ACTION-RATING teaches language agents when to ask for clarification

A new arXiv paper proposes ACTION-RATING, a framework for teaching hierarchical language agents when to ask clarifying questions.

Instead of treating clarification as a separate uncertainty trigger, the method puts asking inside the agent’s action space. That means help-seeking competes directly with acting at each decision point.

The work targets a common agent failure mode: committing to the wrong branch because the system does not recognize that it lacks information soon enough.