rome-counterfact-sparql-neighborhood-prompts
IN premise — summaries/2026/08/24/meng-2022-rome-sR-references-chunk-2.md
Created 2026-08-25T02:58:17+00:00
COUNTERFACT's neighborhood prompts are derived via SPARQL queries for entities sharing the same predicate, chosen for latent-space proximity to test bleedover with linear editing methods.
Summary
COUNTERFACT builds its test prompts by querying a knowledge graph for entities connected by the same kind of relationship, then picking the ones that sit closest together in the model's internal representation. This is a deliberate stress test: if editing one fact in the model accidentally corrupts a nearby, similar fact, the method fails the "bleedover" check.