Paper 2601.22060
Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models
- Published
- Jan 2026
- Research lab
- Independent
- Citations
- 25
- GitHub
- 669 stars
01 In brief
Summary
Vision-DeepResearch introduces a new paradigm for multimodal deep research, addressing the hit-rate problem in image search and the limited reasoning depth and search breadth of existing methods.
The approach enables multi-turn, multi-entity, and multi-scale visual and textual search, supporting dozens of reasoning steps and hundreds of engine interactions.
A highly automated data pipeline synthesizes factual VQA instances and multi-turn trajectories, which are used to train the MLLM via supervised fine-tuning (SFT) and reinforcement learning (RL).
The model outperforms prior multimodal deep-research MLLMs and agent workflows built on strong closed-source models like GPT-5, Gemini-2.5-Pro, and Claude-4-Sonnet across six benchmarks.
Key components include multi-entity and multi-scale visual cropping, text bridging with a deep-research LLM, and an obfuscated termination strategy.
Training uses 30K trajectories for SFT and 15K VQA instances for RL with GRPO, employing an asynchronous rollout architecture for efficiency.
Ablations show that multi-scale cropping and text search are jointly necessary, and RL further improves performance by 3.1% on average.
The model achieves state-of-the-art results, with Vision-DeepResearch-30B-A3B reaching 56.9% average accuracy, a +16.0% improvement over the base agentic model.
The code is released at https://github.com/Osilly/Vision-DeepResearch.
The work highlights the importance of long-horizon, trial-and-error search behavior for robust multimodal reasoning in noisy web environments.
The authors expect further gains from scaling RL training.
The paper includes detailed ablations on pipeline components and training data, confirming the effectiveness of the proposed approach.
The model's success suggests that deep-research capabilities can be effectively internalized into MLLMs through careful data synthesis and training.
The work provides…
02 From the paper
Abstract
Multimodal large language models (MLLMs) have achieved remarkable success across a broad range of vision tasks. However, constrained by the capacity of their internal world knowledge, prior work has proposed augmenting MLLMs by ``reasoning-then-tool-call'' for visual and textual search engines to obtain substantial gains on tasks requiring extensive factual information. However, these approaches typically define multimodal search in a naive setting, assuming that a single full-level or entity-level image query and few text query suffices to retrieve the key evidence needed to answer the question, which is unrealistic in real-world scenarios with substantial visual noise. Moreover, they are often limited in the reasoning depth and search breadth, making it difficult to solve complex questions that require aggregating evidence from diverse visual and textual sources. Building on this, we propose Vision-DeepResearch, which proposes one new multimodal deep-research paradigm, i.e., performs multi-turn, multi-entity and multi-scale visual and textual search to robustly hit real-world search engines under heavy noise. Our Vision-DeepResearch supports dozens of reasoning steps and hundreds of engine interactions, while internalizing deep-research capabilities into the MLLM via cold-start supervision and RL training, resulting in a strong end-to-end multimodal deep-research MLLM. It substantially outperforming existing multimodal deep-research MLLMs, and workflows built on strong closed-source foundation model such as GPT-5, Gemini-2.5-pro and Claude-4-Sonnet. The code will be released in https://github.com/Osilly/Vision-DeepResearch.