Research map / Reinforcement learning / RL from feedback
Preference Optimization via Contrastive Divergence: Your Policy Is Secretly an NLL Estimator
Zhuotong Chen, Fang Liu, Xuan Zhu, Haozhu Wang, Jiayu Li, Yanjun Qi, Mohammad Ghavamzadeh
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 MC-kernel rejected sampling instead of random sampling (MC-PO vs DPO) | DPO · AlpacaEval 2 / Arena-Hard, Mistral-7B-SFT | win-rate: 25.07 -> 30.86; 42.01 -> 52.75 | Improved |
| use MC-kernel rejected sampling instead of random sampling | DPO · AlpacaEval 2 / Arena-Hard, Llama-3.1-8B-SFT | win-rate: 33.74 -> 35.84; 60.25 -> 63.77 | Improved |
| use MC-kernel rejected sampling in instruct setup | DPO · AlpacaEval 2 / Arena-Hard, Llama-3.1-8B-Instruct | win-rate: 64.22 -> 66.90; 75.88 -> 76.71 | Mixed |
| add noisy near-duplicate sample to candidates: MC-PO vs DPO (ablation) | DPO · Noisy candidates, Llama-3.1-8B | win-rate: 23.62 -> 28.98 (Alpaca); 50.51 -> 58.09 (Arena) | Improved |
| select the noisy sample as rejected response (DPO(-)) (ablation) | DPO · Noisy candidates, Llama-3.1-8B | win-rate: 23.62 -> 1.08 (Alpaca) | Hurt |
| increase rejected samples M from 1 to 3 with MC kernel (ablation) | MC-PO · AlpacaEval 2, Llama-3.1-8B-SFT | win-rate: 35.84 -> 37.40 | Improved |
| increase rejected samples M from 1 to 3 with random sampling (ablation) | DPO · AlpacaEval 2, Llama-3.1-8B-SFT | win-rate: 33.74 -> 34.36 | No clear effect |
| replace DPO 4x augmented pairs with MC-PO on original data (ablation) | DPO · Llama-3.1-8B-SFT | win-rate: 34.18 -> 35.84 (Alpaca); 59.62 -> 63.77 (Arena) | Improved |
| use Min sampling instead of MC kernel (ablation) | MC-PO · Llama-3.1-8B-SFT Figure 2 | win-rate | Hurt |
| remove high-ranked candidates from the candidate set (ablation) | MC-PO · Llama-3.1-8B-SFT Nectar x-negs | win-rate | Hurt |
Closest papers by meaning
- Statistical Rejection Sampling Improves Preference Optimization (ICLR 2024 · similarity 0.89)
- Preference Optimization by Estimating the Ratio of the Data Distribution (NeurIPS 2025 · similarity 0.88)
- The Importance of Online Data: Understanding Preference Fine-tuning via Coverage (NeurIPS 2024 · similarity 0.88)
- Preference Fine-Tuning of LLMs Should Leverage Suboptimal, On-Policy Data (ICML 2024 · similarity 0.88)
- Token-level Direct Preference Optimization (ICML 2024 · similarity 0.88)
- Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints (ICLR 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.
