A phenomenon where the model learns to place its initial output near the fixed point, allowing inference without iteration.