This isn't really surprising. LLM reasoning and before that, chain of thought prompting are essentially forms of test-time compute scaling.