anthology-of-the-sota

LIT-577: Evaluating the Ripple Effects of Knowledge Editing in Language Models

2 min read

Properties7
StatusActive
Tagsanalysis-and-evaluation, adaptation-and-tuning, in-context-learning
Compared againstLIT-576 — Mass-Editing Memory in a Transformer
Cited as “contested by” bySOTA-347 — To write many facts into a model's weights, update a range of early-to-mid MLP layers jointly in one batched solve (MEMIT), choosing the range by measured edit success rather than causal tracing
Cited as “introduced by” bySOTA-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
Cited as “paper” byNOTE-314 — RippleEdits
Cited as “source” bySOTA-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

Graph View

Cohen, Biran, Yoran, Globerson, Geva, Tel Aviv University and Google (2023) — ARXIV-2307.12976

Key takeaways

  • The criteria (§3): Logical Generalization (symmetric or implied relations), Compositionality I and II (multi-hop through the edited fact), Subject Aliasing, Preservation (other objects of a one-to-many relation survive), and Relation Specificity (the subject’s other facts don’t change)
  • The benchmark: about 5K edits in three subsets. RECENT holds facts newer than the models, RANDOM holds random Wikidata facts, and POPULAR holds popular subjects. Each comes with generated test queries, filtered to those the model answered correctly before editing
  • Results (Tables 3–5): parameter editors average 38–66. On POPULAR, logical generalization is 5.5–7.0 for ROME and MEMIT on GPT-2 and GPT-J. Subject aliasing is high everywhere (86–100), so edits reach the name and stop there
  • In-context editing (ICE): prefix “Imagine that ” to the query. It is best overall. LLaMA-7B with ICE averages 81–83 against 49–61 for ROME, and GPT-3 with ICE averages 78–90

Standing in the anthology

Filed from #163 (LARQL). It is compared_against MEMIT (LIT-576), contests SOTA-347, and sources SOTA-348.

Relevance to the tool that prompted it. LARQL’s INSERT defaults to a KNN retrieval override (MODE KNN), not a weight edit, although the write-up introducing it calls it a direct weight-level edit. The default is closer to the retrieval approach this paper’s results favor than to the editors it evaluates.

Read — NOTE-314.

Backlinks