Seeing Is Not Remembering: PersistBench Exposes 4D Models' Weak Visual Memory
zhenjun_zhao · x · 2026-09-19
The arXiv paper "Can 4D Foundation Models Remember?" introduces PersistBench, a dataset and metric suite that evaluates visual memory in 4D foundation models such as camera-controllable video and 4D reconstruction models.
- Motivation: existing benchmarks rely on pixel-level metrics and lack ground truth once objects leave the field of view.
- Method: uses 360° videos as omniscient ground truth, evaluating object permanence, motion continuity, and appearance preservation.
- Findings: current models maintain only short-term consistency that degrades sharply once objects leave view — "seeing is not remembering." Dataset and code are open-sourced.
More from Research
- Resolution-Inspired Pretraining Experiments 'Pretty Much a Disaster,' Researcher Says — gordic_aleksa · 2026-09-19
- Auto-Formalizing a 76-Page Paper With Opus 5 High Would Take ~40 Days — kfountou · 2026-09-19
- Anthropic quietly built a wet lab to let Claude run robotic experiments — ChrisGPT · 2026-09-19
- ICLR Submissions Near 45,000 as Researcher Decries AI-Generated Paper Slop — DimitrisPapail · 2026-09-19
- Quanta video explains why Navier-Stokes is so hard and how an OpenAI proof could resolve it — eigensteve · 2026-09-19
- BBFM conjectures see formalization progress building on Axiom Math's prior work — ctjlewis · 2026-09-19