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
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 |
|---|---|---|---|
| use on-policy rather than off-policy DPO | DPO · General chat AlpacaEval 2.0 | LC win-rate: 41.26 -> 45.44 | Improved |
| use on-policy rather than off-policy preference distillation (PDFR) | PDFR · General chat AlpacaEval 2.0 | LC win-rate: 36.45 -> 43.99 | Improved |
| run more on-policy DPO iterations | On-policy DPO · General chat AlpacaEval 2.0 | LC win-rate: 39.83 -> 45.44 | Improved |
| run more on-policy PDFR iterations | On-policy PDFR · General chat AlpacaEval 2.0 | LC win-rate: 37.89 -> 43.99 | Improved |
| use on-policy rather than off-policy REBEL | REBEL · General chat AlpacaEval 2.0 | LC win-rate: 33.85 -> 43.81 | Improved |
| run REBEL beyond iteration 3 | On-policy REBEL · General chat AlpacaEval 2.0 | LC win-rate: 43.81 -> 42.90 | Hurt |
| add reward calibration (RDRC) to REBEL, continue to iteration 5 | On-policy REBEL · General chat AlpacaEval 2.0 | LC win-rate: 42.90 -> 44.07 | Improved |
| run more RDRC iterations | On-policy RDRC · General chat AlpacaEval 2.0 | LC win-rate: 39.02 -> 44.07 | Improved |
| use on-policy RDRC instead of off-policy REBEL | RDRC vs REBEL · General chat AlpacaEval 2.0 | win-rate: 33.13 -> 42.61 | Improved |
| use PDFR instead of DPO on-policy | On-policy · General chat AlpacaEval 2.0 | LC win-rate: 45.44 -> 43.99 | Mixed |
| on-policy DPO vs on-policy preference distillation: effect on OpenLLM alignment tax | On-policy PDFR/RDRC · OpenLLM leaderboard | MMLU: 66.76 -> 67.13 | No clear effect |
| use on-policy DPO instead of off-policy DPO on TL;DR | DPO (Pythia-1.4B) · TL;DR summarization | win-rate | Improved |
Closest papers by meaning
- The Importance of Online Data: Understanding Preference Fine-tuning via Coverage (NeurIPS 2024 · similarity 0.91)
- Preference Fine-Tuning of LLMs Should Leverage Suboptimal, On-Policy Data (ICML 2024 · similarity 0.89)
- AlphaDPO: Adaptive Reward Margin for Direct Preference Optimization (ICML 2025 · similarity 0.89)
- Asymmetric REINFORCE for off-Policy Reinforcement Learning: Balancing positive and negative rewards (NeurIPS 2025 · similarity 0.89)
- Towards Efficient Exact Optimization of Language Model Alignment (ICML 2024 · similarity 0.89)
- Statistical Rejection Sampling Improves Preference Optimization (ICLR 2024 · similarity 0.89)
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.
