Jingwen Zhang

Forget complex regularizations – stable sequential knowledge editing in LLMs boils down to correctly managing accumulated constraints, unlocking simpler and more reliable updates.

Short videos that are *too* sensational actually decrease user engagement, revealing an inverted U-shaped relationship between "Message Sensation Value" and behavioral engagement.

LLMs can now automatically validate smart contract vulnerabilities by generating and refining profit-driven exploits, significantly reducing the burden of manual verification.