Research map / Vision-language and multimodal / Multimodal LLMs

Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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.

Builds Vision-DeepResearch, a multimodal deep-research MLLM trained with cold-start SFT on synthesized multi-turn visual+text search trajectories followed by GRPO RL, reporting SoTA among open models on six factual VQA benchmarks.

Problem

Existing multimodal deep-research MLLMs use naive full-image or entity-level visual queries that fail under noisy real-world search engines, and produce short trajectories that limit reasoning depth and search breadth on fact-intensive VQA.

Approach

Multi-entity, multi-scale visual cropping with hit-signal-controlled search depth, bridged to a text deep-research LLM via image descriptions to synthesize long (dozens of steps, hundreds of engine calls) multimodal trajectories, plus fuzzy multi-hop VQA synthesis via answer/entity obfuscation, then SFT and GRPO RL.

Key results

• Vision-DeepResearch-30B-A3B reaches 56.9 average across six benchmarks vs 43.2 for Qwen3-VL-30B-A3B-Thinking agentic and 50.7 for Gemini-2.5 Pro agentic.
• Vision-DeepResearch-8B reaches 50.5 average, +10.4 over Qwen3-VL-8B-Instruct agentic baseline.
• Ablation: full pipeline CIS+TS gives 40.0 avg vs 12.0 for direct answer and 16.0 for whole-image visual search only; RL adds 3.1 points on three benchmarks over SFT.

Surprising findings

• Whole-image visual search alone slightly degrades BC-VL performance relative to direct answering (27.6 to 26.1).
• Claude-3.7-Sonnet agentic reaches 27.2 on VDR, above GPT-5 agentic at 20.4.

Limitations

• RL was not scaled exhaustively due to API cost and wall-clock constraints.
• Judge and synthesis pipeline depend on proprietary or closed MLLM/LLM components, with unspecified judge model.
• Evaluation uses random 300-sample subsets for FVQA and LiveVQA, adding sampling variance.

Experimental setup

Datasets: VDR-Bench FVQA MMSearch-Plus MMSearch LiveVQA BrowseComp-VL

Benchmarks: VDR-Bench FVQA MMSearch-Plus MMSearch LiveVQA BrowseComp-VL

Metrics: accuracy

Baselines: GPT-5 Gemini-2.5 Pro Claude-4-Sonnet Qwen3-VL-30B-A3B-Thinking WebWatcher-32B MMSearch-R1-7B

Builds on: ReAct GRPO rLLM WebWatcher MMSearch-R1 DeepMMSearch-R1

Closest papers by meaning

• Accelerating Multimodal Large Language Models by Searching Optimal Vision Token Reduction (CVPR 2025 · similarity 0.90)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.90)
• CVSearch: Empowering Multimodal LLMs with Cognitive Visual Search for High-Resolution Image Perception (ICML 2026 · similarity 0.90)
• Mordal: Automated Pretrained Model Selection for Vision Language Models (ICLR 2026 · similarity 0.90)
• SearchLVLMs: A Plug-and-Play Framework for Augmenting Large Vision-Language Models by Searching Up-to-Date Internet Knowledge (NeurIPS 2024 · similarity 0.89)
• LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.89)