Benchmarking Knowledge Editing using Logical Rules

In three linesNew benchmark to evaluate knowledge editing techniques in LLMs. Existing methods (ROME, FT) successfully insert direct facts but fail on logical consequences: up to 24% performance gap between directly edited and entailed knowledge.

Summary generated by Claude — human-verified