llm reasoning without tokens