There has been much recent interest in distributed demand response-where loads adjust in response to shortfalls (or excesses) in power supply-in order to maintain power quality, reduce dispatch costs, and even prevent or ameliorate blackouts. Some of the most promising candidates for demand response are thermal loads, such as heaters, air conditioners, and refrigeration units, which can often be switched OFF temporarily without inconveniencing the user. However, naive attempts to control such loads have experienced difficulties due to devices becoming synchronized in an undesirable way. Recent work using stochastic controls has been shown to be very promising, but still has certain practical limitations. In particular, these controls exhibit poor robustness and undesirable nominal device behavior. In this paper, we propose a novel stable closed-loop stochastic thermal controller that shows much greater robustness than the state of the art algorithm and provides nominal behavior very similar to that of ordinary hysteretic control. Our controller allows a large population of devices to react to the same demand response signal while maintaining an even distribution of device states. This allows them to act in unison to closely approximate a single deterministically controlled load.