Research map / Large language models / Alignment and preferences
Normalized Rewards for Preference Optimization
Shawn Im, Federico Danieli, Skyler Seto, Barry-John Theobald, Katherine Metcalf
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 probability-mass regularization and length normalization to DPO (N-DPO) | DPO Llama-3.1-8B-Instruct · AlpacaEval 2 (LC) | win rate: 21.72 -> 24.97 | Improved |
| add N-DPO regularization | DPO Mistral-7B-Instruct · AlpacaEval 2 (LC) | win rate: 23.66 -> 25.48 | Improved |
| add N-DPO regularization | DPO OLMo-7B-SFT · AlpacaEval 2 (LC) | win rate: 8.96 -> 9.88 | Improved |
| add N-SimPO regularization | SimPO OLMo-7B-SFT · AlpacaEval 2 (WR) | win rate: 3.66 -> 10.02 | Improved |
| add N-SimPO regularization | SimPO Llama-3.1-8B-Instruct · AlpacaEval 2 (LC) | win rate: 26.74 -> 25.54 | Hurt |
| add N-SimPO regularization | SimPO Mistral-7B-Instruct · AlpacaEval 2 (LC) | win rate: 26.34 -> 26.05 | No clear effect |
| add regularization lambda>0 to length-normalized DPO (ablation) | N-DPO Mistral-7B-Instruct · AlpacaEval 2 (LC) | win rate: 20.21 -> 22.10 | Improved |
| add regularization lambda>0 to length-normalized DPO (ablation) | N-DPO Llama-3.1-8B-Instruct · AlpacaEval 2 (LC) | win rate: 28.50 -> 29.41 | Improved |
| add regularization lambda>0 to length-normalized DPO (ablation) | N-DPO OLMo-7B-SFT · AlpacaEval 2 (LC) | win rate: 7.34 -> 8.02 | Improved |
| add N-DPO regularization | DPO Llama-3.1-8B-Instruct · Common sense and reasoning avg | score change vs reference: -3.25 -> +3.04 | Improved |
| add N-DPO regularization | DPO Mistral-7B-Instruct · Common sense and reasoning avg | score change vs reference: -2.80 -> -0.98 | Improved |
| add N-SimPO regularization | SimPO OLMo-7B-SFT · Common sense and reasoning avg | score change vs reference: -0.73 -> -4.81 | 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.
- SimPO — SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024) · better on 4 of 7 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
- Mitigating Reward Over-optimization in Direct Alignment Algorithms with Importance Sampling (NeurIPS 2025 · similarity 0.91)
- Proximalized Preference Optimization for Diverse Feedback Types: A Decomposed Perspective on DPO (NeurIPS 2025 · similarity 0.90)
- AlphaPO: Reward Shape Matters for LLM Alignment (ICML 2025 · similarity 0.90)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.90)
- What Matters in Data for DPO? (NeurIPS 2025 · similarity 0.90)
- Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 2025 · 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.
