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

TL;DR

Vision-DeepResearch tackles the limitations of existing multimodal deep-research by introducing a long-horizon, multi-turn, multi-scale visual–text search paradigm that operates robustly in noisy real-world environments. It combines a data pipeline that generates high-quality, verifiable multimodal trajectories with supervised fine-tuning and reinforcement learning to internalize deep-research capabilities. The approach leverages multi-entity visual cropping, iterative tool calls, and a vision–text bridge to distal search reasoning, achieving state-of-the-art results among open models and competitive performance against closed-source agents on multiple factual benchmarks. The work provides extensive ablations and a scalable training framework, highlighting the importance of retrieval breadth, long-horizon planning, and end-to-end optimization for robust multimodal reasoning.

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.

Figures (4)

• Figure 1: Panel A: We identify two key limitations of existing multimodal deep-research paradigms for image search. First, prior multimodal deep-research MLLMs largely ignore the search engine hit-rate problem. In image retrieval, a single full-image or even entity-level query often fails to retrieve the required evidence; moreover, querying different-scale crops of the same entity can yield highly variable results. Second, existing methods are constrained in both reasoning depth and retrieval breadth, typically producing only short trajectories. In contrast, our approach supports dozens of reasoning steps and hundreds of engine interactions, leading to substantially stronger performance. Panel B: Pipeline Overview. We synthesize high-quality VQA instances and multi-turn trajectories, and then integrate multimodal deep-research capabilities into an MLLM via SFT and RL training. This enables long-horizon reasoning that performs multi-turn, multi-entity, and multi-scale visual and textual search. Bottom Image: Performance Comparison. Our model achieves the SoTA performance on six benchmarks with a comparatively smaller parameter. The our "Large" and "Small" models correspond to the 30B-A3B and 8B parameter scales, respectively, while "Qwen3-VL" and "WebWatcher" refer to Qwen3-VL-30B-A3B-Thinking and WebWatcher-32B, respectively. All models are evaluated fairly under the same agentic-reasoning setting.
• Figure 2: Our Data Pipeline. As shown in the top panel, we construct a complete multimodal deep-research synthesis pipeline. Leveraging the capabilities of an MLLM and a text-based DeepResearch foundation LLM, we generate long-horizon, multi-tool trajectories. As shown in the bottom panel, we obtain high-quality factual VQA instances via a rigorous verification and obfuscation procedure, which are then used for trajectory synthesis and RL training.
• Figure 3: RL Curves of Mean Trajectory Length and Reward.
• Figure 4: Our Data Pipeline. As shown in the top panel, we construct a complete multimodal deep-research synthesis pipeline. Leveraging the capabilities of an MLLM and a text-based DeepResearch foundation LLM, we generate long-horizon, multi-tool trajectories. As shown in the bottom panel, we obtain high-quality factual VQA instances via a rigorous verification and obfuscation procedure, which are then used for trajectory synthesis and RL training.