LLMs and the Madness of Crowds

TL;DR

Analysis of patterns of incorrect answers produced by large language models during evaluation reveals that the incorrect responses are not randomly distributed but systematically correlated across models, providing new insights into the underlying structures and relationships among LLMs.

Abstract

We investigate the patterns of incorrect answers produced by large language models (LLMs) during evaluation. These errors exhibit highly non-intuitive behaviors unique to each model. By analyzing these patterns, we measure the similarities between LLMs and construct a taxonomy that categorizes them based on their error correlations. Our findings reveal that the incorrect responses are not randomly distributed but systematically correlated across models, providing new insights into the underlying structures and relationships among LLMs.

Figures (6)

• Figure 1: Histograms of answers selected by gpt-4o-2024-08-06 for nine multiple-choice problems without context.
• Figure 2: Histograms of answers selected by gpt-4o-2024-08-06 for nine multiple-choice problems without questions or context. Labels "correct" and "incorrect" are mirrored from Figure \ref{['fig:histo_by_answer']} for comparability, although they are not strictly meaningful without a question.
• Figure 3: Histograms of answers selected for Question 3 by seven different LLMs.
• Figure 4: LLM similarity measured by correlation of errors on the MMLU-Pro test. Correlation is measured by $z$-score.
• Figure 5: Hierarchical clustering of LLMs based on correlated errors.