Vision-DeepResearch & Vision-DeepResearch Benchmark (VDR-Bench) & Video-DeepResearch

The official repo for "Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models", "Vision-DeepResearch Benchmark: Rethinking Visual and Textual Search for Multimodal Large Language Models" and "Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent".

Project Page

🤗 VQA Benchmark Subset

🤗 Cold-start Dataset (demo)   |   🤗 RL Dataset (demo)   |   🤗 VDR-Bench (full)   |   🤗 VDR-Bench (testmini)

🤗 Vision-DeepResearch-30B-A3B (SFT+RL)   |   🤗 Vision-DeepResearch-8B (SFT-only)

📑Vision-DeepResearch Paper   |   📑 VDR-Bench Paper

🤗 Video-DeepResearch-30B-A3B   |   🤗 Video-DeepResearch-35B-A3B

📑Video-DeepResearch Paper

The datasets, code and weights will be released, stay tuned!

Timeline

• [2026/08/05] We release Video-DeepResearch paper! Video-DR extends Vision-DR to continuous video streams. Code lives under Video-DeepResearch/.

• [2026/06/08] We released Vision-DeepResearch-30B-A3B.

• [2026/05/24] We have released the subset of our used benchmark subset and detailed evaluation guidance! Please see VQA Benchmark Subset and Eval Guidance.

• [2026/05/01] Vision-DeepResearch has been accepted by ICML 2026! Nice to see you at the conference!

• [2026/02/03] We released SFT code, RL code! We will finish the guidance later, stay tune!

• [2026/02/02] We released Cold-start Dataset (demo), RL Dataset (demo), Vision-DeepResearch-8B (SFT-only), VDR-Bench (full), VDR-Bench (testmini)!

Demo (click to watch on YouTube)

Compare to Other Methods

More Cases

Performance

Teaser

Vision-DeepResearch

VDR-Bench

Data Pipeline

Vision-DeepResearch

VDR-Bench

Quickstart

Environment Setup

# 1. Clone the repository
git clone https://github.com/Osilly/Vision-DeepResearch.git
cd Vision-DeepResearch

# 2. Install verl
cd rllm/verl
pip install -e .

# 3. Install Megatron-LM
cd ../../Megatron-LM
pip install -e .

# 4. Install mbridge
cd ../mbridge
pip install -e .

# 5. Install rllm
cd ../rllm
pip install -e .

# 6. Install additional dependencies
pip install requests==2.32.3
pip install oss2

# 7. Return to the project root directory
cd ..

Data Preparation

SFT Data

Download the Cold-start dataset (Demo 1K).

You need to convert the data in Parquet format into the JSONL training format supported by ms-swift. We provide a conversion script for this purpose: ms-swift/run/data_prepare/convert_parquet2jsonl.sh.

You need to provide an --image_dir, where images stored as bytes in the Parquet file will be converted to .png/.jpg files and saved to disk.

RL Data

Download the RL dataset (Demo 1K).

First, you need to convert the data in Parquet format into the JSONL format. We provide a conversion script for this purpose: rllm/vision_deepresearch_async_workflow/data_prepare/convert_parquet2jsonl.sh.

Then, you need to run rllm/vision_deepresearch_async_workflow/data_prepare/register_rl_dataset.sh to register the RL dataset.

SFT Train

cd ms-swift
bash run/vision_deepresearch_SFT_30B_A3B_megatron_lr2e5_2ep.sh

RL Train

First, deploy the Extract model (used to summarize web page contents) and the Judge model:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 vllm serve \
 Qwen/Qwen3-VL-30B-A3B-Instruct \
  --host 0.0.0.0 \
  --port 8001 \
  --tensor-parallel-size 8 \
  --gpu-memory-utilization 0.8 \
  --served-model-name "Qwen3-VL-30B-A3B-Instruct" \
  --max_model_len 160000 \
  --mm-processor-cache-gb 0 \
  --no-enable-prefix-caching

Then, modify the vLLM URL service endpoints for JUDGE_MODEL and EXTRACT_MODEL in rllm/.env, and enter your SERP_API_KEY, JINA_API_KEY, and OSS configuration.

Run RL train.

cd rllm
bash vision_deepresearch_async_workflow/run/vision_deepresearch_30B_A3B_grpo_plus_bfloat16_sglang_megatron_128batch_128mini_8n.sh

Eval

Run the command below to start an OpenAI-compatible API service:

Vision-DeepResearch model:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 vllm serve \
 Osilly/Vision-DeepResearch-8B \
  --host 0.0.0.0 \
  --port 8001 \
  --tensor-parallel-size 8 \
  --gpu-memory-utilization 0.8 \
  --served-model-name "Vision-DeepResearch-8B" \
  --max_model_len 160000 \
  --mm-processor-cache-gb 0 \
  --no-enable-prefix-caching

Extract model (used to summarize web page contents) and Judge model:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 vllm serve \
 Qwen/Qwen3-VL-30B-A3B-Instruct \
  --host 0.0.0.0 \
  --port 8001 \
  --tensor-parallel-size 8 \
  --gpu-memory-utilization 0.8 \
  --served-model-name "Qwen3-VL-30B-A3B-Instruct" \
  --max_model_len 160000 \
  --mm-processor-cache-gb 0 \
  --no-enable-prefix-caching

Modify the vLLM URL service endpoints for JUDGE_MODEL and EXTRACT_MODEL in rllm/.env, and enter your SERP_API_KEY, JINA_API_KEY, and OSS configuration.

Modify the base-url and model (the Vision DeepResearch vLLM service endpoint and model name) in rllm/eval/run_eval.sh. For the data format of test.parquet, refer to rllm/eval/README.md.

Run rllm/eval/run_eval.sh to start inference.

bash rllm/eval/run_eval.sh

Star History

Citation

@article{huang2026vision,
  title={Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models},
  author={Huang, Wenxuan and Zeng, Yu and Wang, Qiuchen and Fang, Zhen and Cao, Shaosheng and Chu, Zheng and Yin, Qingyu and Chen, Shuang and Yin, Zhenfei and Chen, Lin and others},
  journal={arXiv preprint arXiv:2601.22060},
  year={2026}
}

@article{zeng2026vision,
  title={Vision-DeepResearch Benchmark: Rethinking Visual and Textual Search for Multimodal Large Language Models},
  author={Zeng, Yu and Huang, Wenxuan and Fang, Zhen and Chen, Shuang and Shen, Yufan and Cai, Yishuo and Wang, Xiaoman and Yin, Zhenfei and Chen, Lin and Chen, Zehui and others},
  journal={arXiv preprint arXiv:2602.02185},
  year={2026}
}
@misc{fang2026videodeepresearchnextgenerationmultimodaldeepresearch,
      title={Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent}, 
      author={Zhen Fang and Yu Zeng and Wenxuan Huang and Yiming Zhao and Shiting Huang and Tianfei Ren and Qi Lu and Qingnan Ren and Qisheng Su and Lionel Z. Wang and Qingyu Yin and Shuang Chen and Zehui Chen and Lin Chen and Zhenfei Yin and Yao Hu and Shaohui Lin and Wanli Ouyang and Shaosheng Cao and Feng Zhao},
      year={2026},
      eprint={2608.03979},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.03979}, 
}