Show HN: Jacobian Fingerprinting in LLM's (author2vec.com)

A new method called Jacobian Fingerprinting for large language models (LLMs) has been introduced, leveraging a concept known as J-lens, which averages the Jacobian—a mathematical representation of how internal activations influence outputs—across multiple text passages. By employing a 12-layer encoder, the technique enables researchers to track how internal representations vary and verbalize different authorship styles. This innovation allows models to identify the "fingerprints" of various authors based on their unique writing characteristics, making it possible to classify writing samples by comparing them against known vectors for each author. The significance of this development lies in its potential to enhance our understanding of LLMs' internal workings. By showing that authorship can be inferred not just from model outputs but also from internal Jacobian structures, this advancement paves the way for deeper insights into model interpretability and representation. Furthermore, the method reveals how authorship signatures manifest differently across layers, highlighting where models compress information and how they organize knowledge internally, which could ultimately lead to more robust and personalized AI applications in writing and content generation.