designed to be neutral at baseline and systematically modified with gender and age cues. Grounded in the theory of Normative-Narrative Scenarios, our approach provides ethically coherent and socially plausible settings for probing model behavior. Analyzing responses from five leading LLMs—GPT-4o, LLaMA 3.1, Qwen2.5, Phi-4, and Mistral—using Critical Discourse Analysis and quantitative linguistic metrics, we find consistent evidence of bias. Gender emerges as the dominant axis of bias, with intersectional cues (e.g., age and gender combined) further intensifying disparities. Our results underscore the value of dynamic narrative progression for detecting implicit, systemic biases in Large Language Models.
Can Large Language Models (LLMs) simulate humans in making important decisions? Recent research has unveiled the potential of using LLMs to develop role-playing language agents (RPLAs), mimicking mainly the knowledge and tones of various characters. However, imitative decision-making necessitates a more nuanced understanding of personas. In this paper, we benchmark the ability of LLMs in persona-driven decision-making. Specifically, we investigate whether LLMs can predict characters’ decisions provided by the preceding stories in high-quality novels. Leveraging character analyses written by literary experts, we construct a dataset LIFECHOICE comprising 2,512 characters’ decision points from 470 books. Then, we conduct comprehensive experiments on LIFECHOICE with various LLMs and RPLA methodologies. The results demonstrate that state-of-the-art LLMs exhibit promising capabilities in this task, yet substantial room for improvement remains. Hence, we further propose the CHARMAP method, which adopts persona-based memory retrieval and significantly advances RPLAs on this task.
Large Language Model (LLMs) can be used to write or modify documents, presenting a challenge for understanding the intent behind their use. For example, benign uses may involve using LLM on a human-written document to improve its grammar or to translate it into another language. However, a document entirely produced by a LLM may be more likely to be used to spread misinformation than simple translation (, from use by malicious actors or simply by hallucinating). Prior works in Machine Generated Text (MGT) detection mostly focus on simply identifying whether a document was human or machine written, ignoring these fine-grained uses. In this paper, we introduce a HiErarchical, length-RObust machine-influenced text detector (HERO), which learns to separate text samples of varying lengths from four primary types: human-written, machine-generated, machine-polished, and machine-translated. HERO accomplishes this by combining predictions from length-specialist models that have been trained with Subcategory Guidance. Specifically, for categories that are easily confused (, different source languages), our Subcategory Guidance module encourages separation of the fine-grained categories, boosting performance. Extensive experiments across five LLMs and six domains demonstrate the benefits of our HERO, outperforming the state-of-the-art by 2.5-3 mAP on average.
SaCa: A Highly Compatible Reinforcing Framework for Knowledge Graph Embedding via Structural Pattern Contrast
PDF ↗Knowledge Graph Embedding (KGE) seeks to learn latent representations of entities and relations to support knowledge-driven AI systems. However, existing KGE approaches often exhibit a growing discrepancy between the learned embedding space and the intrinsic structural semantics of the underlying knowledge graph. This divergence primarily stems from the over-reliance on geometric criteria for assessing triple plausibility, whose effectiveness is inherently limited by the sparsity of factual triples and the disregard of higher-order structural dependencies in the knowledge graph. To overcome this limitation, we introduce Structure-aware Calibration (SaCa), a versatile framework designed to calibrate KGEs through the integration of global structural patterns. SaCa designs two new components: (i) Structural Proximity Measurement, which captures multi-order structural signals from both entity and entity-relation perspectives; and (ii) KG-Induced Soft-weighted Contrastive Learning (KISCL), which assigns soft weights to hard-to-distinguish positive and negative pairs, enabling the model to better reflect nuanced structural dependencies. Extensive experiments on seven benchmarks demonstrate that SaCa consistently boosts performance across ten KGE models on link prediction and entity classification tasks with minimal overhead.
Concerns about text-to-image (T2I) generative models infringing on privacy, copyright, and safety have led to the development of concept erasure techniques (CETs). The goal of an effective CET is to prohibit the generation of undesired “target” concepts specified by the user, while preserving the ability to synthesize high-quality images of other concepts. In this work, we demonstrate that concept erasure has side effects and CETs can be easily circumvented. For a comprehensive measurement of the robustness of CETs, we present the Side Effect Evaluation (SEE) benchmark that consists of hierarchical and compositional prompts describing objects and their attributes. The dataset and an automated evaluation pipeline quantify side effects of CETs across three aspects: impact on neighboring concepts, evasion of targets, and attribute leakage. Our experiments reveal that CETs can be circumvented by using superclass-subclass hierarchy, semantically similar prompts, and compositional variants of the target. We show that CETs suffer from attribute leakage and a counterintuitive phenomenon of attention concentration or dispersal. We release our benchmark and evaluation tools to aid future work on robust concept erasure.
An LLM-based Temporal-spatial Data Generation and Fusion Approach for Early Detection of Late Onset Alzheimer’s Disease (LOAD) Stagings Especially in Chinese and English-speaking Populations
PDF ↗Alzheimer’s Disease (AD), the 7th leading cause of death globally, demands scalable methods for early detection. While speech-based diagnostics offer promise, existing approaches struggle with temporal-spatial (T-S) challenges in capturing subtle linguistic shifts across different disease stages (temporal) and in adapting to cross-linguistic variability (spatial). This study introduces a novel Large Language Model (LLM)-driven T-S fusion framework that integrates multilingual LLMs, contrastive learning, and interpretable marker discovery to revolutionize Late Onset AD (LOAD) detection. Our key innovations include: (1) T-S Data Imputation: Leveraging LLMs to generate synthetic speech transcripts across different LOAD stages (NC, Normal Control; eMCI, early Mild Cognitive Impairment; lMCI, late Mild Cognitive Impairment; AD) and languages (Chinese, English, Spanish), addressing data scarcity while preserving clinical relevance (expert validation: 86% agreement with LLM-generated labels). (2) T-S Transformer with Contrastive Learning: A multilingual model that disentangles stage-specific (temporal) and language-specific (spatial) patterns, achieving a notable improvement of 10.9–24.7% in F1-score over existing baselines. (3) Cross-Linguistic Marker Discovery: Identifying language-agnostic markers and language-specific patterns to enhance interpretability for clinical adoption. By unifying temporal LOAD stages and spatial diversity, our framework achieves state-of-the-art performance in early LOAD detection while enabling cross-linguistic diagnostics. This study bridges NLP and clinical neuroscience, demonstrating LLMs’ potential to amplify limited biomedical data and advance equitable healthcare AI.
Large language models (LLMs) have been used to synthesize persuasive dialogues for studying persuasive behavior. However, existing approaches often suffer from issues such as stance oscillation and low informativeness. To address these challenges, we propose reinforced instructional prompting, a method that ensures speaker characteristics consistently guide all stages of dialogue generation. We further introduce multilingual prompting, which aligns language use with speakers’ native languages to better capture cultural nuances. Our experiments involving speakers from eight countries show that continually reinforcing speaker profiles and cultural context improves argument diversity, enhances informativeness, and stabilizes speaker stances. Moreover, our analysis of inter-group versus intra-group persuasion reveals that speakers engaging within their own cultural groups employ more varied persuasive strategies than in cross-cultural interactions. These findings underscore the importance of speaker and cultural awareness in LLM-based persuasion modeling and suggest new directions for developing more personalized, ethically grounded, and culturally adaptive LLM-generated dialogues.
Error Typing for Smarter Rewards: Improving Process Reward Models with Error-Aware Hierarchical Supervision
PDF ↗Large Language Models (LLMs) are prone to hallucination, especially during multi‐hop and reasoning-intensive tasks such as mathematical problem solving. While Outcome Reward Models verify only final answers, Process Reward Models (PRMs) score each intermediate step to steer generation toward coherent solutions. We introduce PathFinder‐PRM, a novel hierarchical, error‐aware discriminative PRM that first classifies math and consistency errors at each step, then combines these fine‐grained signals to estimate step correctness. To train PathFinder‐PRM, we construct a 400K‐sample dataset by enriching the human‐annotated PRM800K corpus and RLHFlow Mistral traces with three‐dimensional step‐level labels. On PRMBench, PathFinder‐PRM achieves a new state‐of‐the‐art PRMScore of 67.7, outperforming the prior best (65.5) while using 3× less data. When applied to reward guided greedy search, our model yields prm@8 48.3, a +1.5 point gain over the strongest baseline. These results demonstrate that decoupled error detection and reward estimation not only boost fine‐grained error detection but also substantially improve end‐to‐end, reward‐guided mathematical reasoning with greater data efficiency. Our code is available at https://github.com/declare-lab/PathFinder-PRM.
Large language models (LLMs) have shown impressive performance by generating reasoning paths before final answers, but learning such a reasoning path requires costly human supervision. To address this issue, recent studies have explored self-training methods that improve reasoning capabilities using pseudo-labels generated by the LLMs themselves. Among these, confidence-based self-training fine-tunes LLMs to prefer reasoning paths with high-confidence answers, where confidence is estimated via majority voting. However, such methods exclusively focus on the quality of the final answer and may ignore the quality of the reasoning paths, as even an incorrect reasoning path leads to a correct answer by chance. Instead, we advocate the use of reasoning-level confidence to identify high-quality reasoning paths for self-training, supported by our empirical observations. We then propose a new self-training method, **CORE-PO**, that fine-tunes LLMs to prefer high-**CO**nfidence **RE**asoning paths through **P**olicy **O**ptimization. Our experiments show that CORE-PO improves the accuracy of outputs on four in-distribution and two out-of-distribution benchmarks, compared to existing self-training methods.