Research map / Large language models / Reasoning and chain of thought
Reasoning via Implicit Self-supervised Emergence for Instruction Segmentation
Qing Zhou, Lichang Yang, Yuyu Jia, Junyu Gao, Weiping Ni, Junzheng Wu, Qi Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RISE, a framework that learns compositional reasoning for instruction-guided segmentation via reinforcement learning with GRPO and no ground-truth masks.
Problem
Instruction-guided segmentation tasks are assumed to require complex, explicit mask supervision; the paper challenges this assumption.
Approach
Learns intricate compositional reasoning (spatial relations to world knowledge) for segmentation through self-supervised reinforcement learning with GRPO, without any ground-truth masks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CLIP semantic reward to format rewards (ablation) | RISE (3B) · ReasonSeg test | gIoU: 28.6 -> 52.9 | Improved |
| add format rewards to semantic reward (ablation) | RISE (3B) · ReasonSeg test | gIoU: 51.5 -> 52.9 | Improved |
| add think format reward (ablation) | RISE (3B) · ReasonSeg test | cIoU: 41.1 -> 45.1 | Improved |
| add segmentation format reward (ablation) | RISE (3B) · ReasonSeg test | cIoU: 42.5 -> 45.1 | Improved |
| add explicit chain-of-thought block (ablation) | RISE (3B) · ReasonSeg test | cIoU: 38.5 -> 45.1 | Improved |
| add explicit chain-of-thought block (ablation) | RISE-ClsOnly (3B) · ReasonSeg test | cIoU: 32.0 -> 41.7 | Improved |
| require auxiliary points in output (ablation) | RISE (3B) · ReasonSeg test | cIoU: 42.8 -> 45.1 | Improved |
| require auxiliary points in output (ablation) | RISE (3B) · ReasonSeg test | gIoU: 52.2 -> 52.9 | No clear effect |
| scale reasoning agent from 3B to 7B | RISE · ReasonSeg test | gIoU: 52.9 -> 58.7 | Improved |
| train on 2K ImageNet class labels instead of RefCOCOg expressions | RISE-7B · ReasonSeg test | gIoU: 58.7 -> 54.5 | Hurt |
| RISE self-supervised RL on top of Qwen2.5VL-7B + SAM2 | Qwen2.5VL-7B + SAM2 · RefCOCO+ testA | cIoU: 73.5 -> 77.7 | Improved |
| RISE-3B-ClsOnly vs base Qwen2.5VL-3B | Qwen2.5VL-3B + SAM2 · RefCOCO testA | cIoU: 75.9 -> 76.3 | Improved |
Surprising findings
- Compositional reasoning for segmentation can emerge without any ground-truth masks or explicit complex supervision.
Experimental setup
Builds on: GRPO
Closest papers by meaning
- LENS: Learning to Segment Anything with Unified Reinforced Reasoning (AAAI 2026 · similarity 0.87)
- Incentivizing Reasoning for Advanced Instruction-Following of Large Language Models (NeurIPS 2025 · similarity 0.86)
- Fast Reasoning Segmentation for Images and Videos (CVPR 2026 · similarity 0.86)
- GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.86)
- SAM-R1: Leveraging SAM for Reward Feedback in Multimodal Segmentation via Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- CoPRS: Learning Positional Prior from Chain-of-Thought for Reasoning Segmentation (ICLR 2026 · similarity 0.85)
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.
