Chishui Chen

OCSD reveals how isolating observation effects can lead to more effective token-level updates in reinforcement learning, outperforming traditional methods.

Teacher guidance can be strategically enhanced by targeting high-disagreement states, leading to significant performance gains in agentic tasks.