Reproduce “Idiosyncrasies in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mingjie Sun, Yida Yin, Zhiqiu Xu, J. Zico Kolter, Zhuang Liu
This study demonstrates that Large Language Models (LLMs) exhibit consistent and identifiable "idiosyncrasies" in their generated text, enabling a classifier to reliably predict the source LLM. A neural network achieved 97.1% accuracy in distinguishing outputs from five different chat APIs, showing these unique patterns exist at lexical, formatting, and even semantic levels.
Key insights
- Large Language Models possess inherent, robust, and distinct 'idiosyncrasies' or 'fingerprints' that can be reliably detected even across different prompts and domains.
- These unique patterns are embedded at multiple linguistic levels, including lexical choices, structural formatting (e.g., markdown), and underlying semantic meaning, rather than just surface-level text.
- Idiosyncrasies are propagated during supervised fine-tuning with synthetic data, indicating a potential for inherited traits that could affect model diversity and lineage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.