Natural language explanations (NLEs) are commonly used to provide plausible free-text explanations of a model’s reasoning about its predictions. However, recent work has questioned their faithfulness, as they may not accurately reflect the model’s internal reasoning process regarding its predicted answer. In contrast, highlight explanations–input fragments critical for the model’s predicted answers–exhibit measurable faithfulness. Building on this foundation, we propose G-TEx, a Graph-Guided Textual Explanation Generation framework designed to enhance the faithfulness of NLEs. Specifically, highlight explanations are first extracted as faithful cues reflecting the model’s reasoning logic toward answer prediction. They are subsequently encoded through a graph neural network layer to guide the NLE generation, which aligns the generated explanations with the model’s underlying reasoning toward the predicted answer. Experiments on both encoder-decoder and decoder-only models across three reasoning datasets demonstrate that G-TEx improves NLE faithfulness by up to 12.18% compared to baseline methods. Additionally, G-TEx generates NLEs with greater semantic and lexical similarity to human-written ones. Human evaluations show that G-TEx can decrease redundant content and enhance the overall quality of NLEs. Our work presents a novel method for explicitly guiding NLE generation to enhance faithfulness, serving as a foundation for addressing broader criteria in NLE and generated text.
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
Reasoning-to-Defend: Safety-Aware Reasoning Can Defend Large Language Models from Jailbreaking
PDF ↗Large Reasoning Models (LRMs) have recently demonstrated impressive performances across diverse domains. However, how the safety of Large Language Models (LLMs) benefits from enhanced reasoning capabilities against jailbreak queries remains unexplored. To bridge this gap, in this paper, we propose Reasoning-to-Defend (R2D), a novel training paradigm that integrates a safety-aware reasoning mechanism into LLMs’ generation process. This enables self-evaluation at each step of the reasoning process, forming safety pivot tokens as indicators of the safety status of responses. Furthermore, in order to improve the accuracy of predicting pivot tokens, we propose Contrastive Pivot Optimization (CPO), which enhances the model’s perception of the safety status of given dialogues. LLMs dynamically adjust their response strategies during reasoning, significantly enhancing their safety capabilities defending jailbreak attacks. Extensive experiments demonstrate that R2D effectively mitigates various attacks and improves overall safety, while maintaining the original performances. This highlights the substantial potential of safety-aware reasoning in improving robustness of LRMs and LLMs against various jailbreaks.
Extensive research on LLM-based spoken dialogue systems has significantly advanced the development of intelligent voice assistants. However, the integration of role information within speech remains an underexplored area, limiting its application in real-world scenarios, particularly in multi-party dialogue settings. With the growing demand for personalization, voice assistants that can recognize and remember users establish a deeper connection with them. We focus on enabling LLMs with speaker-awareness capabilities and enhancing their understanding of character settings through synthetic data to generate contextually appropriate responses. We introduce Persona-Dialogue, the first large-scale multi-party spoken dialogue dataset that incorporates speaker profiles. Based on this dataset, we propose PAChat, an architecture that simultaneously models both linguistic content and speaker features, allowing LLMs to map character settings to speaker identities in speech. Through extensive experiments, we demonstrate that PAChat successfully achieves speaker-specific responses, character understanding, and the generation of targeted replies in multi-party dialogue scenarios, surpassing existing spoken dialogue systems.
Logit Space Constrained Fine-Tuning for Mitigating Hallucinations in LLM-Based Recommender Systems
PDF ↗Large language models (LLMs) have gained increasing attention in recommender systems, but their inherent hallucination issues significantly compromise the accuracy and reliability of recommendation results. Existing LLM-based recommender systems predominantly rely on standard fine-tuning methodologies, often ignoring hallucination issues during the fine-tuning process. To address this challenge, we propose Logit Space Constraints Fine-Tuning (LCFT), a novel fine-tuning framework designed to mitigate hallucination in LLM-based recommenders. Specifically, LCFT takes as input semantically positive and negative instruction pairs and incorporates Kullback–Leibler (KL) divergence into the training objective to explicitly maximise their distributional disparity in the logit space. By conducting such logit space-constrained fine-tuning, LCFT encourages more distinguishable and semantically grounded representations, thereby reducing the model’s susceptibility to hallucination. Extensive experiments on two recommendation models with distinct LLM backbones and four real-world datasets demonstrate that LCFT consistently reduces hallucination and enhances recommendation performance.
Language models can be used to provide interactive, personalized student feedback in educational settings. However, real-world deployment faces three key challenges: privacy concerns, limited computational resources, and the need for pedagogically valid responses. These constraints require small, open-source models that can run locally and reliably ground their outputs in correct information. We introduce SCRIBE, a framework for multi-hop, tool-augmented reasoning designed to generate valid responses to student questions about feedback reports. SCRIBE combines domain-specific tools with a self-reflective inference pipeline that supports iterative reasoning, tool use, and error recovery. We distil these capabilities into 3B and 8B models via two-stage LoRA fine-tuning on synthetic GPT-4o-generated data. Evaluation with a human-aligned GPT-Judge and a user study with 108 students shows that 8B-SCRIBE models achieve comparable or superior quality to much larger models in key dimensions such as relevance and actionability, while being perceived on par with GPT-4o and Llama-3.3 70B by students. These findings demonstrate the viability of SCRIBE for low-resource, privacy-sensitive educational applications.
Mechanistic interpretability research faces a gap between analyzing simple circuits in toy tasks and discovering features in large models. To bridge this gap, we propose text-to-SQL generation as an ideal task to study, as it combines the formal structure of toy tasks with real-world complexity. We introduce TinySQL, a synthetic dataset, progressing from basic to advanced SQL operations, and train models ranging from 33M to 1B parameters to establish a comprehensive testbed for interpretability. We apply multiple complementary interpretability techniques, including Edge Attribution Patching and Sparse Autoencoders, to identify minimal circuits and components supporting SQL generation. We compare circuits for different SQL subskills, evaluating their minimality, reliability, and identifiability. Finally, we conduct a layerwise logit lens analysis to reveal how models compose SQL queries across layers: from intent recognition to schema resolution to structured generation. Our work provides a robust framework for probing and comparing interpretability methods in a structured, progressively complex setting.
Multilingual representations embed words with similar meanings to share a common semantic space across languages, creating opportunities to transfer debiasing effects between languages. However, existing methods for debiassing are unable to exploit this opportunity because they operate on individual languages. We present Iterative Multilingual Spectral Attribute Erasure (IMSAE), which identifies and mitigates joint bias subspaces across multiple languages through iterative SVD-based truncation. Evaluating IMSAE across eight languages and five demographic dimensions, we demonstrate its effectiveness in both standard and zero-shot settings, where target language data is unavailable, but linguistically similar languages can be used for debiasing. Our comprehensive experiments across diverse language models (BERT, LLaMA, Mistral) show that IMSAE outperforms traditional monolingual and cross-lingual approaches while maintaining model utility.
Large Language Models (LLMs) are widely used as conversational agents exploiting their capabilities in various sectors such as education, law, medicine, and more. However, LLMs are often subjected to context-shifting behaviour, resulting in a lack of consistent and interpretable personality-aligned interactions. Adherence to psychological traits lacks comprehensive analysis, especially in the case of dyadic (pairwise) conversations. We examine this challenge from two viewpoints, initially using two conversation agents to generate a discourse on a certain topic with an assigned personality from the OCEAN framework (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism) as High/Low for each trait. This is followed by using multiple judge agents to infer the original traits assigned to explore prediction consistency, inter-model agreement, and alignment with the assigned personality. Our findings indicate that while LLMs can be guided toward personality-driven dialogue, their ability to maintain personality traits varies significantly depending on the combination of models and discourse settings. These inconsistencies emphasise the challenges in achieving stable and interpretable personality-aligned interactions in LLMs.
The emergence of large language models (LLMs) has revolutionized AI development, yet their resource demands beyond a single cluster or even datacenter, limiting accessibility to well-resourced organizations. Decentralized training has emerged as a promising paradigm to leverage dispersed resources across clusters, datacenters and even regions, offering the potential to democratize LLM development for broader communities. As the first comprehensive exploration of this emerging field, we present decentralized LLM training as a resource-driven paradigm and categorize existing efforts into community-driven and organizational approaches. We further clarify this through: (1) a comparison with related paradigms, (2) characterization of decentralized resources, and (3) a taxonomy of recent advancements. We also provide up-to-date case studies and outline future directions to advance research in decentralized LLM training.
Effective human communication in social settings is contingent on recognizing subtle cues, such as intents or implications. Without such cues, NLP models risk missing social signals, instead relying on surface patterns. We introduce SOCIAL SCAFFOLDS, an automated framework for facilitating generalization across social reasoning tasks by generating rationales that make these social cues explicit. Grounded in narrative modeling principles, we generate task-agnostic rationales that capture different perspectives, i.e., that of the speaker, the listener, and the general world-view. Our experimental suite showcases that providing rationales as augmentations aids task performance for both supervised fine-tuning and in-context learning paradigms. Notably, providing all three rationale types significantly improves cross-task performance in 44% of cases, and inferred speaker intent in 31.3% of cases. We conduct statistical and ablation analyses that show how rationales complement the input text and are used effectively by models.
Instructing language models with user intent requires large instruction datasets, which are only available for a limited set of languages. In this paper, we explore alternatives to conventional instruction adaptation pipelines in low-resource scenarios. We assume a realistic scenario for low-resource languages, where only the following are available: corpora in the target language, existing open-weight multilingual base and instructed backbone LLMs, and synthetically generated instructions sampled from the instructed backbone. We present a comprehensive set of experiments for Basque that systematically study different combinations of these components evaluated on benchmarks and human preferences from 1,680 participants. Our conclusions show that target language corpora are essential, with synthetic instructions yielding robust models, and, most importantly, that using as backbone an instruction-tuned model outperforms using a base non-instructed model. Scaling up to Llama 3.1 Instruct 70B as backbone, our model comes near frontier models of much larger sizes for Basque, without using any Basque instructions. We release code, models, instruction datasets, and human preferences to support full reproducibility in future research on low-resource language adaptation.
Improving Handshape Representations for Sign Language Processing: A Graph Neural Network Approach
PDF ↗Handshapes serve a fundamental phonological role in signed languages, with American Sign Language employing approximately 50 distinct shapes. However, computational approaches rarely model handshapes explicitly, which limits both recognition accuracy and linguistic analysis. We introduce a novel graph neural network that separates temporal dynamics from static handshape configurations. Our approach combines anatomically informed graph structures with contrastive learning to address key challenges in handshape recognition, including subtle inter-class distinctions and temporal variations. We establish the first benchmark for structured handshape recognition in signing sequences, achieving 46% accuracy across 37 handshape classes, compared to 25% for baseline methods.
LiTransProQA: An LLM-based Literary Translation Evaluation Metric with Professional Question Answering
PDF ↗The impact of Large Language Models (LLMs) has extended into literary domains. However, existing evaluation metrics for literature prioritize mechanical accuracy over artistic expression and tend to overrate machine translation as being superior to human translation from experienced professionals. In the long run, this bias could result in an irreversible decline in translation quality and cultural authenticity. In response to the urgent need for a specialized literary evaluation metric, we introduce LITRANSPROQA, a novel, reference-free, LLM-based question-answering framework designed for literary translation evaluation. LITRANSPROQA integrates humans in the loop to incorporate insights from professional literary translators and researchers, focusing on critical elements in literary quality assessment such as literary devices, cultural understanding, and authorial voice. Our extensive evaluation shows that while literary-finetuned XCOMET-XL yields marginal gains, LITRANSPROQA substantially outperforms current metrics, achieving up to 0.07 gain in correlation and surpassing the best state-of-the-art metrics by over 15 points in adequacy assessments. Incorporating professional translator insights as weights further improves performance, highlighting the value of translator inputs. Notably, LITRANSPROQA reaches an adequacy performance comparable to trained linguistic student evaluators, though it still falls behind experienced professional translators. LITRANSPROQA shows broad applicability to open-source models like LLaMA3.3-70b and Qwen2.5-32b, indicating its potential as an accessible and training-free tool for evaluating literary translations that require local processing due to copyright or ethical considerations.
How Much Do LLMs Hallucinate across Languages? On Realistic Multilingual Estimation of LLM Hallucination
PDF ↗In the age of misinformation, hallucination—the tendency of Large Language Models (LLMs) to generate non-factual or unfaithful responses—represents the main risk for their global utility. Despite LLMs becoming increasingly multilingual, the vast majority of research on detecting and quantifying LLM hallucination are (a) English-centric and (b) focus on machine translation (MT) and summarization, tasks that are less common in realistic settings than open information seeking. In contrast, we aim to quantify the extent of LLM hallucination across languages in knowledge-intensive long-form question answering (LFQA). To this end, we train a multilingual hallucination detection model and conduct a large-scale study across 30 languages and 6 open-source LLM families. We start from an English hallucination detection dataset and rely on MT to translate-train a detection model. We also manually annotate gold data for five high-resource languages; we then demonstrate, for these languages, that the estimates of hallucination rates are similar between silver (LLM-generated) and gold test sets, validating the use of silver data for estimating hallucination rates for other languages. For the final rates estimation, we build open-domain QA dataset for 30 languages with LLM-generated prompts and Wikipedia articles as references. Our analysis shows that LLMs, in absolute terms, hallucinate more tokens in high-resource languages due to longer responses, but that the actual hallucination rates (i.e., normalized for length) seems uncorrelated with the sizes of languages’ digital footprints. We also find that smaller LLMs hallucinate more, and significantly, LLMs with broader language support display higher hallucination rates.
Retrieval of previously fact-checked claims is a well-established task, whose automation can assist professional fact-checkers in the initial steps of information verification. Previous works have mostly tackled the task monolingually, i.e., having both the input and the retrieved claims in the same language. However, especially for languages with a limited availability of fact-checks and in case of global narratives, such as pandemics, wars, or international politics, it is crucial to be able to retrieve claims across languages. In this work, we examine strategies to improve the multilingual and crosslingual performance, namely selection of negative examples (in the supervised) and re-ranking (in the unsupervised setting). We evaluate all approaches on a dataset containing posts and claims in 47 languages (283 language combinations). We observe that the best results are obtained by using LLM-based re-ranking, followed by fine-tuning with negative examples sampled using a sentence similarity-based strategy. Most importantly, we show that crosslinguality is a setup with its own unique characteristics compared to the multilingual setup.
Do LLMs Behave as Claimed? Investigating How LLMs Follow Their Own Claims using Counterfactual Questions
PDF ↗Large Language Models (LLMs) require robust evaluation. However, existing frameworks often rely on curated datasets that, once public, may be accessed by newer LLMs. This creates a risk of data leakage, where test sets inadvertently become part of training data, compromising evaluation fairness and integrity. To mitigate this issue, we propose Behave as Claimed (BaC), a novel evaluation framework inspired by counterfactual reasoning. BaC constructs a “what-if” scenario where LLMs respond to counterfactual questions about how they would behave if the input were manipulated. We refer to these responses as claims, which are verifiable by observing the LLMs’ actual behavior when given the manipulated input. BaC dynamically generates and verifies counterfactual questions using various few-shot in-context learning evaluation datasets, reducing their susceptibility to data leakage. Moreover, BaC provides a more challenging evaluation paradigm for LLMs. LLMs must thoroughly understand the prompt, the task, and the consequences of their responses to achieve better performance. We evaluate several state-of-the-art LLMs and find that, while most perform well on the original datasets, they struggle with BaC. This suggests that LLMs usually fail to align their claims with their actual behavior and that high performance on standard datasets may be less stable than previously assumed.
Multimodal knowledge editing is an important method for modifying outdated or incorrect knowledge in Multimodal Large Language Models (MLLMs). However, existing datasets for multimodal knowledge editing lack multi-granularity knowledge. In this paper, we present a more realistic dataset called M2Edit, which includes three distinct types of knowledge: entity, relation, and action. Additionally, existing knowledge editing methods for MLLMs lack the ability to handle multi-granularity knowledge and generalize to multimodal data. To address these limitations, we propose the multimodal knowledge editing method MLE. This approach identifies key knowledge layers within different components and collaboratively edits the various components of MLLMs. As a result, we observe significant improvements in visual generality performance, ranging from 4.8 to 10.8, and achieve the best overall performance on knowledge data of different granularities.
Recent methods address Chinese Spelling Correction (CSC) with either BERT-based models or large language models (LLMs) independently. However, both of them face challenges. BERT-based models are efficient for this task but struggle with limited generalizability to error patterns, thus failing in open-domain CSC. LLMs are advantageous in their extensive knowledge but fall into low efficiency in character-level editing. To address this dilemma, we propose Automatic Corrector Iteration (ACI), a novel model collaboration pipeline to iteratively optimize a BERT-based corrector. This pipeline is free of human annotation, by leveraging an LLM verifier to provide useful signals for the corrector. Experimental results demonstrate that our pipeline consistently improves the model performance across iterations and significantly outperforms existing data augmentation methods, achieving comparable performance with human annotation.
The Good, the Bad and the Constructive: Automatically Measuring Peer Review’s Utility for Authors
PDF ↗Providing constructive feedback to paper authors is a core component of peer review. With reviewers increasingly having less time to perform reviews, automated support systems are required to ensure high reviewing quality, thus making the feedback in reviews useful for authors. To this end, we identify four key aspects of review comments (individual points in weakness sections of reviews) that drive the utility for authors: Actionability, Grounding & Specificity, Verifiability, and Helpfulness. To enable evaluation and development of models assessing review comments, we introduce the RevUtil dataset. We collect 1,430 human-labeled review comments and scale our data with 10k synthetically labeled comments for training purposes. The synthetic data additionally contains rationales, i.e., explanations for the aspect score of a review comment. Employing the RevUtil dataset, we benchmark fine-tuned models for assessing review comments on these aspects and generating rationales. Our experiments demonstrate that these fine-tuned models achieve agreement levels with humans comparable to, and in some cases exceeding, those of powerful closed models like GPT-4o. Our analysis further reveals that machine-generated reviews generally underperform human reviews on our four aspects.
With the widespread adoption of open-source code language models (code LMs), intellectual property (IP) protection has become an increasingly critical concern. While current watermarking techniques have the potential to identify the code LM to protect its IP, they have limitations when facing the more practical and complex demand, i.e., offering the individual user-level tracing in the black-box setting. This work presents CLMTracing, a black-box code LM watermarking framework employing the rule-based watermarks and utility-preserving injection method for user-level model tracing. CLMTracing further incorporates a parameter selection algorithm sensitive to the robust watermark and adversarial training to enhance the robustness against watermark removal attacks. Comprehensive evaluations demonstrate CLMTracing is effective across multiple state-of-the-art (SOTA) code LMs, showing significant harmless improvements compared to existing SOTA baselines and strong robustness against various removal attacks.