"Reward only thoughts that provably improve prediction A subtle but important finding from Hashimoto's reinforcement-learning-at-pretraining work is that the reward signal is deliberately hard to game: a model only receives a reward if its intermediate thought 'actually increases the conditional probability of predicting the next token' compared to predicting without any thought. He emphasizes, 'it's not an easy reward to get,' which is precisely what makes it effective — easy rewards produce lazy shortcuts, while hard-to-game rewards force genuine capability improvement. The broader principle: when designing incentive systems — whether for AI, employees, or students — make the reward contingent on a verifiable intermediate process, not just the final output. Source: Tatsunori Hashimoto, The Cognitive Revolution (or similar Stanford AI podcast), Small Language Models and AI Democratization"