Research map / Large language models / Alignment and preferences
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
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 |
|---|---|---|---|
| replace PPO-based RLHF with DPO | GPT-J SFT · TL;DR summarization | GPT-4 win rate: 57 -> 61 | Helped |
| replace PPO with DPO | GPT-J TL;DR policy · CNN/DailyMail OOD | GPT-4 win rate: 0.26 -> 0.36 | Helped |
| replace PPO with DPO | GPT-2-large · IMDb controlled sentiment | reward-KL frontier | Helped |
| use DPO instead of Best of 128 / Preferred-FT | Pythia-2.8B · Anthropic HH single-turn | GPT-4 win rate | Helped |
| add per-example importance weighting (vs naive unlikelihood objective) (ablation) | DPO · TL;DR / dialogue | generation quality | Helped |
| Preferred-FT on chosen completions | GPT-J SFT · TL;DR summarization | GPT-4 win rate | No clear effect |
Closest papers by meaning
- Explicit Preference Optimization: No Need for an Implicit Reward Model (ICML 2025 · similarity 0.93)
- Would I Lie To You? Inference Time Alignment of Language Models using Direct Preference Heads (NeurIPS 2024 · similarity 0.93)
- Active Preference Learning for Large Language Models (ICML 2024 · similarity 0.92)
- Online Preference Alignment for Language Models via Count-based Exploration (ICLR 2025 · similarity 0.92)
- Multi-Reference Preference Optimization for Large Language Models (AAAI 2025 · similarity 0.92)
- Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference (ICLR 2025 · similarity 0.92)
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.