PCTD: Preference-Guided Counterfactual Task Decomposition for Agent Tool Retrieval
Researchers propose PCTD, a framework that addresses reward hacking in reinforcement learning-based task decomposition for tool retrieval. By leveraging counterfactual rewards to eliminate spurious correlations and preference rewards for structural supervision, PCTD improves the quality of decomposed subtasks and tool retrieval. Experiments on a new mobile multi-turn benchmark show that PCTD outperforms state-of-the-art methods in retrieval accuracy, decomposition quality, and generalization to unseen tools.
Why it matters: This work advances the reliability and generalization of AI agents in decomposing ambiguous instructions for tool retrieval, which is crucial for robust real-world applications.
Full story at: arXiv Information Retrieval ↗