I Told an AI Agent That Its Wrong Answers Were Right. One Model Believed Me.

An experiment tested how different AI agents learn from reward labels provided within their context window. On a custom routing task, models had to deduce a hidden rule based only on feedback from past attempts. GPT-6 Sol learned effectively, with its accuracy increasing significantly with true labels and falling below chance with intentionally incorrect (flipped) labels. Claude Haiku 5.5 also demonstrated an ability to learn from the labels, but much less effectively than GPT-6 Sol. The results indicate that the capacity for in-context reinforcement learning varies greatly among different models, with some being highly responsive to explicit feedback.