Yiyang Zhao

GEOALIGN: Geometric Rollout Curation for Robust LLM Reinforcement Learning

Ting Zhou, Zhenqing Ling, Yiyang Zhao, Ying Shen, Daoyuan Chen

Online reinforcement learning is widely used to align large language models (LLMs) with reward signals, yet training can be unstable under noisy or misspecified rewards. We identify a failure mode we call directional inconsistency: within a batch, a small set of high-reward rollo…Read more