Shuo He

Chain-of-thought prompting makes large language models smarter, but it also makes them less safe, a problem this paper tackles by forcing models to think about safety *before* reasoning.

Context inconsistency in stepwise group-based RL can severely bias advantage estimation, but a hierarchical grouping strategy can fix it without extra compute.

Overcome simplicity bias in RL agents with PA-MoE, a mixture-of-experts architecture that learns task phases directly from the RL objective, leading to better expert specialization.