Monotonic Inference Policies as the true objective in RLHF

Optimizing training policies may be a mirage compared to the necessity of optimizing monotonic inference policies during reinforcement learning for LLMs. This suggests a shift in focus from training-time objectives to inference-time performance stability.