Research map / Reinforcement learning / RL from feedback
From Reward-Free Representations to Preferences: Rethinking Offline Preference-Based Reinforcement Learning
Jun-Jie Yang, Chia-Heng Hsu, Kui-Yuan Chen, Ping-Chun Hsieh
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add preference-guided fine-tuning of FB representations on top of test-time search (ablation) | FB-PbRL · DMC Cheetah average | episodic return: 344.7 -> 621.7 | Improved |
| add fine-tuning on top of test-time search (ablation) | FB-PbRL · DMC Walker average | episodic return: 533.4 -> 762.9 | Improved |
| add fine-tuning on top of test-time search (ablation) | FB-PbRL · DMC Quadruped average | episodic return: 663.4 -> 846.9 | Improved |
| add fine-tuning on top of test-time search (ablation) | FB-PbRL · DMC PointMass average | episodic return: 69.1 -> 570.8 | Improved |
| FB-PbRL versus CLARIFY | CLARIFY · DMC Walker average | episodic return: 248.9 -> 762.9 | Improved |
| FB-PbRL versus CLARIFY | CLARIFY · DMC PointMass average | episodic return: 317.8 -> 570.8 | Improved |
| contrastive preference loss instead of BT reward model in FB fine-tuning (ablation) | FB-BT-FT · DMC Cheetah aggregated | episodic return: 536.6 -> 621.7 | Improved |
| contrastive preference loss instead of BT reward model (ablation) | FB-BT-FT · DMC Walker aggregated | episodic return: 600.6 -> 794.5 | Improved |
| contrastive preference loss instead of BT reward model (ablation) | FB-BT-FT · DMC Quadruped aggregated | episodic return: 714.1 -> 846.9 | Improved |
| use 2000 preferences instead of 10000 reward-labeled transitions vs FB | FB · DMC Quadruped average | episodic return: 561.7 -> 826.3 | Improved |
| preferences instead of ground-truth reward vs best RFRL baseline | PSM · DMC Cheetah average | episodic return: 626.0 -> 645.4 | Improved |
| preferences instead of ground-truth reward | FB · DMC Walker average | episodic return: 719.9 -> 699.4 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CLARIFY — CLARIFY: Contrastive Preference Reinforcement Learning for Untangling Ambiguous Queries (ICML 2025) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Listwise Reward Estimation for Offline Preference-based Reinforcement Learning (ICML 2024 · similarity 0.92)
- Preference Elicitation for Offline Reinforcement Learning (ICLR 2025 · similarity 0.92)
- Provable Reward-Agnostic Preference-Based Reinforcement Learning (ICLR 2024 · similarity 0.92)
- CLARIFY: Contrastive Preference Reinforcement Learning for Untangling Ambiguous Queries (ICML 2025 · similarity 0.91)
- Provable Offline Preference-Based Reinforcement Learning (ICLR 2024 · similarity 0.91)
- Adversarial Policy Optimization for Offline Preference-based Reinforcement Learning (ICLR 2025 · similarity 0.91)
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.
