safety
The Corrected Graph Quietly Pushed Other Answers Down
Direct promotion reached the top ten every time, but avoided collateral rank damage in only about 23% of edits.
Summary
Direct promotion reached the top ten every time, but avoided collateral rank damage in only about 23% of edits.
Editing a knowledge-graph embedding to promote one desired answer can displace other correct answers even when ordinary locality tests look clean. The authors audit three scopes: facts sharing the edited parameter, alternative correct answers to the target query and correct answers across the same relation. On FB15k-237 with DistMult and ComplEx, direct promotion always moved the target into the top ten but caused no measured damage in only 23.0–23.2% of edits. Strict preservation avoided damage but succeeded in only 1.3–1.4%. Support regularization delivered the highest joint success, while rank-truncated preservation sharply reduced the number of displaced answers.
Why it matters
Direct promotion reached the top ten every time, but avoided collateral rank damage in only about 23% of edits.
Limits and context
- On FB15k-237 with DistMult and ComplEx, direct promotion always moved the target into the top ten but caused no measured damage in only 23.0–23.2% of edits.
- Strict preservation avoided damage but succeeded in only 1.3–1.4%.
Key claims
Direct promotion reached the top ten every time, but avoided collateral rank damage in only about 23% of edits.
Qualification: On FB15k-237 with DistMult and ComplEx, direct promotion always moved the target into the top ten but caused no measured damage in only 23.0–23.2% of edits.
Evidence: source-2026-09-14-005
Sources
- arXiv preprint 2609.12116arXiv · primary research
Corrections
No corrections have been recorded for this story.