Wanlei Zhou

RISA achieves a notable improvement in refusal reliability for LLMs while preserving their response utility, transforming how we approach inference-time alignment.

Membership inference attacks can now be effectively executed across multiple generative modalities with a single, unified framework.

Poisoning attacks got you down? This defense flips the script by using the attacker's own clustering behavior against them, achieving near-perfect attack mitigation with minimal accuracy loss.

GPT-5 isn't always the smartest student: Qwen-Plus outshines it in Chinese CS certification exams, revealing critical cross-lingual performance gaps in LLMs.

LLM agents can now selectively forget sensitive information without sacrificing overall performance, thanks to a new framework that translates natural language unlearning requests into actionable prompts.

Semantic segmentation models, even recent transformer-based architectures like SAM, are surprisingly vulnerable to new backdoor attacks that current defenses can't reliably stop.

A shockingly small number of poisoned, synthetically distilled data points can completely hijack a model during transfer learning, turning it into an unwitting accomplice.