Research map / Reinforcement learning / RL from feedback
Deep Reinforcement Learning from Human Preferences
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 changed | Where | Measured | Effect |
|---|---|---|---|
| Increase synthetic preference labels from 700 to 1400 | Preference-learned reward with RL · MuJoCo tasks | true reward: 700 -> 1400 labels | Helped |
| Use real human feedback instead of synthetic oracle | Preference-learned reward · MuJoCo tasks | true reward | Mixed |
| Ask humans to prefer upright robot on Ant | Human feedback · Ant | true reward | Helped |
| Real human feedback on Atari vs synthetic labels | Preference-learned reward · Atari | game score: comparable to synthetic with 40% fewer labels | Hurt |
| Human feedback on Enduro instead of A3C true reward | Preference-learned reward · Enduro | game score | Helped |
| Pick queries uniformly at random instead of by ensemble disagreement (ablation) | Preference reward learning · MuJoCo and Atari | true reward | Mixed |
| Train only one predictor instead of ensemble (ablation) | Preference reward learning · MuJoCo and Atari | true reward | Mixed |
| Gather queries only at start of training (offline) (ablation) | Preference reward learning · Pong | agent behavior | Hurt |
| Predict absolute scores (target) instead of comparisons (ablation) | Reward predictor · Continuous control | true reward | Hurt |
| Predict absolute scores (target) instead of comparisons (ablation) | Reward predictor · Atari | game score | No clear effect |
| Use single frames instead of trajectory segments (no segments) (ablation) | Reward predictor · MuJoCo | true reward | Hurt |
| Remove l2 regularization, keep only dropout (ablation) | Reward predictor · MuJoCo and Atari | true reward | Mixed |
Closest papers by meaning
- Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning (NeurIPS 2024 · similarity 0.90)
- DUO: Diverse, Uncertain, On-Policy Query Generation and Selection for Reinforcement Learning from Human Feedback (AAAI 2025 · similarity 0.90)
- Leveraging Sub-Optimal Data for Human-in-the-Loop Reinforcement Learning (ICLR 2025 · similarity 0.89)
- Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference (ICLR 2025 · similarity 0.89)
- Reward Model Learning vs. Direct Policy Optimization: A Comparative Analysis of Learning from Human Preferences (ICML 2024 · similarity 0.88)
- Promptable Behaviors: Personalizing Multi-Objective Rewards from Human Preferences (CVPR 2024 · similarity 0.88)
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.