「Our work addresses compounding uncertainty, heterogeneous signals, and data scarcity, yielding three key takeaways: (1) calibration relies on a hierarchy of diagnostic signals; (2) HTC features capture a transferable “uncertainty patterns” enabling strong cross-task generalization while exposing limits under distribution shift; and (3) a pretrained General Agent Calibrator (GAC) achieves the best ECE (zero-shot) on unseen tasks like GAIA, providing a plug-and-play foundation.」とエージェントの信頼度を総合的に評価していくフレームワーク