Partially Correlated Verifier Cascades in LLM Harnesses: Concave Log-Odds, Polynomial Reliability, and Blind-Spot Ceilings
The paper argues that repeated LLM verifier gates stop buying exponential reliability once their errors are correlated.
Han models false accepts as a latent per-instance rate, making the cascade’s evidence gain concave rather than linear. In the Beta case, failure falls only polynomially, and blind spots can impose a hard ceiling no matter how many gates are added. Synthetic tests in the note show independence-based extrapolation badly understating failure at deeper cascades. The stated practical fix is decorrelating the checks by changing model family, modality, or evidence source, not simply adding more gates. HF Daily Papers' note
Han models false accepts as a latent per-instance rate, making the cascade’s evidence gain concave rather than linear. In the Beta case, failure falls only polynomially, and blind spots can impose a hard ceiling no matter how many gates are added. Synthetic tests in the note show independence-based extrapolation badly understating failure at deeper cascades. The stated practical fix is decorrelating the checks by changing model family, modality, or evidence source, not simply adding more gates. HF Daily Papers' note
score 4