Why static fingerprints miss the fine-tuning story
Most lineage detection methods compare model architectures or weight-space geometry, but that only captures a snapshot. A 2026 study on LLM 'biometrics' found that spectral energy—the overall magnitude pattern of weights—can reliably tell independently trained models apart, but it fails to distinguish models that share a base and then diverged through fine-tuning [4]. To catch those finer differences, you need subspace alignment, which looks at the directional geometry of weights [4]. So a test that only checks static similarity might wrongly conclude two models are unrelated when they actually share a lineage through fine-tuning.
Another 2026 paper argues that true lineage is about the joint trajectory of knowledge evolution and parameter modification during fine-tuning [1]. They propose using model editing to quantify parameter-level changes and then verify the arithmetic consistency of knowledge vectors across models [1]. This suggests that lineage tests that ignore how knowledge evolves during training will miss the dynamic aspect that defines lineage.
The data provenance blind spot
Lineage detection for models often focuses on the model itself, but it can overlook the data that shaped it. A 2024 paper on AI-driven data provenance highlights the need to track and verify data lineage, especially in healthcare and finance [2]. It shows that systems must balance processing speed against accuracy—faster processing improved accuracy in healthcare but increased false positives in finance [2]. For language models, this means a lineage test that doesn't account for training data provenance could miss critical influences, like whether a model was trained on biased or proprietary data.
This is a different kind of lineage than model-to-model relationships. While weight-space fingerprints can tell you if two models share a base, they can't tell you what data went into that base. The 2024 study emphasizes that data lineage is domain-specific and requires tailored solutions [2]. So a comprehensive lineage verification would need to combine model-level and data-level signals.
When lineage traces are incomplete or too large
Real-world lineage is messy: training pipelines are complex, and traces are often incomplete. Meta's Unified Lineage System (ULS), described in a 2025 paper, tackles this by stitching together incomplete lineage traces and processing a lineage graph with billions of nodes and edges [3]. This shows that at scale, lineage detection can't rely on perfect records—it has to infer and fill gaps, which introduces uncertainty.
For language models, this means a lineage test might fail to reveal the full chain of development if some intermediate steps aren't recorded. ULS also customizes precision and recall, meaning you can trade off between catching all possible lineage links and avoiding false ones [3]. That trade-off is a limitation: a test optimized for precision might miss true lineage, while one optimized for recall might flag false relationships.
About These Sources
This answer is built on 4 studies (2 peer-reviewed, 2 preprints) — published from 2024 to 2026, 4 from 2024 or later — selected as the most relevant from 4 studies that passed quality screening, drawn from 39 papers retrieved from a database of over 500 million.
Sources used in this answer
Attesting Model Lineage by Consisted Knowledge Evolution with Fine-Tuning Trajectory
Proposes a lineage attestation framework that uses model editing and knowledge vectorization to verify the joint trajectory of knowledge evolution and parameter modification, tested across classifiers, diffusion models, and LLMs.
AI-Driven Data Provenance: Tracking and Verifying Data Lineage
In a study of AI-driven data provenance for healthcare and finance, found that processing speed improves accuracy in healthcare but increases false positives in finance, highlighting the need for domain-specific tuning.
Unified Lineage System: Tracking Data Provenance at Scale
Describes Meta's Unified Lineage System, which tracks data flows at scale with a graph of billions of nodes and edges, and uses techniques to stitch incomplete lineage traces and customize precision/recall.
Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space
Across 110+ open-weight LLM pairs, found that spectral energy in weight space reliably distinguishes independent models, while subspace alignment is needed for fine-grained discrimination among closely related models.