AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Beyond Euclidean Clipping: Overcoming Exploration Collapse in LLM RL via Riemannian Isometric Policy Optimization

Zhicheng Cai, Xinyuan Guo, Hanlin Wu, Mingxuan Wang, Wei-Ying Ma, Ya-Qin Zhang, Hao Zhou

ICML 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards · 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
replace GRPO clipping with RIPO isometric clippingGRPO · 7 math benchmarks average, Qwen3-8B-BaseAvg@8 (%): 28.5 -> 38.5Improved
replace GRPO clipping with RIPOGRPO · 7 math benchmarks average, Qwen3-1.7B-BaseAvg@8 (%): 11.2 -> 15.4Improved
replace GRPO clipping with RIPOGRPO · 7 math benchmarks average, Llama3.2-3B-InstructAvg@8 (%): 6.4 -> 8.6Improved
replace GRPO clipping with RIPOGRPO · 7 math benchmarks average, Qwen3-4B-BaseAvg@8 (%): 25.7 -> 30.1Improved
replace DCPO clipping with RIPODCPO · 7 math benchmarks average, Qwen3-8B-BaseAvg@8 (%): 34.5 -> 38.5Improved
replace DCPO clipping with RIPODCPO · 7 math benchmarks average, Llama3.2-3B-InstructAvg@8 (%): 6.9 -> 8.6Improved
use symmetric delta 0.05 instead of 0.02 (ablation)RIPO · AIME24, Qwen3-8B-BaseAvg@8 (%): 40.8 -> 43.8Improved
make delta_high larger than delta_low (0.05/0.02 vs 0.05/0.05) (ablation)RIPO · AIME24, Qwen3-8B-BaseAvg@8 (%): 43.8 -> 28.8Hurt
use asymmetric delta_low 0.08, delta_high 0.04 instead of symmetric 0.08 (ablation)RIPO · AIME24, Qwen3-8B-BaseAvg@8 (%): 42.1 -> 27.9Hurt
vary symmetric delta from 0.02 to 0.08 (ablation)RIPO · AIME24, Qwen3-8B-BaseAvg@8 (%): 40.8, 43.8, 42.1No clear effect
use RIPO-Clip instead of PPO-Clip in PPOPPO · GSM8K, Qwen2.5-0.5B-InstructAvg@1 (%): 58.0 -> 61.1Improved
use RIPO-Clip instead of PPO-Clip in PPOPPO · GSM8K, Qwen2.5-14B-InstructAvg@1 (%): 93.2 -> 94.4Improved

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