Zicheng Li

The paper demonstrates that while distilling large language models for medical QA can significantly improve final answer accuracy, this gain often comes at the cost of factual accuracy and detailed reasoning within the generated thought trace.

The paper proposes Predictive Routing Replay (PR2) to stabilize reinforcement learning on Mixture of Experts (MoE) LLMs by predicting and incorporating short-horizon router evolution during training and rollout.

The paper proposes S2L-PO, a framework that uses smaller, naturally diverse models as structured explorers to enhance the policy-level diversity and performance of larger language models during training.

The paper introduces pause-and-think-T, a reasoning-centric dataset and benchmark that enables compact Vision-Language Models to perform visually grounded, context-aware action suggestion, matching large models like GPT-4o.