Identifying linguistic bias in text demands the identification not only of explicitly asserted content but also of implicit content including presuppositions. Large language models (LLMs) offer a promising automated approach to detecting presuppositions, yet the extent to which their judgments align with human intuitions remains unexplored. Moreover, LLMs may inadvertently reflect societal biases when identifying presupposed content. To empirically investigate this, we prompt multiple large language models to evaluate presuppositions across diverse textual domains, drawing from three distinct datasets annotated by human raters. We calculate the agreement between LLMs and human raters, and find several linguistic factors associated with fluctuations in human-model agreement. Our observations reveal discrepancies in human-model alignment, suggesting potential biases in LLMs, notably influenced by gender and political ideology.
论文检索
输入标题、作者或关键词，从 3,086 篇学术成果中精准定位
Moderation Matters: Measuring Conversational Moderation Impact in English as a Second Language Group Discussion
PDF ↗English as a Second Language (ESL) speakers often struggle to engage in group discussions due to language barriers. While moderators can facilitate participation, few studies assess conversational engagement and evaluate moderation effectiveness. To address this gap, we develop a dataset comprising 17 sessions from an online ESL conversation club, which includes both moderated and non-moderated discussions. We then introduce an approach that integrates automatic ESL dialogue assessment and a framework that categorizes moderation strategies. Our findings indicate that moderators help improve the flow of topics and start/end a conversation. Interestingly, we find active acknowledgement and encouragement to be the most effective moderation strategy, while excessive information and opinion sharing by moderators has a negative impact. Ultimately, our study paves the way for analyzing ESL group discussions and the role of moderators in non-native conversation settings. Code and data are available at https://github.com/RenaGao/L2Moderator.
SafeRoute: Adaptive Model Selection for Efficient and Accurate Safety Guardrails in Large Language Models
PDF ↗Deploying large language models (LLMs) in real-world applications requires robust safety guard models to detect and block harmful user prompts. While large safety guard models achieve strong performance, their computational cost is substantial. To mitigate this, smaller distilled models are used, but they often underperform on “hard” examples where the larger model provides accurate predictions. We observe that many inputs can be reliably handled by the smaller model, while only a small fraction require the larger model’s capacity. Motivated by this, we propose SafeRoute, a binary router that distinguishes hard examples from easy ones. Our method selectively applies the larger safety guard model to the data that the router considers hard, improving efficiency while maintaining accuracy compared to solely using the larger safety guard model. Experimental results on multiple benchmark datasets demonstrate that our adaptive model selection significantly enhances the trade-off between computational cost and safety performance, outperforming relevant baselines.
Knowledge Editing (KE) has gained increasing attention, yet current KE tasks remain relatively simple. Under current evaluation frameworks, many editing methods achieve exceptionally high scores, sometimes nearing perfection. However, few studies integrate KE into real-world application scenarios (e.g., recent interest in LLM-as-agent). To support our analysis, we introduce a novel script-based benchmark – ScEdit (Script-based Knowledge Editing Benchmark) – which encompasses both counterfactual and temporal edits. We integrate token-level and text-level evaluation methods, comprehensively analyzing existing KE techniques. The benchmark extends traditional fact-based (“What”-type question) evaluation to action-based (“How”-type question) evaluation. We observe that all KE methods exhibit a drop in performance on established metrics and face challenges on text-level metrics, indicating a challenging task. Our benchmark is available at https://github.com/asdfo123/ScEdit.
Alleviating Hallucinations in Large Language Models via Truthfulness-driven Rank-adaptive LoRA
PDF ↗Improving the truthfulness of LLMs to alleviate hallucinations has become critical for promoting the practical deployment of LLMs. Current fine-tuning-based methods ignore the intrinsic discrepancy in the truthfulness correlations across LLM internal modules, and instead treat them equally, which may potentially decrease the performance of truthfulness improvement. In this paper, we propose a truthfulness-driven rank-adaptive LoRA method to improve LLM truthfulness (RaLFiT), which adaptively allocates the ranks in LoRA training according to the truthfulness correlations of modules within LLM. Specifically, it first measures the truthfulness correlation of each LLM module by a probing process, and allocates higher ranks to strongly correlated modules, which means a larger update subspace during training. Experimental results on TruthfulQA show that RaLFiT consistently outperforms previous state-of-the-art methods across the Llama LLM family, verifying its effectiveness and superiority, and for the first time makes the performance of 7B Llama LLMs exceed GPT-4.
Low-Rank Adaptation (LoRA) has gained popularity for fine-tuning large foundation models, leveraging low-rank matrices \mathbf A and \mathbf B to represent weight changes (i.e., \Delta \mathbf W = \mathbf B \mathbf A). This method reduces trainable parameters and mitigates heavy memory consumption associated with full delta matrices by sequentially multiplying \mathbf A and \mathbf B with the activation. Despite its success, the intrinsic low-rank characteristic may limit its performance. Although several variants have been proposed to address this issue, they often overlook the crucial computational and memory efficiency brought by LoRA. In this paper, we propose Circular Convolution Adaptation (C^3A), which not only achieves high-rank adaptation with enhanced performance but also excels in both computational power and memory utilization. Extensive experiments demonstrate that C^3A consistently outperforms LoRA and its variants across various fine-tuning tasks.
Chain-Talker: Chain Understanding and Rendering for Empathetic Conversational Speech Synthesis
PDF ↗Conversational Speech Synthesis (CSS) aims to align synthesized speech with the emotional and stylistic context of user-agent interactions to achieve empathy. Current generative CSS models face interpretability limitations due to insufficient emotional perception and redundant discrete speech coding. To address the above issues, we present Chain-Talker, a three-stage framework mimicking human cognition: Emotion Understanding derives context-aware emotion descriptors from dialogue history; Semantic Understanding generates compact semantic codes via serialized prediction; and Empathetic Rendering synthesizes expressive speech by integrating both components. To support emotion modeling, we develop CSS-EmCap, an LLM-driven automated pipeline for generating precise conversational speech emotion captions. Experiments on three benchmark datasets demonstrate that Chain-Talker produces more expressive and empathetic speech than existing methods, with CSS-EmCap contributing to reliable emotion modeling. The code and demos are available at: https://github.com/AI-S2-Lab/Chain-Talker.
Large language models (LLMs) have made significant advancements across various tasks, but their safety alignment remains a major concern. Exploring jailbreak prompts can expose LLMs’ vulnerabilities and guide efforts to secure them. Existing methods primarily design sophisticated instructions for the LLM to follow, or rely on multiple iterations, which could hinder the performance and efficiency of jailbreaks. In this work, we propose a novel jailbreak paradigm, Simple Assistive Task Linkage (SATA), which can effectively circumvent LLM safeguards and elicit harmful responses. Specifically, SATA first masks harmful keywords within a malicious query to generate a relatively benign query containing one or multiple [MASK] special tokens. It then employs a simple assistive task—such as a masked language model task or an element lookup by position task—to encode the semantics of the masked keywords. Finally, SATA links the assistive task with the masked query to jointly perform the jailbreak. Extensive experiments show that SATA achieves state-of-the-art performance and outperforms baselines by a large margin. Specifically, on AdvBench dataset, with mask language model (MLM) assistive task, SATA achieves an overall attack success rate (ASR) of 85% and harmful score (HS) of 4.57, and with element lookup by position (ELP) assistive task, SATA attains an overall ASR of 76% and HS of 4.43.
Don’t Miss the Forest for the Trees: Attentional Vision Calibration for Large Vision Language Models
PDF ↗Large Vision Language Models (LVLMs) demonstrate strong capabilities in visual understanding and description, yet often suffer from hallucinations, attributing incorrect or misleading features to images. We observe that LVLMs disproportionately focus on a small subset of image tokens—termed blind tokens—which are typically irrelevant to the query (e.g., background or non-object regions). We hypothesize that such attention misalignment plays a key role in generating hallucinated responses. To mitigate this issue, we propose Attentional Vision Calibration (AvisC), a test-time approach that dynamically recalibrates the influence of blind tokens without modifying the underlying attention mechanism. AvisC first identifies blind tokens by analyzing layer-wise attention distributions over image tokens, then employs a contrastive decoding strategy to balance the influence of original and blind-token-biased logits. Experiments on standard benchmarks, including POPE, MME, and AMBER, demonstrate that AvisC effectively reduces hallucinations in LVLMs.
The rapid advancement of AI-based music generation tools is revolutionizing the music industry but also posing challenges to artists, copyright holders, and providers alike. This necessitates reliable methods for detecting such AI-generated content. However, existing detectors, relying on either audio or lyrics, face key practical limitations: audio-based detectors fail to generalize to new or unseen generators and are vulnerable to audio perturbations; lyrics-based methods require cleanly formatted and accurate lyrics, unavailable in practice. To overcome these limitations, we propose a novel, practically grounded approach: a multimodal, modular late-fusion pipeline that combines automatically transcribed sung lyrics and speech features capturing lyrics related information within the audio. By relying on lyrical aspects directly from audio, our method enhances robustness, mitigates susceptibility to low-level artifacts, and enables practical applicability. Experiments show that our method, DE-detect, outperforms existing lyrics-based detectors while also being more robust to audio perturbations. Thus, it offers an effective, robust solution for detecting AI-generated music in real-world scenarios. Our code is available at https://github.com/deezer/robust-AI-lyrics-detection.
Retrieval-Augmented Generation (RAG) systems often struggle with imperfect retrieval, as traditional retrievers focus on lexical or semantic similarity rather than logical relevance. To address this, we propose HopRAG, a novel RAG framework that augments retrieval with logical reasoning through graph-structured knowledge exploration. During indexing, HopRAG constructs a passage graph, with text chunks as vertices and logical connections established via LLM-generated pseudo-queries as edges. During retrieval, it employs a retrieve-reason-prune mechanism: starting with lexically or semantically similar passages, the system explores multi-hop neighbors guided by pseudo-queries and LLM reasoning to identify truly relevant ones. Experiments on multiple multi-hop benchmarks demonstrate that HopRAG’s retrieve-reason-prune mechanism can expand the retrieval scope based on logical connections and improve final answer quality.
Large language models (LLMs) often generate hallucinated content, making it crucial to identify and quantify inconsistencies in their outputs. We introduce HaluMap, a post-hoc framework that detects hallucinations by mapping entailment and contradiction relations between source inputs and generated outputs using a natural language inference (NLI) model. To improve reliability, we propose a calibration step leveraging intra-text relations to refine predictions. HaluMap outperforms state-of-the-art NLI-based methods by five percentage points compared to other training-free approaches, while providing clear, interpretable explanations. As a training-free and model-agnostic approach, HaluMap offers a practical solution for verifying LLM outputs across diverse NLP tasks. The resources of this paper are available at https://github.com/caisa-lab/acl25-halumap.
The CUTE benchmark showed that LLMs struggle with character understanding in English. We extend it to more languages with diverse scripts and writing systems, introducing EXECUTE. Our simplified framework allows easy expansion to any language. Tests across multiple LLMs reveal that challenges in other languages are not always on the character level as in English. Some languages show word-level processing issues, some show no issues at all. We also examine sub-character tasks in Chinese, Japanese, and Korean to assess LLMs’ understanding of character components.
Event Pattern-Instance Graph: A Multi-Round Role Representation Learning Strategy for Document-Level Event Argument Extraction
PDF ↗For document-level event argument extraction, existing role-based span selection strategies suffer from several limitations: (1) ignoring interrelations among arguments within an event instance; (2) relying on pre-trained language models to capture role semantics at either the event pattern or document, without leveraging pattern-instance associations. To address these limitations, this paper proposes a multi-round role representation learning strategy. First, we construct an event pattern-instance graph (EPIG) to comprehensively capture the role semantics embedded in various direct and indirect associations, including those among roles within event patterns, arguments within event instances, and the alignments between patterns and instances. Second, to enhance the learning of role node representation in the graph, we optimize the update mechanisms for both node and edge representations in the EPIG graph. By leveraging the graph attention network, we iteratively update the representations of role nodes and role edges. The role representations learned from the EPIG are then integrated into the original role representations, further enriching their semantic information. Finally, a role representation memory module and a multi-round learning strategy is proposed to retain and refine role representations learned from previously analyzed documents. This memory mechanism enhances the prediction performance in subsequent rounds of span selection. Extensive experiments on three datasets verify the effectiveness of the model.
‘No’ Matters: Out-of-Distribution Detection in Multimodality Multi-Turn Interactive Dialogue Download PDF
PDF ↗Out-of-distribution (OOD) detection in multimodal contexts is essential for identifying deviations in different modalities, particularly for interactive dialogue systems in real-life interactions, where the systems are usually infeasible to deploy large language models (LLMs) to generate dialogue responses due to data privacy and ethical issues. This paper aims to improve label detection that involves multi-round long dialogues by efficiently detecting OOD dialogues and images. We introduce a novel scoring framework named Dialogue Image Aligning and Enhancing Framework (DIAEF) that integrates the visual language models with the novel proposed scores that detect OOD in two key scenarios (1) mismatches between the dialogue and image input pair and (2) input pairs with previously unseen labels. Our experimental results, derived from various benchmarks, demonstrate that integrating image and multi-round dialogue OOD detection is more effective with previously unseen labels than using either modality independently. In the presence of mismatched pairs, our proposed score effectively identifies these mismatches and demonstrates strong robustness in long dialogues. This approach enhances domain-aware, adaptive conversational agents and establishes baselines for future studies.
Logical Consistency is Vital: Neural-Symbolic Information Retrieval for Negative-Constraint Queries
PDF ↗Information retrieval plays a crucial role in resource localization. Current dense retrievers retrieve the relevant documents within a corpus via embedding similarities, which compute similarities between dense vectors mainly depending on word co-occurrence between queries and documents, but overlook the real query intents. Thus, they often retrieve numerous irrelevant documents. Particularly in the scenarios of complex queries such as negative-constraint queries, their retrieval performance could be catastrophic. To address the issue, we propose a neuro-symbolic information retrieval method, namely NS-IR, that leverages first-order logic (FOL) to optimize the embeddings of naive natural language by considering the logical consistency between queries and documents. Specifically, we introduce two novel techniques, logic alignment and connective constraint, to re-rank candidate documents, thereby enhancing retrieval relevance. Furthermore, we construct a new dataset NegConstraint including negative-constraint queries to evaluate our NS-IR’s performance on such complex IR scenarios. Our extensive experiments demonstrate that NS-IR not only achieves superior zero-shot retrieval performance on web search and low-resource retrieval tasks, but also performs better on negative-constraint queries. Our scource code and dataset are available at https://github.com/xgl-git/NS-IR-main.
In psycholinguistic modeling, surprisal from larger pre-trained language models has been shown to be a poorer predictor of naturalistic human reading times. However, it has been speculated that this may be due to data leakage that caused language models to see the text stimuli during training. This paper presents two studies to address this concern at scale. The first study reveals relatively little leakage of five naturalistic reading time corpora in two pre-training datasets in terms of length and frequency of token n-gram overlap. The second study replicates the negative relationship between language model size and the fit of surprisal to reading times using models trained on ‘leakage-free’ data that overlaps only minimally with the reading time corpora. Taken together, this suggests that previous results using language models trained on these corpora are not driven by the effects of data leakage.
Systematic generalization remains challenging for current language models, which are known to be both sensitive to semantically similar permutations of the input and to struggle with known concepts presented in novel contexts. Although benchmarks exist for assessing compositional behavior, it is unclear how to measure the difficulty of a systematic generalization problem. In this work, we show how one aspect of systematic generalization can be described by the entropy of the distribution of component parts in the training data. We formalize a framework for measuring entropy in a sequence-to-sequence task and find that the performance of popular model architectures scales with the entropy. Our work connects systematic generalization to information efficiency, and our results indicate that success at high entropy can be achieved even without built-in priors, and that success at low entropy can serve as a target for assessing progress towards robust systematic generalization.
Critic-CoT: Boosting the Reasoning Abilities of Large Language Model via Chain-of-Thought Critic
PDF ↗Self-critic has become a crucial mechanism for enhancing the reasoning performance of LLMs. However, current approaches mainly involve basic prompts for intuitive instance-level feedback, which resembles System-1 processes and limits the reasoning capabilities. Moreover, there is a lack of in-depth investigations into the relationship between LLM’s ability to criticize and its task-solving performance. To address these issues, we propose Critic-CoT, a novel framework that pushes LLMs toward System-2-like critic capability. Through a step-wise CoT reasoning paradigm and the automatic construction of weak-supervision data without human annotation, Critic-CoT enables LLMs to engage in slow, analytic self-critique and refinement, thereby improving their reasoning abilities. Experiments on GSM8K and MATH and out-of-domain evaluation demonstrate that our enhanced model significantly boosts task-solving performance by filtering out invalid solutions or iterative refinement. Furthermore, we investigate the intrinsic correlation between critique and task-solving abilities within LLMs, discovering that these abilities can mutually reinforce each other rather than conflict.
Emotions are fundamental to conversational understanding. While significant advancements have been achieved in conversational emotion recognition and emotional response generation, recognizing the causes of eliciting emotions is less explored. Previous studies have primarily focused on identifying the causes of emotions by understanding verbal contextual utterances, overlooking that non-verbal emotional cues can elicit emotions. To address this issue, we develop an Emotional Contagion Graph Network (ECGN) that simulates the impact of non-verbal implicit emotions on the counterpart’s emotions. To achieve this, we construct a heterogeneous graph that simulates the transmission of non-verbal emotions alongside verbal influences. By applying message passing between nodes, the constructed graph effectively models both the implicit emotional dynamics and explicit verbal interactions. We evaluate ECGN’s performance through extensive experiments on the benchmark datasets and compare it against multiple state-of-the-art models. Experimental results demonstrate the effectiveness of the proposed model. Our code is available at https://github.com/Yu-Fangxu/ECGN.