R-Quest Qwen3-4B Validity-RL

This checkpoint initializes the Solver in R-Quest: Questioning the Questions: Sustaining Self-Evolution in Reasoning Models.

Paper · Project Page · Code · Training Data

Training

The model is initialized from Qwen3-4B-Base and trained with Validity-RL for 15 GRPO steps on the clean-v1 dataset. Valid questions are paired with verified mathematical answers; invalid questions have the target INVALID.

These are the merged BF16 weights from global step 15, before alternating Questioner/Solver training.

Use with R-Quest

Download the model:

huggingface-cli download jinyuan222/R-Quest-Qwen3-4B-Validity-RL \
  --local-dir /path/to/R-Quest-Qwen3-4B-Validity-RL

From the R-Quest training repository:

export RQUEST_INITIAL_SOLVER=/path/to/R-Quest-Qwen3-4B-Validity-RL
bash train_rquest.sh

Citation

@misc{li2026rquest,
  title={Questioning the Questions: Sustaining Self-Evolution in Reasoning Models},
  author={Jinyuan Li and Chengsong Huang and Langlin Huang and Donghong Cai and Shiping Gao and Yuyi Yang and Jiaxin Huang},
  year={2026},
  eprint={2610.04299},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2610.04299}
}