Research map / Reinforcement learning / RL from feedback
IRPM: Intergroup Relative Preference Modeling for Pointwise Generative Reward Models
Haonan Song, Qingchen Xie, Huan Zhu, Feng Xiao, Luxi Xing, Liu Kang, Fuzhen Li, Zhiyong Zheng, Feng Jiang, Ziheng Li, Kun Yan, Qingyi Si and 3 more
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 CoT critique before the score (vs score-only output) (ablation) | IRPM-Mean Qwen3-8B · RM-Bench + JudgeBench | overall accuracy: 71.6 -> 74.0 | Improved |
| filter out weak-preference samples (scores 0-1) (ablation) | IRPM-Mean Qwen3-8B · RM-Bench + JudgeBench | overall accuracy: 70.7 -> 74.0 | Improved |
| add adaptive margin scaled by preference strength (ablation) | IRPM-Mean Qwen3-8B · RM-Bench + JudgeBench | overall accuracy: 72.9 -> 74.0 | Improved |
| increase rollout group size from 2 to 4 (ablation) | IRPM-Mean Qwen3-8B · RM-Bench + JudgeBench | overall accuracy: 71.6 -> 74.0 | Improved |
| increase rollout group size from 4 to 8 (ablation) | IRPM-Mean Qwen3-8B · RM-Bench + JudgeBench | overall accuracy: 74.0 -> 73.7 | No clear effect |
| vary rollout temperature (0.3, 0.7, 1.5 vs 1.0) (ablation) | IRPM-Mean Qwen3-8B · RM-Bench + JudgeBench | overall accuracy: 74.0 -> 72.1 (0.3), 72.5 (0.7), 73.1 (1.5) | Hurt |
| use IRPM-Preference (soft preference strength) as the intergroup reward instead of IRPM-Mean (ablation) | IRPM Qwen3-32B · RM-Bench, JudgeBench, RewardBench, PPE | accuracy | Hurt |
| use IRPM-Interval (confidence-interval rule reward) instead of IRPM-Mean (ablation) | IRPM Qwen3-32B · four reward benchmarks | accuracy | Hurt |
| add inference-time voting (average sampled scores) | IRPM-Mean-32B · four reward benchmarks (average) | average accuracy: 77.2 -> 79.6 (voting@2); 77.2 -> 81.4 (voting@8) | Improved |
| train pointwise GRM with IRPM instead of the backbone alone | Llama-3.1-8B-Instruct · four reward benchmarks (average) | average accuracy: 40.1 -> 64.8 | Improved |
| use IRPM pointwise GRM vs Bradley-Terry scalar model on same 21k data | Qwen3-32B · RM-Bench + JudgeBench | overall accuracy: 71.6 -> 77.1 | Improved |
| use pairwise GRM instead of IRPM pointwise on same data | Qwen3-32B · RM-Bench + JudgeBench | overall accuracy: 77.1 -> 77.7 | No clear effect |
Closest papers by meaning
- Capturing Individual Human Preferences with Reward Features (NeurIPS 2025 · similarity 0.89)
- Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning (NeurIPS 2024 · similarity 0.89)
- Data with High and Consistent Preference Difference Are Better for Reward Model (AAAI 2025 · similarity 0.89)
- PILAF: Optimal Human Preference Sampling for Reward Modeling (ICML 2025 · similarity 0.89)
- Bayesian Preference Learning for Test-Time Steerable Reward Models (NeurIPS 2026 · similarity 0.88)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (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.
