AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

EGM: Efficient Visual Grounding Language Models

Guanqi Zhan, Changye Li, Zhijian Liu, Yao Lu, Yi Wu, Song Han, Ligeng Zhu

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 3 citations (OpenAlex)

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 changedWhereMeasuredEffect
add EGM reasoning SFT+RL trainingQwen3-VL-2B-Thinking · RefCOCO/+/g (8-split average)accuracy: 83.6 -> 89.6Improved
add EGM reasoning SFT+RL trainingQwen3-VL-8B-Thinking · RefCOCO/+/g (8-split average)accuracy: 87.8 -> 91.4Improved
add EGM reasoning SFT+RL trainingInternVL-3-1B · RefCOCO/+/g (8-split average)accuracy: 81.6 -> 86.8Improved
add EGM reasoning SFT+RL trainingInternVL-3-8B · RefCOCO/+/g (8-split average)accuracy: 89.6 -> 90.7Improved
add the reasoning process to EGM (ablation)InternVL-3-8B · RefCOCO/+/g (8-split average)accuracy: 87.7 -> 90.7Improved
add the reasoning process to EGM (ablation)Qwen3-VL-8B-Thinking · RefCOCO/+/g (8-split average)accuracy: 88.6 -> 91.4Improved
add the reasoning process to EGM (ablation)Qwen3-VL-2B-Thinking · RefCOCO/+/g (8-split average)accuracy: 88.7 -> 89.6Improved
add the reasoning process to EGM (ablation)InternVL-3-1B · RefCOCO/+/g (8-split average)accuracy: 83.6 -> 86.8Improved
add RL stage after SFT (ablation)Qwen3-VL-8B-Thinking · RefCOCO/+/g (8-split average)accuracy: 88.8 -> 91.4Improved
add SFT stage on reasoning data (ablation)Qwen3-VL-4B-Thinking · RefCOCO/+/g (8-split average)accuracy: 87.2 -> 87.9Improved
use Thinking instead of Instruct variantQwen3-VL 2B/4B/8B · RefCOCO/+/g (8-split average)accuracy: -0.6/-0.1/-0.5No clear effect
add EGM amodal reasoning trainingInternVL-3-8B · Amodal grounding (COCO-Amodal)accuracy: 62.0 -> 73.5Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy