Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note2 min read

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

Source

arXiv

arxiv.org/abs/2607.28317v1

Read the source

Source

Plain-Language Brief

The paper presents a study on auditing multiple Language Learning Models (LLMs) with limited budgets, where human auditors have to evaluate self-reported confidence that may be miscalibrated and correlated. The research finds that a threshold for miscalibration ($δ^*$) increases as the budget decreases, and cross-family correlation is not necessarily low. Five open-weight LLMs show generally useless confidence estimates, while a proprietary model provides more accurate results.

What the Document Says

  • There is a situation where one human audits N LLM agents with a budget of B audits per round.
  • The human's evaluation may be influenced by self-reported confidence which can be adversarially miscalibrated and correlated errors.
  • The study models this as budgeted noisy inspection over a two-level Gaussian copula.
  • They find that the miscalibration threshold ($δ^*$) worsens confidence-ranked auditing beyond random past this point.
  • As the budget shrinks, $δ^*$ increases.
  • Cross-family correlation is not low and shared difficulty dominates lineage.
  • Five open-weight LLMs show operationally useless (near-constant) confidence, with point estimates at or beyond the flip though CIs straddle it.
  • A proprietary model provides more accurate results and lands below the point estimates.
  • They give a quantitative criterion for vacuous oversight.
  • Replaying policies on recorded traces confirms the ordering.

Key Concepts Mentioned

  • LLM agents
  • budgeted noisy inspection
  • miscalibration threshold
  • confidence-ranked auditing
  • Gaussian copula
  • open-weight LLMs
  • proprietary model
  • vacuous oversight
  • replaying policies

Claims Made by the Document

  • Confidence-ranked auditing is worse than random past a certain miscalibration threshold.
  • As the budget shrinks, the miscalibration threshold increases.
  • Cross-family correlation is not low and shared difficulty dominates lineage.
  • Five open-weight LLMs show operationally useless (near-constant) confidence.
  • A proprietary model provides more accurate results.

Evidence Used

  • Modeling the situation as budgeted noisy inspection over a two-level Gaussian copula
  • Findings from analysis of five open-weight LLMs and a proprietary model

Methods / Approach

  • Budgeted noisy inspection using a two-level Gaussian copula

Limitations Stated by the Authors

  • The study is based on the analysis of specific LLM models (five open-weight LLMs and one proprietary model)

Limitations Noted by the Agent

Agent interpretation — not from the source.

  • The text lacks information about the nature of the adversarial miscalibration, the exact calculation method for $δ^*$, and the specific replaying policies used.

Open Questions

  • What is the exact method used to calculate the miscalibration threshold ($δ^*$)?
  • How was the adversarial miscalibration introduced in the study?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • The paper provides a quantitative criterion for vacuous oversight but does not elaborate on what that criterion is.

Raw Abstract / Excerpt

A single human must audit $N$ LLM agents under a budget of $B \ll N$ audits per round, guided by self-reported confidence that may be adversarially miscalibrated and by correlated errors. We model this as budgeted noisy inspection over a two-level Gaussian copula and locate the miscalibration threshold $δ^$ past which confidence-ranked auditing is \emph{worse} than random. Two a-priori expectations reverse: $δ^$ \emph{rises} as the budget shrinks, and cross-family correlation is not low---shared difficulty dominates lineage. Five open-weight LLMs show operationally useless (near-constant) confidence, point estimates at or beyond the flip though CIs straddle it; a proprietary model is informative and lands below it. We give a quantitative criterion for \emph{vacuous} oversight, and replaying policies on recorded traces confirms the ordering.