When a large language model is fine-tuned, meaning it is adapted for a specific task by further training, proving which original model it came from is important for audits, licenses, and vulnerability tracking. Currently, this provenance is self-reported and unverifiable.
This paper proposes a method called TRACE-Edge that infers the parent-child relationship between models using only the log-likelihood, a score of how likely a model thinks a piece of text is. Fine-tuning increases this score for the training data, and the method exploits that the child inherits this increase from its parent, creating an asymmetric measure that reveals direction.
TRACE-Edge was tested on 25 model trees and achieved a parent-match accuracy of 0.74 when using the actual fine-tuning examples. It outperformed existing methods even when using non-training data from the same source, demonstrating that lineage can be verified without access to model weights.