SimpleSearch-VL is a simple agentic system for multimodal deep search, training visual-language agents to actively search the web, perform region-level reverse image search, visit evidence pages, verify retrieved evidence, and produce grounded answers. Beyond the agent recipe, this repository provides a practical multimodal agentic RL training infrastructure: tool-interleaved rollout generation, cache-aware online retrieval, and Factorized Adaptive Rollout (FAR) for efficient signal-aware RL training. With Qwen3-VL backbones, SimpleSearch-VL obtains strong performance across six multimodal search benchmarks with only 5K SFT trajectories and 2K RL prompts. Instead of simply scaling data, adding more tools, or relying on extra auxiliary models, we build higher-quality evidence-aware trajectories, make visual search results directly verifiable, and use FAR to extract stronger RL signal from a small training budget.

• Factorized Adaptive Rollout (FAR): factorizes rollout budget into prompt expansion and rollout allocation, spending extra attempts on hard prompt groups while skipping redundant tail rollouts once useful reward variation appears.
• Evidence-verified reasoning: trains the agent to verify textual evidence and to check matched thumbnails from region-level reverse image search before trusting titles, URLs, or webpage evidence.
• Self-summarized visit: keeps webpage summarization inside the policy model, so training and inference do not require deploying an additional summary model.
• Minimal engineering dependencies: relies only on Serper, JINA Reader, and OSS for online search, webpage reading, and image-region upload, yet can train a strong multimodal search agent.
• Data- and compute-efficient training: trains 8B and 30B-A3B variants with only 5K SFT trajectories and 2K RL prompts, requiring roughly 1-2 days on 8xH200; the paper reports +15.8 and +16.0 average points over corresponding Qwen3-VL agentic baselines.
• Strong benchmark results: SimpleSearch-VL-30B-A3B outperforms agentic Gemini-3-Pro on MMSearch, MMSearch+, BrowseComp-VL, FVQA, and LiveVQA.

• July 1, 2026: We are happy to share that the SimpleSearch-VL paper is now available on arXiv. Thank you for your interest, and we will keep updating this as more resources are ready.

The following long-form visualizations show representative SimpleSearch-VL agent rollouts. Click either thumbnail to open the full-resolution image.

MMSearch+ rollout demo

LiveVQA research rollout demo

This preview repository currently releases only the README and public figures. The following release items are TODO:

• Release the paper
• Release SimpleSearch-VL model weights
• Release training and evaluation data
• Release training and evaluation code
• Release complete training and inference scripts

If you find SimpleSearch-VL useful, please cite:

@misc{dai2026simplesearchvlsimplerecipemultimodal,
      title={SimpleSearch-VL: A Simple Recipe for Multimodal Agentic Deep Search},
      author={Ming Dai and Zhihong Lu and Jinjie Gu and Jiedong Zhuang and Yefeng Liu and Wankou Yang and Jian Wang and Chunhua Shen},
      year={2026},
      eprint={2606.31504},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.31504},
}