Self-supervised reinforcement learning (RL) presents a promising avenue for enhancing the reasoning capabilities of LLMs without reliance on expensive human-annotated data. However, existing methods are prone to "policy collapse," a phenomenon where the learning process becomes unstable during extended training, leading to a sharp degradation in both reward and task performance. This paper diagnoses this instability, attributing it to the lack of a stable target in self-rewarding systems. To address this, we introduce M-GRPO, a momentum-anchored method that leverages a slowly evolving momentum model to provide a consistent and reliable training signal, stabilizing the generation of pseudo-labels for policy optimization. Our experiments, conducted on the MATH dataset, demonstrate that M-GRPO effectively prevents policy collapse, maintaining a stable training reward and consistently high validation accuracy.

First, download the MATH dataset and prepare it using the following Python script:

python examples/data_preprocess/math_dataset_ours.py --model Qwen2.5-3B

Then, run the following command to start the training (Modify the WANDB_KEY in the math_intuitor.sh script to your own WANDB key.):

bash math_intuitor.sh

This project builds upon the following open-source repositories:

• intuitor License: Apache License 2.0

• open-r1 License: Apache License 2.0

• verl License: Apache License 2.0