There's a common assumption about reasoning models. Once the answer stops changing mid reasoning, the rest of the thinking is a waste. I tested it on DeepSeek-R1-Distill-Qwen, 1.5B and 7B, with 400…