This model is a fine-tuned version of
Qwen/Qwen2.5-7B trained on the MATH dataset for one epoch using the
Intuitor method.
Intuitor is an implementation of
Reinforcement Learning from Internal Feedback (RLIF), presented in the paper
Learning to Reason without External Rewards. RLIF enables large language models to learn from intrinsic signals—specifically "self-certainty"—without relying on external rewards, gold labels, or verifiers. Intuitor replaces external rewards in Group Relative Policy Optimization (GRPO) with self-certainty scores, allowing for effective unsupervised learning across reasoning domains like mathematics and code generation.
1@article{zhao2025learning,
2 title = {Learning to Reason without External Rewards},
3 author = {Zhao, Xuandong and Kang, Zhewei and Feng, Aosong and Levine, Sergey and Song, Dawn},
4 journal = {arXiv preprint arXiv:2505.19590},
5 year = {2025}
6}
