Video Large Language Models (VideoLLMs) are increasingly deployed on numerous critical applications, where users rely on auto-generated summaries while casually skimming the video stream. We show that this interaction hides a critical safety gap: if harmful content is embedded in a video, either as full-frame inserts or as small corner patches, state-of-the-art VideoLLMs rarely mention the harmful content in the output, despite its clear visibility to human viewers. A root-cause analysis reveals three compounding design flaws: (1) insufficient temporal coverage resulting from the sparse, uniformly spaced frame sampling used by most leading VideoLLMs, (2) spatial information loss introduced by aggressive token downsampling within sampled frames, and (3) encoder-decoder disconnection, whereby visual cues are only weakly utilized during text generation. Leveraging these insights, we craft three zero-query black-box attacks, aligning with these flaws in the processing pipeline. Our large-scale evaluation across five leading VideoLLMs shows that the harmfulness omission rate exceeds 90\% in most cases. Even when harmful content is clearly present in all frames, these models consistently fail to identify it. These results underscore a fundamental vulnerability in current VideoLLMs' designs and highlight the urgent need for sampling strategies, token compression, and decoding mechanisms that guarantee semantic coverage rather than speed alone.
论文检索
输入标题、作者或关键词，从 11,272 篇学术成果中精准定位
When Safe Unimodal Inputs Collide: Optimizing Reasoning Chains for Cross-Modal Safety in Multimodal Large Language Models
PDF ↗Multimodal Large Language Models (MLLMs) are susceptible to the implicit reasoning risk, wherein innocuous unimodal inputs synergistically assemble into risky multimodal data that produce harmful outputs. We attribute this vulnerability to the difficulty of MLLMs maintaining safety alignment through long-chain reasoning.To address this issue, we introduce Safe-Semantics-but-Unsafe-Interpretation (SSUI), the first dataset featuring interpretable reasoning paths tailored for such a cross-modal challenge.A novel training framework, Safety-aware Reasoning Path Optimization (SRPO), is also designed based on the SSUI dataset to align the MLLM's internal reasoning process with human safety values. Experimental results show that our SRPO-trained models achieve state-of-the-art results on key safety benchmarks, including the proposed Reasoning Path Benchmark (RSBench), significantly outperforming both open-source and top-tier commercial MLLMs.
Your Prompts Are Not Safe: Output-Free Membership Inference via Prompt Vectors in Vision-Language Tuning
PDF ↗Prompt tuning enables Vision-Language Models (VLMs) to efficiently adapt to new tasks through learnable prompt vectors. This naturally raises a question: do these prompts leak private information about their training data? While Membership Inference Attacks (MIAs) can quantify this risk, current methods rely on access to model outputs or internal gradients. This limitation prevents a clear assessment of a prompt’s standalone privacy leakage, particularly in deployment scenarios where such information is inaccessible. In this paper, we propose Prompt Intrinsic Privacy Risk Analyzer (PIPRA) to address this gap. As the first output-free MIA, PIPRA leverages open-source pre-trained VLMs to extract features from both prompts and samples within a shared cross-modal semantic space. By employing a contrastive learning-based feature projector to enhance these representations, PIPRA enables a subsequent discriminator to effectively perform membership inference. Extensive experiments across nine benchmark datasets and multiple VLMs show PIPRA achieves an average AUC of 87.58%, significantly outperforming traditional output-dependent methods (77.05%). These findings reveal that prompts pose a substantially greater privacy risk than previously recognized, highlighting the urgent need for prompt-level privacy protection.
Large language models (LLMs) have significantly transformed natural language understanding and generation, but they raise privacy concerns due to potential exposure of sensitive information. Studies have highlighted the risk of information leakage, where adversaries can extract sensitive information embedded in the prompts. In this work, we introduce a novel private prediction framework for generating high-quality synthetic text with strong privacy guarantees. Our approach leverages the Differential Privacy (DP) framework to ensure worst-case theoretical bounds on information leakage without requiring any fine-tuning of the underlying models. The proposed method performs inference on private records and aggregates the resulting per-token output distributions. This enables the generation of longer and coherent synthetic text while maintaining privacy guarantees. Additionally, we propose a simple blending operation that combines private and public inference to further enhance utility. Empirical evaluations demonstrate that our approach outperforms previous state-of-the-art methods on in-context-learning (ICL) tasks, making it a promising direction for privacy-preserving text generation while maintaining high utility.
Textual open intent classification is crucial for real-world dialogue systems, enabling robust detection of unknown user intents without prior knowledge and contributing to the robustness of the system. While adaptive decision boundary methods have shown great potential by eliminating manual threshold tuning, existing approaches assume isotropic distributions of known classes, restricting boundaries to balls and overlooking distributional variance along different directions. To address this limitation, we propose EliDecide, a novel method that learns ellipsoid decision boundaries with varying scales along different feature directions. First, we employ supervised contrastive learning to obtain a discriminative feature space for known samples. Second, we apply learnable matrices to parameterize ellipsoids as the boundaries of each known class, offering greater flexibility than spherical boundaries defined solely by centers and radii. Third, we optimize the boundaries via a novelly designed dual loss function that balances empirical and open-space risks: expanding boundaries to cover known samples while contracting them against synthesized pseudo-open samples. Our method achieves state-of-the-art performance on multiple text intent benchmarks and further on a question classification dataset. The flexibility of the ellipsoids demonstrates superior open intent detection capability and strong potential for generalization to more text classification tasks in diverse complex open-world scenarios.
Contextual priming, where earlier stimuli covertly bias later judgments, offers an unexplored attack surface for large language models (LLMs). We uncover a contextual priming vulnerability in which the previous response in the dialogue can steer its subsequent behavior toward policy-violating content. While existing jailbreak attacks largely rely on single-turn or multi-turn prompt manipulations, or inject static in-context examples, these methods suffer from limited effectiveness, inefficiency, or semantic drift. We introduce Response Attack (RA), a novel framework that strategically leverages intermediate, mildly harmful responses as contextual primers within a dialogue. By reformulating harmful queries and injecting these intermediate responses before issuing a targeted trigger prompt, RA exploits a previously overlooked vulnerability in LLMs. Extensive experiments across eight state-of-the-art LLMs show that RA consistently achieves significantly higher attack success rates than nine leading jailbreak baselines. Our results demonstrate that the success of RA is directly attributable to the strategic use of intermediate responses, which induce models to generate more explicit and relevant harmful content while maintaining stealth, efficiency, and fidelity to the original query.
Medical Visual Question Answering (Med-VQA) aims to generate accurate answers for clinical questions grounded in medical images, which has attracted increasing research attention due to its potential to streamline diagnostics and reduce clinical burden. Recent advances in Large Vision-Language Models (LVLMs) have shown great promise for Med-VQA, but still suffer from two inference-time issues: (1) attention shift, where the LVLM over-relies on textual priors; and (2) attention dispersion, where it fails to focus on critical diagnostic regions. To tackle these issues, we propose Contrastive Mutual Information Decoding (CMID), a training-free inference-time intervention grounded in information theory for Med-VQA. Concretely, CMID first identifies the Principal Focus Area (PFA) from decoder attention maps, then constructs focus-preserving and focus-excluding views to derive dual contrastive signals that simultaneously amplify salient visual cues and suppress background noise. Crucially, these corrective signals are adaptively scaled by a reliability-gated self-correction mechanism, based on the distributional shift induced by the PFA. Extensive experiments on three Med-VQA benchmarks demonstrate the effectiveness of CMID. Further analyses showcase its robust generalizability across diverse medical architectures and tasks.
S³-MSD: Large Vision-Language Model for Explainable and Generalizable Multi-modal Sarcasm Detection
PDF ↗Multimodal sarcasm detection (MSD) aims to identify sarcasm polarity from diverse modalities (i.e., image–text pairs), a task that has received increasing attention. While significant progress has been made, existing approaches still face two major issues: lack of explainability and weak generalizability. In this paper, we introduce a new large vision–language model (LVLM) dubbed S³-MSD for explainable and generalizable MSD through three key components. For explainability, we develop (1) a self-training paradigm that automatically bootstraps answers with explanations, and (2) a self-calibrating mechanism that rectifies flawed explanations. For generalizability, we design (3) a self-focusing module that amplifies visual semantic entities through preference optimization, thereby mitigating textual over-reliance. Experimental results on both in-distribution and out-of-distribution (OOD) benchmarks demonstrate that S³-MSD consistently outperforms state-of-the-art methods in detection performance. Furthermore, the proposed S³-MSD provides persuasive explanations, as verified by both quantitative metrics and human evaluations.
Large Reasoning Models (LRMs) have demonstrated promising performance in complex tasks. However, the resource-consuming reasoning processes may be exploited by attackers to maliciously occupy the resources of the servers, leading to a crash, like the DDoS attack in cyber. To this end, we propose a novel attack method on LRMs termed ExtendAttack to maliciously occupy the resources of servers by stealthily extending the reasoning processes of LRMs. Concretely, we systematically obfuscate characters within a benign prompt, transforming them into a complex, poly-base ASCII representation. This compels the model to perform a series of computationally intensive decoding sub-tasks that are deeply embedded within the semantic structure of the query itself. Extensive experiments demonstrate the effectiveness of our proposed ExtendAttack. Remarkably, it significantly increases response length and latency, with the former increasing by over 2.7 times for the o3 model on the HumanEval benchmark. Besides, it preserves the original meaning of the query and achieves comparable answer accuracy, showing the stealthiness.
Unsupervised Semantic Discovery via Global and Local Semantic Alignment in Multimodal Clustering
PDF ↗Unsupervised multimodal semantic discovery aims to learn discriminative representations from multimodal data. However, existing methods suffer from two key limitations. First, they only align instances across modalities without modeling semantic-level consistency, which fails to mitigate semantic bias caused by the gaps among feature distributions of multiple modalities. Second, they inevitably generate incorrect negative pairs during contrastive learning, pushing semantically similar samples apart. To address these challenges, we propose GLAD (Global and Local semantic Alignment for unsupervised multimodal semantic Discovery), which aligns multimodal data at both global and local semantic levels. At the global level, GSA integrates multi-modal features into a shared space and employs joint clustering via optimal transport to capture common semantic patterns while mitigating cross-modality semantic bias. At the local level, LSA adaptively weights samples within each cluster based on their semantic importance, alleviating the effect of incorrect negative pairs. Through the joint optimization of GSA and LSA, GLAD effectively captures both the global semantic structure and the local semantic nuances of multimodal data. Extensive experiments on three benchmark datasets demonstrate GLAD significantly outperforms state-of-the-art methods, with an average improvement of 3.22%.
Large Language Models (LLMs) hold promise in automating data analysis tasks, yet open-source models face significant limitations in these kinds of reasoning-intensive scenarios. In this work, we investigate strategies to enhance the data analysis capabilities of open-source LLMs. By curating a seed dataset of diverse, realistic scenarios, we evaluate models across three dimensions: data understanding, code generation, and strategic planning. Our analysis reveals three key findings: (1) Strategic planning quality serves as the primary determinant of model performance; (2) Interaction design and task complexity significantly influence reasoning capabilities; (3) Data quality demonstrates a greater impact than diversity in achieving optimal performance. We leverage these insights to develop a data synthesis methodology, demonstrating significant improvements in open-source LLMs' analytical reasoning capabilities.
REAP: Enhancing RAG with Recursive Evaluation and Adaptive Planning for Multi-Hop Question Answering
PDF ↗Retrieval-augmented generation (RAG) has been extensively employed to mitigate hallucinations in large language models (LLMs). However, existing methods for multi-hop reasoning tasks often lack global planning, increasing the risk of falling into local reasoning impasses. Insufficient exploitation of retrieved content and the neglect of latent clues fail to ensure the accuracy of reasoning outcomes. To overcome these limitations, we propose **R**ecursive **E**valuation and **A**daptive **P**lanning (REAP), whose core idea is to explicitly maintain structured sub-tasks and facts related to the current task through the Sub-task Planner (SP) and Fact Extractor (FE) modules. SP maintains a global perspective, guiding the overall reasoning direction and evaluating the task state based on the outcomes of FE, enabling dynamic optimization of the task-solving trajectory. FE performs fine-grained analysis over retrieved content to extract reliable answers and clues. These two modules incrementally enrich a logically coherent representation of global knowledge, enhancing the reliability and the traceability of the reasoning process. Furthermore, we propose a unified task paradigm design that enables effective multi-task fine-tuning, significantly enhancing SP's performance on complex, data-scarce tasks. We conduct extensive experiments on multiple public multi-hop datasets, and the results demonstrate that our method significantly outperforms existing RAG methods in both in-domain and out-of-domain settings, validating its effectiveness in complex multi-hop reasoning tasks.
Beyond Chains: Bridging Large Language Models and Knowledge Bases in Complex Question Answering
PDF ↗Knowledge Base Question Answering (KBQA) aims to answer natural language questions using structured knowledge from KBs. While LLM-only approaches offer generalization, they suffer from outdated knowledge, hallucinations, and lack of transparency. Chain-based KG-RAG methods address these issues by incorporating external KBs, but are limited to simple chain-structured questions due to the absence of planning and logical structuring. Inspired by semantic parsing methods, we propose PDRR: a four-stage framework consisting of Predict, Decompose, Retrieve, and Reason. Our method first predicts the question type and decomposes the question into structured triples. Then retrieves relevant information from KBs and guides the LLM as an agent to reason over and complete the decomposed triples. Experimental results show that our proposed KBQA model, PDRR, consistently outperforms existing methods across different LLM backbones and achieves superior performance on various types of questions.
Large Reasoning Language Models (LRMs) have recently shown remarkable performance in complex reasoning tasks, but their extensive reasoning chains incur substantial computational overhead. To address this challenge, we propose Outlier-aware Reasoning Conciseness Adaptive Merge (ORCA), a novel plug-and-play model merging framework that leverages outlier activation patterns to fuse base models with reasoning models. Our ORCA introduces three key innovations: (1) adaptive alignment that reduces conflicts between disparate activation patterns during merging, (2) outlier-guided allocation that assigns merging coefficients proportional to each layer's reasoning importance as indicated by outlier concentrations, and (3) dynamic probe-based adjustment that adapts merging coefficients during inference based on input-specific activation characteristics. These strategies allow seamless integration into existing merging pipelines while creating unified models that maintain reasoning accuracy with significantly reduced response verbosity. Comprehensive evaluation across six benchmarks using Qwen and LLaMA models shows ORCA reduces average response length by 55% while improving accuracy by 2.4∼5.7% over existing methods.
Graph-based retrieval-augmented generation (RAG) enables large language models (LLMs) to incorporate structured knowledge via graph retrieval as contextual input, enhancing more accurate and context-aware reasoning. We observe that for different queries, it could retrieve similar subgraphs as prompts, and thus we propose SubGCache, which aims to reduce inference latency by reusing computation across queries with similar structural prompts (i.e., subgraphs). Specifically, SubGCache clusters queries based on subgraph embeddings, constructs a representative subgraph for each cluster, and pre-computes the key-value (KV) cache of the representative subgraph. For each query with its retrieved subgraph within a cluster, it reuses the pre-computed KV cache of the representative subgraph of the cluster without computing the KV tensors again for saving computation. Extensive experiments on three datasets across multiple LLM backbones and graph-based RAG frameworks demonstrate that SubGCache consistently reduces inference latency with comparable and even improved generation quality, achieving up to 6.68x reduction in time-to-first-token (TTFT).
In-Token Rationality Optimization: Towards Accurate and Concise LLM Reasoning via Self-Feedback
PDF ↗Training Large Language Models (LLMs) for chain-of-thought reasoning presents a significant challenge: supervised fine-tuning on a single "golden" rationale hurts generalization as it penalizes equally valid alternatives, whereas reinforcement learning with verifiable rewards struggles with credit assignment and prohibitive computational cost. To tackle these limitations, we introduce InTRO (In-Token Rationality Optimization), a new framework that enables both token-level exploration and self-feedback for accurate and concise reasoning. Instead of directly optimizing an intractable objective over all valid reasoning paths, InTRO leverages correction factors—token-wise importance weights estimated by the information discrepancy between the generative policy and its answer-conditioned counterpart, for informative next-token selection. This approach allows the model to perform token-level exploration and receive self-generated feedback within a single forward pass, ultimately encouraging accurate and concise rationales. Across six math-reasoning benchmarks, InTRO consistently outperforms other baselines, raising solution accuracy by up to 20% relative to the base model. Its chains of thought are also notably more concise, exhibiting reduced verbosity. Beyond this, InTRO enables cross-domain transfer, successfully adapting to out-of-domain reasoning tasks that extend beyond the realm of mathematics, demonstrating robust generalization.
Effective customer support requires not only accurate problem-solving but also structured and empathetic communication aligned with professional standards. However, existing dialogue datasets often lack strategic guidance, and real-world service data is difficult to access and annotate. To address this, we introduce the task of Customer Support Conversation (CSC), aimed at training customer service supporters to respond using well-defined support strategies. We propose a structured CSC framework grounded in COPC guidelines, defining five conversational stages and twelve strategies to guide high-quality interactions. Based on this, we construct CSConv, an evaluation dataset of 1,855 real-world customer–agent conversations rewritten using LLMs to reflect deliberate strategy use, and annotated accordingly. Additionally, we develop a role-playing approach that simulates strategy-rich conversations using LLM-powered roles aligned with the CSC framework, resulting in the training dataset RoleCS. Experiments show that fine-tuning strong LLMs on RoleCS significantly improves their ability to generate high-quality, strategy-aligned responses on CSConv. Human evaluations further confirm gains in problem resolution.
Large Language Model-based Multi-Agent Systems (LLM-based MAS), where multiple LLM agents collaborate to solve complex tasks, have shown impressive performance in many areas. However, MAS are typically distributed across different devices or environments, making them vulnerable to perturbations such as agent failures. While existing works have studied the adversarial attacks and corresponding defense strategies, they mainly focus on reactively detecting and mitigating attacks after they occur rather than proactively designing inherently resilient systems. In this work, we study the resilience of LLM-based MAS under perturbations and find that both the communication topology and prompt design significantly influence system resilience. Motivated by these findings, we propose ResMAS: a two-stage framework for enhancing MAS resilience. First, we train a reward model to predict the MAS’s resilience, based on which we train a topology generator to automatically design resilient topology for specific tasks through reinforcement learning. Second, we introduce a topology-aware prompt optimization method that refines each agent’s prompt based on its connections and interactions with other agents. Extensive experiments across a range of tasks show that our approach substantially improves MAS resilience under various constraints. Moreover, our framework demonstrates strong generalization ability to new tasks and models, highlighting its potential for building resilient MASs.
Comprehension of ancient texts plays an important role in archaeology and understanding of Chinese history and civilization. The rapid development of large language models needs benchmarks that can evaluate their comprehension of ancient characters. Existing Chinese benchmarks are mostly targeted at modern Chinese and transmitted documents in ancient Chinese, but the part of excavated documents in ancient Chinese is not covered. To meet this need, we propose the AncientBench, which aims to evaluate the comprehension of ancient characters, especially in the scenario of excavated documents. The AncientBench is divided into four dimensions, which correspond to the four competencies of ancient character comprehension: glyph comprehension, pronunciation comprehension, meaning comprehension, and contextual comprehension. The benchmark also contains ten tasks, including radical, phonetic radical, homophone, cloze, translation, and more, providing a comprehensive framework for evaluation. We convened archaeological researchers to conduct experimental evaluations, proposed an ancient model as baseline, and conducted extensive experiments on the currently best-performing large language models. The experimental results reveal the great potential of large language models in ancient textual scenarios as well as the gap with humans. Our research aims to promote the development and application of large language models in the field of archaeology and ancient Chinese language.
Existing large language models (LLMs) occasionally generate plausible yet factually incorrect responses, known as hallucinations. Two main approaches have been proposed to mitigate hallucinations: retrieval-augmented language models (RALMs) and refusal post-training. However, current research predominantly focuses on their individual effectiveness while overlooking the evaluation of the refusal capability of RALMs. Ideally, if RALMs know when they do not know, they should refuse to answer. In this study, we ask the fundamental question: Do RALMs know when they don’t know? Specifically, we investigate three questions. First, are RALMs well calibrated with respect to different internal and external knowledge states? We examine the influence of various factors. Contrary to expectations, when all retrieved documents are irrelevant, RALMs still tend to refuse questions they could have answered correctly. Next, given the model's pronounced over-refusal behavior, we raise a second question: How does a RALM's refusal ability align with its calibration quality? Our results show that the over-refusal problem can be mitigated through in-context fine-tuning. However, we observe that improved refusal behavior does not necessarily imply better calibration or higher overall accuracy. Finally, we ask: Can we combine refusal-aware RALMs with uncertainty-based answer abstention to mitigate over-refusal? We develop a simple yet effective refusal mechanism for refusal-post-trained RALMs that improves their overall answer quality by balancing refusal and correct answers. Our study provides a more comprehensive understanding of the factors influencing RALM behavior. Meanwhile, we emphasize that uncertainty estimation for RALMs remains an open problem deserving deeper investigation.