AI Paper Atlas

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

ICML 2026 · Reinforcement learning › RL from feedback · approach: Theory of RLHF · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add R2M to RLOORLOO, LLaMA3-8B-Instruct · AlpacaEval 2LC win rate: 28.4 -> 34.5Improved
add R2M to GRPOGRPO, LLaMA3-8B-Instruct · AlpacaEval 2LC win rate: 29.5 -> 35.6Improved
add R2M to RLOORLOO, Qwen2.5-3B-Instruct · AlpacaEval 2raw win rate: 26.0 -> 31.2Improved
add R2M to RLOORLOO, Pythia-2.8B · TL;DR summarizationraw win rate: 75.3 -> 81.6Improved
train R2M iteratively instead of using it frozen (w/o Train) (ablation)RLOO + R2M, LLaMA3-8B-Instruct · AlpacaEval 2LC win rate: 24.4 -> 34.5Improved
use real policy feedback instead of Gaussian noise (ablation)RLOO + R2M, LLaMA3-8B-Instruct · AlpacaEval 2LC win rate: 25.4 -> 34.5Improved
add BT loss term (ablation)RLOO + R2M, LLaMA3-8B-Instruct · AlpacaEval 2LC win rate: 31.5 -> 34.5Improved
add GRE loss term (ablation)RLOO + R2M, LLaMA3-8B-Instruct · AlpacaEval 2LC win rate: 32.3 -> 34.5Improved
use R2M instead of offline pretrained RMGRPO, Qwen2.5-3B-Instruct · AlpacaEval 2LC win rate: 22.9 -> 25.8Improved
iterative update of reward modelReward model · UltraFeedback test (Qwen2.5)RM accuracy: 72.3 -> 77.4Improved
update only scoring head (Iterative RM Head) instead of R2M (ablation)GRPO, LLaMA3-8B-Instruct · AlpacaEval 2LC win rate: 32.0 -> 35.6Improved
add R2M w/o Train to GRPOGRPO, Qwen2.5-3B-Instruct · MT-BenchGPT-4 score: 6.3 -> 6.2Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy