Current vision-language-action (VLA) models such as OpenVLA, π0, RT-2, and RDT-1B are monolithic—they generate raw motor commands or very short action sequences without reusable abstractions. This limits performance on long-horizon tasks and makes learned behavior hard to interpret. Existing skill-discovery methods rely on contrastive clustering (AtomicVLA, AtomSkill) or LLM-based equivalence judgments (BLADE, LRLL), but the latter are not calibrated to the robot’s own dynamics.

REFACTOR-VLA addresses these limitations with a wake/sleep architecture. In the sleep phase, the system clusters segments of motor programs using a Behavioral-Equivalence Kernel (BEK) that rolls out actions in a learned latent world model M_φ. Only action sequences that yield similar outcomes under the world model are considered equivalent. In the wake phase, the system generates typed lambda terms (simple, structured programs inspired by the Hindley–Milner type system) and uses a library-conditioned rectified-flow action decoder to produce actions. New abstractions are accepted as skills only if they pass both a Minimum Description Length (MDL) criterion and a return-preservation gate.

Training follows a three-phase schedule: Phase A warms up the latent world model M_φ; Phase B optimizes the policy that leverages the library of skills; Phase C discovers reusable skills by clustering action fragments. Evaluations on the full LIBERO benchmark show two key results. First, increasing the world model size from 188M to 430M parameters worsened performance on all four benchmark suites—contradicting the assumption that larger world models are always better. Second, adding an auxiliary supervised contrastive loss (specifically InfoNCE loss) during the world-model warmup (Phase A) significantly improved skill clustering quality in the sleep phase (Phase C). Measured via Normalized Mutual Information (NMI) under n = 3 multi-seeding, the scores were: Object suite 0.462 ± 0.021, Spatial suite 0.867 ± 0.025, Goal suite 0.915 ± 0.013, and LIBERO-10 suite 0.754 ± 0.010.

The results indicate that REFACTOR-VLA can learn reusable, typed motor programs without manual supervision, and that the quality of the latent world model—particularly when trained with contrastive objectives—is crucial for discovering behaviorally meaningful skills. The paper does not report end-to-end task success rates on LIBERO, so the practical impact of these skills on downstream task completion remains a boundary that is not addressed in the text.