AnyStep-WAM for Tunable-Budget World Action Models
September 26, 2026
AnyStep-WAM introduces a framework for adaptive inference in world-action models using budget-aligned teacher-trajectory distillation. It allows models to scale from one-step predictions to multi-step refinement based on the required precision of a specific action.
HOW THIS AFFECTS YOU
●
builderYou can implement scene-dependent computation, saving latency on simple actions while maintaining precision for critical ones.
●
researcherThis demonstrates how flow-map distillation can support variable-step denoising in robotics.