Research map / Large language models / Alignment and preferences
GRAM-R²: Self-Training Generative Foundation Reward Models for Reward Reasoning
Chenglong Wang, Yongyu Mu, Hang Zhou, Yifu Huo, Ziming Zhu, Jiali Zeng, Murun Yang, Bei Li, Xiaoyang Hao, Chunliang Zhang, Fandong Meng, Jingbo Zhu and 1 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 reward reasoning with self-training on unlabeled data (vs Generative RM) | LLaMA-3.1-8B-Instruct reward model · RM-Bench Overall | accuracy: 79.2 -> 85.7 | Improved |
| add reward reasoning with self-training (vs Generative RM) | LLaMA-3.1-8B-Instruct reward model · JudgeBench Overall | accuracy: 76.9 -> 81.0 | Improved |
| add reward reasoning with self-training (vs Generative RM) | LLaMA-3.2-3B-Instruct reward model · RM-Bench Overall | accuracy: 77.1 -> 83.8 | Improved |
| add voting@16 at test time | GRAM-R2 8B · RM-Bench Overall | accuracy: 85.7 -> 86.1 | Improved |
| add reward reasoning (preference-proving rationales) (ablation) | GRAM-R2 · RM-Bench Overall | accuracy: 80.4 -> 85.7 | Improved |
| add preference-proving model during self-training on unlabeled data (ablation) | GRAM-R2 · RM-Bench Overall | accuracy: 84.1 -> 85.7 | Improved |
| add rationale-free labeled data pre-training (ablation) | GRAM-R2 · RM-Bench Overall | accuracy: 82.6 -> 85.7 | Improved |
| use trained preference-proving model Qwen3-14B instead of prompting GPT-4o (ablation) | GRAM-R2-100k · RM-Bench | accuracy: 64.7 -> 74.3 | Improved |
| scale preference-proving model from Qwen3-14B to Qwen3-32B (ablation) | GRAM-R2-100k · RM-Bench | accuracy: 74.3 -> 74.8 | No clear effect |
| use GRAM-R2 as PPO reward model instead of generative baseline | Llama-3.1-8B policy · AlpacaEval2 WinRate | win rate: 11.62 -> 15.62 | Improved |
| use GRAM-R2 as PPO reward model instead of Skywork-Reward-8B | Llama-3.1-8B policy · AlpacaEval2 LC-WinRate | length-controlled win rate: 8.03 -> 13.80 | Improved |
| fine-tune GRAM-R2 instead of vanilla generative RM on 1K STEM samples | GRAM-R2 · HelpSteer3 STEM | accuracy: +17.1 points | Improved |
Later papers that compared with this method
1 later accepted paper reports GRAM-r2 as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Bayesian Preference Learning for Test-Time Steerable Reward Models (NeurIPS 2026 · better on 1 of 1 shared result)
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
- GRAM: A Generative Foundation Reward Model for Reward Generalization (ICML 2025 · similarity 0.93)
- RM-R1: Reward Modeling as Reasoning (ICLR 2026 · similarity 0.93)
- Think-RM: Enabling Long-Horizon Reasoning in Generative Reward Models (NeurIPS 2025 · similarity 0.92)
- Reward Reasoning Models (NeurIPS 2025 · similarity 0.90)
- Active Reward Modeling: Adaptive Preference Labeling for Large Language Model Alignment (ICML 2025 · similarity 0.90)
- Tool-Augmented Reward Modeling (ICLR 2024 · similarity 0.90)
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.
