This is the Qwen2.5-7B model trained by the Co-rewarding method using the MATH training set, as presented in the paper Co-rewarding: Stable Self-supervised RL for Eliciting Reasoning in Large Language Models.

While reinforcement learning with verifiable rewards (RLVR) is effective to improve the reasoning ability of large language models (LLMs), its reliance on human-annotated labels leads to the scaling up dilemma, especially for complex tasks. Recent self-rewarding methods investigate a label-free alternative to unlock the reasoning capabilities of LLMs, yet they frequently encounter the non-negligible training collapse issue, as the single-view supervision signal easily forms the self-consistent illusion, yielding the reward hacking.

Co-rewarding is a novel self-supervised RL framework that improves training stability by seeking complementary supervision from another views. Specifically, Co-rewarding is instantiated in two ways:

Intuitively, such instantiations introduce different levels of discrepancy to increase the difficulty of training collapse on trivial reasoning solutions. Empirically, Co-rewarding exhibits stable training across various setups, and outperforms other self-rewarding baselines by $+3.31%$ improvements on average on multiple mathematical reasoning benchmarks, especially by $+7.49%$ on Llama-3.2-3B-Instruct. Notably, Co-rewarding reaches or even surpasses RLVR with ground-truth (GT) label in several cases, such as a Pass@$1$ of $94.01%$ on GSM8K with Qwen3-8B-Base.

For more details about the Co-rewarding method, including code and training scripts, please refer to the official Github Repository.

If you use our datasets or models, please cite our paper: