Reasoning models: chain of thought, RL with verifiable rewards, test-time compute

A · 10-second intuition Let the model write out its working before answering, then use reinforcement learning on problems with checkable answers to make that working better.