"Teach models to think before answering, not after Hashimoto's team discovered that injecting a reinforcement learning objective into pre-training — rewarding the model for generating an intermediate 'thought' that genuinely improves its next-token prediction — produced models that performed better not just on reasoning benchmarks but on all downstream tasks after post-training. The insight is that 'the model is forced to be completely passive in the way that it learns to predict which token comes next,' but when encouraged to think first, it builds a fundamentally stronger foundation. He likens this to how humans benefit from learning math and logical thinking early in life, before later specialization. For anyone designing learning systems, training programs, or problem-solving workflows, the lesson is to build in a structured reflection step before the response, not after. Source: Tatsunori Hashimoto, The Cognitive Revolution (or similar Stanford AI podcast), Small Language Models and AI Democratization"