Open-weight language models are frequently fine-tuned, quantized, pruned, or merged, yet their provenance often remains undocumented. This raises a critical question: can the weights alone reveal whether two compatible checkpoints share a common ancestry? A new research paper introduces a data-free, white-box approach to lineage verification.
The method, called Centered Residual Signatures, removes the identity-aligned component that arises from residual training, which alone cannot establish ancestry. Instead, it compares checkpoint-specific structure across residual blocks, producing a symmetric lineage score calibrated against independent checkpoints.
On residual-MLP and GPT-2 benchmarks, the score perfectly separates fine-tuned, LoRA-merged, pruned, and quantized descendants from independent and distilled models (AUROC=1.0), distinguishing weight ancestry from behavioral similarity. Under function-preserving checkpoint laundering experiments, weight-space baselines lose margin or fail, while the proposed score remains unchanged and runs 76x faster than the nearest robust baseline on GPT-2.
The projection-pairing signal appears across six language-model families and beyond. A case study correctly identifies 3 related and 7 unrelated LLaMA-2 public checkpoints, establishing a passive, data-free provenance signal for compatible open-weight models.