Research map / Reinforcement learning / RL from feedback
Real-Time Aligned Reward Model beyond Semantics
Zixuan Huang, Xin Xia, Yuxi Ren, Jianbin Zheng, Xuefeng Xiao, Hongyan Xie, Li Huaqiu, Songshi Liang, Zhongxiang Dai, Fuzhen Zhuang, Jianxin Li, Yikun Ban 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 R2M to RLOO | RLOO, LLaMA3-8B-Instruct · AlpacaEval 2 | LC win rate: 28.4 -> 34.5 | Improved |
| add R2M to GRPO | GRPO, LLaMA3-8B-Instruct · AlpacaEval 2 | LC win rate: 29.5 -> 35.6 | Improved |
| add R2M to RLOO | RLOO, Qwen2.5-3B-Instruct · AlpacaEval 2 | raw win rate: 26.0 -> 31.2 | Improved |
| add R2M to RLOO | RLOO, Pythia-2.8B · TL;DR summarization | raw win rate: 75.3 -> 81.6 | Improved |
| train R2M iteratively instead of using it frozen (w/o Train) (ablation) | RLOO + R2M, LLaMA3-8B-Instruct · AlpacaEval 2 | LC win rate: 24.4 -> 34.5 | Improved |
| use real policy feedback instead of Gaussian noise (ablation) | RLOO + R2M, LLaMA3-8B-Instruct · AlpacaEval 2 | LC win rate: 25.4 -> 34.5 | Improved |
| add BT loss term (ablation) | RLOO + R2M, LLaMA3-8B-Instruct · AlpacaEval 2 | LC win rate: 31.5 -> 34.5 | Improved |
| add GRE loss term (ablation) | RLOO + R2M, LLaMA3-8B-Instruct · AlpacaEval 2 | LC win rate: 32.3 -> 34.5 | Improved |
| use R2M instead of offline pretrained RM | GRPO, Qwen2.5-3B-Instruct · AlpacaEval 2 | LC win rate: 22.9 -> 25.8 | Improved |
| iterative update of reward model | Reward model · UltraFeedback test (Qwen2.5) | RM accuracy: 72.3 -> 77.4 | Improved |
| update only scoring head (Iterative RM Head) instead of R2M (ablation) | GRPO, LLaMA3-8B-Instruct · AlpacaEval 2 | LC win rate: 32.0 -> 35.6 | Improved |
| add R2M w/o Train to GRPO | GRPO, Qwen2.5-3B-Instruct · MT-Bench | GPT-4 score: 6.3 -> 6.2 | Hurt |
Closest papers by meaning
- Sequence to Sequence Reward Modeling: Improving RLHF by Language Feedback (AAAI 2025 · similarity 0.91)
- Provably Efficient Online RLHF with One-Pass Reward Modeling (NeurIPS 2025 · similarity 0.91)
- Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms (NeurIPS 2024 · similarity 0.91)
- Generative RLHF-V: Learning Principles from Multi-modal Human Preference (NeurIPS 2025 · similarity 0.91)
- Mitigating Reward Over-Optimization in RLHF via Behavior-Supported Regularization (ICLR 2025 · similarity 0.91)
- Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs (NeurIPS 2024 · 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.
