论文检索

Zhihao Yao, Yuxuan Gu, Xiachong Feng, Weitao Ma, Bo Li, Xiaocheng Feng, Bing Qin

The privacy leakage problem has become a critical topic in large language models, especially in the scenario of retrieval augmented generation.Current defense methods mitigate privacy leakage but are still suffering from the trade-off between privacy protection and response availability.To address the problem, we propose to explicitly capture the latent leakage tendency of LLM during the generation process, which is able to protect privacy from a more fundamental perspective.In detail, we propose ABack, a training-free mechanism that synchronously monitors the decoding steps, derives the initial leakage intention via modeling mental states, and rewrites the response with privacy awareness. In addition, we construct a new benchmark especially for personally identifiable information, considering the lack of formal privacy datasets.Experiments show that ABack improves privacy by up to 14% over strong baselines against adversarial attacks, avoiding the degradation of response utility.

Zhuoran Li, Rui Xu, Jian Yang, Junnan Liu, Zhijun Chen, Qianren Mao, Hongcheng Guo, Jiaheng Liu, Likang Xiao, Ming LI 等

Model merging is an effective technique for composing the capabilities of a multilingual model and a reasoning model. It has achieved promising generalization in multilingual reasoning tasks by aligning feature spaces of different models. However, the merged single model often fails to address the conflicts between source models, leading to suboptimal performance. In other words, the one-size-fits-all merging strategy may not align with the characteristics of different inputs which may require prioritizing certain models over others. To this end, we propose a Steerable Model Merging (**ST-Merge**) framework to modulate the contribution of each source model. To realize this idea, we introduce a gated cross-attention mechanism to weight or filter the two attended source models in an adaptive manner. Extensive experiments demonstrate that ST-Merge consistently outperforms multiple strong baselines on four multilingual reasoning benchmarks across 21 different languages.

Xingyu Lu, Yumeng Ma, Xiang Zhou, Shengli Gan, Guiying Deng, Yang Wen, Yanbing Liu

The widespread application of LLMs has made MGT detection increasingly important in cyberspace security and governance. The existing detection paradigms mainly focus on statistical likelihood or deep embeddings. However, in complex applications such as short texts, derivative works, and cross-domain content, the discriminative capabilities fragility of these conventional methods increases significantly with the development of LLMs. Conversely, our research reveals that LLMs exhibit inherent style inertia. To address these limitations, this study attempts to synergize stylometrics and semantics for identifying MGT. This approach draws from the forensic perspective of experts who detect human imitation by focusing on stylistic nuances. Based on the above inspiration, we propose Stylometric-Semantic LLM Attribution (SSLA), a framework that extracts model-specific stylistic fingerprints across lexical, syntactic, and structural dimensions. SSLA employs a dual-path attention fusion architecture to dynamically integrate explicit stylistic signals with implicit semantic encodings. Extensive experiments across six LLM families demonstrate that our method achieves state-of-the-art performance. Notably, SSLA achieves a Macro-F1 score of 95.6% on the challenging Wikipedia dataset, demonstrating exceptional robustness and surpassing state-of-the-art baselines like OTBDetector.

Berk Atıl, Namrata Sureddy, Rebecca J. Passonneau

Toxic language includes content that is offensive, abusive, or that promotes harm. Progress in preventing toxic output from large language models (LLMs) is hampered by inconsistent definitions of toxicity. We introduce TRuST, a large-scale dataset that unifies and expands prior resources through a carefully synthesized definition of toxicity, and corresponding annotation scheme. It consists of ∼300k annotations, with high-quality human annotation on ∼11k. To ensure high-quality, we designed a rigorous, multi-stage human annotation process, and evaluated the diversity of the annotators. Then we benchmarked state-of-the-art LLMs and pre-trained models on three tasks: toxicity detection, identification of the target group, and of toxic words. Our results indicate that fine-tuned PLMs outperform LLMs on the three tasks, and that current reasoning models do not reliably improve performance. TRuST constitutes one of the most comprehensive resources for evaluating and mitigating LLM toxicity, and other research in socially-aware and safer language technologies.

Xu Ma, Yitian Zhang, Yun Fu

The recent surge of interest in unified Multimodal Large Language Models (MLLMs) has catalyzed rapid progress toward general-purpose generation and understanding across different modalities. Despite the remarkable advancements, the field lacks a systematic and cohesive framework that connects these developments, revisits the motivations, and situates current trends within a broader landscape. In this survey, we present a comprehensive and in-depth review of unified MLLMs, offering both a methodology taxonomy and unique perspectives on the field. We begin by outlining the foundational concepts and prerequisites for understanding unified MLLMs. We then delve into designs from different aspects, including model architectures, loss functions, alignment techniques, and different representation strategies. Furthermore, we discuss persistent challenges and identify promising directions for future research. By bridging scattered progress and providing a consolidated view, this survey aims to foster a deeper and systematical understanding of unified MLLMs and inspire future innovations in building truly general multimodal intelligence.

Josue Torres-Fonseca, Naihao Deng, Yinpei Dai, Shane Storks, Yichi Zhang, Rada Mihalcea, Casey Kennington, Joyce Chai

Multimodal Large Language Models (MLLMs) are increasingly adopted as autonomous agents in interactive environments, yet their ability to proactively address safety hazards remains insufficient. We introduce SafetyALFRED, built upon the embodied agent benchmark ALFRED, augmented with six categories of real-world kitchen hazards. While existing safety evaluations focus on hazard recognition through disembodied question answering (QA) settings, we evaluate eleven state-of-the-art models from the Qwen, Gemma, and Gemini families on not only hazard recognition, but also active risk mitigation through embodied task planning. Our experimental results reveal a significant alignment gap: while models can accurately recognize hazards in QA settings, average mitigation success rates for these hazards are low in comparison. Our findings demonstrate that static evaluations through QA are insufficient for physical safety, advocating for a paradigm shift toward benchmarks that prioritize multi-step corrective actions in embodied context.

Yichen Xu, Jianzhe Ma, Chuhan Wang, Zhonghao Cao, Liangyu Chen, Wenxuan Wang, Qin Jin

Sports have witnessed growing global enthusiasm in recent years, serving as a vital force for physical health, cultural exchange, social connection, and economic growth. The rapid advancement of large models, particularly (multimodal) large language models (M)LLMs, has demonstrated transformative potential to reshape sports understanding, analysis, and interaction across diverse domains. This paper presents a comprehensive survey of large models in sports, including (i) an overview of tasks and applications across different participant groups; (ii) a detailed analysis of sports-related datasets and benchmarks; and (iii) a critical discussion of current challenges and future directions. Our goal is to establish a foundation for advancing research and practical development of large-model-driven sports intelligence. An open-source GitHub repository is maintained at: https://github.com/Road2Redemption/Awesome_Large_Models_In_Sports1.

Yutao Mou, Zhangchi Xue, Lijun Li, Peiyang Liu, Shikun Zhang, Wei Ye, Jing Shao

While LLM-based agents can interact with environments via invoking external tools, their expanded capabilities also amplify security risks. Monitoring step-level tool invocation behaviors in real time and proactively intervening before unsafe execution is critical for agent deployment, yet remains underexplored. In this work, we first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents. We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning. The model proactively detects unsafe tool invocation actions before execution by reasoning over the interaction history. It assesses request harmfulness and action–attack correlations, producing interpretable and generalizable safety judgments and feedback. Furthermore, We introduce TS-Flow, a guardrail-feedback-driven reasoning framework for LLM agents, which reduces harmful tool invocations of ReAct-style agents by 65% on average and improves benign task completion by approximately 10% under prompt injection attacks.

Yixiao Huang, Lan Zhang, Chaoran Wang

Learning unknown knowledge through ICL and RAG can enhance LLM capabilities in specialized fields. While most research focuses on how to identify and utilize such knowledge, little work examines what factors lead LLMs to trust and adopt it, leaving models prone to errors and harmful content. Grounded in extensive pre-experiments, we design five pairs of trust-enhancing and trust-diminishing transformations on unknown knowledge to experimentally identify the key trust factors. These findings are further substantiated through a detailed theoretical analysis grounded in the epistemological framework of evidentialism. Based on these insights, we challengingly propose a completely unrestricted and fully randomized jailbreak attack that embeds malicious queries within trust-enhanced unknown knowledge. In both defended and undefended scenarios, our method achieves 99% to 100% ASR on all tested LLMs, including the latest GPT-5.1, and becomes SOTA. This attack confirms the trust mechanism and exposes a critical and hard-to-defend security risk. Our conclusions provide valuable guidance for understanding trust mechanism of unknown knowledge and for future research.

Yize Cheng, Arshia Soltani Moakhar, Chenrui Fan, Parsa Hosseini, Kazem Faghih, Zahra Sodagar, Wenxiao Wang, Soheil Feizi

Large language model (LLM) agents are increasingly used to interact with and execute tasks in dynamic environments. However, a critical yet overlooked limitation of these agents is that they, by default, assume a stationary context, failing to account for the real-world time elapsed between messages. We refer to this as "temporal blindness". This limitation hinders decisions about when to invoke tools, leading agents to either over-rely on stale context and skip needed tool calls, or under-rely on it and redundantly repeat tool calls. To study this challenge, we constructed TicToc, a diverse dataset of multi-turn user–agent message trajectories across 76 scenarios, spanning dynamic environments with high, medium, and low time sensitivity. We collected human preferences between "calling a tool" and "directly answering" on each sample, and evaluated how well LLM tool-calling decisions align with human preferences under varying amounts of elapsed time. Our analysis reveals that existing models display poor alignment with human temporal perception, with no models achieving a normalized alignment rate better than 65% when given time stamp information. We also show that naive, prompt-based alignment techniques have limited effectiveness for most models, but specific post-training alignment can be a viable way to align multi-turn LLM tool use with human temporal perception. Our data and findings provide a first step toward understanding and mitigating temporal blindness, offering insights to foster the development of more time-aware and human-aligned agents.

Kaijie Mo, Siddhartha Venkatayogi, Chantal Shaib, Ramez Kouzy, Wei Xu, Byron C Wallace, Junyi Jessy Li

In high-stakes domains like medicine, it may be generally desirable for models to faithfully adhere to the context provided. But what happens if the context does not align with model priors or safety protocols? In this paper, we investigate how LLMs behave and reason when presented with counterfactual (or even adversarial) medical evidence. We first construct MedCounterFact, a counterfactual medical QA dataset that requires the models to answer clinical comparison questions (i.e., judge the efficacy of certain treatments, with evidence consisting of randomized controlled trials provided as context). In MedCounterFact, real-world medical interventions within the questions and evidence are systematically replaced with four types of counterfactual stimuli, ranging from unknown words to toxic substances. Our evaluation across multiple frontier LLMs on MedCounterFact reveals that in the presence of counterfactual evidence, existing models overwhelmingly accept such "evidence" at face value even when it is dangerous or implausible, and provide confident and uncaveated answers. While it may be prudent to draw a boundary between faithfulness and safety, our findings suggest that models arguably overemphasize the former.

Jingxuan Liu, Zhi Qu, Jin Tei, Hidetaka Kamigaito, Lemao Liu, Taro Watanabe

Automatic evaluation metrics are essential for building multilingual translation systems. The common practice of evaluating these systems is averaging metric scores across languages, yet this is suspicious since metrics may suffer from cross-lingual scoring bias, where translations of equal quality receive different scores across languages. This problem has not been systematically studied because no benchmark exists that provides parallel-quality instances across languages, and expert annotation is not realistic. In this work, we propose XQ-MEval, a semi-automatically built dataset covering nine translation directions, to benchmark translation metrics. Specifically, we inject MQM-defined errors into gold translations automatically, filter them by native speakers for reliability, and merge errors to generate pseudo translations with controllable quality. These pseudo translations are then paired with corresponding sources and references to form triplets used in assessing the qualities of translation metrics. Using XQ-MEval, our experiments on nine representative metrics reveal the inconsistency between averaging and human judgment and provide the first empirical evidence of cross-lingual scoring bias. Finally, we propose a normalization strategy derived from XQ-MEval that aligns score distributions across languages, improving the fairness and reliability of multilingual metric evaluation.

Yuqi Kong, Shiyu Liu, Jiaxu Li, Hongtao Liu, Qi Qi, Weiran Shen

Large language models (LLMs) have recently demonstrated strong capability in understanding and simulating humans’ decisions, suggesting a new way to use LLMs as tools to study social systems. We study two-sided-matching markets, such as dating and job matching. Classical matching models assume deterministic, strict preferences, which violate real-world setting. We focus on stable matching under stochastic decision behavior and use LLMs to simulate human-like preferences and probabilistic choice patterns. Based on this, we introduce Expected Blocking Pairs (EBP), a continuous measure to quantify stability that generalizes the classic blocking pair notion. We further propose a Hybrid GS–LLM matching method that integrates the celebrated Gale–Shapley (GS) algorithm with probabilistic acceptance decisions. Experiments show that the proposed hybrid method outperforms classical baselines in terms of stability, suggesting that LLMs provide a principled tool for modeling human decisions and for improving robustness of matching under uncertainty.

Chan-Jan Hsu, Liang-Hsuan Tseng, Yi-Cheng Lin, Yen-Chun Kuo, Ju-Chieh Chou, Kai-Wei Chang, Hung-yi Lee, Carlos Busso

Generative spoken language models pretrained on large-scale raw audio can continue a speech prompt with appropriate content while preserving attributes like speaker and emotion, serving as foundation models for spoken dialogue. In prior literature, these models are often evaluated using “global token perplexity”, which directly applies the text perplexity formulation to speech tokens. However, this practice overlooks fundamental differences between speech and text modalities, possibly leading to an underestimation of the speech characteristics. In this work, we propose a variety of likelihood- and generative-based evaluation methods that serve in place of naive global token perplexity. We demonstrate that the proposed evaluations more faithfully reflect perceived generation quality, as evidenced by stronger correlations with human-rated mean opinion scores (MOS). When assessed under the new metrics, the relative performance landscape of spoken language models is reshaped, revealing a significantly reduced gap between the best-performing model and the human topline. Together, these results suggest that appropriate evaluation is critical for accurately assessing progress in spoken language modeling.

Pinlong Zhao, Zike Ding, Zengshu Ye, Zhou Zhaoting

Multimodal Large Language Models (MLLMs) often suffer from modality bias, where the model disproportionately relies on one modality while neglecting critical information from others. Existing debiasing methods via modality masking create biased responses by completely removing an entire modality, forming an extreme and static training environment. However, real-world multimodal bias often emerges under subtle perturbations (e.g., mild occlusion, noisy instructions), where both modalities are present but the model is tempted to rely on spurious shortcuts. We propose B-APO (Bias-Targeted Adversarial Preference Optimization), which casts debiasing as a bias-targeted min-max game: we generate hard negatives by applying small adversarial perturbations in the latent space to maximally induce language-vision-prior reliance, and then perform preference alignment to enlarge the margin between clean and adversarial responses. This encourages the model to anchor on true cross-modal evidence even under the most adversarial conditions. Extensive experiments on bias and hallucination benchmarks demonstrate that B-APO achieves superior debiasing performance while maintaining general capabilities.

Sriram Padmanabhan, Siyuan Song, Kanishka Misra

Language places subtle constraints on how we make inductive inferences. Developmental evidence by Gelman et al. (2002) has shown children (4 years and older) to differentiate among generic statements ("Bears are daxable"), universally quantified NPs ("all bears are daxable") and indefinite plural NPs ("some bears are daxable") in extending novel properties to a specific member (all > generics > some), suggesting that they represent these types of propositions differently. We test if these subtle differences arise in general purpose statistical learners like Vision Language Models, by replicating the original experiment. On tasking them through a series of precondition tests (robust identification of categories in images and sensitivities to all and some), followed by the original experiment, we find behavioral alignment between models and humans. Post-hoc analyses on their representations revealed that these differences are organized based on inductive constraints and not surface-form differences.

Pinlong Zhao, Huijun Tang, Pengfei Jiao, Mengyang Li

Sampling methods for large language models select candidate tokens based on logit statistics, implicitly assuming that high logits indicate desirable outputs. We identify the Logit Conflation Problem, where a token’s logit aggregates prompt-independent factors, including linguistic fluency and parametric associations, with prompt-relevance. However, only prompt-relevance determines instruction-following quality. We propose SEAL-Sampling (Signal Extraction for Active ReLevance) to isolate this component through attention-weighted attribution. Our framework defines prompt-relevance as the causal effect of prompt content on token logits and establishes attention patterns as an efficient proxy. Experiments on LLaMA-3 demonstrate significant improvements over top-nσ, with gains of 1.8% on AlpacaEval 2.0 and 2.2% on IFEval. Furthermore, attribution scores correlate weakly with raw logits, confirming the extraction of an orthogonal signal. The method is training-free and introduces minimal latency, adding less than 12ms overhead per token.

Tengfei Wen, Xuanang Chen, Ben He, Xiaoliang Cong, Le Sun

Large Language Models (LLMs) struggle with code generation for Ultra Low-Resource Programming Languages (ULRPLs) due to the scarcity of training data. Existing synthetic data generation methods fail in this context, suffering from a severe cold-start problem and resulting in samples that lack diversity. To overcome these challenges, we propose CodeRise, a novel two-stage framework that autonomously generates a high-quality, diverse, and progressively complex curriculum for ULRPLs. The framework first tackles the cold-start and distribution issues by leveraging the full formal syntax of the target language as structural guidance and applying a biased sampling strategy over library modules. Building on this foundation, we fine-tune the model to generate increasingly complex code without explicit syntax input, using an adaptive curriculum and multi-turn self-debugging to progressively improve code quality.We evaluate on two ULRPLs, Tengo and Janet, using migrated HumanEval-Tengo and MBPP-Tengo, as well as our new benchmarks, TengoEval and JanetEval. Experiments show that CodeRise significantly outperforms both training-free and training-based baselines in ultra low-resource environments.

Runxin Cai, Jingtan Wang, Bryan Kian Hsiang Low

In Large Language Model post-training, high-quality data effectively enhances model performance with fine-tuning, highlighting the need to identify high-quality and beneficial fine-tuning data. However, one of the most popular data valuation paradigms, influence function and its variants, are computationally expensive due to their reliance on inverse Hessian-Vector Products (iHVP) computations that scale poorly with increasing model size. To examine whether influence values correlate with efficiently computable intrinsic features, we empirically investigate the distribution of top influential data for the model in fine-tuning, and observe that data with high influence tend to be those with high predictive uncertainty. Yet such highly uncertain samples exhibit a dual nature, which can be either beneficial or detrimental noisy data. Unlike traditional methods that treat uncertainty as a standalone criterion, we introduce a directional indicator to rigorously disentangle these opposing effects. Formally, we propose EULoInf (Entropy-based Uncertainty-aware Lookahead Influence), a computationally efficient valuation framework. By approximating influence via uncertainty and gradient based validation loss lookahead, EULoInf avoids iHVP computation, effectively reducing the iHVP-induced quadratic complexity in model parameters to linear time. We rigorously derive our framework from the influence function. Empirically, it matches or even outperforms prior methods across diverse data valuation tasks and LLM architectures, including mislabel detection and data selection, while reducing computational time and memory usage by over 50%.

Bowen Dong, Wenjun Wang, Xueli Liu, Quanlin Qiu

Understanding whether one event increases or decreases the likelihood of another is critical for real-life applications. Unlike other relationships, promoting and inhibiting relationships capture directional, probabilistic, and context-dependent shifts in event likelihood. A central challenge is to estimate this relative influence from observational data: naive conditional probabilities conflate influence with correlation and are easily distorted by shared contextual confounders. We propose EPIR, a unified framework for estimating promoting and inhibiting relationships from observed event data. EPIR formulates influence as a relative directional effect under comparable contextual conditions, and models event context using : (i) observable history captured and (ii) latent multi-hop propagation mechanisms. EPIR combines context-conditioned predictive evidence with schema-based structural evidence to produce a single signed influence score, where the sign determines promotion versus inhibition. Experiments on real-world datasets show that EPIR outperforms state-of-the-art baselines in accuracy.