Skip to content
Book Open access

NexusQA: Spatio-Temporal Interlacement Grounded VideoQA

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 50 references

Abstract

Video question answering (VideoQA) requires sophisticated interpretation of spatio-temporal dynamics. While grounded VideoQA has emerged to provide verifiable evidence, existing benchmarks are predominantly confined to Coarse-grained Spatio-Temporal (CST) evidence, such as event or object-level localization. Consequently, they fail to scrutinize the Fine-grained Spatio-Temporal (FST) evidence essential for clarifying subtle semantic states, allowing Multimodal Large Language Models (MLLMs) to exploit language priors instead of performing genuine visual grounding. To address this, we introduce NexusBench, a novel benchmark requiring joint reasoning across Multi-grained Spatio-Temporal evidence. Built upon a Hierarchical Video Scene Graph (HVSG), it explicitly models FST interactions to build logic-intensive queries. To diagnose current model limitations, we propose a metric termed NexusScore, assessing a model's capability to utilize evidence. Our analysis reveals that existing MLLMs exhibit inadequate FST grounding abilities, resulting in suboptimal performance, yet supplying them with the correct evidence drastically improves accuracy. Motivated by this, we present NexusQA, a training-free multi-agent baseline that accumulates evidence through collaborative reasoning. Extensive experiments demonstrate that NexusQA achieves state-of-the-art performance on both the proposed NexusBench and the public NExT-GQA dataset. Our code and benchmark will be released at https://github.com/Daria8976/NexusQA.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.