Xudong Jiang

ARENA uncovers vulnerabilities in large audio-language models that traditional text-based red-teaming methods miss, achieving near-perfect safety metrics across multiple systems.

Forget disjointed approaches: OneHOI's unified diffusion transformer generates and edits human-object interactions with unprecedented control and realism.

MLLMs are more vulnerable than we thought: imperceptible visual prompts can effectively hijack their behavior.