State-dependent error correlations shape voting thresholds in committees of AI agents

Summary

The one paper that explicitly marries Sah-Stiglitz screening theory to LLM committees: it models error dependence that can differ between good and bad cases via an exchangeable Gaussian copula, shows shared errors create a positive asymptotic error floor for majority voting, and shows dependence changes the approval threshold that minimizes expected loss - with the framework relating organizational form to cost asymmetry (costlier false acceptances push toward hierarchy, cheaper ones toward polyarchy). Estimated on 174,384 votes from 28 language models over four binary-screening benchmarks with odd-to-even-item out-of-sample validation, identity-line R-squared rises from 0.840 under independence to 0.967, and cost-sensitive thresholding cuts scaled loss 15.73% below majority rule. Outside the requested window but reported because it is the nearest theoretical antecedent to this paper's structure-times-aggregation argument and is absent from the knowledge base - leaving it uncited would expose a priority problem.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-09-07 literature sweep as a high-priority candidate for the Architecture as Epistemology (ISR) paper (angle: MAS topology / structure vs. aggregation rule; aggregating LLM judgments; correlated errors / homogeneity across LLM agents; organization theory applied to AI agents; epistemic quality of LLM outputs; MAS failure and cost analyses). Verified against: Verified by opening https://arxiv.org/abs/2607.23931 on 2026-09-07: full abstract read verbatim, 2 authors (Haifeng Li, Mo Hai), submitted 27 July 2026, cs.CY. Pre-window; flagged as centrally missing rather than as a window find.

Critique / open questions

Queued — not yet assessed.

Key quotes

The aggregation benefit of a committee of artificial intelligence (AI) agents comes from complementary information across members. (abstract)

Classical voting guarantees assume independent errors. (abstract)