Comment by kangalioo

Thinking mode is exactly this. The LLM is allowed to dump thoughts, correcting itself along the way, and only surfaces the cleaned-up final answer to the user