Too Consistent to Detect: A Study of Self-Consistent Errors in LLMs

TL;DR

This paper investigates the challenge of detecting self-consistent errors in large language models, revealing their persistence and the limitations of current detection methods, and proposes a cross-model probing approach to improve detection accuracy.

Contribution

It formally defines self-consistent errors in LLMs, evaluates existing detection methods on them, and introduces a novel cross-model probe method that significantly improves detection performance.

Findings

Self-consistent errors remain stable or increase with LLM scale.

Current detection methods struggle to identify self-consistent errors.

The proposed cross-model probe method enhances detection accuracy across LLMs.

Abstract

As large language models (LLMs) often generate plausible but incorrect content, error detection has become increasingly critical to ensure truthfulness. However, existing detection methods often overlook a critical problem we term as self-consistent error, where LLMs repeatedly generate the same incorrect response across multiple stochastic samples. This work formally defines self-consistent errors and evaluates mainstream detection methods on them. Our investigation reveals two key findings: (1) Unlike inconsistent errors, whose frequency diminishes significantly as the LLM scale increases, the frequency of self-consistent errors remains stable or even increases. (2) All four types of detection methods significantly struggle to detect self-consistent errors. These findings reveal critical limitations in current detection methods and underscore the need for improvement. Motivated by the observation that self-consistent errors often differ across LLMs, we propose a simple but effective cross-model probe method that fuses hidden state evidence from an external verifier LLM. Our method significantly enhances performance on self-consistent errors across three LLM families.