Training AI Without Writing A Reward Function, with Reward Modelling

3-Second Blueprint (English)

1. Hand-coded RL reward functions fail on complex tasks due to specification gaming and unpenalized negative side effects.
2. Reward modeling trains a supervised neural predictor from human pairwise comparisons of short agent trajectory clips.
3. Decoupling async human labeling, continuous reward-model updates, and fast RL rollouts scales alignment per human second.