Open data infrastructure teaches AI object permanence: 150 tasks, 1.5M samples, 16B model
du_yilun · x · 2026-09-26
Researchers released a full-stack data infrastructure for teaching object permanence—the understanding that objects persist behind occluders—to AI world models. It's one of the earliest pillars of biological intelligence and underpins intuitive physics and mechanical reasoning.
What's included:
- 150 cognitive tasks (e.g., two-balls-three-cups occlusion and swap scenarios) with prompt descriptions
- A Blender generator per task, scaling to 10K+ diverse samples
- A 1.5M-sample training set with train/eval splits
- A fine-tuned 16B model validating the data, plus a leaderboard
Everything is open-sourced: website, paper, code, data, model, and benchmark.
Related event: WROP: Open-Source Suite Trains Object Permanence in World Models(3 posts)→
More from Research
- Jev-Mem paper: a fast scorer makes agent memory building 6.6x faster on LoCoMo — rohanpaul_ai · 2026-10-25
- Bayer Taps Stanford AI Startup Phylo (Biomni Lab) for Drug Discovery After $13.5M Seed — RichardsonDx · 2026-10-25
- 96 Humanoid Robots Learn From Each Other, Lifting Success Rates From 81% to 97.3% — CyberRobooo · 2026-10-25
- HCI researcher bets AI can finally make Bayesian stats usable for small-N studies — IanArawjo · 2026-10-25
- Classic short paper explains the linear algebra behind Google's PageRank — HankYeomans · 2026-10-25
- Two code-LLM papers accepted at EMNLP2026: behavior benchmark and self-verifying codegen — algo_diver · 2026-10-25