ROME and causal tracing for factual language-model edits

What changed

Uses corruption and restoration of internal activations to locate factual-recall pathways, then applies a constrained rank-one MLP update; CounterFact tests edited facts, paraphrases and nearby associations.

Scientific scope

Uses corruption and restoration of internal activations to locate factual-recall pathways, then applies a constrained rank-one MLP update; CounterFact tests edited facts, paraphrases and nearby associations. The localized-memory picture is a tested hypothesis for the studied GPT models, not a universal theorem. Rank-one associative-memory algebra inherits earlier work; single-edit success does not guarantee long-run consistency or zero collateral changes.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. The localized-memory picture is a tested hypothesis for the studied GPT models, not a universal theorem. Rank-one associative-memory algebra inherits earlier work; single-edit success does not guarantee long-run consistency or zero collateral changes.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Locating and Editing Factual Knowledge in GPT

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction