Knowing When to Ask: Self-Gated Clarification for Hierarchical Language Agents

The paper introduces ACTION-RATING, a novel framework for hierarchical language agents that integrates clarification requests into the agent's action space, allowing them to compete with decision-making at critical points. This approach reveals two information-seeking modes—mandatory and opportunistic—and demonstrates a significant improvement in Information-Seeking Effectiveness (ISE) from 50% to 74% across three benchmarks using 9 LLMs. The findings highlight the importance of structured help-seeking in enhancing decision-making accuracy, suggesting that better localization could further improve performance.