From Passive Metric to Active Signal: The Evolving Role of Uncertainty Quantification in Large Language Models

TL;DR

This survey explores how uncertainty quantification in Large Language Models has evolved from passive diagnostics to active control signals, enhancing reliability and trustworthiness in AI systems.

Contribution

It provides a comprehensive overview of the shift towards using uncertainty as an active control mechanism in LLMs, grounded in Bayesian and conformal frameworks.

Findings

Uncertainty guides advanced reasoning for computation optimization and self-correction.

Uncertainty informs autonomous agents for better decision-making and tool use.

In reinforcement learning, uncertainty helps mitigate reward hacking and supports self-improvement.

Abstract

While Large Language Models (LLMs) show remarkable capabilities, their unreliability remains a critical barrier to deployment in high-stakes domains. This survey charts a functional evolution in addressing this challenge: the evolution of uncertainty from a passive diagnostic metric to an active control signal guiding real-time model behavior. We demonstrate how uncertainty is leveraged as an active control signal across three frontiers: in \textbf{advanced reasoning} to optimize computation and trigger self-correction; in \textbf{autonomous agents} to govern metacognitive decisions about tool use and information seeking; and in \textbf{reinforcement learning} to mitigate reward hacking and enable self-improvement via intrinsic rewards. By grounding these advancements in emerging theoretical frameworks like Bayesian methods and Conformal Prediction, we provide a unified perspective on this transformative trend. This survey provides a comprehensive overview, critical analysis, and practical design patterns, arguing that mastering the new trend of uncertainty is essential for building the next generation of scalable, reliable, and trustworthy AI.