Saved in:
Bibliographic Details
Main Authors: Yang, Eddie, Wang, Dashun
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.11898
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917270629908480
author Yang, Eddie
Wang, Dashun
author_facet Yang, Eddie
Wang, Dashun
contents Benchmarks underpin how progress in large language models (LLMs) is measured and trusted. Yet our analyses reveal that apparent convergence in benchmark accuracy can conceal deep epistemic divergence. Using two major reasoning benchmarks - MMLU-Pro and GPQA - we show that LLMs achieving comparable accuracy still disagree on 16-66% of items, and 16-38% among top-performing frontier models. These discrepancies suggest distinct error profiles for different LLMs. When such models are used for scientific data annotation and inference, their hidden disagreements propagate into research results: in re-analyses of published studies in education and political science, switching the annotation model can change estimated treatment effects by more than 80%, and in some cases reverses their sign. Together, these findings illustrate a benchmark illusion, where equal accuracy may conceal disagreement, with model choice becoming a hidden yet consequential variable for scientific reproducibility.
format Preprint
id arxiv_https___arxiv_org_abs_2602_11898
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Benchmark Illusion: Disagreement among LLMs and Its Scientific Consequences
Yang, Eddie
Wang, Dashun
Computation and Language
Benchmarks underpin how progress in large language models (LLMs) is measured and trusted. Yet our analyses reveal that apparent convergence in benchmark accuracy can conceal deep epistemic divergence. Using two major reasoning benchmarks - MMLU-Pro and GPQA - we show that LLMs achieving comparable accuracy still disagree on 16-66% of items, and 16-38% among top-performing frontier models. These discrepancies suggest distinct error profiles for different LLMs. When such models are used for scientific data annotation and inference, their hidden disagreements propagate into research results: in re-analyses of published studies in education and political science, switching the annotation model can change estimated treatment effects by more than 80%, and in some cases reverses their sign. Together, these findings illustrate a benchmark illusion, where equal accuracy may conceal disagreement, with model choice becoming a hidden yet consequential variable for scientific reproducibility.
title Benchmark Illusion: Disagreement among LLMs and Its Scientific Consequences
topic Computation and Language
url https://arxiv.org/abs/2602.11898