Zhihao Dou

Adversarial purification can be dramatically improved by focusing on patch-level semantics, leading to state-of-the-art performance in defending against adversarial attacks.

Aha-moment-driven backdoor attacks can redirect reasoning in VLMs while preserving output coherence, making them harder to detect.

LLM agents struggle to generalize from experience to reusable skills, often performing worse than simply replaying past trajectories, revealing a critical gap in current abstraction methods.

Agent evaluation is bottlenecked by environment interaction overhead, but ACE-Bench slashes this by using static JSON files, enabling fast and reproducible training-time validation.