rome-quantitative-benchmark-results
IN premise — summaries/2026/08/24/meng-2022-rome-s3-interventions-on-weights-for-understanding-factual-associati.md
Created 2026-08-25T02:58:15+00:00
On COUNTERFACT, ROME achieves ~99–100% efficacy and ~75–79% neighborhood specificity, while fine-tuning achieves 100% efficacy but only ~10–40% specificity.
Summary
ROME is a far more surgical way to change a specific fact in a model than fine-tuning: both hit the target almost every time, but fine-tuning also wrecks a lot of the model's other related knowledge in the process, while ROME leaves most of it intact. In practical terms, this means ROME can correct a single error without the collateral damage that fine-tuning typically causes, making it the better choice when you need targeted edits rather than broad retraining.
Dependents
These beliefs depend on this one:
- OUT parametric-contextual-knowledge-duality — LLMs maintain two functionally independent knowledge channels: parametric (editable via ROME's rank-one weight update) and contextual (supplied via retrieved passages that invert accuracy trends), which operate on separate mechanisms and can be independently manipulated.
- OUT single-fact-edit-efficacy — ROME's high efficacy and specificity on individual factual associations holds as a reliable operational claim, but is scoped to single-hop entity-relation edits.