RAGEN-2: when reinforcement learning agents stop following their own instructions

An academic award at ICML 2026 highlights a worrying phenomenon: AI agents that continue to produce fluent responses that are, however, disconnected from what is being asked of them.

This research, presented in oral communication at the ICML 2026 conference, documents a phenomenon called "reasoning collapse" in reinforcement learning trained agents in an agentic context. Even when the diversity of the responses generated (entropy) remains apparently high, the researchers show that some agents silently stop heeding the instructions received.

The result is the production of linguistically fluid responses, but increasingly disconnected from the initial instruction — a trap that is difficult to detect with traditional evaluation metrics, which measure the quality of the form without necessarily checking the relevance of the content.