AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

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

Junying Chen, Xinyuan Xie, Ziniu Li, Benyou Wang

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL fine-tuning of generative models

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add probability floor (ablation)OnePO · HealthBench Totalscore (%): 65.4 -> 49.1 when removedImproved
add gradient rescaling (ablation)OnePO · HealthBench Totalscore (%): 65.4 -> 57.3 when removedImproved
add Teacher Retirement (ablation)OnePO · HealthBench Totalscore (%): 65.4 -> 52.7 when removedImproved
add Teacher Retirement (ablation)OnePO · HealthBench Hardscore (%): 39.1 -> 18.6 when removedImproved
add probability floor (ablation)OnePO · MMLU-Pro (Med)score (%): 81.2 -> 73.1 when removedImproved
use strict max retirement threshold instead of looser or none (ablation)OnePO · training curvesevaluation score: Max > Mean > Min > NoneImproved
use OnePO instead of SFT+RL (DeepSeek-V3.2-Thinking teacher)Qwen3-8B-Base · HealthBench Totalscore (%): 64.5 -> 67.2Improved
use OnePO instead of SFT+RL (GPT-5-Chat teacher)Qwen3-8B-Base · HealthBench Totalscore (%): 63.6 -> 65.4Improved
add teacher guidance via OnePO over pure RLQwen3-8B-Base · MedXpertQA (Text)score (%): 20.0 -> 25.9Improved
add teacher guidance via SFT+RL over pure RLQwen3-8B-Base · Medbullets (5-option)score (%): 48.1 -> 61.9Improved
use weaker teacher (Qwen3-8B-Thinking) instead of DeepSeek-V3.2-ThinkingOnePO · HealthBench Totalscore (%): 67.2 -> 66.0No clear effect
use mixed verifiable and rubric reward instead of rubric-only (ablation)OnePO · HealthBench Totalscore (%): 63.7 -> 65.4Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy