Research map / Vision-language and multimodal / Multimodal LLMs
DGSeg: Dynamic Gating of Semantic-Spatial Guided Predictions for Reasoning Segmentation
Ruizhe Zeng, Siyu Cao, Lu Zhang, Zhi-yong Liu
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 |
|---|---|---|---|
| fuse separate semantic and spatial branch predictions with dynamic gating instead of joint cue encoding (no fusion) (ablation) | DGSeg (Qwen2.5-VL-3B) · ReasonSeg val | gIoU: 60.8 -> 66.0 | Improved |
| use learnable fusion instead of average fusion (ablation) | DGSeg · ReasonSeg val | gIoU: 63.8 -> 66.0 | Improved |
| learned scalar weight instead of non-learnable confidence weighting (ablation) | DGSeg · ReasonSeg val/test | gIoU: val 64.1 -> 64.0; test 57.2 -> 57.8 | Mixed |
| replace dynamic gating with MoE fusion (ablation) | DGSeg · ReasonSeg val/test | gIoU: 66.0 -> 65.6 (val), 60.0 -> 60.6 (test) | No clear effect |
| add dual-branch gating on top of single best branch (ablation) | DGSeg · ReasonSeg val | gIoU: 63.4 (spatial-only) -> 66.0 | Improved |
| add RL finetuning (Stage 1) of the MLLM (ablation) | DGSeg · ReasonSeg val | gIoU: 36.0 -> 66.0 | Improved |
| add semantic reward based on SAM3 mask IoU (ablation) | Stage 1 RL · ReasonSeg val | gIoU: 62.1 -> 66.0 | Improved |
| use segmentation-model preference reward instead of category-label reward (ablation) | Stage 1 RL · ReasonSeg val | gIoU: 64.7 -> 66.0 | Improved |
| add spatial (box IoU) reward (ablation) | Stage 1 RL · ReasonSeg val | gIoU: 64.1 -> 66.0 | Improved |
| add format reward (ablation) | Stage 1 RL · ReasonSeg val | gIoU: 64.6 -> 66.0 | Improved |
| add dual-branch dynamic gating overhead | DGSeg · efficiency | FLOPs / FPS: +0.3% FLOPs, -2.0% FPS, +4.2% memory | No clear effect |
| scale MLLM from 3B to 7B | DGSeg · ReasonSeg test | gIoU: 60.0 -> 67.3 | 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.
- SAM-Veteran — SAM-Veteran: An MLLM-Based Human-like SAM Agent for Reasoning Segmentation (ICLR 2026) · better on 2 of 2 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$^2$Seg: Decomposed Two-Stage Rollouts for Efficient Reasoning Segmentation in Multimodal Large Language Models (ICML 2026 · similarity 0.90)
- Seg-ReSearch: Segmentation with Interleaved Reasoning and External Search (ICML 2026 · similarity 0.88)
- LISA: Reasoning Segmentation via Large Language Model (CVPR 2024 · similarity 0.88)
- SegCompass: Exploring Interpretable Alignment with Sparse Autoencoders for Enhanced Reasoning Segmentation (CVPR 2026 · similarity 0.88)
- CoPRS: Learning Positional Prior from Chain-of-Thought for Reasoning Segmentation (ICLR 2026 · similarity 0.88)
- 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.
