DocumentWorkflowTesting
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
We conduct the first comprehensive empirical study examining how input format (TXT, JSON, MD, CSV) affects the behaviors of large language models across four real-world task domains. Our findings reveal that the same LLM given identical data in different formats can produce contradictory answers, with accuracy swings of up to 56 percentage points, and measurable format-induced fairness violations across demographic attributes. This repository contains the prototype of both the dataset construction pipeline and the LLM query pipeline, as well as our datasets and analysis results.
Fields of Study & Subjects
Topic Modeling
Ethics and Social Impacts of AI
Natural Language Processing Techniques
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 27 March 2026
- Language
- EN