Qwen-C3PO-7B

📄 Paper: Parameter Exploration for RLVR via Variational Learning · arXiv:2608.09805

📦 Code: insait-institute/c3po

Qwen2.5-Math 7B fine-tuned with C3PO, from the paper "Parameter Exploration for RLVR via Variational Learning".

3PO is a family of parameter-space exploration strategies for Reinforcement Learning with Verifiable Rewards (RLVR). Instead of relying only on action-space heuristics (temperature, clipping, entropy bonuses), 3PO samples model weights from an approximate posterior learned with the variational optimizer IVON; the amount of weight noise becomes an extra control lever for exploration.

C3PO splits each GRPO group of G rollouts across N independent weight perturbations (G/N rollouts each). Advantages are computed over the full, more-diverse group, with a Seq-MIS (sequence-level multiple-importance-sampling) correction to account for the differing perturbations. This gives extra parameter-space diversity at the same rollout budget as GRPO.

Training

Evaluation

Evaluated on AIME 2024–2026, MATH-500, AMC 2023, and Minerva. See the paper for full results.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("BayesRL/Qwen-C3PO-7B")
tok = AutoTokenizer.from_pretrained("BayesRL/Qwen-C3PO-7B")

Citation

@misc{venkatkrishna2026parameterexploration,
      title={Parameter Exploration for RLVR via Variational Learning}, 
      author={Vatsal Venkatkrishna and Nico Daheim and Iryna Gurevych},
      year={2026},
      eprint={2608.09805},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2608.09805}, 
}