Hidden Reliability Risks in Large Language Models: Systematic Identification of Precision-Induced Output Disagreements DOI: XXXXXXX.XXXXXXX ISBN: 978-1-4503-XXXX-X/26/10 CCS: Software and its engineering Software testing and analysis CCS: Security and privacy Software and application security

Yifei Wang, Tianlin Li ∗, Xiaohan Zhang ∗, Xiaoyu Zhang, Wei Ma, Mingfei Cheng, Li Pan · 2026

Read Hidden Reliability Risks in Large Language Models: Systematic Identification of Precision-Induced Output Disagreements DOI: XXXXXXX.XXXXXXX ISBN: 978-1-4503-XXXX-X/26/10 CCS: Software and its engineering Software testing and analysis CCS: Security and privacy Software and application security by Yifei Wang, Tianlin Li ∗, Xiaohan Zhang ∗, Xiaoyu Zhang, Wei Ma, Mingfei Cheng, Li Pan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-19790.txt.

Description

Wang and six colleagues present PrecisionDiff, an automated differential testing framework. It hunts precision-induced behavioral disagreements in large language models. Models now run under varied formats to meet efficiency and resource limits, from standard float formats like bfloat16 and float16 to quantized integer formats like int16 and int8. Small cross-format inconsistencies are hard to spot and often missed by standard tests. PrecisionDiff generates test inputs and compares outputs across formats to expose hidden divergences.

Catalog record

CREATORYifei Wang, Tianlin Li ∗, Xiaohan Zhang ∗, Xiaoyu Zhang, Wei Ma, Mingfei Cheng, Li Pan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53876
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.19790

Browse the full library · About the Tower of Babel