One Human, $N$ Agents: Audit-Budget Allocation for... | AI Research

This paper addresses the challenge of human oversight for fleets of LLM agents. As organizations deploy multiple agents that generate outputs faster than a human can verify, regulators require effective oversight. The authors model this as a "budgeted noisy inspection" problem, where a single supervisor has a limited audit budget ($B$) to check a much larger number of agents ($N$). The core difficulty is that agents provide self-reported confidence scores that are often adversarially miscalibrated, and their errors are correlated due to shared task difficulty rather than just their underlying lineage.

The Audit Flip Threshold

The researchers identified a "flip threshold" ($\delta^*$) for confidence-ranked auditing. When the miscalibration of an agent's confidence exceeds this threshold, using that confidence to prioritize which outputs to audit becomes less effective than random selection.
Contrary to initial expectations, the study found that this threshold rises as the audit budget shrinks. This means that when budgets are very tight, confidence-ranked auditing remains useful because the supervisor only checks the most extreme low-confidence cases. When budgets are more generous, the supervisor is forced to audit mid-range confidence outputs, which are more likely to be "poisoned" by miscalibration, causing the strategy to fail sooner.

Correlation and Oversight

The paper introduces a two-level Gaussian copula model to account for how agent errors correlate. The authors found that shared item difficulty—rather than the specific family or lineage of the model—is the primary driver of cross-family error correlation.
Because errors are clustered, the researchers propose a "diversity-Bayes" policy that uses audit results from one agent to update beliefs about its siblings. This approach helps mitigate the risk of "vacuous" oversight, a state where no policy performs significantly better than doing nothing. The study provides a quantitative criterion for identifying when oversight has become "rubber-stamping," finding that at very low budgets, oversight is often vacuous regardless of the agents' confidence levels.

Performance of Current Models

The authors tested six LLMs—five open-weight models (Qwen3 variants, Mistral-7B, and Phi-4-mini) and one proprietary model (gpt-4o-mini)—on GSM8K and HotpotQA benchmarks. The open-weight models exhibited near-constant confidence scores, rendering their self-reported confidence operationally useless for auditing. Their performance placed them at or beyond the flip threshold. In contrast, the proprietary model provided more informative confidence scores and remained below the flip threshold, making it more suitable for confidence-ranked auditing.

Limitations

The authors note several constraints on their findings:

• Persistence: The advantage of correlation-aware auditing collapses if errors are not persistent (i.e., if the agents are frequently fine-tuned).

• Verification: The effectiveness of posterior transfer depends on having a sufficiently accurate verifier or strong correlation between agents.

• Scope: The study used a small fleet of 15 agents for policy replay, and the results for the proprietary model serve as a single robustness point rather than a comprehensive survey of hosted models.

• Elicitation: The results are based on a specific, locked elicitation protocol; different methods of prompting or measuring confidence could yield different operating points.