Efficient Process Reward Modeling via Contrastive Mutual Information

This project represents a timely but highly vulnerable contribution to the field of LLM reasoning. As Process Reward Models (PRMs) have become central to the success of 'o1-style' reasoning models, the search for efficient training methods that bypass expensive step-level human labeling is a primary research frontier. While the use of Contrastive Mutual Information (CMI) is a clever methodological pivot to reduce compute/labeling costs, the project currently lacks any significant moat beyond the initial mathematical insight. With 0 stars and only 2 days of public existence, it is a 'paper-first' repository. Frontier labs (OpenAI, Anthropic, Google) and well-funded labs (DeepSeek, 01.AI) are aggressively iterating on PRMs and likely have internal variants of contrastive or self-supervised reward signals. The risk of platform domination is maximum here because reward modeling is being integrated directly into the training and inference pipelines of the major foundation models. This code is a reference implementation of an algorithm that is likely to be superseded or absorbed into larger RLHF/RLAIF frameworks (like axolotl or trl) within months.