AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

From Absolute to Relative: Rethinking Reward Shaping in Group-Based Reinforcement Learning

Wenzhe Niu, Wei He, Zongxia Xie, Jinpeng Ou, Huichuan Fan, Yuchen Ge, Yanru Sun, Ziyin Wang, Yizhao Sun, Chengshun Shi, Jiuchong Gao, Jinghua Hao and 1 more

ICML 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 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 correctness to hierarchical re-ranking (ablation)RLRR · Math (6 benchmarks)average accuracy (%): 45.7 -> 47.1Improved
add length bins to hierarchical re-ranking (ablation)RLRR · Math (6 benchmarks)average accuracy (%): 47.2 -> 47.1Hurt
add length bins to hierarchical re-ranking (ablation)RLRR · Math (6 benchmarks)average response length (tokens): 6317 -> 6013Improved
replace random ranking with Ranking RM within correctness groups (ablation)RLRR · Math avg (AIME24, AIME25, Minerva, AMC)average accuracy (%): 37.5 -> 38.0Improved
add correctness re-ranking with random order to GRPO (ablation)GRPO · Math avg (AIME24, AIME25, Minerva, AMC)average accuracy (%): 36.3 -> 37.5Improved
replace GRPO with RLRR (HRR) on 1.5BDeepSeek-R1-Distill-Qwen-1.5B · Logic reasoning avgaverage accuracy (%): 12.5 -> 24.8Improved
replace GRPO with RLRR (HRR) on 8BDeepSeek-R1-Distill-Llama-8B · Math reasoning avgaverage accuracy (%): 62.9 -> 65.9Improved
use PRR instead of HRR (no rule-based supervision) (ablation)RLRR · Math reasoning avg, 8Baverage accuracy (%): 65.9 -> 64.1Hurt
add RLRR to CISPOCISPO · Math avgaverage accuracy (%): 36.1 -> 37.1Improved
add RLRR to RLOORLOO · Math avgaverage accuracy (%): 36.1 -> 37.1Improved
add RLRR to GRPO at group size 4GRPO · Math avg, G=4average accuracy (%): 54.1 -> 54.5Improved
add RLRR to GRPO at group size 16GRPO · Math avg, G=16average accuracy (%): 55.4 -> 55.4No clear effect

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