Agentic Uncertainty Quantification

TL;DR

The paper presents AUQ, a training-free Dual-Process Agentic UQ framework that converts verbalized uncertainty into bidirectional control to counteract the Spiral of Hallucination in long-horizon agents. It jointly implements System 1 forward uncertainty propagation via Uncertainty-Aware Memory (UAM) and System 2 inverse uncertainty calibration via Uncertainty-Aware Reflection (UAR), enabling dynamic trade-offs between efficient execution and deep deliberation. Across embodied (ALFWorld), web (WebShop), and open-ended (DeepResearch) tasks, AUQ improves trajectory-level calibration and task success, achieving superior performance and discrimination compared with strong baselines. The framework introduces forward and inverse uncertainty formulations, trajectory-level calibration metrics, and a memory-expansion mechanism, offering a principled, scalable path toward more reliable autonomous agents with adaptive compute budgeting.

Abstract

Although AI agents have demonstrated impressive capabilities in long-horizon reasoning, their reliability is severely hampered by the ``Spiral of Hallucination,'' where early epistemic errors propagate irreversibly. Existing methods face a dilemma: uncertainty quantification (UQ) methods typically act as passive sensors, only diagnosing risks without addressing them, while self-reflection mechanisms suffer from continuous or aimless corrections. To bridge this gap, we propose a unified Dual-Process Agentic UQ (AUQ) framework that transforms verbalized uncertainty into active, bi-directional control signals. Our architecture comprises two complementary mechanisms: System 1 (Uncertainty-Aware Memory, UAM), which implicitly propagates verbalized confidence and semantic explanations to prevent blind decision-making; and System 2 (Uncertainty-Aware Reflection, UAR), which utilizes these explanations as rational cues to trigger targeted inference-time resolution only when necessary. This enables the agent to balance efficient execution and deep deliberation dynamically. Extensive experiments on closed-loop benchmarks and open-ended deep research tasks demonstrate that our training-free approach achieves superior performance and trajectory-level calibration. We believe this principled framework AUQ represents a significant step towards reliable agents.

Figures (9)

• Figure 1: Overview of Dual-Process Agentic UQ Framework. To address the spiral of hallucination challenges in long-horizon agents, AUQtransforms verbalized uncertainty into active, bi-directional control signals, comprising two complementary mechanisms: System 1 (Uncertainty-Aware Memory, UAM), which implicitly propagates verbalized confidence and semantic explanations to prevent blind decision-making; and System 2 (Uncertainty-Aware Reflection, UAR), which utilizes these explanations as rational cues to trigger targeted inference-time resolution only when necessary. This allows the agent to balance efficient execution and deep deliberation dynamically.
• Figure 2: Generalization on Deep Research. (Left) AUQoutperforms the Enterprise Baseline (EDR) across four diverse LLM backends. AUQconsistently outperforms the Enterprise Baseline (EDR). (Right) Sensitivity analysis of confidence threshold $\tau$. The performance is robust across models in the high-confidence regime.
• Figure 3: Internal Belief Dynamics on ALFWorld. (Left) Trajectory confidence evolution comparing UAM-Only (Dashed) and Dual-ProcessAUQ(Solid). (Right) Scatter plots of Pre- vs. Post-Reflection confidence (GPT-4o & GPT-5.1). Yellow boxes show summary statistics.
• Figure 4: Effect of memory length on AUQand ReAct.
• Figure 5: Model generalization and memory expansion.