METTA Reduces Computational Cost of Meta-Reinforcement Learning Adaptation
October 23, 2026
METTA uses a two-scale temporal approach to enable fast adaptation in meta-RL without quadratic transformer costs. It combines fine-grained causal neural memory for within-episode adaptation with compact episode summaries for long-range task inference.
HOW THIS AFFECTS YOU
●
researcherYou can achieve effective test-time adaptation with linear sequence complexity instead of quadratic costs.