AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Offline Preference-Based Value Optimization

Hyungkyu Kang, Min-hwan Oh

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Preference-based RL reward learning

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
replace TD loss with value alignment loss (ablation)IQL (giving PVO) · Meta-World medium-replay, 8 tasks, 1000 feedbacksuccess rate (%): IQL 58.76 -> PVO 69.14 (average)Improved
replace TD loss with value alignment loss (ablation)TD3+BC · Meta-World medium-replay, 1000 feedbacksuccess rate (IQM)Improved
replace TD loss with value alignment loss (ablation)XQL · Meta-World medium-replay, 1000 feedbacksuccess rate (IQM)Improved
use a learned transition model instead of dataset-sample approximation (ablation)PVO · Meta-World medium-replay/expert, 500 and 1000 feedbacksuccess rate (%): medium-replay 1000: 69.14 vs 71.18; medium-expert 1000: 73.00 vs 71.16No clear effect
reduce preference feedback to about 100 samples (ablation)PVO · Meta-World medium-replay dial-turnsuccess rate (%)No clear effect
increase proportion of random trajectories in dataset (ablation)PVO, IQL, APPO · Meta-World dial-turn mixture datasetssuccess rate (%)Hurt
use PVO instead of IQL with learned reward on button-press-topdownPVO · Meta-World medium-replaysuccess rate (%)Improved
use PVO instead of Flow to Better (FTB)PVO · Meta-World medium-replay, 1000 feedbacksuccess rate (%): FTB 0.00 on 6 tasks; sweep-into PVO 27.84 vs FTB 62.67Mixed

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