Why negative cases are the real test of a lineage check
A lineage verification method that only reports positive matches is like a smoke detector that only tells you when there's a fire—it's the false alarms that reveal whether it actually works. In the papers here, the most convincing results are the ones that explicitly show how often the method says 'no' when it should. For example, modelDNA tested 15 real models against 8 candidate bases, which included 13 true parent relationships and 107 hard negatives—models that look similar but are not actually related. It achieved zero false positives at its reporting threshold, meaning it never wrongly flagged an unrelated model as a parent [1]. That's the kind of evidence that makes a lineage claim trustworthy.
Similarly, the AWM fingerprinting method was tested on 60 positive and 90 negative model pairs, and it achieved near-zero false positives across all six post-training modifications [2]. Without those negative cases, you'd have no idea whether the method was just matching everything. The authors explicitly highlight this as a strength, which is exactly why negative cases should be front and center in any lineage verification paper.
The gap between best-case and typical-case evidence
There's a big difference between a carefully controlled benchmark and the messy reality of open-weight models. The best studies here report perfect or near-perfect separation—AUROC of 1.0, meaning the method perfectly distinguishes related from unrelated models [1][3]. But that's under ideal conditions: known parentage, controlled merges, and a limited set of candidate bases. In practice, the lineage graph on Hugging Face is self-reported and often missing—over 60% of models document no parentage at all [1]. So the real-world challenge isn't just detecting lineage; it's doing so when you don't have a clean list of suspects.
The papers also differ in how they handle negative cases. modelDNA uses 'hard negatives'—models that are similar but unrelated—which is a tougher test than just random models [1]. The residual signature method also tests against independent and distilled models, and it explicitly distinguishes weight ancestry from behavioral similarity [3]. But not all methods are equally robust: under 'laundering' attacks that try to hide ancestry, some baselines lose accuracy, while the residual method stays unchanged [3]. This suggests that reporting negative cases isn't just about quantity—it's about how challenging those negatives are. A paper that only tests against obviously different models is less convincing than one that tests against near-misses.
What this means for you as a reader or practitioner
If you're evaluating a lineage verification paper, look for three things: how many negative cases were tested, how hard those negatives were, and whether the method's false-positive rate is reported. A paper that reports AUROC 1.0 with 107 hard negatives [1] is far more convincing than one that only shows positive matches. The same logic applies to any tool you might use—if a tool can't show you its false-positive rate on unrelated models, you should be skeptical.
The takeaway is simple: negative cases aren't a footnote; they're the core evidence that a lineage test actually works. The papers here that report them prominently—modelDNA with its 107 hard negatives [1], AWM with its 90 negative pairs [2], and the residual signature method with its independent and distilled models [3]—are the ones that give you confidence. So yes, papers should report negative cases more prominently, and as a reader, you should demand it.
About These Sources
This answer is built on 4 studies (1 peer-reviewed, 3 preprints) — published from 2024 to 2026, 4 from 2024 or later, collectively cited 133 times — selected as the most relevant from 4 studies that passed quality screening, drawn from 47 papers retrieved from a database of over 500 million.
Sources used in this answer
modelDNA: Calibrated Lineage Verification and Merge Decomposition from Sampled Weight Fingerprints
modelDNA reports a benchmark with 13 positive and 107 hard negative parent relationships, achieving AUROC 1.0 and zero false positives at its reporting threshold, and it emphasizes honest abstention over confident error.
AWM: Accurate Weight-Matrix Fingerprint for Large Language Models
AWM tests on 60 positive and 90 negative model pairs across six post-training categories, achieving near-zero false positives and perfect classification metrics, establishing a strong basis for reliable lineage verification.
Training Leaves Traces: Centered Residual Signatures for Language Model Lineage Verification
The residual signature method separates fine-tuned, LoRA-merged, pruned, and quantized descendants from independent and distilled models with AUROC 1.0, and remains unchanged under laundering attacks where baselines fail.
Chain-of-Verification Reduces Hallucination in Large Language Models
Chain-of-Verification reduces hallucinations across tasks by having the model plan and answer verification questions independently, but it does not address lineage verification or negative cases directly.