ripple-edits-ke-methods-fail

IN premise — summaries/2026/08/24/cohen-2023-ripple-effects-s0-abstract.md

Created 2026-08-25T02:57:56+00:00

Prominent parameter-editing methods (ROME, MEMIT, Hernandez et al., Si et al.) fail to introduce consistent changes across logically related facts in the ripple graph

Summary

When you surgically edit a single fact inside a language model using the most well-known parameter-editing techniques, the model's other related facts don't update along with it, leaving the model internally inconsistent. This matters because it means these editing methods can create new contradictions rather than cleanly patching knowledge, so any system relying on them must verify that downstream, logically connected claims were also corrected.