Video question answering has improved with multimodal large language models, but these models still struggle with the temporal structure that makes a video coherent. According to the abstract of a new arXiv paper, they may merge neighboring events or miss a single event entirely, undermining their understanding of what actually happens on screen.
The paper introduces EventLens, a method that uses event-structure reinforcement learning to address this gap. Rather than treating video understanding as a flat sequence of frames, EventLens is designed to recover the underlying event structure, helping the model distinguish between distinct moments and notice isolated events that might otherwise be overlooked.
The abstract does not provide experimental details or quantitative results, so the reported effectiveness of EventLens is not yet clear. But the problem it targets is well-defined: current MLLMs are strong at answering questions about video content yet weak at preserving the temporal organization that gives that content meaning.