The loss paradox: SmolVLA loss: ~0.14 ACT loss: ~0.20 Yet ACT was far more reliable in practice. Lower training loss ≠ better real-world behavior when the inference mechanism itself introduces…