AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Tackling Length Inflation Without Trade-offs: Group Relative Reward Rescaling for Reinforcement Learning

Zichao Li, Jie Lou, Fangchen Dong, Zhiyuan Fan, Mengjie Ren, Hongyu Lin, Xianpei Han, Debing Zhang, Le Sun, Yaojie Lu, XingYu Li

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 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
use GR3 multiplicative reward rescaling instead of GRPOGRPO / GR3 · AIME24, R1-Distill-7BAvg@32: 57.1 -> 60.1Improved
use GR3 instead of GRPO, tokensGRPO / GR3 · AIME24, R1-Distill-7Btokens: 11079 -> 7923Improved
use GR3 instead of initial modelGRPO / GR3 · AIME25, R1-Distill-7BAvg@32 and tokens: 39.4 -> 46.9; 14032 -> 8582Improved
use GR3 instead of GRPO on 1.5BGRPO / GR3 · AIME24, R1-Distill-1.5BAvg@32: 39.6 -> 45.2Improved
use GR3 instead of GRPO on codeGRPO / GR3 · LiveCodeBench v6, R1-Distill-7Bscore: 42.4 -> 41.6; tokens 10956 -> 7504Mixed
use GR3 instead of GRPO in RLHFGRPO / GR3 · Arena-Hard-Auto, Qwen3-8Btokens: 2343 -> 1178Improved
use GR3 instead of GRPO in RLHFGRPO / GR3 · AlpacaEval, Qwen3-4Bscore: 33.9 -> 44.1Improved
use threshold truncation (4k) instead of GRPOGRPO / GR3 · AIME24, R1-Distill-1.5B step 800score: 41.0 -> 32.5Hurt
use Efficiently group-relative penalty alpha 0.2 instead of GRPOGRPO / GR3 · MATH500, R1-Distill-1.5Bscore: 87.2 -> 60.5Hurt
increase length penalty alpha from 0.33 to 1.00 (ablation)GRPO / GR3 · AIME24, R1-Distill-1.5BAvg@32: 45.2 -> 34.9Hurt
decrease alpha from 0.33 to 0.10 (ablation)GRPO / GR3 · AIME24, R1-Distill-1.5BAvg@32: 45.2 -> 43.2Hurt
decrease alpha from 0.33 to 0.20 (ablation)GRPO / GR3 · AIME24, R1-Distill-1.5BAvg@32: 45.2 -> 45.1No 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