How can protection against evolving jailbreaks, harmful content, and hallucinations preserve useful task performance?
Research / Project detail
LLM safety & trustworthy generation
Research and engineering across dynamic attack–defense, closed-loop training, and detection–response separation; the related project received the Beijing Science and Technology Progress Award, Second Prize.
AlignmentSafetyTrustworthy AI
Problem
Method
Attack generation → evaluation → safety training → online protection → experience feedback. Related work includes TriPlay-RL and Magic-Token safety control.
Outcomes
The work includes TriPlay-RL, Magic Token, detection–response separation, and trustworthy generation. The related project received the Beijing Science and Technology Progress Award, Second Prize. “Key Technologies and Applications of Large-Model Safety” passed a CAAI technology achievement appraisal on June 19, 2025.
Related publications
Related papers
[ACL 2026]Safety & alignment
[AAAI 2026]Safety & alignment
[arXiv 2026]Safety & alignment
Related coverage

