In LLMs (Large Language Models), serial verification gates are a core reliability primitive: a candidate answer is returned only if $k$ verifiers all accept it. Under conditionally independent gates, the recent Odds Law shows that posterior log-odds grow linearly in $k$, leading to an exponential decay of failure rates. However, a tight theory of partially correlated verifier cascades remains open. This note provides a minimal such theory.
First, modeling the false-accept rate per instance of the generator’s own errors as a latent variable $\alpha \sim G$ (de Finetti), the exact cascade posterior is given by $\ell_k = \ell_0 - \ln m_k$, where $m_k$ is the $k$-th moment of $G$. The key points are:
- $\ell_k$ is concave in $k$ for every non-degenerate $G$ — the Odds Law is its tangent at the first gate and serves as an upper bound;
- For Beta$(a,b)$ latents, failure decays polynomially as $1-r_k \asymp k^{-b}$, with correlation parameter $\rho_v = 1/(a+b+1)$;
- A blind-spot atom of mass $1-\pi$ at $\alpha=1$ caps the evidence extractable from any number of gates at $-\ln(1-\pi)$ nats, causing reliability to saturate below 1;
- Allowing the true-accept rate to vary ($\beta \sim H$) results in a trichotomy — gates may eventually always help, plateau, or actively harm — determined by the upper-tail exponents of $G$ and $H$, with a closed-form crossover at $k^\dagger$.
The mechanism is survivorship: errors that survive the gates are those with high $\alpha$. This theory is measurable: $R$ repeated verdicts per instance identify the first $R$ moments of $G$, thus two verdicts can identify $\rho_v$; beta-binomial likelihood and NPMLE recover the reliability curve and the ill-posed ceiling.
In synthetic tests, independence-based extrapolation underestimated failure by 20x at $k=5$ and about 3000x at $k=10$; the correlated fit at $R=8$ accurately tracks held-out depths. The practical lever is decorrelation — changing model family, modality, or evidence source — rather than simply adding more gates.
Blogger's Review: This study delves into the verifier cascade issue in LLMs, proposing a theoretical framework for understanding partial correlation, which provides new insights into optimizing model reliability. By modeling latent variables, the research reveals the complexities of error propagation and emphasizes the importance of decorrelation strategies, making it a noteworthy contribution to the field.