Datasets:

anomyousMMDR
/

MMdeepresearch

license: cc-by-4.0
language:
  - en
  - zh
  - fr
  - de
  - fi
size_categories:
  - n<1K
task_categories:
  - text-generation
tags:
  - agent
  - deep-research
  - multimodal
  - benchmark
  - evaluation
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

MMDR-Bench

Anonymous release for peer review.

MMDR-Bench is a benchmark for evaluating multimodal deep research agents. It consists of 140 expert-crafted tasks across 19 domains, organized into two regimes:

Daily (40 tasks, 10 domains): loosely structured visual inputs such as screenshots, photographs, and UI captures.
Research (100 tasks, 9 domains): information-dense visuals including charts, diagrams, and tables that require multi-source synthesis.

Each task is packaged as an image–text bundle. Models are expected to produce a citation-grounded, multimodal research report.

Dataset fields

Field	Type	Description
`id`	int64	Task identifier (0–139)
`caption`	string	Short task title
`body`	string	Full task prompt (query + instructions)
`image_url`	list of images	Visual inputs bundled with the task
`tags`	list of strings	Domain tags (e.g. `HCS`, `EBS`, `CDS`, `SPS`, `EES`, `MES`, `LHS`, `XDIS`, `Exp`)
`language`	string	Task language (`en`, `zh`, `fr`, `de`, `fi`)
`difficulty`	string	`easy`, `hard`, or `complex`

Intended use

MMDR-Bench is designed for process-oriented evaluation of multimodal deep research agents: it measures whether an agent faithfully grounds its outputs in the provided images and in retrieved web evidence, rather than matching against fixed gold reports. See the accompanying paper for the VEF / MOSAIC / TRACE evaluation modules and their assumptions and limitations.

Limitations

Task distribution is dominated by English and Chinese; other languages are in the long tail.
Absolute scores are not comparable across re-runs separated by months because models, judge LLMs, and retrieval toolchains drift over time.
The benchmark certifies groundedness, not absolute factual correctness.

License

CC BY 4.0.

Citation

Anonymized during review. Citation will be provided upon acceptance.