Qwen-M3PO-7B

Name: Qwen-M3PO-7B
Availability: InStock

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

📦 Code: insait-institute/c3po

Qwen2.5-Math 7B fine-tuned with M3PO, 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.

M3PO draws M Monte-Carlo weight perturbations from the IVON posterior per gradient step; rollouts and advantages are computed per sample and the gradients are averaged. To keep compute roughly matched to GRPO, the group size is reduced (GROUP_SIZE = G/M).

Training


Base / warm-start	`BayesRL/Qwen2.5Math-IVON-SFT-7B`
Foundation model	`Qwen/Qwen2.5-Math-7B`
Algorithm	M3PO (GRPO + IVON, `M` MC perturbations per step, equal-compute)
RL data	DAPO-Math-17k
Optimizer	IVON, lr `1.0`, ESS (λ) `1e10`
Hardware	8× NVIDIA H200 (144 GB)

Evaluation

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

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("BayesRL/Qwen-M3PO-7B")
tok = AutoTokenizer.from_pretrained("BayesRL/Qwen-M3PO-7B")

Citation

bibtex

@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}, 
}