AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Grasp Any Region: Prompting MLLM to Understand the Dense World

Haochen Wang, Yuhao Wang, Tao Zhang, Yikang Zhou, Yanwei Li, Jiacong Wang, Jiani zheng, Ye Tian, Jiahao Meng, Zilong Huang, Guangcan Mai, Anran Wang and 4 more

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Grasp Any Region (GAR), a region-level MLLM that models given regions together with global context to understand dense scenes; details beyond the abstract are not available.

Problem

Region-level MLLMs understand given regions in isolation and neglect global context, which hurts fine-grained understanding of dense scenes with intricate details and object inter-relationships.

Approach

Prompts the MLLM with regions while explicitly incorporating global context, so region understanding accounts for object inter-relationships rather than isolated regions. Full method details are not available from the abstract provided.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add RoI-aligned feature replay instead of local-only image (ablation)GAR-1B (PerceptionLM-1B) · GAR-Bench-Capaccuracy: 20.1 -> 57.5Improved
replace DAM-like gated cross-attention with RoI-aligned feature replay (ablation)GAR-1B · GAR-Bench-VQAaccuracy: 40.0 -> 50.6Improved
replace global+cropped local image with RoI-aligned feature replay (ablation)GAR-1B · GAR-Bench-Capaccuracy: 28.4 -> 57.5Improved
replace global+cropped local image with RoI replay (ablation)GAR-1B · DLC-Bench (positive)accuracy: 70.1 -> 66.2Hurt
RoI replay across base models (ablation)GAR · GAR-Bench-Cap (Qwen2.5-VL-3B)accuracy: 34.3 -> 41.2Improved
RoI replay across base models (ablation)GAR · GAR-Bench-Cap (InternVL3-2B)accuracy: 32.8 -> 43.1Improved
add Fine-Grained Dataset-456K to seed data (ablation)GAR-1B · DLC-Benchaccuracy: 74.4 -> 77.5Improved
add Relation Dataset-414K (ablation)GAR-1B · GAR-Bench-Capaccuracy: 14.2 -> 57.5Improved
add Relation Dataset-414K (ablation)GAR-1B · DLC-Benchaccuracy: 77.5 -> 77.1No clear effect
fine-tune PerceptionLM-8B into GAR-8B for region understandingPerceptionLM-8B · V* (general)accuracy: 69.1 -> 59.2Hurt
add 600K general VQA data to GAR-8BGAR-8B · MMStaraccuracy: 43.9 -> 51.6Improved
use GAR-8B zero-shot vs in-domain VideoRefer-7BGAR-8B · VideoRefer-Bench-Qaccuracy: 71.9 -> 72.0Improved

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