Entity · paper

Partially Correlated Verifier Cascades in LLM Harnesses: Concave Log-Odds, Polynomial Reliability, and Blind-Spot Ceilings

paperactiveprovisional

partially-correlated-verifier-cascades-in-llm-harnesses-concave-log-odds-polynomial-reliability-and-blind-spot-ceilings-904bfdcf

·1 events·first seen 5d ago

Aliases: Partially Correlated Verifier Cascades in LLM Harnesses: Concave Log-Odds, Polynomial Reliability, and Blind-Spot Ceilings

Co-occurring entities

Odds Law

More like this (12)

LLM-as-a-Verifier Scaling LLM Reasoning from Minimal Labels: A Semi-Supervised Framework with a Lightweight Verifier When Good Verifiers Go Bad: Self-Improving VLMs Can Regress on New Tasks Operadic consistency: a label-free signal for compositional reasoning failures in LLMs The Count Is There, but Misaligned: Understanding and Correcting Counting Failures in VLMs Which Models Are Our Models Built On? Auditing Invisible Dependencies in Modern LLMs Continual LLM Upcycling: A Predictor-Gated Bank-Wise Sparsity Training Recipe for Dense-to-Sparse LLMs The Illusion of Equivalency: Statistical Characterization of Quantization Effects in LLMs Surrogate Fidelity: When Can Open LLMs Explain Closed Ones?Knowledge Knows, Verbalization Tells: Disentangling Latent Directions for Mathematical Solvability in LLMs When are likely answers right? On Sequence Probability and Correctness in LLMs Words Speak Louder Than Code: Investigating Cognitive Heuristics in LLM-Based Code Vulnerability Detection

Recent events (1)

6arXiv · cs.AI·5d ago·source ↗

Theory of partially correlated verifier cascades in LLM harnesses: concave log-odds and reliability ceilings

A new arXiv preprint provides a formal theory of serial verification gates in LLM harnesses when verifiers are partially correlated, closing an open problem noted in the recent Odds Law paper (arXiv:2606.15712). Using a de Finetti latent-variable model for per-instance false-accept rates, the authors derive that cascade reliability grows only polynomially (not exponentially) under correlation, that blind-spot atoms impose hard ceilings on extractable evidence, and that adding more gates can sometimes actively harm reliability. Synthetic experiments show independence-based extrapolation underestimates failure by 20x at k=5 and ~3000x at k=10, with the practical implication that decorrelation (changing model family, modality, or evidence source) is the correct lever, not adding more gates.

Evaluation and Benchmarking Agent and Tool Ecosystem Odds Law Partially Correlated Verifier Cascades in LLM Harnesses: Concave Log-Odds, Polynomial Reliability, and Blind-Spot Ceilings