C0ldSmi1e··on Stealing Reasoning Traces from Proprietary LLM APIsWhy they use different models to decode the reasoning content? Can the the model decode it?