Read the original at HF Daily Papers
The paper analyzes training-inference mismatch in large language model reinforcement learning, identifying numerical discrepancy and policy lag as distinct sources of off-policy bias.
Carried by: arXiv, HF Daily Papers. First seen: .