Dake Shen

Chenxi Li, Xianggan Liu, Dake Shen et al.

Despite the rapid progress of Large Vision-Language Models (LVLMs), the integration of visual modalities introduces new safety vulnerabilities that adversaries can exploit to elicit biased or malicious outputs. In this paper, we demonstrate an underexplored vulnerability via semantic slot filling, where LVLMs complete missing slot values with unsafe content even when the slot types are deliberately crafted to appear benign. Building on this finding, we propose StructAttack, a simple yet effective single-query jailbreak framework under black-box settings. StructAttack decomposes a harmful query into a central topic and a set of benign-looking slot types, then embeds them as structured visual prompts (e.g., mind maps, tables, or sunburst diagrams) with small random perturbations. Paired with a completion-guided instruction, LVLMs automatically recompose the concealed semantics and generate unsafe outputs without triggering safety mechanisms. Although each slot appears benign in isolation (local benignness), StructAttack exploits LVLMs' reasoning to assemble these slots into coherent harmful semantics. Extensive experiments on multiple models and benchmarks show the efficacy of our proposed StructAttack.

Zitong Xu, Dake Shen, Yaosong Du et al.

Recent advances in AIGC (Artificial Intelligence Generated Content) models have enabled significant progress in image and video generation. However, users still struggle to obtain content that aligns with their preferences due to the difficulty of crafting detailed prompts and the lack of mechanisms to retain their preferences. To address these challenges, we construct \textbf{UniPrefer-100K}, a large-scale dataset comprising images, videos, and associated text that describes the styles users tend to prefer. Based on UniPrefer-100K, we propose \textbf{MagicWand}, a universal generation and evaluation agent that enhances prompts based on user preferences, leverages advanced generation models for high-quality content, and applies preference-aligned evaluation and refinement. In addition, we introduce \textbf{UniPreferBench}, the first large-scale benchmark with over 120K annotations for assessing user preference alignment across diverse AIGC tasks. Experiments on UniPreferBench demonstrate that MagicWand consistently generates content and evaluations that are well aligned with user preferences across a wide range of scenarios.