One Human, $N$ Agents: Audit-Budget Allocation for LLM Agent Fleets under Miscalibrated, Correlated Confidence

TL;DR

This paper analyzes audit-budget allocation for large language model (LLM) agent fleets under miscalibrated confidence. It models the problem as budgeted noisy inspection and locates the miscalibration threshold past which confidence-ranked auditing is worse than random. The study finds that this threshold rises as the budget shrinks.

Why It Matters

The research has practical significance as it helps in understanding how to effectively allocate audit budgets for LLMs, which is crucial for ensuring their reliability and trustworthiness. It also highlights the importance of considering miscalibrated confidence and correlated errors in LLMs. This can inform the development of more robust and trustworthy AI systems.

Key Findings

• miscalibration threshold rises as budget shrinks
• confidence-ranked auditing can be worse than random past a certain threshold
• cross-family correlation is not low and shared difficulty dominates lineage