anthology-of-the-sota

SOTA-348: Evaluate a knowledge edit on the facts it implies — logical consequences, multi-hop compositions, aliases and the subject's other facts — not only on the edited prompt, its paraphrases and unrelated neighbours

2 min read

Properties5
StatusActive
Consensusunreplicated
Tagsanalysis-and-evaluation, adaptation-and-tuning
SourceLIT-577 — Evaluating the Ripple Effects of Knowledge Editing in Language Models
Introduced byLIT-577 — Evaluating the Ripple Effects of Knowledge Editing in Language Models

Graph View

Source

Cohen et al. (2023), LIT-577. Read as NOTE-314.

The practice

Cohen et al. introduced this evaluation in LIT-577, with RippleEdits, a benchmark of six criteria for what an edit should change besides the queried triple. Measured there, ROME, MEMIT and MEND average 38–66 across subsets and models. Logical generalization falls to 5.5–7.0 on the POPULAR subset for GPT-2 and GPT-J while subject aliasing scores 86–100, and in-context editing averages 81–83 on LLaMA-7B against ROME’s 49–61. The first bullet answers the gap between aliases and consequences, and the third is the in-context result.

  • Test the consequences. For an edit (s, r, o → o′), query relations implied by it (inverse and symmetric relations), two-hop questions that pass through o′, the subject under its aliases, and s’s other facts, which should be unchanged
  • Filter to what the model knew before, so a failure is the edit’s and not ignorance
  • Include an in-context baseline. Put the new fact in the prompt and score it on the same queries. On RippleEdits it beat every weight editor

Conditions

  • Template-generated queries from Wikidata, scored by alias match
  • The in-context baseline needs control of the prompt. It is a benchmark reference, and the right production choice only where that control exists

Backlinks