Language AI agents learn to generate their own reward signals instead of waiting for external feedback

Researchers built a method where language AI agents create internal guidance signals during both training and inference, creating a feedback loop where better performance produces better guidance, which then improves performance further. This means agents can improve faster by learning to steer themselves rather than waiting for external rewards from their environment.

AI training has been bottlenecked by sparse, delayed feedback from the environment — the agent has to act many times before knowing if it did something right. This research shows agents can close that loop themselves by generating intermediate guidance, which is faster and doesn't require external reward models. The practical effect is that language agents trained this way improve 8% over agents trained only on environmental feedback, and the improvement doesn't fade at inference time.