AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Segment-Aligned Policy Optimization for Multi-Modal Reasoning

Lei Gao, Zhuoming Li, Mengxi Jia, Jiakang Yuan, Hongbo Sun, Hao Sun, Xuelong Li

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
use step-wise importance sampling instead of token-level IS (ablation)SAPO (Qwen2.5-VL-7B) · Geo3K testaccuracy: 47.92 -> 52.75Improved
use step-wise importance sampling instead of token-level IS (ablation)SAPO (Qwen2.5-VL-7B) · LogicVistaaccuracy: 44.07 -> 47.30Improved
use step-wise importance sampling instead of token-level IS (ablation)SAPO (Qwen2.5-VL-7B) · MathVerseaccuracy: 43.91 -> 45.81Improved
use entropy-based segmentation instead of newline-based (ablation)SAPO (Qwen2.5-VL-7B) · Geo3K testaccuracy: 41.60 -> 52.75Improved
use entropy-based segmentation instead of probability-based (ablation)SAPO (Qwen2.5-VL-7B) · Geo3K testaccuracy: 44.43 -> 52.75Improved
use entropy-based segmentation instead of step-based (ablation)SAPO (Qwen2.5-VL-7B) · LogicVistaaccuracy: 48.10 vs 47.30Hurt
set entropy top-k from 10 up to 30 percent, then beyond 30 (ablation)SAPO · Geo3K testaccuracyMixed
replace PPO with SAPO (ablation)Qwen2.5-VL-3B · Geo3K testaccuracy: 21.80 -> 46.26Improved
replace GRPO with SAPOQwen2.5-VL-3B · WeMathaccuracy: 29.43 -> 36.19Improved
replace GRPO with SAPOQwen2.5-VL-7B · DynaMathaccuracy: 22.75 -> 21.96Hurt
replace GRPO with SAPOQwen2.5-VL-3B · LogicVistaaccuracy: 40.72 -> 39.60Hurt
apply SAPO to general reasoning without retuning kQwen2.5-VL-7B · avg of ScienceQA, MMMU, MMBench, MMStaraccuracy: 71.24 -> 72.36Improved

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