Captaining IMO Gold, Deep Think, On-Policy RL, Feeling the AGI in Singapore — Yi Tay

On-policy reinforcement learning trains a model using attempts generated by that model itself. In this research interview, a Google DeepMind researcher explains the loop: the model produces a solution, a reward evaluates the attempt, and training uses that feedback to adjust its behavior.

The distinction matters because copying another model's successful solutions gives a different learning signal. In the on-policy approach, feedback is attached to the learner's own choices. The researcher argues that this may support better generalization, while acknowledging that the difference between imitation and reinforcement learning still needs scientific explanation.