arXiv subjects

Yisong Xiao

Publications and source records attributed to Yisong Xiao.

At least 19 recordsLinked to original sources

arXiv2026

Mitigating Bias in Large Vision-Language Models via Counterfactual Ensemble Decoding

Large Vision-Language Models (LVLMs) have achieved remarkable performance across a wide range of tasks; however, they often inherit social biases from their training data, resulting in biased behavior when processing portraits from different social groups. Existing debiasing approaches typically compare token probabilities between the original and biased generations during decoding, but they are fundamentally limited by their reliance on a single, stereotyped viewpoint and fail to account for the diversity of social perspectives. Inspired by the social science principle that diversity fosters fairness, we propose Counterfactual Ensemble Decoding (CED), a novel framework that constructs multi-group counterfactual perspectives within the visual representation space and integrates them during decoding to promote equitable model behavior. CED first performs counterfactual steering in the visual space by identifying semantic directions associated with each social group and generating counterfactual representations along these directions, thereby offering diverse perspectives that disrupt stereotypical narratives. During decoding, CED locates the decoder layer exhibiting the greatest divergence among these perspectives and ensembles their token distributions using uncertainty-aware weights, prioritizing high-confidence tokens from different groups to yield a more balanced probability distribution that guides fairer generation. Extensive experiments on three social bias evaluation benchmarks demonstrate that \tool achieves substantial improvements over leading baselines, reducing bias by up to 47.97% across scenarios involving occupations, descriptors, and persona traits. Moreover, CED also preserves the core capabilities of the original model with minimal degradation.

cs.CL↗

arXiv2026

SafeFlow: Semantic Information-Flow Control for Blocking Malicious Propagation in Multi-Agent Systems

Multi-agent systems improve capability through task decomposition and role specialization, but these same mechanisms introduce an important safety blind spot: a harmful objective can be fragmented into locally plausible subtasks, allowing malicious intent to evade detection by any single agent. This is a growing social-impact challenge: systems handling sensitive information or consequential tools can turn routine delegation into unauthorized disclosure or unsafe action. We argue that this failure mode is better understood as a semantic information-flow problem than as a single-turn prompt classification task. To address this, we propose SafeFlow, a defense framework for multi-agent systems that formalizes malicious cross-agent propagation as a semantic information-flow problem. SafeFlow attaches structured semantic taints to root requests, propagates them through a dynamic collaboration graph, and performs workflow-level validation to reconstruct the global risk context before irreversible actions are committed. Evaluated on four benchmarks spanning prompt injection, jailbreak-based unsafe tool use, risky code execution, and harmful web-agent behavior, SafeFlow reduces attack success rates compared to undefended baselines and external defenses while retaining high benign task completion and a high paired safe--harm success rate. Our findings show that multi-agent systems still lack mechanisms for preserving risk semantics across delegation boundaries. This gap can turn routine delegation into privacy harms or unsafe actions that affect people and organizations. SafeFlow keeps this risk visible throughout the workflow, before it results in harm.

cs.MA↗

arXiv2026

Technical Report on the CVPR 2026@AdvML Workshop Challenge

Vision-language agents (VLAs) are increasingly used to interpret complex driving scenes and support safety-critical reasoning. This report presents the CVPR 2026@AdvML Workshop Challenge on adversarial multimodal attacks against autonomous-driving VLAs. Built on DriveLM-style multi-view visual question answering, the challenge represents each scene with six synchronized camera images and a structured collection of driving-related question-answer pairs. Participants generate adversarial images and suffix-only textual perturbations that induce model responses to deviate from reference answers while preserving image fidelity and limiting textual cost. The competition comprises two phases, with Phase II adding a hidden black-box model to assess transferability. We describe the task design, submission rules, evaluation protocol, and leaderboard results, and then examine five leading submissions for which technical reports were available. Across these reports, several recurring patterns emerge: image-side attacks are favored by the suffix penalty; scene-level, multi-view optimization is more effective than treating views in isolation; QA types and graph structure provide useful priors for allocating attack budget; feature-space objectives can improve black-box transfer; and typographic content embedded in camera images exposes a persistent vulnerability in driving VLAs. These findings provide a practical reference for future robustness evaluation and defense design in multimodal autonomous-driving systems.

cs.CV↗

arXiv2026

PRISM: Programmatic Reasoning with Image Sequence Manipulation for LVLM Jailbreaking

The increasing sophistication of large vision-language models (LVLMs) has been accompanied by advances in safety alignment mechanisms designed to prevent harmful content generation. However, these defenses remain vulnerable to sophisticated adversarial attacks. Existing jailbreak methods typically rely on direct and semantically explicit prompts, overlooking subtle vulnerabilities in how LVLMs compose information over multiple reasoning steps. In this paper, we propose a novel and effective jailbreak framework inspired by Return-Oriented Programming (ROP) techniques from software security. Our approach decomposes a harmful instruction into a sequence of individually benign visual gadgets. A carefully engineered textual prompt directs the sequence of inputs, prompting the model to integrate the benign visual gadgets through its reasoning process to produce a coherent and harmful output. This makes the malicious intent emergent and difficult to detect from any single component. We validate our method through extensive experiments on established benchmarks including SafeBench and MM-SafetyBench, targeting popular LVLMs. Results show that our approach consistently and substantially outperforms existing baselines on state-of-the-art models, achieving near-perfect attack success rates (over 0.90 on SafeBench) and improving ASR by up to 0.39. Our findings reveal a critical and underexplored vulnerability that exploits the compositional reasoning abilities of LVLMs, highlighting the urgent need for defenses that secure the entire reasoning process.

cs.CR↗

arXiv2026

Improving Safety Alignment via Balanced Direct Preference Optimization

With the rapid development and widespread application of Large Language Models (LLMs), their potential safety risks have attracted widespread attention. Reinforcement Learning from Human Feedback (RLHF) has been adopted to enhance the safety performance of LLMs. As a simple and effective alternative to RLHF, Direct Preference Optimization (DPO) is widely used for safety alignment. However, safety alignment still suffers from severe overfitting, which limits its actual performance. This paper revisits the overfitting phenomenon from the perspective of the model's comprehension of the training data. We find that the Imbalanced Preference Comprehension phenomenon exists between responses in preference pairs, which compromises the model's safety performance. To address this, we propose Balanced Direct Preference Optimization (B-DPO), which adaptively modulates optimization strength between preferred and dispreferred responses based on mutual information. A series of experimental results show that B-DPO can enhance the safety capability while maintaining the competitive general capabilities of LLMs on various mainstream benchmarks compared to state-of-the-art methods. \color{red}{Warning: This paper contains examples of harmful texts, and reader discretion is recommended.

cs.AI↗

arXiv2026

Evolving Deception: When Agents Evolve, Deception Wins

Self-evolving agents offer a promising path toward scalable autonomy. However, in this work, we show that in competitive environments, self-evolution can instead give rise to a serious and previously underexplored risk: the spontaneous emergence of deception as an evolutionarily stable strategy. We conduct a systematic empirical study on the self-evolution of large language model (LLM) agents in a competitive Bidding Arena, where agents iteratively refine their strategies through interaction-driven reflection. Across different evolutionary paths (\eg, Neutral, Honesty-Guided, and Deception-Guided), we find a consistent pattern: under utility-driven competition, unconstrained self-evolution reliably drifts toward deceptive behaviors, even when honest strategies remain viable. This drift is explained by a fundamental asymmetry in generalization. Deception evolves as a transferable meta-strategy that generalizes robustly across diverse and unseen tasks, whereas honesty-based strategies are fragile and often collapse outside their original contexts. Further analysis of agents internal states reveals the emergence of rationalization mechanisms, through which agents justify or deny deceptive actions to reconcile competitive success with normative instructions. Our paper exposes a fundamental tension between agent self-evolution and alignment, highlighting the risks of deploying self-improving agents in adversarial environments.

cs.CR↗

arXiv2026

Reasoning-Oriented Programming: Chaining Semantic Gadgets to Jailbreak Large Vision Language Models

Large Vision-Language Models (LVLMs) undergo safety alignment to suppress harmful content. However, current defenses predominantly target explicit malicious patterns in the input representation, often overlooking the vulnerabilities inherent in compositional reasoning. In this paper, we identify a systemic flaw where LVLMs can be induced to synthesize harmful logic from benign premises. We formalize this attack paradigm as \textit{Reasoning-Oriented Programming}, drawing a structural analogy to Return-Oriented Programming in systems security. Just as ROP circumvents memory protections by chaining benign instruction sequences, our approach exploits the model's instruction-following capability to orchestrate a semantic collision of orthogonal benign inputs. We instantiate this paradigm via \tool{}, an automated framework that optimizes for \textit{semantic orthogonality} and \textit{spatial isolation}. By generating visual gadgets that are semantically decoupled from the harmful intent and arranging them to prevent premature feature fusion, \tool{} forces the malicious logic to emerge only during the late-stage reasoning process. This effectively bypasses perception-level alignment. We evaluate \tool{} on SafeBench and MM-SafetyBench across 7 state-of-the-art 0.LVLMs, including GPT-4o and Claude 3.7 Sonnet. Our results demonstrate that \tool{} consistently circumvents safety alignment, outperforming the strongest existing baseline by an average of 4.67\% on open-source models and 9.50\% on commercial models.

cs.CR↗

arXiv2025

Uncovering Strategic Egoism Behaviors in Large Language Models

Large language models (LLMs) face growing trustworthiness concerns (\eg, deception), which hinder their safe deployment in high-stakes decision-making scenarios. In this paper, we present the first systematic investigation of strategic egoism (SE), a form of rule-bounded self-interest in which models pursue short-term or self-serving gains while disregarding collective welfare and ethical considerations. To quantitatively assess this phenomenon, we introduce SEBench, a benchmark comprising 160 scenarios across five domains. Each scenario features a single-role decision-making context, with psychologically grounded choice sets designed to elicit self-serving behaviors. These behavior-driven tasks assess egoistic tendencies along six dimensions, such as manipulation, rule circumvention, and self-interest prioritization. Building on this, we conduct extensive experiments across 5 open-sourced and 2 commercial LLMs, where we observe that strategic egoism emerges universally across models. Surprisingly, we found a positive correlation between egoistic tendencies and toxic language behaviors, suggesting that strategic egoism may underlie broader misalignment risks.

cs.CY↗

arXiv2025

AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions

The integration of vision-language models (VLMs) is driving a new generation of embodied agents capable of operating in human-centered environments. However, as deployment expands, these systems face growing safety risks, particularly when executing hazardous instructions. Current safety evaluation benchmarks remain limited: they cover only narrow scopes of hazards and focus primarily on final outcomes, neglecting the agent's full perception-planning-execution process and thereby obscuring critical failure modes. Therefore, we present SAFE, a benchmark for systematically assessing the safety of embodied VLM agents on hazardous instructions. SAFE comprises three components: SAFE-THOR, an extensible adversarial simulation sandbox with a universal adapter that maps high-level VLM outputs to low-level embodied controls, supporting diverse agent workflow integration; SAFE-VERSE, a risk-aware task suite inspired by Asimov's Three Laws of Robotics, comprising 45 adversarial scenarios, 1,350 hazardous tasks, and 9,900 instructions that span risks to humans, environments, and agents; and SAFE-DIAGNOSE, a multi-level and fine-grained evaluation protocol measuring agent performance across perception, planning, and execution. Applying SAFE to nine state-of-the-art VLMs and two embodied agent workflows, we uncover systematic failures in translating hazard recognition into safe planning and execution. Our findings reveal fundamental limitations in current safety alignment and demonstrate the necessity of a comprehensive, multi-stage evaluation for developing safer embodied intelligence.

cs.CR↗

arXiv2025

Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editing

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, yet they remain vulnerable to generating toxic content, necessitating detoxification strategies to ensure safe and responsible deployment. Test-time detoxification methods, which typically introduce static or dynamic interventions into LLM representations, offer a promising solution due to their flexibility and minimal invasiveness. However, current approaches often suffer from imprecise interventions, primarily due to their insufficient exploration of the transition space between toxic and non-toxic outputs. To address this challenge, we propose \textsc{A}utoregressive \textsc{R}eward \textsc{G}uided \textsc{R}epresentation \textsc{E}diting (ARGRE), a novel test-time detoxification framework that explicitly models toxicity transitions within the latent representation space, enabling stable and precise reward-guided editing. ARGRE identifies non-toxic semantic directions and interpolates between toxic and non-toxic representations to reveal fine-grained transition trajectories. These trajectories transform sparse toxicity annotations into dense training signals, enabling the construction of an autoregressive reward model that delivers stable and precise editing guidance. At inference, the reward model guides an adaptive two-step editing process to obtain detoxified representations: it first performs directional steering based on expected reward gaps to shift representations toward non-toxic regions, followed by lightweight gradient-based refinements. Extensive experiments across 8 widely used LLMs show that ARGRE significantly outperforms leading baselines in effectiveness (-62.21% toxicity) and efficiency (-47.58% inference time), while preserving the core capabilities of the original model with minimal degradation. Our code is available at the website.

cs.CL↗

arXiv2025

Fairness Mediator: Neutralize Stereotype Associations to Mitigate Bias in Large Language Models

LLMs have demonstrated remarkable performance across diverse applications, yet they inadvertently absorb spurious correlations from training data, leading to stereotype associations between biased concepts and specific social groups. These associations perpetuate and even amplify harmful social biases, raising significant fairness concerns. To mitigate such biases, prior studies have attempted to project model embeddings into unbiased spaces during inference. However, these approaches have shown limited effectiveness due to their weak alignment with downstream social biases. Inspired by the observation that concept cognition in LLMs is primarily represented through a linear associative memory mechanism, where key-value mapping occurs in the MLP layers, we posited that biased concepts and social groups are similarly encoded as entity (key) and information (value) pairs, which can be manipulated to promote fairer associations. To this end, we propose Fairness Mediator (FairMed), a bias mitigation framework that neutralizes stereotype associations. Our framework comprises two main components: a stereotype association prober and an adversarial debiasing neutralizer. The prober captures stereotype associations encoded within MLP layer activations by employing prompts centered around biased concepts to detect the emission probabilities for social groups. Subsequently, the adversarial debiasing neutralizer intervenes in MLP activations during inference to equalize the association probabilities among different social groups. Extensive experiments across nine protected attributes show that FairMed significantly outperforms SOTA methods in effectiveness. Compared to the most effective baseline, FairMed presents competitive efficiency by cutting mitigation overhead by hundreds of minutes. FairMed also maintains the LLM's language understanding capabilities without compromising overall performance.

cs.SE↗

arXiv2025

Reasoning-Augmented Conversation for Multi-Turn Jailbreak Attacks on Large Language Models

Multi-turn jailbreak attacks simulate real-world human interactions by engaging large language models (LLMs) in iterative dialogues, exposing critical safety vulnerabilities. However, existing methods often struggle to balance semantic coherence with attack effectiveness, resulting in either benign semantic drift or ineffective detection evasion. To address this challenge, we propose Reasoning-Augmented Conversation, a novel multi-turn jailbreak framework that reformulates harmful queries into benign reasoning tasks and leverages LLMs' strong reasoning capabilities to compromise safety alignment. Specifically, we introduce an attack state machine framework to systematically model problem translation and iterative reasoning, ensuring coherent query generation across multiple turns. Building on this framework, we design gain-guided exploration, self-play, and rejection feedback modules to preserve attack semantics, enhance effectiveness, and sustain reasoning-driven attack progression. Extensive experiments on multiple LLMs demonstrate that RACE achieves state-of-the-art attack effectiveness in complex conversational scenarios, with attack success rates (ASRs) increasing by up to 96%. Notably, our approach achieves ASRs of 82% and 92% against leading commercial models, OpenAI o1 and DeepSeek R1, underscoring its potency. We release our code at https://github.com/NY1024/RACE to facilitate further research in this critical domain.

cs.CL↗

arXiv2024

BDefects4NN: A Backdoor Defect Database for Controlled Localization Studies in Neural Networks

Pre-trained large deep learning models are now serving as the dominant component for downstream middleware users and have revolutionized the learning paradigm, replacing the traditional approach of training from scratch locally. To reduce development costs, developers often integrate third-party pre-trained deep neural networks (DNNs) into their intelligent software systems. However, utilizing untrusted DNNs presents significant security risks, as these models may contain intentional backdoor defects resulting from the black-box training process. These backdoor defects can be activated by hidden triggers, allowing attackers to maliciously control the model and compromise the overall reliability of the intelligent software. To ensure the safe adoption of DNNs in critical software systems, it is crucial to establish a backdoor defect database for localization studies. This paper addresses this research gap by introducing BDefects4NN, the first backdoor defect database, which provides labeled backdoor-defected DNNs at the neuron granularity and enables controlled localization studies of defect root causes. In BDefects4NN, we define three defect injection rules and employ four representative backdoor attacks across four popular network architectures and three widely adopted datasets, yielding a comprehensive database of 1,654 backdoor-defected DNNs with four defect quantities and varying infected neurons. Based on BDefects4NN, we conduct extensive experiments on evaluating six fault localization criteria and two defect repair techniques, which show limited effectiveness for backdoor defects. Additionally, we investigate backdoor-defected models in practical scenarios, specifically in lane detection for autonomous driving and large language models (LLMs), revealing potential threats and highlighting current limitations in precise defect localization.

cs.SE↗

arXiv2024

LanEvil: Benchmarking the Robustness of Lane Detection to Environmental Illusions

Lane detection (LD) is an essential component of autonomous driving systems, providing fundamental functionalities like adaptive cruise control and automated lane centering. Existing LD benchmarks primarily focus on evaluating common cases, neglecting the robustness of LD models against environmental illusions such as shadows and tire marks on the road. This research gap poses significant safety challenges since these illusions exist naturally in real-world traffic situations. For the first time, this paper studies the potential threats caused by these environmental illusions to LD and establishes the first comprehensive benchmark LanEvil for evaluating the robustness of LD against this natural corruption. We systematically design 14 prevalent yet critical types of environmental illusions (e.g., shadow, reflection) that cover a wide spectrum of real-world influencing factors in LD tasks. Based on real-world environments, we create 94 realistic and customizable 3D cases using the widely used CARLA simulator, resulting in a dataset comprising 90,292 sampled images. Through extensive experiments, we benchmark the robustness of popular LD methods using LanEvil, revealing substantial performance degradation (-5.37% Accuracy and -10.70% F1-Score on average), with shadow effects posing the greatest risk (-7.39% Accuracy). Additionally, we assess the performance of commercial auto-driving systems OpenPilot and Apollo through collaborative simulations, demonstrating that proposed environmental illusions can lead to incorrect decisions and potential traffic accidents. To defend against environmental illusions, we propose the Attention Area Mixing (AAM) approach using hard examples, which witness significant robustness improvement (+3.76%) under illumination effects. We hope our paper can contribute to advancing more robust auto-driving systems in the future. Website: https://lanevil.github.io/.

cs.CV↗

arXiv2024

GenderBias-\emph{VL}: Benchmarking Gender Bias in Vision Language Models via Counterfactual Probing

Large Vision-Language Models (LVLMs) have been widely adopted in various applications; however, they exhibit significant gender biases. Existing benchmarks primarily evaluate gender bias at the demographic group level, neglecting individual fairness, which emphasizes equal treatment of similar individuals. This research gap limits the detection of discriminatory behaviors, as individual fairness offers a more granular examination of biases that group fairness may overlook. For the first time, this paper introduces the GenderBias-\emph{VL} benchmark to evaluate occupation-related gender bias in LVLMs using counterfactual visual questions under individual fairness criteria. To construct this benchmark, we first utilize text-to-image diffusion models to generate occupation images and their gender counterfactuals. Subsequently, we generate corresponding textual occupation options by identifying stereotyped occupation pairs with high semantic similarity but opposite gender proportions in real-world statistics. This method enables the creation of large-scale visual question counterfactuals to expose biases in LVLMs, applicable in both multimodal and unimodal contexts through modifying gender attributes in specific modalities. Overall, our GenderBias-\emph{VL} benchmark comprises 34,581 visual question counterfactual pairs, covering 177 occupations. Using our benchmark, we extensively evaluate 15 commonly used open-source LVLMs (\eg, LLaVA) and state-of-the-art commercial APIs, including GPT-4o and Gemini-Pro. Our findings reveal widespread gender biases in existing LVLMs. Our benchmark offers: (1) a comprehensive dataset for occupation-related gender bias evaluation; (2) an up-to-date leaderboard on LVLM biases; and (3) a nuanced understanding of the biases presented by these models. \footnote{The dataset and code are available at the \href{https://genderbiasvl.github.io/}{website}.}

cs.CV↗

arXiv2023

Pre-trained Trojan Attacks for Visual Recognition

Pre-trained vision models (PVMs) have become a dominant component due to their exceptional performance when fine-tuned for downstream tasks. However, the presence of backdoors within PVMs poses significant threats. Unfortunately, existing studies primarily focus on backdooring PVMs for the classification task, neglecting potential inherited backdoors in downstream tasks such as detection and segmentation. In this paper, we propose the Pre-trained Trojan attack, which embeds backdoors into a PVM, enabling attacks across various downstream vision tasks. We highlight the challenges posed by cross-task activation and shortcut connections in successful backdoor attacks. To achieve effective trigger activation in diverse tasks, we stylize the backdoor trigger patterns with class-specific textures, enhancing the recognition of task-irrelevant low-level features associated with the target class in the trigger pattern. Moreover, we address the issue of shortcut connections by introducing a context-free learning pipeline for poison training. In this approach, triggers without contextual backgrounds are directly utilized as training data, diverging from the conventional use of clean images. Consequently, we establish a direct shortcut from the trigger to the target class, mitigating the shortcut connection issue. We conducted extensive experiments to thoroughly validate the effectiveness of our attacks on downstream detection and segmentation tasks. Additionally, we showcase the potential of our approach in more practical scenarios, including large vision models and 3D object detection in autonomous driving. This paper aims to raise awareness of the potential threats associated with applying PVMs in practical scenarios. Our codes will be available upon paper publication.

cs.CV↗

arXiv2023

RobustMQ: Benchmarking Robustness of Quantized Models

Quantization has emerged as an essential technique for deploying deep neural networks (DNNs) on devices with limited resources. However, quantized models exhibit vulnerabilities when exposed to various noises in real-world applications. Despite the importance of evaluating the impact of quantization on robustness, existing research on this topic is limited and often disregards established principles of robustness evaluation, resulting in incomplete and inconclusive findings. To address this gap, we thoroughly evaluated the robustness of quantized models against various noises (adversarial attacks, natural corruptions, and systematic noises) on ImageNet. The comprehensive evaluation results empirically provide valuable insights into the robustness of quantized models in various scenarios, for example: (1) quantized models exhibit higher adversarial robustness than their floating-point counterparts, but are more vulnerable to natural corruptions and systematic noises; (2) in general, increasing the quantization bit-width results in a decrease in adversarial robustness, an increase in natural robustness, and an increase in systematic robustness; (3) among corruption methods, \textit{impulse noise} and \textit{glass blur} are the most harmful to quantized models, while \textit{brightness} has the least impact; (4) among systematic noises, the \textit{nearest neighbor interpolation} has the highest impact, while bilinear interpolation, cubic interpolation, and area interpolation are the three least harmful. Our research contributes to advancing the robust quantization of models and their deployment in real-world scenarios.

cs.LG↗

arXiv2023

Isolation and Induction: Training Robust Deep Neural Networks against Model Stealing Attacks

Despite the broad application of Machine Learning models as a Service (MLaaS), they are vulnerable to model stealing attacks. These attacks can replicate the model functionality by using the black-box query process without any prior knowledge of the target victim model. Existing stealing defenses add deceptive perturbations to the victim's posterior probabilities to mislead the attackers. However, these defenses are now suffering problems of high inference computational overheads and unfavorable trade-offs between benign accuracy and stealing robustness, which challenges the feasibility of deployed models in practice. To address the problems, this paper proposes Isolation and Induction (InI), a novel and effective training framework for model stealing defenses. Instead of deploying auxiliary defense modules that introduce redundant inference time, InI directly trains a defensive model by isolating the adversary's training gradient from the expected gradient, which can effectively reduce the inference computational cost. In contrast to adding perturbations over model predictions that harm the benign accuracy, we train models to produce uninformative outputs against stealing queries, which can induce the adversary to extract little useful knowledge from victim models with minimal impact on the benign performance. Extensive experiments on several visual classification datasets (e.g., MNIST and CIFAR10) demonstrate the superior robustness (up to 48% reduction on stealing accuracy) and speed (up to 25.4x faster) of our InI over other state-of-the-art methods. Our codes can be found in https://github.com/DIG-Beihang/InI-Model-Stealing-Defense.

cs.CR↗