Olmo3-B3PO-7B-ESS1e8
📄 Paper: Parameter Exploration for RLVR via Variational Learning · arXiv:2608.09805
📦 Code: insait-institute/c3po
Olmo-3 7B fine-tuned with B3PO (Bayesian Perturbed Parameter Policy Optimization), 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.
B3PO draws a single weight perturbation from the IVON posterior per gradient step and syncs it to the rollout engine, so each step explores from one perturbed parameter sample.
Note: This is the ESS ablation variant with λ =
1e8; the main run is Olmo3-B3PO-7B (λ =1e9).
Training
| Base / warm-start | BayesRL/Olmo3-IVON-SFT-7B |
| Foundation model | allenai/Olmo-3-1025-7B |
| Algorithm | B3PO (GRPO + IVON, single perturbation per step) |
| RL data | DAPO-Math-17k |
| Optimizer | IVON, lr 1.0, ESS (λ) 1e8 |
| 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
Loads as a standard causal LM:
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("BayesRL/Olmo3-B3PO-7B-ESS1e8")
tok = AutoTokenizer.from_pretrained("BayesRL/Olmo3-B3PO-7B-ESS1e8")
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},
}