Yujie Xiong

Deeper reasoning creates an unexpected safety backdoor: extended CoT dilutes attention away from system constraints, making models progressively more vulnerable to jailbreaks the longer they "think."

LLMs can now produce trustworthy clinical diagnoses by explicitly justifying their reasoning steps, rivaling resource-intensive RL methods with a more stable and efficient training approach.

LLMs can achieve state-of-the-art mathematical reasoning accuracy while pruning redundant computations by verifying solutions in a hierarchical, reversible manner.