---
configs:
- config_name: ChangeIt
  data_files:
  - split: test
    path: ChangeIt_preprocessed.jsonl
- config_name: OcclusionDataSet-MM20
  data_files:
  - split: test
    path: OcclusionDataSet-MM20_preprocessed.jsonl
- config_name: coco2014
  data_files:
  - split: test
    path: coco2014_preprocessed.jsonl
- config_name: ImageNet-C
  data_files:
  - split: test
    path: ImageNet-C_preprocessed.jsonl
- config_name: MVP-N
  data_files:
  - split: test
    path: MVP-N_preprocessed.jsonl
- config_name: QuickDraw
  data_files:
  - split: test
    path: QuickDraw_preprocessed.jsonl
- config_name: Realistic-Occlusion-Dataset
  data_files:
  - split: test
    path: Realistic-Occlusion-Dataset_preprocessed.jsonl
license: cc-by-nc-4.0
task_categories:
- question-answering
language:
- en
size_categories:
- 10K<n<100K
tags:
- Proactive
- Reasoning
- Multimodal
- LLM
pretty_name: ProactiveBench
---

<!-- Provide a quick summary of the dataset. -->
This is the benchmark associated with submission 1331 at NeurIPS 2025 - Dataset and Benchmarks track.
The benchmark is intended to be used with the proposed submission environments (see the source code).
The .jsonl files do not contain proper image paths but rather image path templates, as each .jsonl entry is a sample, and each sample corresponds to a different environment with its own images.
See the submitted [code](https://anonymous.4open.science/r/ProactiveBench/README.md) README for information about dataset downloading and preprocessing, and to re-run the evaluations.