Setting the temperature to 0 should give deterministic results but that's not any better - it's just hiding the huge variance by only taking one sample.