From Training Exposure to Membership Evidence: A Controlled Study of OLMo 2

Summary

Researchers re-ran OLMo 2 1B's 50-billion-token mid-training stage, swapping in 7,814 matched document pairs with a coin flip deciding which copy was trained on. A single exposure raises a document's average likelihood but rarely identifies it: the reference-ratio attack reaches an AUC of 0.767 yet catches only 3.8% of members at a 1% FPR. Paired comparisons across 13 matched pairs give 80% power to detect training on a collection.

Related items

• InfoStop counting tokens, start counting what shipsSame vendor · MIT Technology Review
• MediumHundreds of Mundane AI Apps Are Hiding Sexually Abusive Deepfake FeaturesSame vendor · Wired (Security)
• InfoThe agentic web is hereSame vendor · MIT Technology Review
• InfoHumans are still doing the work behind some AI agentsSame vendor · CNBC Technology
• InfoAmazon AWS posts 39% revenue growth as cloud rivals also accelerate Same vendor · CNBC Technology