Preprint Proposes Video AI Answers With Pixel-Level Evidence

What happened

The authors propose Evidence-Backed Video Question Answering, or E-VQA. Under the task, a video model must provide a semantic answer plus the relevant time segments and dense tracked object-segmentation masklets, which are pixel-level outlines that follow objects through footage.

The paper introduces ST-Evidence, which the authors describe as a human-verified benchmark, and ST-Evidence-Instruct, a 160k-scale instruction-tuning dataset. The project repository lists the benchmark, dataset, 3B and 7B models, and code as released resources.

In author-reported tests, fine-tuning a 7B model on the dataset improved temporal-grounding t-mean by 27.2 points and J&F, a segmentation-quality measure, by 13.8 points against size-matched UniPixel baselines.

Why it matters

Video AI can give a correct-looking answer without showing what in the footage supports it. Requiring time-specific, pixel-level evidence could make it easier to inspect whether an answer matches visible events, especially when videos involve occlusions or changing object shapes.

What remains unclear

• The results are reported in a preprint and have not been independently replicated or audited in the reviewed sources.
• The dossier does not establish that the benchmark’s human-verification process, evidence outputs, or reported scores make a model reliable or generalizable in deployment.
• The repository was accessible, but its downloads, model execution, licenses, and reproducibility were not independently tested.

Related claims

• is ai hallucinating too much to be trustworthy

Sources

1. Evidence-Backed Video Question AnsweringPrimary source - arXiv - preprint paper - Jul 13, 2026

   Used for: The proposed task, datasets, and author-reported 7B evaluation results.

   Open source

2. GitHub - SalesforceAIResearch/EVQAPrimary source - Salesforce AI Research - subject-authored code repository - Date not listed

   Used for: The project’s listed benchmark, dataset, models, and code resources.

   Open source

3. [Literature Review] Evidence-Backed Video Question AnsweringMoonlight - independent automated literature review - Date not listed

   Used for: Independent confirmation that the paper reports the 160k dataset scale and stated 7B metrics.

   Open source