Datasets:
| cff-version: 1.2.0 | |
| message: "If you use this dataset, please cite as below." | |
| title: "NExT-CF: A Counterfactual Entity-Removal Benchmark for Video Question Answering" | |
| authors: | |
| - name: "Anonymous Authors (submission under review)" | |
| year: 2026 | |
| version: "1.0.0" | |
| license: "CC-BY-NC-SA-4.0" | |
| type: dataset | |
| abstract: > | |
| NExT-CF is a benchmark of 570 counterfactually edited videos derived from NExT-QA, | |
| used to probe whether Video Question Answering models rely on visual evidence | |
| or language priors. Each video has a target entity removed via SAM3 segmentation | |
| and DiffuEraser diffusion inpainting, with the corresponding QA pairs carried | |
| over from NExT-QA. | |
| keywords: | |
| - video question answering | |
| - counterfactual | |
| - benchmark | |
| - diagnostic benchmark | |
| - visual grounding | |
| - evidence reliance | |
| - diffusion inpainting | |
| references: | |
| - type: conference-paper | |
| authors: | |
| - family-names: Xiao | |
| given-names: Junbin | |
| - family-names: Shang | |
| given-names: Xindi | |
| - family-names: Yao | |
| given-names: Angela | |
| - family-names: Chua | |
| given-names: Tat-Seng | |
| title: "NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions" | |
| conference: | |
| name: "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)" | |
| year: 2021 | |
| start: 9777 | |
| end: 9786 | |
| notes: "arXiv:2105.08276; project repo: https://github.com/doc-doc/NExT-QA" | |
| notes: > | |
| NeurIPS 2026 Evaluations & Datasets track submission (under review). | |
| Author identities will be added upon acceptance. | |