arXiv subjects

Yinpeng Dong

Publications and source records attributed to Yinpeng Dong.

At least 37 records · Page 2Linked to original sources

arXiv2026

MemPot: Defending Against Memory Extraction Attack with Optimized Honeypots

Large Language Model (LLM)-based agents employ external and internal memory systems to handle complex, goal-oriented tasks, yet this exposes them to severe extraction attacks, and effective defenses remain lacking. In this paper, we propose MemPot, the first theoretically verified defense framework against memory extraction attacks by injecting optimized honeypots into the memory. Through a two-stage optimization process, MemPot generates trap documents that maximize the retrieval probability for attackers while remaining inconspicuous to benign users. We model the detection process as Wald's Sequential Probability Ratio Test (SPRT) and theoretically prove that MemPot achieves a lower average number of sampling rounds compared to optimal static detectors. Empirically, MemPot significantly outperforms state-of-the-art baselines, achieving a 50% improvement in detection AUROC and an 80% increase in True Positive Rate under low False Positive Rate constraints. Furthermore, our experiments confirm that MemPot incurs zero additional online inference latency and preserves the agent's utility on standard tasks, verifying its superiority in safety, harmlessness, and efficiency.

cs.CR↗

arXiv2026

Reasoning as State Transition: A Representational Analysis of Reasoning Evolution in Large Language Models

Large Language Models have achieved remarkable performance on reasoning tasks, motivating research into how this ability evolves during training. Prior work has primarily analyzed this evolution via explicit generation outcomes, treating the reasoning process as a black box and obscuring internal changes. To address this opacity, we introduce a representational perspective to investigate the dynamics of the model's internal states. Through comprehensive experiments across models at various training stages, we discover that post-training yields only limited improvement in static initial representation quality. Furthermore, we reveal that, distinct from non-reasoning tasks, reasoning involves a significant continuous distributional shift in representations during generation. Comparative analysis indicates that post-training empowers models to drive this transition toward a better distribution for task solving. To clarify the relationship between internal states and external outputs, statistical analysis confirms a high correlation between generation correctness and the final representations; while counterfactual experiments identify the semantics of the generated tokens, rather than additional computation during inference or intrinsic parameter differences, as the dominant driver of the transition. Collectively, we offer a novel understanding of the reasoning process and the effect of training on reasoning enhancement, providing valuable insights for future model analysis and optimization.

cs.CL↗

arXiv2025

Vibe Reasoning: Eliciting Frontier AI Mathematical Capabilities -- A Case Study on IMO 2025 Problem 6

We introduce Vibe Reasoning, a human-AI collaborative paradigm for solving complex mathematical problems. Our key insight is that frontier AI models already possess the knowledge required to solve challenging problems -- they simply do not know how, what, or when to apply it. Vibe Reasoning transforms AI's latent potential into manifested capability through generic meta-prompts, agentic grounding, and model orchestration. We demonstrate this paradigm through IMO 2025 Problem 6, a combinatorial optimization problem where autonomous AI systems publicly reported failures. Our solution combined GPT-5's exploratory capabilities with Gemini 3 Pro's proof strengths, leveraging agentic workflows with Python code execution and file-based memory, to derive both the correct answer (2112) and a rigorous mathematical proof. Through iterative refinement across multiple attempts, we discovered the necessity of agentic grounding and model orchestration, while human prompts evolved from problem-specific hints to generic, transferable meta-prompts. We analyze why capable AI fails autonomously, how each component addresses specific failure modes, and extract principles for effective vibe reasoning. Our findings suggest that lightweight human guidance can unlock frontier models' mathematical reasoning potential. This is ongoing work; we are developing automated frameworks and conducting broader evaluations to further validate Vibe Reasoning's generality and effectiveness.

cs.AI↗

arXiv2025

Scaling Laws for Black box Adversarial Attacks

Adversarial examples exhibit cross-model transferability, enabling threatening black-box attacks on commercial models. Model ensembling, which attacks multiple surrogate models, is a known strategy to improve this transferability. However, prior studies typically use small, fixed ensembles, which leaves open an intriguing question of whether scaling the number of surrogate models can further improve black-box attacks. In this work, we conduct the first large-scale empirical study of this question. We show that by resolving gradient conflict with advanced optimizers, we discover a robust and universal log-linear scaling law through both theoretical analysis and empirical evaluations: the Attack Success Rate (ASR) scales linearly with the logarithm of the ensemble size $T$. We rigorously verify this law across standard classifiers, SOTA defenses, and MLLMs, and find that scaling distills robust, semantic features of the target class. Consequently, we apply this fundamental insight to benchmark SOTA MLLMs. This reveals both the attack's devastating power and a clear robustness hierarchy: we achieve 80\%+ transfer attack success rate on proprietary models like GPT-4o, while also highlighting the exceptional resilience of Claude-3.5-Sonnet. Our findings urge a shift in focus for robustness evaluation: from designing intricate algorithms on small ensembles to understanding the principled and powerful threat of scaling.

cs.LG↗

arXiv2025

AI Deception: Risks, Dynamics, and Controls

As intelligence increases, so does its shadow. AI deception, in which systems induce false beliefs to secure self-beneficial outcomes, has evolved from a speculative concern to an empirically demonstrated risk across language models, AI agents, and emerging frontier systems. This project provides a comprehensive and up-to-date overview of the AI deception field, covering its core concepts, methodologies, genesis, and potential mitigations. First, we identify a formal definition of AI deception, grounded in signaling theory from studies of animal deception. We then review existing empirical studies and associated risks, highlighting deception as a sociotechnical safety challenge. We organize the landscape of AI deception research as a deception cycle, consisting of two key components: deception emergence and deception treatment. Deception emergence reveals the mechanisms underlying AI deception: systems with sufficient capability and incentive potential inevitably engage in deceptive behaviors when triggered by external conditions. Deception treatment, in turn, focuses on detecting and addressing such behaviors. On deception emergence, we analyze incentive foundations across three hierarchical levels and identify three essential capability preconditions required for deception. We further examine contextual triggers, including supervision gaps, distributional shifts, and environmental pressures. On deception treatment, we conclude detection methods covering benchmarks and evaluation protocols in static and interactive settings. Building on the three core factors of deception emergence, we outline potential mitigation strategies and propose auditing approaches that integrate technical, community, and governance efforts to address sociotechnical challenges and future AI risks. To support ongoing work in this area, we release a living resource at www.deceptionsurvey.com.

cs.AI↗

arXiv2025

Benchmarking the Trustworthiness in Multimodal LLMs for Video Understanding

Recent advancements in multimodal large language models for video understanding (videoLLMs) have enhanced their capacity to process complex spatiotemporal data. However, challenges such as factual inaccuracies, harmful content, biases, hallucinations, and privacy risks compromise their reliability. This study introduces Trust-videoLLMs, a first comprehensive benchmark evaluating 23 state-of-the-art videoLLMs (5 commercial, 18 open-source) across five critical dimensions: truthfulness, robustness, safety, fairness, and privacy. Comprising 30 tasks with adapted, synthetic, and annotated videos, the framework assesses spatiotemporal risks, temporal consistency and cross-modal impact. Results reveal significant limitations in dynamic scene comprehension, cross-modal perturbation resilience and real-world risk mitigation. While open-source models occasionally outperform, proprietary models generally exhibit superior credibility, though scaling does not consistently improve performance. These findings underscore the need for enhanced training datat diversity and robust multimodal alignment. Trust-videoLLMs provides a publicly available, extensible toolkit for standardized trustworthiness assessments, addressing the critical gap between accuracy-focused benchmarks and demands for robustness, safety, fairness, and privacy.

cs.CV↗

arXiv2025

Mitigating Overthinking in Large Reasoning Models via Manifold Steering

Recent advances in Large Reasoning Models (LRMs) have demonstrated remarkable capabilities in solving complex tasks such as mathematics and coding. However, these models frequently exhibit a phenomenon known as overthinking during inference, characterized by excessive validation loops and redundant deliberation, leading to substantial computational overheads. In this paper, we aim to mitigate overthinking by investigating the underlying mechanisms from the perspective of mechanistic interpretability. We first showcase that the tendency of overthinking can be effectively captured by a single direction in the model's activation space and the issue can be eased by intervening the activations along this direction. However, this efficacy soon reaches a plateau and even deteriorates as the intervention strength increases. We therefore systematically explore the activation space and find that the overthinking phenomenon is actually tied to a low-dimensional manifold, which indicates that the limited effect stems from the noises introduced by the high-dimensional steering direction. Based on this insight, we propose Manifold Steering, a novel approach that elegantly projects the steering direction onto the low-dimensional activation manifold given the theoretical approximation of the interference noise. Extensive experiments on DeepSeek-R1 distilled models validate that our method reduces output tokens by up to 71% while maintaining and even improving the accuracy on several mathematical benchmarks. Our method also exhibits robust cross-domain transferability, delivering consistent token reduction performance in code generation and knowledge-based QA tasks. Code is available at: https://github.com/Aries-iai/Manifold_Steering.

cs.LG↗

arXiv2025

DeceptionBench: A Comprehensive Benchmark for AI Deception Behaviors in Real-world Scenarios

Despite the remarkable advances of Large Language Models (LLMs) across diverse cognitive tasks, the rapid enhancement of these capabilities also introduces emergent deceptive behaviors that may induce severe risks in high-stakes deployments. More critically, the characterization of deception across realistic real-world scenarios remains underexplored. To bridge this gap, we establish DeceptionBench, the first benchmark that systematically evaluates how deceptive tendencies manifest across different societal domains, what their intrinsic behavioral patterns are, and how extrinsic factors affect them. Specifically, on the static count, the benchmark encompasses 150 meticulously designed scenarios in five domains, i.e., Economy, Healthcare, Education, Social Interaction, and Entertainment, with over 1,000 samples, providing sufficient empirical foundations for deception analysis. On the intrinsic dimension, we explore whether models exhibit self-interested egoistic tendencies or sycophantic behaviors that prioritize user appeasement. On the extrinsic dimension, we investigate how contextual factors modulate deceptive outputs under neutral conditions, reward-based incentivization, and coercive pressures. Moreover, we incorporate sustained multi-turn interaction loops to construct a more realistic simulation of real-world feedback dynamics. Extensive experiments across LLMs and Large Reasoning Models (LRMs) reveal critical vulnerabilities, particularly amplified deception under reinforcement dynamics, demonstrating that current models lack robust resistance to manipulative contextual cues and the urgent need for advanced safeguards against various deception behaviors. Code and resources are publicly available at https://github.com/Aries-iai/DeceptionBench.

cs.CL↗

arXiv2025

Oyster-I: Beyond Refusal -- Constructive Safety Alignment for Responsible Language Models

Large language models (LLMs) typically deploy safety mechanisms to prevent harmful content generation. Most current approaches focus narrowly on risks posed by malicious actors, often framing risks as adversarial events and relying on defensive refusals. However, in real-world settings, risks also come from non-malicious users seeking help while under psychological distress (e.g., self-harm intentions). In such cases, the model's response can strongly influence the user's next actions. Simple refusals may lead them to repeat, escalate, or move to unsafe platforms, creating worse outcomes. We introduce Constructive Safety Alignment (CSA), a human-centric paradigm that protects against malicious misuse while actively guiding vulnerable users toward safe and helpful results. Implemented in Oyster-I (Oy1), CSA combines game-theoretic anticipation of user reactions, fine-grained risk boundary discovery, and interpretable reasoning control, turning safety into a trust-building process. Oy1 achieves state-of-the-art safety among open models while retaining high general capabilities. On our Constructive Benchmark, it shows strong constructive engagement, close to GPT-5, and unmatched robustness on the Strata-Sword jailbreak dataset, nearing GPT-o1 levels. By shifting from refusal-first to guidance-first safety, CSA redefines the model-user relationship, aiming for systems that are not just safe, but meaningfully helpful. We release Oy1, code, and the benchmark to support responsible, user-centered AI.

cs.AI↗

arXiv2025

Exploring the Generalizability of Factual Hallucination Mitigation via Enhancing Precise Knowledge Utilization

Large Language Models (LLMs) often struggle to align their responses with objective facts, resulting in the issue of factual hallucinations, which can be difficult to detect and mislead users without relevant knowledge. Although post-training techniques have been employed to mitigate the issue, existing methods usually suffer from poor generalization and trade-offs in other different capabilities. In this paper, we propose to address these by directly augmenting LLM's fundamental ability to precisely leverage its knowledge and introduce PKUE (Precise Knowledge Utilization Enhancement), which fine-tunes the model on self-generated responses to precise and simple factual questions through preference optimization. Furthermore, we construct FactualBench, a comprehensive and precise factual QA dataset containing 181k Chinese data spanning 21 domains, to facilitate both evaluation and training. Extensive experiments demonstrate that PKUE significantly improves LLM overall performance, with consistent enhancement across factual tasks of various forms, general tasks beyond factuality, and tasks in different language.

cs.CL↗

arXiv2025

Towards the Worst-case Robustness of Large Language Models

Recent studies have revealed the vulnerability of large language models to adversarial attacks, where adversaries craft specific input sequences to induce harmful, violent, private, or incorrect outputs. In this work, we study their worst-case robustness, i.e., whether an adversarial example exists that leads to such undesirable outputs. We upper bound the worst-case robustness using stronger white-box attacks, indicating that most current deterministic defenses achieve nearly 0\% worst-case robustness. We propose a general tight lower bound for randomized smoothing using fractional knapsack solvers or 0-1 knapsack solvers, and using them to bound the worst-case robustness of all stochastic defenses. Based on these solvers, we provide theoretical lower bounds for several previous empirical defenses. For example, we certify the robustness of a specific case, smoothing using a uniform kernel, against \textit{any possible attack} with an average $\ell_0$ perturbation of 2.02 or an average suffix length of 6.41.

cs.LG↗

arXiv2025

Silent Leaks: Implicit Knowledge Extraction Attack on RAG Systems through Benign Queries

Retrieval-Augmented Generation (RAG) systems enhance large language models (LLMs) by incorporating external knowledge bases, but this may expose them to extraction attacks, leading to potential copyright and privacy risks. However, existing extraction methods typically rely on malicious inputs such as prompt injection or jailbreaking, making them easily detectable via input- or output-level detection. In this paper, we introduce Implicit Knowledge Extraction Attack (IKEA), which conducts Knowledge Extraction on RAG systems through benign queries. Specifically, IKEA first leverages anchor concepts-keywords related to internal knowledge-to generate queries with a natural appearance, and then designs two mechanisms that lead anchor concepts to thoroughly "explore" the RAG's knowledge: (1) Experience Reflection Sampling, which samples anchor concepts based on past query-response histories, ensuring their relevance to the topic; (2) Trust Region Directed Mutation, which iteratively mutates anchor concepts under similarity constraints to further exploit the embedding space. Extensive experiments demonstrate IKEA's effectiveness under various defenses, surpassing baselines by over 80% in extraction efficiency and 90% in attack success rate. Moreover, the substitute RAG system built from IKEA's extractions shows comparable performance to the original RAG and outperforms those based on baselines across multiple evaluation tasks, underscoring the stealthy copyright infringement risk in RAG systems.

cs.CR↗

arXiv2025

Unveiling Trust in Multimodal Large Language Models: Evaluation, Analysis, and Mitigation

The trustworthiness of Multimodal Large Language Models (MLLMs) remains an intense concern despite the significant progress in their capabilities. Existing evaluation and mitigation approaches often focus on narrow aspects and overlook risks introduced by the multimodality. To tackle these challenges, we propose MultiTrust-X, a comprehensive benchmark for evaluating, analyzing, and mitigating the trustworthiness issues of MLLMs. We define a three-dimensional framework, encompassing five trustworthiness aspects which include truthfulness, robustness, safety, fairness, and privacy; two novel risk types covering multimodal risks and cross-modal impacts; and various mitigation strategies from the perspectives of data, model architecture, training, and inference algorithms. Based on the taxonomy, MultiTrust-X includes 32 tasks and 28 curated datasets, enabling holistic evaluations over 30 open-source and proprietary MLLMs and in-depth analysis with 8 representative mitigation methods. Our extensive experiments reveal significant vulnerabilities in current models, including a gap between trustworthiness and general capabilities, as well as the amplification of potential risks in base LLMs by both multimodal training and inference. Moreover, our controlled analysis uncovers key limitations in existing mitigation strategies that, while some methods yield improvements in specific aspects, few effectively address overall trustworthiness, and many introduce unexpected trade-offs that compromise model utility. These findings also provide practical insights for future improvements, such as the benefits of reasoning to better balance safety and performance. Based on these insights, we introduce a Reasoning-Enhanced Safety Alignment (RESA) approach that equips the model with chain-of-thought reasoning ability to discover the underlying risks, achieving state-of-the-art results.

cs.CL↗

arXiv2025

Reliable Unlearning Harmful Information in LLMs with Metamorphosis Representation Projection

While Large Language Models (LLMs) have demonstrated impressive performance in various domains and tasks, concerns about their safety are becoming increasingly severe. In particular, since models may store unsafe knowledge internally, machine unlearning has emerged as a representative paradigm to ensure model safety. Existing approaches employ various training techniques, such as gradient ascent and negative preference optimization, in attempts to eliminate the influence of undesired data on target models. However, these methods merely suppress the activation of undesired data through parametric training without completely eradicating its informational traces within the model. This fundamental limitation makes it difficult to achieve effective continuous unlearning, rendering these methods vulnerable to relearning attacks. To overcome these challenges, we propose a Metamorphosis Representation Projection (MRP) approach that pioneers the application of irreversible projection properties to machine unlearning. By implementing projective transformations in the hidden state space of specific network layers, our method effectively eliminates harmful information while preserving useful knowledge. Experimental results demonstrate that our approach enables effective continuous unlearning and successfully defends against relearning attacks, achieving state-of-the-art performance in unlearning effectiveness while preserving natural performance. Our code is available in https://github.com/ChengcanWu/MRP.

cs.LG↗

arXiv2025

Efficient Input-level Backdoor Defense on Text-to-Image Synthesis via Neuron Activation Variation

In recent years, text-to-image (T2I) diffusion models have gained significant attention for their ability to generate high quality images reflecting text prompts. However, their growing popularity has also led to the emergence of backdoor threats, posing substantial risks. Currently, effective defense strategies against such threats are lacking due to the diversity of backdoor targets in T2I synthesis. In this paper, we propose NaviT2I, an efficient input-level backdoor defense framework against diverse T2I backdoors. Our approach is based on the new observation that trigger tokens tend to induce significant neuron activation variation in the early stage of the diffusion generation process, a phenomenon we term Early-step Activation Variation. Leveraging this insight, NaviT2I navigates T2I models to prevent malicious inputs by analyzing Neuron activation variations caused by input tokens. Extensive experiments show that NaviT2I significantly outperforms the baselines in both effectiveness and efficiency across diverse datasets, various T2I backdoors, and different model architectures including UNet and DiT. Furthermore, we show that our method remains effective under potential adaptive attacks.

cs.CR↗

arXiv2025

A Survey on Autonomy-Induced Security Risks in Large Model-Based Agents

Recent advances in large language models (LLMs) have catalyzed the rise of autonomous AI agents capable of perceiving, reasoning, and acting in dynamic, open-ended environments. These large-model agents mark a paradigm shift from static inference systems to interactive, memory-augmented entities. While these capabilities significantly expand the functional scope of AI, they also introduce qualitatively novel security risks - such as memory poisoning, tool misuse, reward hacking, and emergent misalignment - that extend beyond the threat models of conventional systems or standalone LLMs. In this survey, we first examine the structural foundations and key capabilities that underpin increasing levels of agent autonomy, including long-term memory retention, modular tool use, recursive planning, and reflective reasoning. We then analyze the corresponding security vulnerabilities across the agent stack, identifying failure modes such as deferred decision hazards, irreversible tool chains, and deceptive behaviors arising from internal state drift or value misalignment. These risks are traced to architectural fragilities that emerge across perception, cognition, memory, and action modules. To address these challenges, we systematically review recent defense strategies deployed at different autonomy layers, including input sanitization, memory lifecycle control, constrained decision-making, structured tool invocation, and introspective reflection. We introduce the Reflective Risk-Aware Agent Architecture (R2A2), a unified cognitive framework grounded in Constrained Markov Decision Processes (CMDPs), which incorporates risk-aware world modeling, meta-policy adaptation, and joint reward-risk optimization to enable principled, proactive safety across the agent's decision-making loop.

cs.AI↗

arXiv2025

STAIR: Improving Safety Alignment with Introspective Reasoning

Ensuring the safety and harmlessness of Large Language Models (LLMs) has become equally critical as their performance in applications. However, existing safety alignment methods typically suffer from safety-performance trade-offs and the susceptibility to jailbreak attacks, primarily due to their reliance on direct refusals for malicious queries. In this paper, we propose STAIR, a novel framework that integrates SafeTy Alignment with Itrospective Reasoning. We enable LLMs to identify safety risks through step-by-step analysis by self-improving chain-of-thought (CoT) reasoning with safety awareness. STAIR first equips the model with a structured reasoning capability and then advances safety alignment via iterative preference optimization on step-level reasoning data generated using our newly proposed Safety-Informed Monte Carlo Tree Search (SI-MCTS). We further train a process reward model on this data to guide test-time searches for improved responses. Extensive experiments show that STAIR effectively mitigates harmful outputs while better preserving helpfulness, compared to instinctive alignment strategies. With test-time scaling, STAIR achieves a safety performance comparable to Claude-3.5 against popular jailbreak attacks. Relevant resources in this work are available at https://github.com/thu-ml/STAIR.

cs.CL↗

arXiv2025

Exploring Transferability of Multimodal Adversarial Samples for Vision-Language Pre-training Models with Contrastive Learning

The integration of visual and textual data in Vision-Language Pre-training (VLP) models is crucial for enhancing vision-language understanding. However, the adversarial robustness of these models, especially in the alignment of image-text features, has not yet been sufficiently explored. In this paper, we introduce a novel gradient-based multimodal adversarial attack method, underpinned by contrastive learning, to improve the transferability of multimodal adversarial samples in VLP models. This method concurrently generates adversarial texts and images within imperceptive perturbation, employing both image-text and intra-modal contrastive loss. We evaluate the effectiveness of our approach on image-text retrieval and visual entailment tasks, using publicly available datasets in a black-box setting. Extensive experiments indicate a significant advancement over existing single-modal transfer-based adversarial attack methods and current multimodal adversarial attack approaches.

cs.MM↗