Gen-Verse/ReasonFlux

Implements trajectory-aware process reward models for dense supervision across offline data curation and online RL policy optimization, plus co-evolutionary reinforcement learning where coding and unit-testing agents improve together. Leverages hierarchical thought templates to structure reasoning trajectories, integrating with standard test-time scaling and agentic pipelines while releasing production-ready models (7B–32B) on Hugging Face optimized for mathematics, coding, and long-chain-of-thought reasoning.

524 stars. No commits in the last 6 months.