Symbolic world models strike back: mu0 beats pi0.5 with 1/100 of the data

furongh · x · 2026-09-04

Furong Huang cites Gary Marcus's 2020 argument for symbolic world models, now validated in physical AI: the mu0 model beats pi0.5 on real-world robot tasks using only 1/100 of the training data.

Marcus adds that GPT-6 Astra performs well precisely by inducing and using symbolic world models, leaving open questions: how general is this ability, how does it work, and can such models reliably follow explicit instructions?

Original post →

More from Embodied

Embodied channel →