OriginBlame cuts over-deletion from 101x to 1.3x with record-level data provenance

OriginBlame, a new data provenance system, reduces over-deletion from 101x to 1.3x when honoring data-removal requests by pinpointing exact training records to remove. This enables model trainers to implement unlearning algorithms 42% more effectively, saving weeks of compute and preserving model performance with only 1-4% pipeline overhead.

arXiv https://arxiv.org/abs/2607.13037 OriginBlame cuts over-deletion from 101x to 1.3x with record-level data provenance Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Data provenance systems can now pinpoint exact training records to be removed for a given author, reducing dataset-level over-deletion from 101x to 1.3x, and this capability enables model trainers to implement unlearning algorithms 42% more effectively, directly impacting the efficiency of handling data removal requests in production LLM environments. Record-level provenance cuts over-deletion from 101× to 1.3× when honoring data-removal requests. This means you can now comply with revocation demands without nuking entire datasets or retraining from scratch, saving weeks of compute and preserving model performance—just add 1–4% pipeline overhead.