Can 4D Foundation Models Remember?

What it claims

The paper argues that existing benchmarks for 4D foundation models (camera-controllable video models, 4D reconstruction models) evaluate the wrong thing: they lean on pixel-level metrics and have no ground truth for objects once they leave the field of view, so they cannot measure visual memory in an object-centric way against a reference. PersistBench closes that gap by using 360° videos as omniscient ground truth — the full scene is always captured, so an object that exits the visible frame still has a reference to compare against.

How it works

The suite scores models on three aspects the summary names — object permanence, motion continuity, appearance preservation — but the page's added detail is how it grounds them: comparison against the 360° reference rather than against visible pixels only. The distinction is object-centric-vs-reference versus prior pixel-level metrics; that framing is what lets it probe out-of-view persistence at all.

Evidence

The page is abstract-only and reports the result qualitatively: across "various models" spanning "diverse categories," consistency is only short-term and degrades significantly once objects leave the field of view. No per-model numbers, no baseline table, no count of videos or categories is given on this page.

How to use it

Dataset and code are stated to be available on the project page (linked as "this https URL" in the abstract, not resolved in the fetched text). If you build or evaluate 4D video/reconstruction models, this is a memory-focused evaluation set to run rather than a model to adopt.

Limits

• Only the abstract is available here — no leaderboard, no metric definitions, no dataset size, no list of which models were tested. Verify all of these on the project page before planning against it.
• The project-page URL is not resolved in the fetched text.
• v1 submitted 17 Sep 2026; a v2 (21 Sep 2026) exists — check which version the dataset/code track.