New method judges AI reasoning steps by final success, not correctness

Original: From Correctness to Utility: Gain-Based Prefix Evaluation for LLM Reasoning

Source: arxiv.org ↗