What do LLMs think about on their own, when we let them think freely? We generated 250K “daydream” samples across models 🧠 GPT → coding Qwen → multiple-choice math exams Llama → literature DeepSeek → math, religion, psychology