Research map / Vision-language and multimodal / Multimodal LLMs
Learning Active Perception for Pixel-Space Reasoning via Visual-Intent Stratified GRPO
Mingkang Zhu, Xi Chen, Senqiao Yang, Bei Yu, Hengshuang ZHAO, Jiaya Jia
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 |
|---|---|---|---|
| replace GRPO with VIS-GRPO (depth and intent stratification) | Qwen3-VL-8B · average of four benchmarks | accuracy: 74.2 -> 77.1 | Improved |
| replace GRPO with VIS-GRPO | Qwen3-VL-8B · HR-Bench 8K | accuracy: 77.6 -> 81.6 | Improved |
| replace GRPO with VIS-GRPO | Qwen3-VL-8B · MME-RealWorld | accuracy: 52.7 -> 56.3 | Improved |
| add visual-intent stratification on top of depth-only stratification (ablation) | VIS-GRPO · average of four benchmarks | accuracy: 74.8 -> 77.1 | Improved |
| add depth-only stratification to GRPO (ablation) | GRPO · MME-RealWorld | accuracy: 52.7 -> 55.7 | Improved |
| add depth-only stratification to GRPO (ablation) | GRPO · V* Bench | accuracy: 87.4 -> 84.8 | Hurt |
| add depth-only stratification to GRPO (ablation) | GRPO · average of four benchmarks | accuracy: 74.2 -> 74.8 | Improved |
| use full VIS-GRPO instead of Pixel Reasoner zoom-reward heuristics | Qwen3-VL-8B · average of four benchmarks | accuracy: 74.4 -> 77.1 | Improved |
| apply Pixel Reasoner reward heuristics instead of GRPO | Qwen3-VL-8B · V* Bench | accuracy: 87.4 -> 83.8 | Hurt |
| apply Pixel Reasoner reward heuristics instead of GRPO | Qwen3-VL-8B · HR-Bench 8K | accuracy: 77.6 -> 79.0 | Improved |
| apply GRPO RL training to base Qwen3-VL-8B | Qwen3-VL-8B · average of four benchmarks | accuracy: 71.5 -> 74.2 | Improved |
| add visual-intent stratification (VIS-GRPO) over depth-only on intent balance (ablation) | Depth-Stratified GRPO · MME-RealWorld multi-zoom | explore:verify ratio: 9:1 -> 7:3 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Pixel Reasoner — Pixel Reasoner: Incentivizing Pixel Space Reasoning via Curiosity-Driven Reinforcement Learning (NeurIPS 2025) · better on 5 of 5 shared results
- DeepEyes — DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design (CVPR 2026 · similarity 0.91)
- Mind the GAP: Glimpse-based Active Perception improves generalization and sample efficiency of visual reasoning (ICLR 2025 · similarity 0.89)
- Pixel Reasoner: Incentivizing Pixel Space Reasoning via Curiosity-Driven Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026 · similarity 0.89)
- ACTIVE-o3 : Empowering MLLMs with Active Perception via Pure Reinforcement Learning (ICML 2026 · similarity 0.88)
- The Perceptual Bandwidth Bottleneck in Vision-Language Models: Active Visual Reasoning via Sequential Experimental Design (ICML 2026 · similarity 0.88)
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.
