LLM Reinforcement Learning

Revisiting the Training-Inference Gap in LLM Reinforcement Learning: Origins and Remedies

CIS fixes LLM RL training-inference mismatch with confidence-aware importance sampling, boosting MoE math reas…