PaperlayerView exact revision on arXiv

Source-linked AI summary

Correlated Errors in Large Language Models

Elliot Kim, Avi Garg, Kenny Peng, Nikhil Garg

arXiv:2506.07962v1cs.CLcs.AIcs.CYstat.ML

TL;DR

The paper asks whether different LLMs meaningfully differ in their errors, given limited large-scale empirical evidence. It evaluates error correlations across leaderboard and resume-screening data, finding correlated errors—including among more accurate models—and implications for judge-based evaluation and hiring.

  • Problem

    Large-scale empirical evidence is limited on whether available LLMs differ meaningfully in their errors, a question relevant to understanding ecosystem diversity and multi-agent systems.

    Source: §1 · p. 1

  • Method

    The authors analyze error correlations across leaderboard multiple-choice evaluations and offline numerical resume ratings, covering over 350 models.

    Source: §3.1 · p. 3, §6 · p. 9

  • Results

    Errors are correlated across LLMs, with greater correlation among more accurate models and models sharing providers or architectures; the HELM mean agreement rate when both models are wrong is 0.6.

    Source: §1 · p. 1, §3.2 · p. 4, §6 · p. 9

  • Takeaways & Limitations

    Error correlation has implications for LLM-as-judge evaluation, hiring, and multi-agent systems broadly.

    Source: §6 · p. 9

    The standardized multiple-choice and offline numerical resume evaluations provide only a limited view of LLM capabilities.

    Source: §6 · p. 9

Abstract

from arXiv · show

Diversity in training data, architecture, and providers is assumed to mitigate homogeneity in LLMs. However, we lack empirical evidence on whether different LLMs differ meaningfully. We conduct a large-scale empirical evaluation on over 350 LLMs overall, using two popular leaderboards and a resume-screening task. We find substantial correlation in model errors -- on one leaderboard dataset, models agree 60% of the time when both models err. We identify factors driving model correlation, including shared architectures and providers. Crucially, however, larger and more accurate models have highly correlated errors, even with distinct architectures and providers. Finally, we show the effects of correlation in two downstream tasks: LLM-as-judge evaluation and hiring -- the latter reflecting theoretical predictions regarding algorithmic monoculture.

1. Introduction

The paper addresses limited large-scale evidence on whether available LLMs differ meaningfully in their errors and what that means for multi-agent systems. Across three datasets, it finds substantial error correlation, including among more accurate models, with consequences for LLM-as-judge evaluation and hiring.

Motivation

  • The motivation is that model diversity may support robustness and reduce correlated failure, while its implications for high-stakes settings remain an empirical question.The paper frames this evidence as relevant to understanding ecosystem diversity and engineering multi-agent systems.Source: §1 · p. 1

Approach

  • The authors analyze responses from 349 HuggingFace models, 71 HELM models, and 20 models rating resume-job pairs.The datasets include leaderboard multiple-choice answers and resume-screening evaluations.Source: §1 · p. 1

Main findings

  • On HELM, the mean agreement rate across model pairs when both are wrong is 0.6.Across the three datasets, errors correlate highly across LLMs.Source: §1 · p. 1
  • More accurate models have more correlated errors, even after accounting for provider, architecture, and size.The authors say newer models that differ on the surface may be converging in their outputs.Source: §1 · p. 1, Figure 1 · p. 2

Downstream effects

  • The study examines how correlated errors affect LLM-as-judge evaluation and hiring outcomes, including who is hired and whether workers match with preferred firms.The judge setting uses another model to proxy ground-truth labels.Source: §1 · pp. 1–2

2. Related Work

Related work examines model diversity, shared components, and algorithmic monoculture, but leaves open how strongly errors correlate across different LLMs. This paper measures cross-model error similarity and studies its implications for evaluation and multi-agent settings.

Multi-agent and multi-model use

  • Prior work on multi-model use often assumes that using different LLMs can reduce self-preferencing or provide benefits when models differ [Panickssery et al., 2024; Wataoka et al., 2024].The paper contributes empirical evidence about how correlated different models’ errors are and what predicts that correlation.Source: §2 · p. 2

Algorithmic monoculture and systemic exclusion

  • Work on algorithmic monoculture considers correlated hiring decisions and systemic exclusion, while Peng & Garg (2024a) argue monoculture may benefit applicants on average.The paper simulates firms screening resumes with the same or different LLMs.Source: §2 · p. 2

Ecosystem evaluation

  • The component sharing hypothesis proposes that models sharing components are more likely to have correlated outputs [Toups et al., 2023].This paper studies that hypothesis at large scale.Source: §2 · p. 3

Generative diversity and monoculture in large language models

  • Prior work studies within-model generative monoculture; this paper instead finds cross-model errors more correlated than ground truth, especially within the same company and among larger models.The authors conclude that using multiple different models is not a panacea for within-model diversity concerns.Source: §2 · p. 3

Complementary and concurrent work

  • Compared with Goel et al. (2025), this paper also examines shared developers, base architectures, model size, and hiring-related downstream outcomes.Both works report that more capable or accurate models make more similar mistakes.Source: §2 · p. 3

3. Correlated Errors

The paper measures whether models converge on the same wrong answers and finds substantial error agreement across datasets. Agreement is higher among more accurate models and models with shared characteristics.

3.1. Data and methods

  • The study compares pairwise error agreement on two existing multiple-choice leaderboards and residual correlation in a generated resume-rating dataset.For leaderboard answers, it focuses on agreement conditional on both models being wrong to reduce confounding from accuracy.Source: §3.1 · pp. 3–4
  • The analysis uses HuggingFace and HELM multiple-choice answers and RESUMES ratings, measuring wrong-answer agreement for leaderboards and residual correlation against human labels for resumes.RESUMES contains 1,800 resume-job pairs rated by 20 LLMs; 450 pairs are hand-labeled.Source: §3.1 · pp. 3–4

3.2. Results

  • The mean agreement rate across model pairs when both are wrong is 0.423 on HuggingFace and 0.6 on HELM.Almost all pairs exceed the random-response baselines: 100% on HuggingFace and 97.5% on HELM.Source: §3.2 · p. 4
  • Some model pairs have exceptionally high error agreement: Google’s text-unicorn@001 and Writer’s palmyra-x-v3 agree on 0.9987 of questions where both are incorrect.The passage reports no publicly stated direct relationship between these models.Source: §3.2 · p. 4
  • Models from the same developer, with the same base architecture, or of similar sizes show higher error agreement; individually more accurate models are also more correlated.Included features explain 34% to 62% of variation in error agreement across the three datasets.Source: §3.2 · p. 4

4. Case Study: LLM as judge

The paper examines LLM-as-judge evaluation by comparing judged accuracy with ground truth. Judges inflate the accuracy of models less accurate than themselves and underinflate that of more accurate models.

Results

  • Judges systematically inflate the accuracy of models less accurate than themselves and underinflate the accuracy of models more accurate than themselves.The judge can mark a shared incorrect answer as correct, but cannot reward a model for a correct answer when the judge itself is wrong.Source: §4 · pp. 4–5

Evaluation setup

  • In the HELM plots, each dot represents a judged model; the axes show its true accuracy and accuracy inflation, with a vertical line at the judge’s true accuracy.Models from the judge’s provider or family are shown in red.Source: Figure 2 · p. 5

Implication

  • The authors recommend calibrating each model-judge pair’s error metrics using ground-truth data when comparing models or establishing error rates.They also report significant inflation for some models from the same provider.Source: §4 · p. 5

5. Case Study: LLMs in labor markets

The case study tests how correlated LLM hiring decisions shape exclusion and matching outcomes across labor-market settings. It finds that systemic exclusion and applicant welfare can move in opposite directions as firms shift from monoculture to polyculture.

5. Case Study: LLMs in labor markets

  • The study uses real resumes and job descriptions to compare five firm preference methods spanning same-LLM monoculture to random LLM selection.The market experiments use 60 applicants and 30 firms, each with capacity for one hire.Source: §5 · pp. 5–6, §5.2 · p. 6

5.1. Results: Systemic Exclusion

  • Around 20% of applicants remain systemically excluded even with 20 firms each using a random LLM.Firms interview the top quarter of applicants, and the same job description is used across firms.Source: §5.1 · p. 6, Figure 3 · p. 7
  • Same-company and latest-LLM markets have somewhat more systemic exclusion than random-LLM markets, but less than same-LLM markets.The differences between same-company, latest-LLM, and random-LLM settings are fairly small compared with the same-LLM setting.Source: §5.1 · p. 6, Figure 3 · p. 7

5.2. Results: Matching markets effects

  • Single-LLM markets yield the weakest matching for the highest-rated applicants, while latest-LLM markets maximize firm welfare.Latest LLMs are described as individually more accurate but more correlated than random LLMs.Source: §5.2 · p. 7, Figure 4 · p. 7
  • Single-LLM markets produce the best average applicant rank, while applicant welfare worsens as markets move toward random firm preferences.Average applicant rank is the rank of the matched firm in each applicant’s preference list.Source: §5.2 · p. 8, Figure 5 · p. 8
  • Markets with the highest systemic exclusion also produce the highest-quality applicant matches, revealing a trade-off between exclusion and welfare when matching effects count.Markets with correlated models have outcomes between markets using the same LLM and those using random preferences.Source: Figure 4 · p. 7, §5.2 · p. 8, Figure 5 · p. 8
  • Submitting more applications increases relative match probability, but the advantage is smaller in same-LLM markets than under uniformly random preferences.Applicants submitting 30 applications are approximately 2 times more likely to match than those submitting 1 in same-LLM markets, versus approximately 7 times under uniformly random preferences.Source: §5.2 · pp. 8–9, Figure 5 · p. 8

6. Conclusion

LLM errors correlate substantially, especially among individually accurate models and models sharing developers or base architectures. The paper connects this pattern to downstream evaluation and hiring, while noting limits in its measures and task coverage.

6. Conclusion

  • More accurate models and models sharing a developer or base architecture have more correlated errors, with implications for LLM-as-judge evaluation and hiring.The conclusion says increasing model performance is accompanied by convergence in errors.Source: §6 · p. 9
  • The correlation measure conditions on both models being wrong, but treats all incorrect answers alike and does not account for question difficulty.The authors identify these as limitations of current metrics and call for future metrics robust to them.Source: §6 · p. 9
  • The evaluation uses standardized leaderboard questions and offline numerical resume scores, so it covers only a limited view of model capabilities and tasks.The authors identify open-ended generation and complex reasoning as directions for future work; their analysis does not measure correlations induced by LLMs helping write resumes or job descriptions.Source: §6 · p. 9
  • Leaderboard developers can track model correlations using benchmark data already collected for individual evaluations, revealing pairwise patterns and informing model combinations.The paper notes that benchmark questions are often repeated across models.Source: §6 · p. 9

Impact Statement

The observed error correlations have implications for hiring and LLM-as-judge evaluation, and motivate further ecosystem evaluation and monitoring.

Impact Statement

  • Correlated errors create a risk that applicants are systematically screened out of opportunities in high-stakes settings such as hiring.The impact statement also identifies possible overestimation of model performance in LLM-as-judge settings.Source: Impact Statement · p. 10
  • In LLM-as-judge settings, correlated errors can lead models to inflate estimates of other models’ performance.Source: Impact Statement · p. 10
  • The analysis covers a subset of tasks where homogeneity may be a concern and serves as a foundation for further evaluation and monitoring.Source: Impact Statement · p. 10

A. Dataset Construction

The appendix describes leaderboard datasets and model features, then outlines a resume-screening evaluation built from clustered resumes and job descriptions. It also points to a figure showing the HuggingFace LLM-as-judge analysis.

A.1. HUGGINGFACE, HELM

  • The HuggingFace and HELM sources provide multiple-choice leaderboard responses and model metadata, including company and performance features.HuggingFace also supplies parameter counts and base architectures; the sources include 349 and 71 models, respectively.Source: §A.1 · p. 13
  • 349 HuggingFace models were selected from leaderboard candidates using API availability, developer representation, and accuracy-ranked fine-tuning models.The dataset covers 12,032 questions; HELM covers 14,042 questions across 57 categories, with four answer options per question.Source: §A.2 · p. 13, §A.1 · p. 13

A.2. RESUMES

  • 28 clusters were chosen to maximize similarity between resume and job-description clusters, then three high-similarity pairs supplied 60 resumes and 30 job descriptions.The selected pairs represent general IT, web development, and consulting/finance; Figure 6 depicts this cluster analysis.Source: §A.2 · p. 14, Figure 6 · p. 14
  • The resume evaluation used 20 models from Meta, Mistral AI, Amazon, Anthropic, and OpenAI to score resume-job fit using a structured criterion.Each pair received a single score, using two prompts.Source: §A.2 · p. 14

B.1. LLM-as-judge

  • Figure 8 presents the LLM-as-judge analysis for HuggingFace models.Source: §B.1 · p. 14

B.2. Matching Markets with LLM-Specified Applicant Preferences

The appendix examines downstream evaluations and matching-market experiments, including a replication with LLM-determined applicant preferences. It also reports prompt choices and regression setup.

B.2. Matching Markets with LLM-Specified Applicant Preferences

  • The matching-market experiments replace random applicant preferences with preferences determined by Llama 3.1 405B and examine applicant welfare and differential application access.Figure 10 replicates these experiments; its panels show average applicant welfare and the effect of differential application access.Source: §B.2 · p. 14, Figure 10 · p. 18
  • The main experiments use the “firm rate comb 2” prompt because it gives the LLM specific criteria for rating resume-job fit.The figure contrasts this prompt with “firm rate comb short 1”; applicant preferences use a similarly adapted prompt.Source: Figure 7 · p. 15
  • Each judge tends to inflate the accuracy of models less accurate than itself, especially models from the same provider or family.Figure 8 plots accuracy inflation against each model’s true accuracy, with same-provider or family models shown in red.Source: Figure 8 · p. 16
  • The supplied result fragment reports strong correlation among larger models from the same company.The fragment does not provide a correlation value or further conditions.Source: §B.2 · p. 17

C. Full Regressions

C.1. HUGGINGFACE

The HUGGINGFACE appendix tables report agreement measures and regression fit statistics for model-pair analyses.

C.1. HUGGINGFACE

  • Table 3 reports agreement rate when either model is wrong as the dependent variable; the regression has 60726 observations and reports R2 = 0.300.Source: Table 3 · p. 19, §C.1 · p. 19

C.2. HELM

The HELM appendix tables report agreement measures and regression fit statistics for model-pair analyses.

C.2. HELM

  • For HELM, the regression uses agreement rate when either model is wrong as its dependent variable, with 2,485 observations and R2 = 0.406.Source: Table 6 · p. 20, §C.2 · p. 20
  • Table 7 reports “Agreement on all questions: HELM”; the regression’s dependent variable is overall agreement rate, based on 2,485 observations, with R2 = 0.865.Source: Table 7 · p. 20, §C.2 · p. 20

C.3. RESUMES

The appendix lists tables on RESUMES correlation measures and identifies models analyzed for market, HELM, and HUGGINGFACE analyses.

C.3. RESUMES

  • Table 8 covers correlation in residuals when estimating job-resume fit.Source: Table 8 · p. 21
  • Table 9 covers agreement on job-resume fit ratings.Source: Table 9 · p. 21

D. Models analyzed

  • Table 10 lists models analyzed for market analysis and their correlation scores with the main-text prompt.Source: Table 10 · p. 21
  • The appendix also includes tables listing models analyzed from HELM and HUGGINGFACE.Source: Table 11 · p. 21, Table 12 · p. 23

Related papers

Loading 2506.07962v1…