Research map / Vision-language and multimodal / Multimodal LLMs
Exploring Efficient Reasoning Segmentation with Small Language Models
Changsong Wen, Zelin Peng, Yu Huang, Xiaokang Yang, Wei Shen
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 |
|---|---|---|---|
| add register tokens to the <SEG> token (ablation) | LReSeg · ReasonSeg val | cIoU: 47.8 -> 52.8 | Improved |
| add register tokens to the <SEG> token (ablation) | LReSeg · ReasonSeg test | cIoU: 46.6 -> 50.0 | Improved |
| add <SEG> token alongside register tokens (ablation) | LReSeg · ReasonSeg val | cIoU: 48.2 -> 52.8 | Improved |
| use learnable instead of fixed register tokens (ablation) | LReSeg · ReasonSeg val | cIoU: 48.7 -> 52.8 | Improved |
| select visual tokens by attention instead of randomly (ablation) | LReSeg · ReasonSeg val | cIoU: 50.8 -> 52.8 | Improved |
| selective injection instead of uniform fusion (ablation) | LReSeg · ReasonSeg val | cIoU: 50.6 -> 52.8 | Improved |
| inject at one selected layer instead of every layer (ablation) | LReSeg · ReasonSeg val | cIoU: 52.1 -> 52.8 | Improved |
| keep causal attention mask (ablation) | LReSeg · ReasonSeg val | cIoU: 49.5 -> 52.8 | Improved |
| keep vision-to-text attention mask (ablation) | LReSeg · ReasonSeg val | cIoU: 52.2 -> 52.8 | Improved |
| use DINO unified encoder instead of SigLIP (ablation) | LReSeg · ReasonSeg val | cIoU: 50.5 -> 52.8 | Improved |
| use DINO unified encoder instead of CLIP (ablation) | LReSeg · ReasonSeg val | cIoU: 49.4 -> 52.8 | Improved |
| use unified DINO encoder instead of DINO & SAM (ablation) | LReSeg · ReasonSeg val | cIoU: 51.2 -> 52.8 | Improved |
Closest papers by meaning
- LISA: Reasoning Segmentation via Large Language Model (CVPR 2024 · similarity 0.91)
- DR$^2$Seg: Decomposed Two-Stage Rollouts for Efficient Reasoning Segmentation in Multimodal Large Language Models (ICML 2026 · similarity 0.89)
- CoPRS: Learning Positional Prior from Chain-of-Thought for Reasoning Segmentation (ICLR 2026 · similarity 0.87)
- FineRS: Fine-grained Reasoning and Segmentation of Small Objects with Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Fast Reasoning Segmentation for Images and Videos (CVPR 2026 · similarity 0.87)
- Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design (CVPR 2026 · 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.
