Mid-training is essential for LLM reasoning, IBM study shows

Adding a step between pre-training and post-training matters more than people may think.