Saved in:
Bibliographic Details
Main Authors: Li, Yanhang, Fan, Zhichao, Zhuang, Zexin
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.25492
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913161287827456
author Li, Yanhang
Fan, Zhichao
Zhuang, Zexin
author_facet Li, Yanhang
Fan, Zhichao
Zhuang, Zexin
contents Pairwise model comparisons drawn from foundation-model benchmarks ("A is safer than B") are read as quantitative verdicts but hinge on harness choices benchmark papers under-specify. We close one theory-benchmark loop on this primitive: a finite-envelope proposition tying a measurable pairwise-disagreement rate to whether the strict ordering admits a configuration-pair reversal, paired with a commit-stamped evaluation protocol that operationalises it on widely cited alignment benchmarks. On every benchmark we test, configuration choice alone can flip the pairwise verdict; the proposition isolates this strict-reversal failure mode.
format Preprint
id arxiv_https___arxiv_org_abs_2605_25492
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle SafetyRepro: Configuration-Conditional Rank Instability on Alignment Benchmarks
Li, Yanhang
Fan, Zhichao
Zhuang, Zexin
Machine Learning
Pairwise model comparisons drawn from foundation-model benchmarks ("A is safer than B") are read as quantitative verdicts but hinge on harness choices benchmark papers under-specify. We close one theory-benchmark loop on this primitive: a finite-envelope proposition tying a measurable pairwise-disagreement rate to whether the strict ordering admits a configuration-pair reversal, paired with a commit-stamped evaluation protocol that operationalises it on widely cited alignment benchmarks. On every benchmark we test, configuration choice alone can flip the pairwise verdict; the proposition isolates this strict-reversal failure mode.
title SafetyRepro: Configuration-Conditional Rank Instability on Alignment Benchmarks
topic Machine Learning
url https://arxiv.org/abs/2605.25492