Abstract
Recent advances in Video Large Language Models (Video-LLMs) have yielded promising performance on Video Question Answering (VideoQA). Nevertheless, existing benchmarks are predominantly evaluated through answer correctness, while the faithfulness of predicted evidence supporting those answers remains insufficiently evaluated. This disconnect between answer generation and evidence verification motivates the construction of the Evidence-Grounded Video Question Answering Benchmark (EG-VQA), a large-scale open-ended benchmark in which each QA pair is annotated with temporally localized textual evidence, requiring models to jointly produce answers and verifiable evidence. EG-VQA comprises 2,067 videos and 11,838 QA pairs with fine-grained evidence annotations. To evaluate predicted evidence, Evidence-Grounded F1 (EG-F1) is introduced as a metric that jointly measures temporal alignment and semantic consistency between predicted and ground-truth evidence. Experiments reveal a substantial discrepancy between answer correctness and evidence faithfulness: even strong proprietary models can achieve high answer accuracy while exhibiting gaps in temporal-semantic grounding. To investigate evidence-aware learning under EG-VQA, we develop EG-Reasoner, an evidence-aware model trained with explicit evidence supervision. EG-Reasoner achieves strong evidence grounding performance among open-source models while maintaining competitive answer generation results compared with proprietary systems. These findings highlight the importance of explicitly modeling evidence grounding and suggest that evidence-aware supervision provides an effective direction toward more reliable and interpretable VideoQA systems.
Benchmark Overview
Comparison between our EG-VQA and existing VideoQA benchmarks.
Overview of EG-VQA statistics and representative QA examples. EG-VQA contains 2,067 videos and 11,838 QA pairs, split in a video-disjoint manner into 8,949 training and 2,889 test samples. It covers four reasoning categories: descriptive, temporal, causal, and counterfactual questions, progressing from directly observable video content to event ordering, causal inference, and hypothetical reasoning grounded in specific video segments.
Construction Pipeline. EG-VQA is constructed through a pipeline that combines video data curation, reasoning-oriented question-answer generation, prompt refinement, and two-stage quality control. Each QA pair is paired with temporally localized evidence.
Model
EG-Reasoner Training Framework. Given a video and question, EG-Reasoner learns to produce structured responses that explicitly organize temporal evidence, intermediate reasoning, and the final answer. The model is optimized with reinforcement learning using a composite reward that jointly evaluates response format, answer correctness, and evidence grounding. During training, GRPO updates the policy by comparing multiple sampled responses, encouraging the model to generate faithful answers supported by localized video evidence.
Experiments
We compare the performance of EG-Reasoner with commercial large multimodal models and recent open-source Video-LLMs on EG-VQA. EG-Reasoner outperforms other open-source models in terms of both answer accuracy and evidence grounding.
Qualitative Analysis
As videos are analyzed, EG-Reasoner localizes key evidence regions and decomposes complex events into fine-grained temporal reasoning steps. Compared with the base model, it produces answers that are better grounded in the video content and less affected by loosely related context or hallucinated intermediate steps.
BibTeX
@misc{huang2026egvqa,
title={EG-VQA: Benchmarking Verifiable Video Question Answering with Grounded Temporal Evidence},
author={Huang, Linpeng and Chen, Weixing and Chen, Zexin and Liu, Yang and Lin, Liang},
year={2026},
eprint={2606.24797},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2606.24797}
}