Model-stealing paper author says the original theorem was wrong, but the result can be repaired

ArthurConmy · x · 2026-08-04

The author says the original theorem in the model-stealing paper had a flaw, and the error can be patched with slightly stronger assumptions that hold for nearly all LLMs.

The attached screenshot explains the fix: replace the original assumption with one requiring the queried outer products to span the symmetric matrices and the quadratic design matrix to have the right rank. Under that condition, the orthogonality result still goes through. The thread also points to follow-up work that moves closer to the correct assumption, while noting that no follow-up paper explicitly called out the false inference in Lemma H.3.

Related event: ICML 2024 Paper Authors Acknowledge Error in Model Stealing Theorem(2 posts)→

Original post →

More from Research

Research channel →