Agentic Uncertainty Quantification (AUQ): Training-Free Dual-Process Framework Converts Verbalized Confidence Into Active Control Signals

arXiv·medium signalXBlueskyLinkedInCopy link

ArXiv 2601.15703 proposes a training-free Dual-Process AUQ framework that reframes uncertainty not as a passive diagnostic but as an active bi-directional control signal throughout agent execution. System 1 (Uncertainty-Aware Memory) propagates confidence estimates across agent steps to prevent blind compounding; System 2 (Uncertainty-Aware Reflection) triggers targeted inference-time correction only when confidence thresholds are crossed. Evaluated on closed-loop benchmarks and deep research tasks with superior trajectory-level calibration over baselines.