ROME and MEMIT edits share common knowledge suppression mechanism

A new paper accepted to Findings of ACL 2026 uncovers a hidden similarity beneath popular knowledge editing techniques for transformer models. Methods like ROME and MEMIT modify MLP weights to update factual associations, but have long puzzled researchers due to their inconsistent behavior when editing related facts. The authors, led by Ali Holmov, hypothesized that despite weight changes appearing fact-specific, all edits actually leverage a common functional subspace. To test this, they trained a single binary mask over the set of edited weights. Remarkably, this mask reversed 80% of edits on the training set and over 70% on unseen test edits, proving that diverse edits share a unified underlying mechanism.

The mask's analysis revealed that edits work by suppressing overattention in later transformer layers—effectively hiding the original fact rather than overwriting it. When the mask was injected during the editing process itself, success rates plummeted from 98% to 38%, establishing that this suppression mechanism is necessary for edits to take effect. This explains why ROME and MEMIT fail to propagate changes to related facts: they don't truly modify the model's knowledge but only mask its expression. The discovery opens new avenues for detecting and defending against unwanted or malicious edits, as the common subspace provides a fingerprint for identifying tampered models.