A research preview for steerable LLM reasoning, driven by self-improvement at test-time.