EG-VQA: Benchmarking Verifiable Video Question Answering with Grounded Temporal Evidence

A new benchmark called EG-VQA has been introduced to test whether video question-answering models can ground their answers in specific video evidence, moving beyond simple answer accuracy [1]. The open-ended protocol includes 2,067 videos and 11,838 question-answer pairs, each annotated with supporting temporal evidence [2]. The Evidence-Grounded Video Question Answering Benchmark (EG-VQA) was detailed in a paper submitted on June 23, 2026 [1]. It is designed to address a gap in current evaluations of Video Large Language Models (Video-LLMs), which typically measure only whether an answer is correct, not whether the model identified the correct video segment to support it [2]. A language model benchmark is a standardized test with a dataset and evaluation metrics used to compare model capabilities [3]. To measure performance on this new task, the researchers created a unified metric called Evidence-Grounded F1 (EG-F1), which jointly measures temporal alignment and semantic consistency against ground-truth evidence [2]. Testing revealed that even strong proprietary models struggle to accurately ground their predictions, exposing a fundamental discrepancy between answer correctness and faithful evidence localization [2]. To bridge this gap, the paper proposes EG-Reasoner, an evidence-grounded reasoning model trained with explicit supervision [2]. The model achieved state-of-the-art performance among open-source models, with results competitive against proprietary systems [2]. Particularly pronounced gains were observed on reasoning-intensive tasks such as counterfactual questions [2]. The findings indicate that scaling alone is insufficient for robust video understanding and that structured evidence supervision is essential for developing more reliable and interpretable VideoQA systems [2]. The research was posted on arXiv, an open-access repository for scholarly articles [1]. The work aligns with broader industry efforts to create more trustworthy artificial intelligence. Microsoft Research, for example, has made significant advances in AI and holds a large percentage of global AI patents, with a mission to expand the state of the art and transfer technologies into products [4]. The development of benchmarks like EG-VQA provides a standardized method for tracking progress in areas such as reasoning and evidence-based understanding [3].