Inspired by the AZ-R paper “Absolute Zero: Reinforced Self-Play Reasoning with Zero Data” (2025) by Andrew Zhao Yiran Wu, Tong Wu, Quentin Xu, and collaborators.
This repository is not affiliated with the original authors, but serves as a practical miniature replication and experimental foundation toward that vision.
Train an RL agent from scratch to deduce symbolic mathematical equations through trial and error, guided by curriculum learning and self-play without supervised labels or solvers. From initial failure and debugging to achieving mastery on a curriculum of increasingly difficult mathematical tasks.
✅ Actor-Critic agent with entropy & KL regularization
✅ Reward shaping and log_std tracking for exploration
✅ Curriculum learning pipeline (Easy → Medium → Hard tasks)
✅ Progressive reward & error convergence
✅ Evaluation dashboard & symbolic tagging
✅ Logging for interpretability and training insights
Here are some results which showing a proper improvement in the agent reducing loss and growth in the reward signals
- Add post-hoc error correction
- Integrate code-based symbolic execution
- Diversify tasks to include chain-of-thought
- Add multi-agent curriculum evolution
- Enable multi-step program synthesis reasoning
This project is not affiliated with the original AZR authors.
Inspired by research, built independently as a learning replica.
Open to collaborations, extensions, and feedback. Star ⭐ and fork if useful!