Zenghui Yuan

Jiawen Shi, Zenghui Yuan, Yinuo Liu et al.

LLM-as-a-Judge uses a large language model (LLM) to select the best response from a set of candidates for a given question. LLM-as-a-Judge has many applications such as LLM-powered search, reinforcement learning with AI feedback (RLAIF), and tool selection. In this work, we propose JudgeDeceiver, an optimization-based prompt injection attack to LLM-as-a-Judge. JudgeDeceiver injects a carefully crafted sequence into an attacker-controlled candidate response such that LLM-as-a-Judge selects the candidate response for an attacker-chosen question no matter what other candidate responses are. Specifically, we formulate finding such sequence as an optimization problem and propose a gradient based method to approximately solve it. Our extensive evaluation shows that JudgeDeceive is highly effective, and is much more effective than existing prompt injection attacks that manually craft the injected sequences and jailbreak attacks when extended to our problem. We also show the effectiveness of JudgeDeceiver in three case studies, i.e., LLM-powered search, RLAIF, and tool selection. Moreover, we consider defenses including known-answer detection, perplexity detection, and perplexity windowed detection. Our results show these defenses are insufficient, highlighting the urgent need for developing new defense strategies. Our implementation is available at this repository: https://github.com/ShiJiawenwen/JudgeDeceiver.

Yinuo Liu, Zenghui Yuan, Guiyao Tie et al.

Multimodal retrieval-augmented generation (RAG) enhances the visual reasoning capability of vision-language models (VLMs) by dynamically accessing information from external knowledge bases. In this work, we introduce \textit{Poisoned-MRAG}, the first knowledge poisoning attack on multimodal RAG systems. Poisoned-MRAG injects a few carefully crafted image-text pairs into the multimodal knowledge database, manipulating VLMs to generate the attacker-desired response to a target query. Specifically, we formalize the attack as an optimization problem and propose two cross-modal attack strategies, dirty-label and clean-label, tailored to the attacker's knowledge and goals. Our extensive experiments across multiple knowledge databases and VLMs show that Poisoned-MRAG outperforms existing methods, achieving up to 98\% attack success rate with just five malicious image-text pairs injected into the InfoSeek database (481,782 pairs). Additionally, We evaluate 4 different defense strategies, including paraphrasing, duplicate removal, structure-driven mitigation, and purification, demonstrating their limited effectiveness and trade-offs against Poisoned-MRAG. Our results highlight the effectiveness and scalability of Poisoned-MRAG, underscoring its potential as a significant threat to multimodal RAG systems.

Zhi Luo, Zenghui Yuan, Wenqi Wei et al.

With the remarkable success of Vision-Language Models (VLMs) on multimodal tasks, concerns regarding their deployment efficiency have become increasingly prominent. In particular, the number of tokens consumed during the generation process has emerged as a key evaluation metric.Prior studies have shown that specific inputs can induce VLMs to generate lengthy outputs with low information density, which significantly increases energy consumption, latency, and token costs. However, existing methods simply delay the occurrence of the EOS token to implicitly prolong output, and fail to directly maximize the output token length as an explicit optimization objective, lacking stability and controllability.To address these limitations, this paper proposes a novel verbose-text induction attack (VTIA) to inject imperceptible adversarial perturbations into benign images via a two-stage framework, which identifies the most malicious prompt embeddings for optimizing and maximizing the output token of the perturbed images.Specifically, we first perform adversarial prompt search, employing reinforcement learning strategies to automatically identify adversarial prompts capable of inducing the LLM component within VLMs to produce verbose outputs. We then conduct vision-aligned perturbation optimization to craft adversarial examples on input images, maximizing the similarity between the perturbed image's visual embeddings and those of the adversarial prompt, thereby constructing malicious images that trigger verbose text generation. Comprehensive experiments on four popular VLMs demonstrate that our method achieves significant advantages in terms of effectiveness, efficiency, and generalization capability.

Guiyao Tie, Zenghui Yuan, Zeli Zhao et al.

Self-correction of large language models (LLMs) emerges as a critical component for enhancing their reasoning performance. Although various self-correction methods have been proposed, a comprehensive evaluation of these methods remains largely unexplored, and the question of whether LLMs can truly correct themselves is a matter of significant interest and concern. In this study, we introduce CorrectBench, a benchmark developed to evaluate the effectiveness of self-correction strategies, including intrinsic, external, and fine-tuned approaches, across three tasks: commonsense reasoning, mathematical reasoning, and code generation. Our findings reveal that: 1) Self-correction methods can improve accuracy, especially for complex reasoning tasks; 2) Mixing different self-correction strategies yields further improvements, though it reduces efficiency; 3) Reasoning LLMs (e.g., DeepSeek-R1) have limited optimization under additional self-correction methods and have high time costs. Interestingly, a comparatively simple chain-of-thought (CoT) baseline demonstrates competitive accuracy and efficiency. These results underscore the potential of self-correction to enhance LLM's reasoning performance while highlighting the ongoing challenge of improving their efficiency. Consequently, we advocate for further research focused on optimizing the balance between reasoning capabilities and operational efficiency. Project Page: https://correctbench.github.io/