Article

Limits of Self-Correction in LLMs: An Information-Theoretic Analysis of Correlated Errors

2026-05-25

Abstract excerpt

We develop a diagnostic framework for evaluating when LLM self-evaluation can be trusted. The framework's central results are: (1) under a shared-blind-spot modeling assumption with joint conditional independence of evaluations given the shared failure structure, k rounds of self-critique provide information about correctness bounded by what the shared latent failure variable Z mediates---not by any independent ch...

Topics

Open a Topic to create a Post that cites this publication.

Identifiers and source

Literature Corpus work: ae3b5b29-436f-502c-bba0-8c5d299566e4
DOI: 10.20944/preprints202601.0892.v5

Open publication

Related research

Semantic proximity does not establish scientific evidence.

Click a neighbor to travelStep 1 · 12 closest

Limits of Self-Correction in LLMs: An Information-Theoretic Analysis of Correlated ErrorsDOI 10.20944/preprints202601.0892.v5Read

Protocol-Dependent Resolution in Finite-Sample Model SelectionSemantic proximity Read

Graduated Dissent: Budgeted Disagreement Resolution for Multi-Model InferenceSemantic proximity Read

Measuring What Survives: Pressure-Filtered Persistence and a Masking Index for LLM-Agent Safety EvaluationSemantic proximity Read

Graduated Dissent: Budgeted Disagreement Resolution for Multi-Model InferenceSemantic proximity Read

A Layered Guardrail Architecture for Agentic AI in Cloud-Native DevOps Pipelines: Design, Implementation, and Empirical Evaluation on Live AWS InfrastructureSemantic proximity Read

Human-Centric Zero Trust Identity Architecture for the Fifth Industrial Revolution: A JEPA-Driven Approach to Adaptive Identity GovernanceSemantic proximity Read

Detection and Mitigation of Mythos-Class Frontier Model Capabilities: A Layered Reference ArchitectureSemantic proximity Read

Bounded Recursive Self-Improvement Under Human Governance: Evidence from a Seven-Cycle Reinforcement Learning Contour ExperimentSemantic proximity Read

Reassessing One-Round Test-Time Refinement for Code GenerationSemantic proximity Read

Multi-Agent Human-AI Systems with Low-Code Platforms Enabling Adaptive Web Services and Real-Time Anomaly Remediation in Distributed ArchitecturesSemantic proximity Read

MutT prevents leakiness.Semantic proximity Read

Runtime Monitoring of Static Fairness PropertiesSemantic proximity Read