Most present vision-language-action (VLA) fashions—similar to OpenVLA, π0, RT-2, and RDT-1B—are “monolithic.” This implies they generate uncooked motor instructions or very brief sequences of actions, with out organizing behaviors into reusable, well-defined abstractions. In consequence, these fashions carry out poorly on long-horizon (multi-step) duties, and it’s troublesome to interpret what they’ve realized. Present approaches for locating expertise typically keep away from the core drawback of deciding when two motion sequences are “behaviorally equal.” For instance, AtomicVLA and AtomSkill group motion sequences by clustering their contrastive embeddings. In distinction, BLADE and LRLL depend on a big language mannequin (LLM) to evaluate whether or not two sequences are equal, however these LLMs aren’t calibrated to the robotic’s personal dynamics. We introduce REFACTOR-VLA, a system that learns reusable expertise utilizing a “wake/sleep” structure. Within the sleep section, the system clusters segments of motor applications utilizing a Behavioral-Equivalence Kernel (BEK). This BEK relies on the outcomes of rolling out actions in a realized latent world mannequin, Mφ. Within the wake section, the system generates typed lambda phrases (easy, structured applications) from a vocabulary impressed by the Hindley–Milner kind system. These lambda phrases are then utilized by a library-conditioned rectified-flow motion decoder to provide actions. Solely abstractions that move each a Minimal Description Size (MDL) criterion and a return-preservation gate are accepted as expertise. To coach REFACTOR-VLA, we use a three-phase schedule: • Section A (World-model warmup): The latent world mannequin Mφ is educated. • Section B (Wake-phase coverage optimization): The coverage that makes use of the library of expertise is optimized. • Section C (Sleep-phase ability discovery): The system clusters motion fragments into reusable expertise. We evaluated REFACTOR-VLA on the complete LIBERO benchmark suite. Our outcomes present two predominant findings. First, merely rising the dimensions of the world mannequin—from 188 million to 430 million parameters—worsened efficiency on 4 out of 4 benchmark suites, disproving the concept simply making the world mannequin larger at all times helps. Second, altering the coaching goal makes a giant distinction: including an auxiliary supervised contrastive loss (particularly, InfoNCE loss) in the course of the world-model warmup (Section A) vastly improved the standard of ability clustering within the sleep section (Section C). We measured this utilizing Normalized Mutual Info (NMI) beneath n = 3 multi-seeding: • Object suite: 0.462 ± 0.021 • Spatial suite: 0.867 ± 0.025 • Purpose suite: 0.915 ± 0.013 • LIBERO-10 suite: 0.754 ± 0.010
