Temporal Difference Calibration in Sequential Tasks: Application to Vision-Language-Action Models