Research map / Large language models / Alignment and preferences
Learning Ordinal Probabilistic Reward from Preferences
Longze Chen, Lu Wang, Renke Shan, Ze Gong, Run Luo, Jiaming Li, Jing Luo, Qiyao Wang, Min Yang
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 the OPRM ordinal loss instead of Bradley-Terry loss (ablation) | Qwen-32B reward model · RM benchmarks | accuracy: consistent performance penalty of 1.1% to 2.1% when replaced by BT | Improved |
| use weighted-average decoding instead of Argmax (ablation) | OPRM · RM benchmarks | accuracy: Argmax leads to 8.7% to 17.5% drop | Improved |
| add region flooding tuning (RgFT) | OPRM-32B · RM benchmarks Overall | Overall: 74.2 -> 73.9 | Hurt |
| add region flooding tuning (RgFT) | OPRM-32B · PPE-Correctness | accuracy: 66.1 -> 67.3 | Improved |
| add region flooding tuning (RgFT) | OPRM-72B · RM benchmarks Overall | Overall: 73.1 -> 73.8 | Improved |
| add region flooding tuning (RgFT) | OPRM-32B · RewardBench | ECE-10: 10.62 -> 5.18 | Improved |
| use OPRM instead of base Qwen2.5-32B | Qwen-2.5-32B · RewardBench | accuracy: 69.56 -> 81.04 | Improved |
| add RgFT with unlabeled general-domain data (mixed training) | OPRM-RgFT · Role Play | Overall: 65.4 -> 70.2 | Improved |
| add general-domain data | OPRM · Role Play | Overall: 64.3 -> 68.1 | Improved |
| use region flooding instead of hard classification (ablation) | OPRM-RgFT-32B · RM benchmarks | Overall: 54.3 -> 73.9 | Improved |
| use uniform instead of irregular bin boundaries (ablation) | OPRM-RgFT · RM benchmarks | Overall: 73.8 -> 73.9 | No clear effect |
| scale model from 32B to 72B | OPRM · RM benchmarks | Overall: 74.2 -> 73.1 | Hurt |
Closest papers by meaning
- GRAM: A Generative Foundation Reward Model for Reward Generalization (ICML 2025 · similarity 0.91)
- Scaling Reward Modeling without Human Supervision (NeurIPS 2026 · similarity 0.89)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.89)
- RRM: Robust Reward Model Training Mitigates Reward Hacking (ICLR 2025 · similarity 0.89)
- Bayesian Preference Learning for Test-Time Steerable Reward Models (NeurIPS 2026 · similarity 0.89)
- Active Reward Modeling: Adaptive Preference Labeling for Large Language Model Alignment (ICML 2025 · 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.
