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

Researchers propose a model for auditing a large number of LLM agents with limited budget, where agents' self-reported confidence may be misleading. The study finds that when the confidence is adversarially miscalibrated and correlated, relying on it for auditing can be less effective than random selection. The researchers identify a threshold for miscalibration beyond which this method becomes worse than random. They also find that five open-weight LLMs have confidence levels that are operationally useless and nearly constant, while a proprietary model provides more accurate information. The study presents a quantitative criterion for 'vacuous' oversight and confirms the findings using recorded traces.

What the Document Says

  • A single human audits N LLM agents with budget B, which is much smaller than the number of agents.
  • The human relies on the self-reported confidence from the agents, but these may be adversarially miscalibrated and correlated.
  • The study models this as a budgeted noisy inspection using a two-level Gaussian copula.
  • The researchers find a threshold δ* where confidence-ranked auditing is worse than random when confidence exceeds this threshold.
  • They find that the miscalibration threshold δ* increases as the budget shrinks.
  • Shared difficulty between families of LLMs dominates lineage in cross-family correlation.
  • The study finds that five open-weight LLMs have confidence levels that are operationally useless and almost constant, with point estimates at or beyond the flip but CIs straddling it.
  • A proprietary model provides more accurate information and has lower estimates.
  • The researchers present a quantitative criterion for 'vacuous' oversight.
  • Using recorded traces, they confirm the ordering of the policies.

Key Concepts Mentioned

  • LLM agents
  • budgeted noisy inspection
  • Gaussian copula
  • miscalibrated confidence
  • correlated errors
  • miscalibration threshold δ*
  • open-weight LLMs
  • proprietary model

Claims Made by the Document

  • Confidence-ranked auditing is worse than random when confidence exceeds a certain threshold.
  • The miscalibration threshold δ* increases as the budget shrinks.
  • Shared difficulty between families of LLMs dominates lineage in cross-family correlation.

Evidence Used

  • Modeling audit-budget allocation for LLM agent fleets under adversarially miscalibrated, correlated confidence using a two-level Gaussian copula.
  • Identifying the miscalibration threshold δ* and showing that confidence-ranked auditing is worse than random when it exceeds this threshold.
  • Observing that the miscalibration threshold δ* rises as the budget shrinks.
  • Noticing that shared difficulty between families of LLMs dominates lineage in cross-family correlation.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

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.