Video Reasoning
Overview
Video reasoning asks models to interpret how people, objects, and scenes change over time, rather than rely only on information in individual frames. Common approaches include multimodal language models with temporal grounding, event or scene graphs, memory and retrieval for long videos, and agents that select frames or use tools; reinforcement learning increasingly rewards answers tied to visual evidence. Benchmarks test video question answering, temporal localization, causal and spatial reasoning, egocentric understanding, and cross-view or cross-video inference. A central challenge is ensuring that fluent reasoning reflects correctly perceived and grounded evidence, especially over long horizons, while keeping computation manageable.
Latest papers
211 papers33 papers in the last four weeks, up 267% on the four weeks before. 0.2% of all new papers.