PersistBench (NeurIPS Spotlight): 4D foundation models can see but not remember
weichiuma · x · 2026-10-07
PersistBench, from a Cornell team, earned a Spotlight at NeurIPS 2026 (Evaluations & Datasets Track). It asks: 4D foundation models can perceive and reconstruct dynamic environments, but can they remember what they perceived, as humans do? 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, sidestepping the challenge of knowing what happens to objects off-camera, and evaluates three object-centric aspects of visual memory: object permanence, motion continuity, and appearance preservation.
Findings: Across 12 models from three families, evaluation reveals a substantial gap between synthesizing visible content and remembering objects beyond the input view. Project page, leaderboard, code and dataset are all open for submissions.
More from Research
- After Navier-Stokes, OpenAI's internal math model now claims pieces of two more Millennium Problems — physiopeng · 2026-10-07
- MMM optimizers only exploit: bandit-style exploration recovers lost revenue — twiecki · 2026-10-07
- ADAG automates circuit-tracing interpretation, finds jailbreak circuits in Llama 3.1 — aryaman2020 · 2026-10-07
- Researcher confirms OpenAI results include Unique Games proof and rational Hodge over abelian varieties — aran_nayebi · 2026-10-07
- OpenAI's unreleased model reportedly proves quasi-Riemann hypothesis with Lean proof — ChrisGPT · 2026-10-07
- Mathematician celebrates AI model cracking a problem he chased for a decade — ctjlewis · 2026-10-07