Narrative Aligned Long Form Video Question Answering

A team of researchers has published a new paper, "Narrative Aligned Long Form Video Question Answering," introducing a critical benchmark called NA-VQA. This benchmark is designed to test multimodal large language models (MLLMs) on their ability to perform deep narrative reasoning across entire movies, not just localized scene recognition. It contains 88 full-length films and 4.4K open-ended question-answer pairs, with questions specifically tagged by the distance between evidence spans (Short, Medium, Far). The results expose a significant weakness: current SOTA models perform poorly on questions requiring 'Far' evidence, showing they rely on shallow pattern matching rather than truly understanding and connecting distant narrative events.

To address this gap, the researchers propose Video-NaRA, a new narrative-centric framework. Video-NaRA works by constructing event-level chains from the video and storing them in a structured memory system for retrieval during reasoning. This explicit modeling of narrative structure allows the AI to track intentions, connect causally distant events, and reconstruct story arcs. In extensive experiments, Video-NaRA improved performance on long-range reasoning questions by up to 3 percentage points, demonstrating a clear path forward for more sophisticated video understanding AI.

The paper, available on arXiv, argues that most existing video benchmarks fail to capture true narrative comprehension. By requiring generative, multi-scene answers, NA-VQA pushes models to integrate dispersed information. The team plans to release the full NA-VQA benchmark upon publication, providing a crucial tool for the community to develop AI that can genuinely understand complex, long-form visual stories.