Research map / Large language models / Alignment and preferences
A Regret Minimization Framework on Preference Learning in Large Language Models
Suhwan Kim, Taehyun Cho, Geon-Hyeong Kim, Yu Jin Kim, Youngsoo Jang, Moontae Lee, Jungwoo Lee
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 masked-solution data augmentation (30K) (ablation) | DPO on Qwen3-1.7B · GSM8K | accuracy: 72.55 -> 77.71 (+5.16) | Improved |
| add masked-solution data augmentation (30K) (ablation) | DPO on Qwen3-4B · GSM8K | accuracy: 85.22 -> 88.63 (+3.41) | Improved |
| add masked-solution data augmentation (30K) (ablation) | DPO on Qwen3-1.7B · MATH | accuracy: 52.14 -> 53.52 (+1.38) | Improved |
| add masked-solution data augmentation (30K) (ablation) | DPO on Qwen3-4B · MATH | accuracy: 63.70 -> 64.46 (+0.72) | Improved |
| add masked-solution data augmentation (30K) (ablation) | RePO on Qwen3-1.7B · GSM8K | accuracy: 80.52 -> 80.29 (-0.23) | No clear effect |
| add masked-solution data augmentation (30K) (ablation) | RePO on Qwen3-4B · GSM8K | accuracy: 90.60 -> 90.98 (+0.38) | No clear effect |
| add masked-solution data augmentation (30K) (ablation) | RePO on Qwen3-1.7B · MATH | accuracy: 54.50 -> 55.26 (+0.76) | No clear effect |
| add masked-solution data augmentation (30K) (ablation) | RePO on Qwen3-4B · MATH | accuracy: 65.54 -> 65.46 (-0.08) | No clear effect |
| add masked augmentation to DPO lowers MSE to RePO scores (ablation) | DPO on Qwen3-1.7B · token log-prob MSE vs RePO | MSE: 0.073 -> 0.056 | Improved |
| replace DPO with RePO | Qwen3-1.7B-Base · AlpacaEval 2 LC | LC win rate: 23.90 -> 36.61 | Improved |
| replace DPO with RePO | Qwen3-4B-Base · AlpacaEval 2 LC | LC win rate: 32.89 -> 55.08 | Improved |
| replace DPO with RePO | Qwen3-4B-Base · GSM8K | accuracy: 87.87 -> 90.60 | Improved |
Closest papers by meaning
- Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHF (ICLR 2025 · similarity 0.92)
- Online Preference Alignment for Language Models via Count-based Exploration (ICLR 2025 · similarity 0.92)
- Greedy Sampling Is Provably Efficient For RLHF (NeurIPS 2025 · similarity 0.92)
- Explicit Preference Optimization: No Need for an Implicit Reward Model (ICML 2025 · similarity 0.92)
- Capturing Individual Human Preferences with Reward Features (NeurIPS 2025 · similarity 0.91)
- PILAF: Optimal Human Preference Sampling for Reward Modeling (ICML 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.
