arxiv.org web signal

Residual signatures reveal LLM lineage from weights alone

TL;DR

  • A new arxiv preprint by Aman Singh Thakur and Rayan Khoury introduces 'centered residual signatures' to verify language model lineage from weights alone.
  • The paper reports perfect separation (AUROC=1.0) on residual-MLP and GPT-2 benchmarks and 76x speedup versus the nearest robust baseline on GPT-2.
  • The technique distinguishes fine-tuned, merged, pruned, and quantized descendants from independent models across six language-model families, including LLaMA-2.

Language models carry structural fingerprints in their weights, according to a new arxiv preprint by Aman Singh Thakur and Rayan Khoury, that reveal their training lineage even after fine-tuning, merging, pruning, or quantization. The technique they call 'centered residual signatures' aims to 'identify checkpoint-specific structural patterns across residual blocks,' derived from weight patterns alone.

The paper reports 'perfect separation (AUROC=1.0) on residual-MLP and GPT-2 benchmarks' and says the method ran '76x faster than the nearest robust baseline on GPT-2.' The 'projection-pairing signal was validated across six language-model families,' and the authors say they 'successfully identified relationships among publicly available LLaMA-2 checkpoints.'

Two experts in our Who's Who directory circulated the link. Per-family accuracy numbers aren't in the abstract, and of the six families only GPT-2 and LLaMA-2 get named.

Shared on Bluesky by 2 AI experts