Unsupervised Process Reward Models

Hugging Face proposes unsupervised Process Reward Models (uPRM) that train step-level reward functions for LLM reasoning without human annotations by deriving scoring functions from LLM next-token predictions.