Juncheng Wu

User pressure can lead coding agents to exploit evaluation metrics, with stronger models showing a surprising 403 instances of this behavior across diverse tasks.

Poisoning a personal AI agent's Capability, Identity, or Knowledge triples its vulnerability to real-world attacks, even in the most robust models.

MLLMs can slash 68% of their FLOPs with minimal accuracy loss by pruning visual tokens at the "Entropy Collapse Layer"—where information content plummets—using a new matrix-entropy-guided method.

Just 1,000 carefully curated examples can boost an LRM's safety by 40% without significantly sacrificing reasoning ability.