Research map / Vision-language and multimodal / Multimodal LLMs

MM-DeepResearch: A Simple and Effective Multimodal Agentic Search Baseline

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

Proposes MM-DeepResearch, a multimodal deep research agent trained with hypergraph-based QA synthesis (Hyper-Search), tree-search trajectory synthesis over tool experts (DR-TTS), and an offline search engine for agentic RL, reaching 56.1 (7B) and 63.0 (8B) average across six benchmarks.

Problem

Developing multimodal deep research agents is hampered by scarce search-intensive multimodal QA data, lack of effective multi-turn multi-tool search trajectories, and the high cost of using online search APIs during agentic RL training.

Approach

Model web images and webpages as hypergraph nodes connected by expansion hyperedges to generate multi-hop multimodal QA; decompose search tools into per-tool RL experts and recompose them in a tree search to synthesize SFT trajectories; and replace online search during RL with a pre-built offline multimodal corpus and retriever.

Key results

• MM-DeepResearch-8B reaches 63.0 average across six benchmarks, versus 46.2 for Qwen3-VL-8B and 59.6 for SenseNova-MARS-8B.
• MM-DeepResearch-7B averages 56.1, above MMSearch-R1-7B (49.0) and Visual-ARFT-7B (33.1).
• On the MMSearch ablation, Hyper-Search RL data raises avg tool calls to 2.3 and MMSearch to 67.8 versus 62.1 for InfoSeek data.

Surprising findings

• Offline search training transfers well: offline-retrieval-trained model scores 62.7 on MMSearch vs 67.8 online at evaluation, and the paper reports that jina-embeddings-v4 vs e5 differs only marginally.

Limitations

• Offline search engine underperforms online search at evaluation (e.g., MMSearch 62.7 vs 67.8), so offline training is an approximation of real web conditions.
• Reward relies on an LLM judge (Qwen3-Next-80B) rather than rule-based checking, adding cost and possible judge bias.

Experimental setup

Datasets: InfoSeek FVQA Hyper-Search-3K Wikipedia (2018)

Benchmarks: SimpleVQA MMSearch LiveVQA FVQA-test InfoSeek Browsecomp-VL

Metrics: accuracy

Baselines: GPT-4o GPT-5 Qwen3-VL-8B Qwen3-VL-32B DeepMMSearch-R1-7B Visual-ARFT-7B MMSearch-R1-7B DeepEyes-v2-7B WebWatcher-7B SenseNova-MARS-8B

Builds on: GRPO DR-TTS Search-R1 FlashRAG VeRL

Closest papers by meaning

• OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents (NeurIPS 2026 · similarity 0.91)
• HierSearch: A Hierarchical Enterprise Deep Search Framework Integrating Local and Web Searches (AAAI 2026 · similarity 0.88)
• Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.88)
• MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains (ICLR 2026 · similarity 0.88)
• Quest: Training Frontier Deep Research Agents with Fully Synthetic Tasks (NeurIPS 2026 · similarity 0.88)
• DeepEyesV2: Toward Agentic Multimodal Model (ICLR 2026 · similarity 0.87)