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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use step-wise importance sampling instead of token-level IS (ablation) | SAPO (Qwen2.5-VL-7B) · Geo3K test | accuracy: 47.92 -> 52.75 | Improved |
| use step-wise importance sampling instead of token-level IS (ablation) | SAPO (Qwen2.5-VL-7B) · LogicVista | accuracy: 44.07 -> 47.30 | Improved |
| use step-wise importance sampling instead of token-level IS (ablation) | SAPO (Qwen2.5-VL-7B) · MathVerse | accuracy: 43.91 -> 45.81 | Improved |
| use entropy-based segmentation instead of newline-based (ablation) | SAPO (Qwen2.5-VL-7B) · Geo3K test | accuracy: 41.60 -> 52.75 | Improved |
| use entropy-based segmentation instead of probability-based (ablation) | SAPO (Qwen2.5-VL-7B) · Geo3K test | accuracy: 44.43 -> 52.75 | Improved |
| use entropy-based segmentation instead of step-based (ablation) | SAPO (Qwen2.5-VL-7B) · LogicVista | accuracy: 48.10 vs 47.30 | Hurt |
| set entropy top-k from 10 up to 30 percent, then beyond 30 (ablation) | SAPO · Geo3K test | accuracy | Mixed |
| replace PPO with SAPO (ablation) | Qwen2.5-VL-3B · Geo3K test | accuracy: 21.80 -> 46.26 | Improved |
| replace GRPO with SAPO | Qwen2.5-VL-3B · WeMath | accuracy: 29.43 -> 36.19 | Improved |
| replace GRPO with SAPO | Qwen2.5-VL-7B · DynaMath | accuracy: 22.75 -> 21.96 | Hurt |
| replace GRPO with SAPO | Qwen2.5-VL-3B · LogicVista | accuracy: 40.72 -> 39.60 | Hurt |
| apply SAPO to general reasoning without retuning k | Qwen2.5-VL-7B · avg of ScienceQA, MMMU, MMBench, MMStar | accuracy: 71.24 -> 72.36 | Improved |
Closest papers by meaning
- DAPO : Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage-Based Policy Optimization (NeurIPS 2025 · similarity 0.88)
- Segment Policy Optimization: Effective Segment-Level Credit Assignment in RL for Large Language Models (NeurIPS 2025 · similarity 0.88)
- Agentic Reinforced Policy Optimization (ICLR 2026 · similarity 0.88)
- Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.88)
- Hybrid Latent Reasoning via Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Accelerating RL for LLM Reasoning with Optimal Advantage Regression (NeurIPS 2025 · similarity 0.87)
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.
