Apple AI Updates: September 3, 2026

1. REFACTOR-VLA Learns a Reusable Skill Library Instead of Raw Motor Commands

Apple. The paper argues that monolithic vision-language-action models such as OpenVLA and RT-2 emit raw motor commands or very short action sequences without organizing behavior into reusable abstractions, which limits multi-step tasks and makes learned behavior hard to interpret. REFACTOR-VLA uses a three-phase wake/sleep architecture: a Behavioral-Equivalence Kernel clusters motor programs by outcome under a learned world model, the system emits typed lambda terms borrowed from functional programming type systems, and only abstractions passing Minimum Description Length and return-preservation tests are accepted as skills. On LIBERO, scaling the world model from 188M to 430M parameters hurt performance across all four suites, while an auxiliary contrastive loss improved skill clustering to normalized mutual information scores of 0.462 to 0.915 depending on suite. Source