AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Unbiased Principles, Robust Rewards

Qingnan Ren, Zhen Fang, Shiting Huang, Yu Zeng, Lin Chen, Zehui Chen, Feng Zhao

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

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
generate principles from question only (Q->P) instead of Q+R->PGRM-16B (fine-tuned) · RewardBench (pointwise)overall accuracy (%): 61.65 -> 68.22Improved
decouple principle generation from responseGRM-16B · RewardBench Safetyaccuracy (%): 67.41 -> 78.22Improved
decouple principle generation from responseGRM-16B · RewardBench Reasoningaccuracy (%): 60.04 -> 65.67Improved
decouple principle generation from responseGRM-16B · RewardBench Chat Hardaccuracy (%): 55.83 -> 56.48Improved
condition principles on chosen response R+ (biased principles) (ablation)DeepSeek-GRM-16B · RewardBenchscore of R+ vs unbiased P0: +6.6%Hurt
condition principles on rejected response R- (ablation)DeepSeek-GRM-16B · RewardBenchscore of R- vs unbiased P0: +4.6%Hurt
train actor with IP-GRM reward instead of GRMQwen3-8B · WritingBench (all)Avg@3 score: 85.2 -> 87.6Improved
train actor with IP-GRM reward instead of GRMQwen3-8B · Creative Writing v3Avg@3 score: 74.2 -> 77.8Improved
train actor with IP-GRM reward instead of GRMQwen3-8B-base · WritingBench (all)Avg@3 score: 76.9 -> 79.7Improved
train actor with IP-GRM reward instead of GRMQwen3-8B-base · Creative Writing v3Avg@3 score: 68.7 -> 74.5Improved
train actor with IP-GRM reward instead of scalar RMQwen3-8B · WritingBench (all)Avg@3 score: 85.0 -> 87.6Improved
use IP-GRM instead of GRM as GRPO rewardQwen3-8B · WritingBench held-outtest score (0-10): 8.05 -> 8.75Improved

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