OnePO: Direct One-stage Policy Optimization for SFT-free Domain Adaptation

Junying Chen, Xinyuan Xie, Ziniu Li, Benyou Wang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:17665-17691, 2026.

Abstract

Domain adaptation typically follows a two-stage pipeline: Supervised Fine-Tuning (SFT) then Reinforcement Learning (RL). However, does RL necessarily require a pre-SFT phase for domain adaptation? SFT confines the model to an imitation distribution, limiting RL exploration, while the two-stage transition causes capability regression and extra engineering. We propose One-stage Policy Optimization (OnePO), an SFT-free paradigm that adapts pretrained LLMs to target domains in a single RL stage. OnePO uses teacher outputs as transient guidance to overcome the slow convergence of pure RL, while avoiding two failures of naive teacher-output integration: gradient starvation for low-probability teacher tokens and distribution anchoring from persistent teacher signals. It introduces two mechanisms: (1) Adaptive Objective Evolution, reshaping the RL objective for rapid absorption of teacher-provided knowledge; and (2) Teacher Retirement, automatically discarding teacher outputs once the model surpasses them. On medical adaptation, OnePO achieves 67.2 on HealthBench with only 20K training samples, outperforming SFT+RL by +2.7 and pure RL by +7.4 points. Scaling the same recipe produces HuatuoGPT-3, an open-source medical LLM series whose 32B variant reaches 70.3 on HealthBench. Additional writing and legal-domain experiments show that OnePO extends beyond medicine. Models and code are available at https://github.com/FreedomIntelligence/HuatuoGPT-3.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-chen26fi, title = {{O}ne{PO}: Direct One-stage Policy Optimization for {SFT}-free Domain Adaptation}, author = {Chen, Junying and Xie, Xinyuan and Li, Ziniu and Wang, Benyou}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {17665--17691}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/chen26fi/chen26fi.pdf}, url = {https://proceedings.mlr.press/v306/chen26fi.html}, abstract = {Domain adaptation typically follows a two-stage pipeline: Supervised Fine-Tuning (SFT) then Reinforcement Learning (RL). However, does RL necessarily require a pre-SFT phase for domain adaptation? SFT confines the model to an imitation distribution, limiting RL exploration, while the two-stage transition causes capability regression and extra engineering. We propose One-stage Policy Optimization (OnePO), an SFT-free paradigm that adapts pretrained LLMs to target domains in a single RL stage. OnePO uses teacher outputs as transient guidance to overcome the slow convergence of pure RL, while avoiding two failures of naive teacher-output integration: gradient starvation for low-probability teacher tokens and distribution anchoring from persistent teacher signals. It introduces two mechanisms: (1) Adaptive Objective Evolution, reshaping the RL objective for rapid absorption of teacher-provided knowledge; and (2) Teacher Retirement, automatically discarding teacher outputs once the model surpasses them. On medical adaptation, OnePO achieves 67.2 on HealthBench with only 20K training samples, outperforming SFT+RL by +2.7 and pure RL by +7.4 points. Scaling the same recipe produces HuatuoGPT-3, an open-source medical LLM series whose 32B variant reaches 70.3 on HealthBench. Additional writing and legal-domain experiments show that OnePO extends beyond medicine. Models and code are available at https://github.com/FreedomIntelligence/HuatuoGPT-3.} }
Endnote
%0 Conference Paper %T OnePO: Direct One-stage Policy Optimization for SFT-free Domain Adaptation %A Junying Chen %A Xinyuan Xie %A Ziniu Li %A Benyou Wang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-chen26fi %I PMLR %P 17665--17691 %U https://proceedings.mlr.press/v306/chen26fi.html %V 306 %X Domain adaptation typically follows a two-stage pipeline: Supervised Fine-Tuning (SFT) then Reinforcement Learning (RL). However, does RL necessarily require a pre-SFT phase for domain adaptation? SFT confines the model to an imitation distribution, limiting RL exploration, while the two-stage transition causes capability regression and extra engineering. We propose One-stage Policy Optimization (OnePO), an SFT-free paradigm that adapts pretrained LLMs to target domains in a single RL stage. OnePO uses teacher outputs as transient guidance to overcome the slow convergence of pure RL, while avoiding two failures of naive teacher-output integration: gradient starvation for low-probability teacher tokens and distribution anchoring from persistent teacher signals. It introduces two mechanisms: (1) Adaptive Objective Evolution, reshaping the RL objective for rapid absorption of teacher-provided knowledge; and (2) Teacher Retirement, automatically discarding teacher outputs once the model surpasses them. On medical adaptation, OnePO achieves 67.2 on HealthBench with only 20K training samples, outperforming SFT+RL by +2.7 and pure RL by +7.4 points. Scaling the same recipe produces HuatuoGPT-3, an open-source medical LLM series whose 32B variant reaches 70.3 on HealthBench. Additional writing and legal-domain experiments show that OnePO extends beyond medicine. Models and code are available at https://github.com/FreedomIntelligence/HuatuoGPT-3.
APA
Chen, J., Xie, X., Li, Z. & Wang, B.. (2026). OnePO: Direct One-stage Policy Optimization for SFT-free Domain Adaptation. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:17665-17691 Available from https://proceedings.mlr.press/v306/chen26fi.html.

Related Material