论文检索

Sergey Pletenev, Maria Marina, Nikolay Ivanov, Daria Galimzianova, Nikita Krayko, Mikhail Salnikov, Vasily Konovalov, Alexander Panchenko, Viktor Moskvoretskii

Large Language Models (LLMs) often hallucinate in question answering (QA) tasks. A key yet underexplored factor contributing to this is the temporality of questions – whether they are evergreen (answers remain stable over time) or mutable (answers change). In this work, we introduce EverGreenQA, the first multilingual QA dataset with evergreen labels, supporting both evaluation and training. Using EverGreenQA, we benchmark 12 modern LLMs to assess whether they encode question temporality explicitly (via verbalized judgments) or implicitly (via uncertainty signals). We also train EG-E5, a lightweight multilingual classifier that achieves SoTA performance on this task. Finally, we demonstrate the practical utility of evergreen classification across three applications: improving self-knowledge estimation, filtering QA datasets, and explaining GPT-4o’s retrieval behavior.

Hao Fang, Jiawei Kong, Tianqu Zhuang, Yixiang Qiu, Kuofeng Gao, Bin Chen, Shu-Tao Xia, Yaowei Wang, Min Zhang

The misuse of large language models (LLMs), such as academic plagiarism, has driven the development of detectors to identify LLM-generated texts. To bypass these detectors, paraphrase attacks have emerged to purposely rewrite these texts to evade detection. Despite the success, existing methods require substantial data and computational budgets to train a specialized paraphraser, and their attack efficacy greatly reduces when faced with advanced detection algorithms. To address this, we propose Contrastive Paraphrase Attack (CoPA), a training-free method that effectively deceives text detectors using off-the-shelf LLMs. The first step is to carefully craft instructions that encourage LLMs to produce more human-like texts. Nonetheless, we observe that the inherent statistical biases of LLMs can still result in some generated texts carrying certain machine-like attributes that can be captured by detectors. To overcome this, CoPA constructs an auxiliary machine-like word distribution as a contrast to the human-like distribution generated by the LLM. By subtracting the machine-like patterns from the human-like distribution during the decoding process, CoPA is able to produce sentences that are less discernible by text detectors. Our theoretical analysis suggests the superiority of the proposed attack. Extensive experiments validate the effectiveness of CoPA in fooling text detectors across various scenarios.

Yuezhang Peng, Yuxin Liu, Fei Wen, Xie Chen

Fine-tuning pre-trained large language models (LLMs) on downstream tasks has achieved significant success across various domains. However, as model sizes grow, traditional first-order fine-tuning algorithms incur substantial memory overhead due to the need for activation storage for back-propagation (BP). The BP-free Memory-Efficient Zeroth-Order Optimization (MeZO) method estimates gradients through finite differences, avoiding the storage of activation values, and has been demonstrated as a viable approach for fine-tuning large language models. This work proposes the Multiple-query Memory Efficient Zeroth-Order (MUZO) method, which is based on variance-reduced multiple queries to obtain the average of gradient estimates. When combined with Adam optimizer, MUZO-Adam demonstrates superior performance in fine-tuning various LLMs. Furthermore, we provide theoretical guarantees for the convergence of the MUZO-Adam optimizer. Extensive experiments empirically demonstrate that MUZO-Adam converges better than MeZO-SGD and achieves near first-order optimizer performance on downstream classification, multiple-choice, and generation tasks.

Michael Sejr Schlichtkrull

When AI agents retrieve and reason over external documents, adversaries can manipulate the data they receive to subvert their behaviour. Previous research has studied indirect prompt injection, where the attacker injects malicious instructions. We argue that injection of instructions is not necessary to manipulate agents – attackers could instead supply biased, misleading, or false information. We term this an *attack by content*. Existing defenses, which focus on detecting hidden commands, are ineffective against attacks by content. To defend themselves and their users, agents must critically evaluate retrieved information, corroborating claims with external evidence and evaluating source trustworthiness. We argue that this is analogous to an existing NLP task, automated fact-checking, which we propose to repurpose as a cognitive self-defense tool for agents.

Yu Zhang, Dong Guo, Fang Wu, Guoliang Zhu, Dian Ding, Yiming Zhang

Large Language Models (LLMs) with extended context lengths face significant computational challenges during the pre-filling phase, primarily due to the quadratic complexity of self-attention. Existing methods typically employ dynamic pattern matching and block-sparse low-level implementations. However, their reliance on local information for pattern identification fails to capture global contexts, and the coarse granularity of blocks leads to persistent internal sparsity, resulting in suboptimal accuracy and efficiency. To address these limitations, we propose AnchorAttention, a difference-aware, dynamic sparse attention mechanism that efficiently identifies critical attention regions at a finer stripe granularity while adapting to global contextual information, achieving superior speed and accuracy. AnchorAttention comprises three key components: (1) Pattern-based Anchor Computation, leveraging the commonalities present across all inputs to rapidly compute a set of near-maximum scores as anchor; (2) Difference-aware Stripe Sparsity Identification, performing difference-aware comparisons with anchor to quickly obtain discrete coordinates of significant regions in a stripe-like sparsity pattern; (3) Fine-grained Sparse Computation, replacing the traditional contiguous loading strategy with a discrete key-value loading approach to maximize sparsity rates while preserving hardware computational potential. Additionally, we integrate the identification strategy into a single operator to maximize parallelization potential. With its finer-grained sparsity strategy, AnchorAttention achieves higher sparsity rates at the same recall level, significantly reducing computation time. Compared to previous state-of-the-art methods, at a text length of 128k, it achieves a speedup of 1.44\times while maintaining higher recall rates.

Xinhao Huang, Zhibo Ren, Yipeng Yu, Ying Zhou, Zulong Chen, Zeyi Wen

In long structured document retrieval, existing methods typically fine-tune pre-trained language models (PLMs) using contrastive learning on datasets lacking explicit structural information. This practice suffers from two critical issues: 1) current methods fail to leverage structural features and element-level semantics effectively, and 2) the lack of datasets containing structural metadata. To bridge these gaps, we propose SEAL, a novel contrastive learning framework. It leverages structure-aware learning to preserve semantic hierarchies and masked element alignment for fine-grained semantic discrimination. Furthermore, we release StructDocRetrieval, a long structured document retrieval dataset with rich structural annotations. Extensive experiments on both the released and industrial datasets across various modern PLMs, and online A/B testing demonstrate consistent improvements, boosting NDCG@10 from 73.96% to 77.84% on BGE-M3. The resources are available at https://github.com/xinhaoH/SEAL.

Arghodeep Nandi, Megha Sundriyal, Euna Mehnaz Khan, Jikai Sun, Emily K. Vraga, Jaideep Srivastava, Tanmoy Chakraborty

Misinformation remains one of the most significant issues in the digital age. While automated fact-checking has emerged as a viable solution, most current systems are limited to evaluating factual accuracy. However, the detrimental effect of misinformation transcends simple falsehoods; it takes advantage of how individuals perceive, interpret, and emotionally react to information. This underscores the need to move beyond factuality and adopt more human-centered detection frameworks. In this survey, we explore the evolving interplay between traditional fact-checking approaches and psychological concepts such as cognitive biases, social dynamics, and emotional responses. By analyzing state-of-the-art misinformation detection systems through the lens of human psychology and behavior, we reveal critical limitations of current methods and identify opportunities for improvement. Additionally, we outline future research directions aimed at creating more robust and adaptive frameworks, such as neuro-behavioural models that integrate technological factors with the complexities of human cognition and social influence. These approaches offer promising pathways to more effectively detect and mitigate the societal harms of misinformation.

Yingming Wang, Pepa Atanasova

With the rapid development of Large Language Models (LLMs), Natural Language Explanations (NLEs) have become increasingly important for understanding model predictions. However, these explanations often fail to faithfully represent the model’s actual reasoning process. While existing work has demonstrated that LLMs can self-critique and refine their initial outputs for various tasks, this capability remains unexplored for improving explanation faithfulness. To address this gap, we introduce Self-critique and Refinement for Natural Language Explanations (SR-NLE), a framework that enables models to improve the faithfulness of their own explanations – specifically, post-hoc NLEs – through an iterative critique and refinement process without external supervision. Our framework leverages different feedback mechanisms to guide the refinement process, including natural language self-feedback and, notably, a novel feedback approach based on feature attribution that highlights important input words. Our experiments across three datasets and four state-of-the-art LLMs demonstrate that SR-NLE significantly reduces unfaithfulness rates, with our best method achieving an average unfaithfulness rate of 36.02%, compared to 54.81% for baseline – an absolute reduction of 18.79%. These findings reveal that the investigated LLMs can indeed refine their explanations to better reflect their actual reasoning process, requiring only appropriate guidance through feedback without additional training or fine-tuning.

Nisrine Rair, Alban Goupil, Valeriu Vrabie, Emmanuel Chochoy

Language models are often evaluated with scalar metrics like accuracy, but such measures fail to capture how models internally represent ambiguity, especially when human annotators disagree. We propose a topological perspective to analyze how fine-tuned models encode ambiguity and more generally instances.Applied to RoBERTa-Large on the MD-Offense dataset, Mapper, a tool from topological data analysis, reveals that fine-tuning restructures embedding space into modular, non-convex regions aligned with model predictions, even for highly ambiguous cases. Over 98% of connected components exhibit \geq 90\% prediction purity, yet alignment with ground-truth labels drops in ambiguous data, surfacing a hidden tension between structural confidence and label uncertainty.Unlike traditional tool such as PCA or UMAP, Mapper captures this geometry directly uncovering decision regions, boundary collapses, and overconfident clusters. Our findings position Mapper as a powerful diagnostic tool for understanding how models resolve ambiguity. Beyond visualization, it also enables topological metrics that may inform proactive modeling strategies in subjective NLP tasks.

Yizheng Sun, Hao Li, Chang Xu, Hongpeng Zhou, Chenghua Lin, Riza Batista-Navarro, Jingyuan Sun

Vision-Language Models (VLMs) are powerful yet computationally intensive for widespread practical deployments. To address such challenge without costly re-training, post-training acceleration techniques like quantization and token reduction are extensively explored. However, current acceleration evaluations primarily target minimal overall performance degradation, overlooking a crucial question: does the accelerated model still give the same answers to the same questions as it did before acceleration? This is vital for stability-centered industrial applications where consistently correct answers for specific, known situations are paramount, such as in AI-based disease diagnosis. We systematically investigate this for accelerated VLMs, testing four leading models (LLaVA-1.5, LLaVA-Next, Qwen2-VL, Qwen2.5-VL) with eight acceleration methods on ten multi-modal benchmarks. Our findings are stark: despite minimal aggregate performance drops, accelerated models changed original answers up to 20% of the time. Critically, up to 6.5% of these changes converted correct answers to incorrect. Input perturbations magnified these inconsistencies, and the trend is confirmed by case studies with the medical VLM LLaVA-Med. This research reveals a significant oversight in VLM acceleration, stressing an urgent need for instance-level stability checks to ensure trustworthy real-world deployment.

Jiawei Chen, Xinyan Guan, Qianhao Yuan, Guozhao Mo, Weixiang Zhou, Yaojie Lu, Hongyu Lin, Ben He, Le Sun, Xianpei Han

Current instruction data synthesis methods primarily focus on single-turn instructions and often neglect cross-turn coherence, resulting in context drift and reduced task completion rates in extended conversations. To address this limitation, we propose Skeleton-Guided Multi-Turn Dialogue Generation, a framework that constrains multi-turn instruction synthesis by explicitly modeling human conversational intent. It operates in two stages: (1) Intent Modeling, which captures the global structure of human dialogues by assigning each conversation to one of nine well-defined intent trajectories, ensuring a coherent and goal-oriented information flow; and (2) Skeleton Generation, which constructs a structurally grounded sequence of user queries aligned with the modeled intent, thereby serving as a scaffold that constrains and guides the downstream instruction synthesis process. Based on this process, we construct ConsistentChat, a multi-turn instruction dataset with approximately 15,000 multi-turn conversations and 224,392 utterances. Experiments on the Light, Topdial, and MT-Eval benchmarks show that models fine-tuned on ConsistentChat achieve a 20–30% improvement in chat consistency and up to a 15% increase in task success rate, significantly outperforming models trained on existing single-turn and multi-turn instruction datasets.

Ahmet Yavuz Uluslu, Tannon Kew, Tilia Ellendorff, Gerold Schneider, Rico Sennrich

Large language models (LLMs) often achieve high performance in native language identification (NLI) benchmarks by leveraging superficial contextual clues such as names, locations, and cultural stereotypes, rather than the underlying linguistic patterns indicative of native language (L1) influence. To improve robustness, previous work has instructed LLMs to disregard such clues. In this work, we demonstrate that such a strategy is unreliable and model predictions can be easily altered by misleading hints. To address this problem, we introduce an agentic NLI pipeline inspired by forensic linguistics, where specialized agents accumulate and categorize diverse linguistic evidence before an independent final overall assessment. In this final assessment, a goal-aware coordinating agent synthesizes all evidence to make the NLI prediction. On two benchmark datasets, our approach significantly enhances NLI robustness against misleading contextual clues and performance consistency compared to standard prompting methods.

Min Hyuk Kim, Changheon Kim, Seok Bong Yoo

Medical visual question answering (VQA) and federated learning (FL) have emerged as vital approaches for enabling privacy-preserving, collaborative learning across clinical institutions. However, both these approaches face significant challenges in cross-modal FL scenarios, where each client possesses unpaired images from only one modality. To address this limitation, we propose X-FLoRA, a cross-modal FL framework that uses modality-expert low-rank adaptation (LoRA) for medical VQA. Specifically, X-FLoRA enables the synthesis of images from one modality to another without requiring data sharing between clients. This is achieved by training a backward translation model within a federated asymmetric translation scheme that integrates clinical semantics from textual data. Additionally, X-FLoRA introduces modality-expert LoRA, which fine-tunes separate LoRA modules to strengthen modality-specific representations in the VQA task. The server aggregates the trained backward translation models and fine-tuned LoRA modules using discriminator quality scores and expert-aware weighting, which regulate the relative contributions from different clients. Experiments were conducted on VQA datasets encompassing different medical modalities, and the results demonstrate that X-FLoRA outperforms existing FL methods in terms of VQA performance.

Jonghyun Hong, Sungyoon Lee

Attention-based language models commonly rely on the softmax function to convert attention logits into probability distributions. However, this softmax re-weighting can lead to *attention entropy collapse*, in which attention disproportionately concentrates on a single token, ultimately causing training instability. In this work, we identify the high *variance sensitivity* of softmax as a primary cause of this collapse. We show that *entropy-stable* attention methods, which either control or are insensitive to the variance of attention logits, can prevent entropy collapse and enable more stable training. We provide empirical evidence of this effect in both large language models (LLMs) and a small Transformer model composed solely of self-attention and support our findings with theoretical analysis. Moreover, we identify that the concentration of attention probabilities increases the probability matrix norm, leading to the gradient exploding.

Hanbo Huang, Yihan Li, Bowen Jiang, Bo Jiang, Lin Liu, Zhuotao Liu, Ruoyu Sun, Shiyu Liang

Privacy-sensitive users require deploying large language models (LLMs) within their own infrastructure (on-premises) to safeguard private data and enable customization. However, vulnerabilities in local environments can lead to unauthorized access and potential model theft. To address this, prior research on small models has explored securing only the output layer within hardware-secured devices to balance model confidentiality and customization. Yet this approach fails to protect LLMs effectively. In this paper, we discover that (1) query-based distillation attacks targeting the secured top layer can produce a functionally equivalent replica of the victim model; (2) securing the same number of layers, bottom layers before a transition layer provide stronger protection against distillation attacks than top layers, with comparable effects on customization performance; and (3) the number of secured layers creates a trade-off between protection and customization flexibility. Based on these insights, we propose SOLID, a novel deployment framework that secures a few bottom layers in a secure environment and introduces an efficient metric to optimize the trade-off by determining the ideal number of hidden layers. Extensive experiments on five models (1.3B to 70B parameters) demonstrate that SOLID outperforms baselines, achieving a better balance between protection and downstream customization.

Houxing Ren, Zimu Lu, Weikang Shi, Haotian Hou, Yunqiao Yang, Ke Wang, Aojun Zhou, Junting Pan, Mingjie Zhan, Hongsheng Li

The code generation capabilities of Large Language Models (LLMs) have advanced applications like tool invocation and problem-solving. However, improving performance in code-related tasks remains challenging due to limited training data that is verifiable with accurate test cases. While Direct Preference Optimization (DPO) has shown promise, existing methods for generating test cases still face limitations. In this paper, we propose a novel approach that splits code snippets into smaller, granular blocks, creating more diverse DPO pairs from the same test cases. Additionally, we introduce the Abstract Syntax Tree (AST) splitting and curriculum training method to enhance the DPO training. Our approach demonstrates significant improvements in code generation tasks, as validated by experiments on benchmark datasets such as HumanEval (+), MBPP (+), APPS, LiveCodeBench, and BigCodeBench. Code and data are available at https://github.com/SenseLLM/StructureCoder.

Tatiana Anikina, Jan Cegin, Jakub Simko, Simon Ostermann

Large Language Models (LLMs) are increasingly used to generate synthetic textual data for training smaller specialized models. However, a comparison of various generation strategies for low-resource language settings is lacking. While various prompting strategies have been proposed—such as demonstrations, label-based summaries, and self-revision—their comparative effectiveness remains unclear, especially for low-resource languages. In this paper, we systematically evaluate the performance of these generation strategies and their combinations across 11 typologically diverse languages, including several extremely low-resource ones. Using three NLP tasks and four open-source LLMs, we assess downstream model performance on generated versus gold-standard data. Our results show that strategic combinations of generation methods — particularly target-language demonstrations with LLM-based revisions — yield strong performance, narrowing the gap with real data to as little as 5% in some settings. We also find that smart prompting techniques can reduce the advantage of larger LLMs, highlighting efficient generation strategies for synthetic data generation in low-resource scenarios with smaller models.

Tarun Tater, Diego Frassinelli, Sabine Schulte im Walde

Abstract concepts like mercy and peace often lack clear visual grounding, and thus challenge humans and models to provide suitable image representations. To address this challenge, we introduce AbsVis – a dataset of 675 images annotated with 14,175 concept–explanation attributions from humans and two Vision-Language Models (VLMs: Qwen and LLaVA), where each concept is accompanied by a textual explanation. We compare human and VLM attributions in terms of diversity, abstractness, and alignment, and find that humans attribute more varied concepts. AbsVis also includes 2,680 human preference judgments evaluating the quality of a subset of these annotations, showing that overlapping concepts (attributed by both humans and VLMs) are most preferred. Explanations clarify and strengthen the perceived attributions, both from humans and VLMs. Explanations clarify and strengthen the perceived attributions, both from human and VLMs. Finally, we show that VLMs can approximate human preferences and use them to fine-tune VLMs via Direct Preference Optimization (DPO), yielding improved alignments with preferred concept–explanation pairs.

Yini Wang, Xian Zhou, Shengan Zheng, Linpeng Huang, Zhunchen Luo, Wei Luo, Xiaoying Bai

Understanding the underlying argumentative flow in analytic argumentative writing is essential for discourse comprehension, especially in complex argumentative discourse such as think-tank commentary. However, existing structure modeling approaches often rely on surface-level topic segmentation, failing to capture the author’s rhetorical intent and reasoning process. To address this limitation, we propose a Question-Focus discourse structuring framework that explicitly models the underlying argumentative flow by anchoring each argumentative unit to a guiding question (reflecting the author’s intent) and a set of attentional foci (highlighting analytical pathways). To assess its effectiveness, we introduce an argument reconstruction task in which the modeled discourse structure guides both evidence retrieval and argument generation. We construct a high-quality dataset comprising 600 authoritative Chinese think-tank articles for experimental analysis. To quantitatively evaluate performance, we propose two novel metrics: (1) Claim Coverage, measuring the proportion of original claims preserved or similarly expressed in reconstructions, and (2) Evidence Coverage, assessing the completeness of retrieved supporting evidences. Experimental results show that our framework uncovers the author’s argumentative logic more effectively and offers better structural guidance for reconstruction, yielding up to a 10% gain in claim coverage and outperforming strong baselines across both curated and LLM-based metrics.

Minh Duc Bui, Carolin Holtermann, Valentin Hofmann, Anne Lauscher, Katharina von der Wense

Dialects represent a significant component of human culture and are found across all regions of the world. In Germany, more than 40% of the population speaks a regional dialect (Adler and Hansen, 2022). However, despite cultural importance, individuals speaking dialects often face negative societal stereotypes. We examine whether such stereotypes are mirrored by large language models (LLMs). We draw on the sociolinguistic literature on dialect perception to analyze traits commonly associated with dialect speakers. Based on these traits, we assess the dialect naming bias and dialect usage bias expressed by LLMs in two tasks: association task and decision task. To assess a model’s dialect usage bias, we construct a novel evaluation corpus that pairs sentences from seven regional German dialects (e.g., Alemannic and Bavarian) with their standard German counterparts. We find that: (1) in the association task, all evaluated LLMs exhibit significant dialect naming and dialect usage bias against German dialect speakers, reflected in negative adjective associations; (2) all models reproduce these dialect naming and dialect usage biases in their decision making; and (3) contrary to prior work showing minimal bias with explicit demographic mentions, we find that explicitly labeling linguistic demographics—German dialect speakers—amplifies bias more than implicit cues like dialect usage.