Video models no longer read only the current frame: paired entities across time, future trajectories and events predicted together

Scene Evolution Understanding is defined as a task of its own: given observed video, a model estimates entity histories and predicts one scene continuation made of future trajectories and Events, accompanied by Visual Evidence from observed motion and a scene description. Large multimodal models previously needed to understand how entities move and interact as a scene develops, which requires following the same entities over time and relating observed motion to subsequent changes. The SceneForesee benchmark evaluates entity continuity and scene evolution reasoning through paired observed and future scenes with shared entity identities, containing 144.57K temporal pairs, 1.05M entity correspondences, and 272.49K Events paired with observed evidence.

EvoSee learns future geometry and Event semantics through shared candidate states and joint assignment, selecting one continuation per entity from observation and using the selected changes for entity refinement, retention, and scene representation. The framework brings predicted entity changes into spatiotemporal reasoning rather than treating evolution as an auxiliary prediction target alone. On SceneForesee, EvoSee outperforms Qwen3.5-9B by 10.45 points in Combined Point-F1 and 17.24 points in Event Global F1; on the 4D reasoning benchmark Dyn-Bench, it improves question answering accuracy and grounding J&F by 4.8 and 3.58 points over their respective base systems.

The work targets video understanding settings that require following entities over time and predicting a scene continuation, evaluated on SceneForesee's paired observed and future scenes and on Dyn-Bench's 4D reasoning question answering and grounding. A project link is provided for readers to inspect the implementation and data.

The material is a summary and does not state how EvoSee performs beyond SceneForesee and Dyn-Bench, nor does it give cross-domain or real deployment validation; the supported scope of scene evolution as a dynamic representation is limited to the two reported benchmarks.