AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Coverage Improvement and Fast Convergence of On-policy Preference Learning

Juno Kim, Jihun Yun, Jason D. Lee, Kwang-Sung Jun

ICML 2026 · Reinforcement learning › RL from feedback · approach: Reward modeling versus direct optimization · 0 citations (OpenAlex)

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
use on-policy rather than off-policy DPODPO · General chat AlpacaEval 2.0LC win-rate: 41.26 -> 45.44Improved
use on-policy rather than off-policy preference distillation (PDFR)PDFR · General chat AlpacaEval 2.0LC win-rate: 36.45 -> 43.99Improved
run more on-policy DPO iterationsOn-policy DPO · General chat AlpacaEval 2.0LC win-rate: 39.83 -> 45.44Improved
run more on-policy PDFR iterationsOn-policy PDFR · General chat AlpacaEval 2.0LC win-rate: 37.89 -> 43.99Improved
use on-policy rather than off-policy REBELREBEL · General chat AlpacaEval 2.0LC win-rate: 33.85 -> 43.81Improved
run REBEL beyond iteration 3On-policy REBEL · General chat AlpacaEval 2.0LC win-rate: 43.81 -> 42.90Hurt
add reward calibration (RDRC) to REBEL, continue to iteration 5On-policy REBEL · General chat AlpacaEval 2.0LC win-rate: 42.90 -> 44.07Improved
run more RDRC iterationsOn-policy RDRC · General chat AlpacaEval 2.0LC win-rate: 39.02 -> 44.07Improved
use on-policy RDRC instead of off-policy REBELRDRC vs REBEL · General chat AlpacaEval 2.0win-rate: 33.13 -> 42.61Improved
use PDFR instead of DPO on-policyOn-policy · General chat AlpacaEval 2.0LC win-rate: 45.44 -> 43.99Mixed
on-policy DPO vs on-policy preference distillation: effect on OpenLLM alignment taxOn-policy PDFR/RDRC · OpenLLM leaderboardMMLU: 66.76 -> 67.13No clear effect
use on-policy DPO instead of off-policy DPO on TL;DRDPO (Pythia-1.4B) · TL;DR summarizationwin-rateImproved

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