Tracing LLM Lineage Using Spectral Fingerprints in Weight Space.

The complex development pipelines of open-weight large language models (LLMs) often obscure their origins, ownership, and evolutionary paths, posing challenges for provenance and governance. This study explores the concept of LLM "biometrics," investigating whether intrinsic fingerprints exist within a model's weight space that can reveal its lineage without access to training data. The problem is framed as lineage discrimination, aiming to differentiate between independently originated, same-series, and shared-base models. A unified geometric fingerprinting framework is proposed, which examines weight matrices from two complementary angles: spectral energy, captured by singular value distributions to reflect global magnitude patterns, and subspace alignment, quantified by subspace deviations to capture directional geometry. Experiments on over 110 diverse LLM pairs demonstrate a clear hierarchy of structural similarity. Spectral energy reliably distinguishes independently trained models and different model families, while subspace alignment enables more granular discrimination among closely related models, even identifying variations due to dataset scale or post-training procedures. This indicates that weight-space geometry offers a robust and interpretable signal for understanding LLM lineage.