AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Sequential Preference Ranking for Efficient Reinforcement Learning from Human Feedback

Minyoung Hwang, Gunmin Lee, Hogun Kee, Chan Woo Kim, Kyungjae Lee, Songhwai Oh

NeurIPS 2023 · Reinforcement learning › RL from feedback · approach: Theory of RLHF · 5 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 root pairwise comparison instead of pairwise (ablation)MRN · DMControl locomotion (avg)reward after convergence: 29.0%Improved
use sequential pairwise comparison instead of pairwise (ablation)MRN · DMControl locomotion (avg)reward after convergence: 10.3%Improved
use root pairwise comparison instead of pairwise (ablation)MRN · Meta-World manipulation (avg)success rate: 25.0%Improved
use sequential pairwise comparison instead of pairwise (ablation)MRN · Meta-World manipulation (avg)success rate: 1.2%No clear effect
use root pairwise comparison instead of sequential pairwise (ablation)MRN · Meta-World Door Opensuccess rate: 77.4 ± 40.8 -> 97.8 ± 5.0Improved
use sequential pairwise instead of pairwise (ablation)MRN · Meta-World Sweep Intosuccess rate: 68.4 ± 35.2 -> 55.7 ± 48.1Hurt
use sequential pairwise instead of pairwise (ablation)MRN · Meta-World Hammersuccess rate: 31.0 ± 24.7 -> 30.8 ± 35.4No clear effect
use root pairwise instead of pairwise (ablation)MRN · DMControl Quadruped Walkreward after convergence: 353.5 ± 183.9 -> 728.8 ± 274.1Improved
use root pairwise instead of pairwise (ablation)MRN · DMControl Hopper Hopreward after convergence: 26.2 ± 37.0 -> 100.1 ± 70.7Improved
increase number of feedbacks (0.1K to 4K) (ablation)sequential and root pairwise · DMControl Quadruped Walkaverage rewardImproved
use sequential pairwise instead of pairwise, real robot (ablation)DQN with MRN · Real UR-5 block placingsuccess rate: 34.4 -> 41.0Improved
use root pairwise instead of pairwise, real robot (ablation)DQN with MRN · Real UR-5 block placingreward model accuracy: 75.8 -> 84.9Improved

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