PersistBench wins NeurIPS Spotlight, finds 4D foundation models lack visual memory
weichiuma · x · 2026-10-09
The team behind PersistBench, a NeurIPS 2026 E&D Spotlight, tests whether foundation models modeling a persistent 4D world actually have visual memory — across camera-controllable video models, 4D reconstruction models, and video-to-360° models. TL;DR: not really.
The key evaluation challenge is obtaining real-world ground truth for what happens to an object after it leaves the camera's view; PersistBench introduces a simple way to sidestep this. Project page, leaderboard, GitHub repo, and Hugging Face dataset are all public.
More from Research
- ML-guided chemistry cuts capacitor temperature sensitivity sixfold — bravo_abad · 2026-10-09
- Cohere Labs: discrete diffusion gives lossless LLM inference speedup — Cohere_Labs · 2026-10-09
- MIT/Sakana AI researcher on artificial life, evolution beyond SGD — burny_tech · 2026-10-09
- Compliant spined footpad tackles the one-shot landing problem of hoppers — lukas_m_ziegler · 2026-10-09
- StarkWare founder Eli Ben-Sasson: AI solved the Erdős Unit Distance problem, all bets are off — jamestagg · 2026-10-09
- Claude Science produces first complete ultraviolet map of the sky, ~10% deviation — The Decoder · 2026-10-09