Research map / Vision-language and multimodal / Multimodal LLMs
StAR: Segment Anything Reasoner
Seokju Yun, Dongheon Lee, Noori Bae, Jaesung Jun, Chanseul Cho, Youngmin Ro
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 |
|---|---|---|---|
| LoRA (rank 16, lr 1e-5, no KL) then rank 64 instead of full tuning | VisionReasoner-7B · ReasonSeg-X test / ReasonSeg-R | gIoU: 42.2 -> 45.4 / 64.8 -> 66.1 | Improved |
| Add SAM-level tiered mask IoU reward and scale rollouts 8 to 16 | StAR stage-1 · ReasonSeg-X test / ReasonSeg-R | gIoU: 45.4 -> 47.1 / 66.1 -> 66.6 | Improved |
| Stage-2 training with REST on ReasonSeg-X train | StAR-7B · ReasonSeg-X / ReasonSeg-R | gIoU: +1.5 / +1.6 | Improved |
| Add label prediction before coordinates (ablation) | StAR-7B · ReasonSeg-X / ReasonSeg-R | gIoU: 48.6/68.2 -> 49.2/69.7 | Improved |
| Predict label last instead of first (ablation) | StAR-7B · ReasonSeg-X test | gIoU: 49.2 -> 46.3 | Hurt |
| Predict label last instead of first (ablation) | StAR-8B · ReasonSeg-R | cIoU: 69.2 -> 70.8 cIoU; gIoU 73.8 -> 73.1 | Mixed |
| Increase REST rollouts n from 16 to 256 (ablation) | StAR-8B + MV · ReasonSeg-X test | gIoU: 58.1 -> 59.6 | Improved |
| Increase REST rollouts n from 16 to 256 (ablation) | StAR-7B (no MV) · ReasonSeg-X test | gIoU: 48.4 -> 49.2 (peak 49.4 at n=128) | No clear effect |
| Add mask-level majority voting (32 samples) | StAR-7B · ReasonSeg-X test | gIoU: 49.2 -> 50.3 | Improved |
| Add mask-level majority voting | StAR-Qwen3-VL 32B · ReasonSeg-X test | gIoU: 61.7 -> 64.1 | Improved |
| Add mask-level majority voting | StAR-Qwen2.5-VL 7B · ReasonSeg-X C/MH | gIoU: 28.0 -> 27.9 | No clear effect |
| Scale base model from 8B to 32B | StAR · ReasonSeg-X test | gIoU: 57.9 -> 61.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DR2Seg — DR$^2$Seg: Decomposed Two-Stage Rollouts for Efficient Reasoning Segmentation in Multimodal Large Language Models (ICML 2026) · better on 1 of 1 shared result
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
- LISA: Reasoning Segmentation via Large Language Model (CVPR 2024 · similarity 0.87)
- Fast Reasoning Segmentation for Images and Videos (CVPR 2026 · similarity 0.87)
- No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers (ICLR 2026 · similarity 0.87)
- HoneyBee: Data Recipes for Vision-Language Reasoners (CVPR 2026 · similarity 0.86)
- LENS: Learning to Segment Anything with Unified Reinforced Reasoning (AAAI 2026 · similarity 0.86)
- DR$^2$Seg: Decomposed Two-Stage Rollouts for Efficient Reasoning Segmentation in Multimodal Large Language Models (ICML 2026 · similarity 0.86)
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.
