VIS-Ground gains 10.3 points on interactive video task

A new method called VIS-Ground reports a 10.3-point average absolute improvement over the strongest per-backbone baselines on interactive video storytelling, the task of letting a viewer steer a video as it unfolds. The arXiv paper argues the hard part is not rendering frames but handling mid-stream requests that carry hidden assumptions about everything already on screen.

The authors state the problem plainly. "The viewer's request can have latent dependencies on both the grounding source and the current rendered video state. These dependencies may not be explicitly stated in any individual input, but emerge only when the source, rendered history, and new viewer intent are considered jointly."

Their fix is a three-stage pipeline: Structured Context Abstraction to recover grounded states and cross-context dependencies, Generation Constraints Induction to project those dependencies into candidate-specific constraints, and Constrained Video Generation to enforce them through planning, verification, revision, and rendering. The 10.3-point gain is reported as an average across three video generation backbones. The abstract does not publish per-backbone numbers or name the specific baselines beaten.