Blender-generated datasets teach world models object permanence at 10k samples per task
_akhaliq · x · 2026-09-26
The paper "Training Object Permanence in World Models" (arXiv:2609.28654) argues object permanence — a foundation of human cognition — is missing from current video and world models. The authors build a data infrastructure of diverse object-permanence cognitive tasks, each with a Blender-based generator scalable to at least 10,000 diverse samples per task. They show this infrastructure effectively trains video models and world models that acquire object permanence, offering a systematic path toward human-like physical common sense in world models.
Related event: WROP: Open-Source Suite Trains Object Permanence in World Models(3 posts)→
More from Research
- Diffusion Week 2026 releases all 10 lectures on video generation, world models and more — LucaAmb · 2026-11-11
- Most models mention hidden tests on MaxxBench, hinting at possible training contamination — dejavucoder · 2026-11-11
- Hill climbing isn't RL: dev untangles benchmarking vs benchmaxxing vs metric optimization — dejavucoder · 2026-11-11
- Train 27B models on a 24GB GPU: ternary optimizer cuts memory to 1 byte/weight — oguzeldereli · 2026-11-11
- METR report reignites 2027 intelligence explosion debate; Ord pushes back — tobyordoxford · 2026-11-11
- Nearly 40% of new arXiv preprints are AI-written, raising training-data pollution fears — MaxUnfried · 2026-11-11