GRPO is Secretly a Process Reward Model

This project provides a critical theoretical bridge between two major trends in LLM training: Group Relative Policy Optimization (popularized by DeepSeek-R1) and Process Reward Models (favored by OpenAI's o1). The defensibility is low (3) because this is primarily a mathematical insight and a reference implementation; once the 'secret' is out, any lab can (and will) incorporate this into their training pipelines. Frontier risk is high because labs like OpenAI, Anthropic, and DeepSeek are the primary users of these algorithms and are actively looking for ways to simplify PRM training. The lack of stars and forks is typical for a specialized research paper repository, but the impact of the insight is high. The primary value is reducing the need for expensive step-wise labeling by proving that group-relative outcome rewards can achieve similar mathematical results. This will likely be absorbed into standard RLHF libraries (like TRL or vLLM) within months, rendering a standalone project obsolete.