Research map / Vision-language and multimodal / Multimodal LLMs
Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding
Shihao Wang, Shilong Liu, Yuanguo Kuang, Xinyu Wei, Yangzhou Liu, Zhiqi Li, Yunze Man, Guo Chen, Andrew Tao, Guilin Liu, Jan Kautz, Lei Zhang and 1 more
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 |
|---|---|---|---|
| use box-aligned parallel box decoding (Slow mode) instead of quantized coordinate tokens (ablation) | LocateAnything-3B · COCO ablation | F1: 50.1 -> 52.1 | Improved |
| use quantized coordinate tokens instead of textual digits (ablation) | NTP baseline · COCO ablation | F1: 49.1 -> 50.1 | Improved |
| switch from Slow (NTP) to Fast (MTP) decoding (ablation) | LocateAnything-3B · COCO ablation | F1 and throughput: 52.1 -> 49.6 F1, 3.9 -> 16.9 BPS | Mixed |
| add hybrid NTP fallback on top of Fast mode (ablation) | LocateAnything-3B · COCO ablation | F1 and throughput: 49.6 -> 51.6 F1, 16.9 -> 13.2 BPS | Improved |
| add joint NTP and block losses instead of NTP loss only (Slow mode) (ablation) | LocateAnything-3B · COCO ablation | F1: 50.1 -> 52.1 | Improved |
| add NTP loss to block-loss-only training (Fast mode) (ablation) | LocateAnything-3B · COCO ablation | F1: 47.2 -> 49.6 | Improved |
| replace generic block MTP (SDLM-B6) with box-aligned PBD (ablation) | MTP formulation · COCO ablation | F1 and throughput: 46.1 -> 49.6 F1, 5.5 -> 16.9 BPS | Improved |
| increase generic MTP block size from 4 to 8 (SDLM) (ablation) | SDLM · COCO ablation | F1 and throughput: 46.5 -> 45.8 F1, 5.2 -> 6.7 BPS | Hurt |
| order boxes by top-left distance instead of center, area or random (ablation) | LocateAnything-3B · COCO ablation | F1 | Improved |
| use parallel box decoding instead of NTP when target boxes grow from 20 to 300 | LocateAnything-3B · Throughput vs box count | throughput: 12 BPS to nearly 25 BPS | Improved |
| train LocateAnything on 138M-query data with PBD versus Rex-Omni | LocateAnything-3B · LVIS (zero-shot) | mean F1: 46.9 -> 50.7 | Improved |
| LocateAnything-3B versus Rex-Omni-3B on Dense200 and VisDrone | LocateAnything-3B · VisDrone | mean F1: 35.8 -> 39.9 | Improved |
Closest papers by meaning
- Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding (CVPR 2025 · similarity 0.89)
- GroundingME: Exposing the Visual Grounding Gap in MLLMs through Multi-Dimensional Evaluation (CVPR 2026 · similarity 0.89)
- IVC-Prune: Revealing the Implicit Visual Coordinates in LVLMs for Vision Token Pruning (ICLR 2026 · similarity 0.88)
- Grounded 3D-Aware Spatial Vision-Language Modeling (CVPR 2026 · similarity 0.88)
- From Objects to Anywhere: A Holistic Benchmark for Multi-level Visual Grounding in 3D Scenes (NeurIPS 2025 · similarity 0.88)
- VGent: Visual Grounding via Modular Design for Disentangling Reasoning and Prediction (CVPR 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.
