anthology-of-the-sota

LIT-576: Mass-Editing Memory in a Transformer

2 min read

Properties7
StatusActive
Tagsadaptation-and-tuning, analysis-and-evaluation
ExtendsLIT-578 — Locating and Editing Factual Associations in GPT
Compared againstLIT-577 — Evaluating the Ripple Effects of Knowledge Editing in Language Models
Cited as “introduced 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 “paper” byNOTE-316 — MEMIT
Cited as “source” 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

Graph View

Meng, Sen Sharma, Andonian, Belinkov, Bau, MIT, Northeastern and Technion (2023) — ARXIV-2210.07229

Key takeaways

  • The method (§4): compute a target hidden state for each fact at the last critical layer. Distribute the needed change across the range of critical MLP layers, R = {3…8} for GPT-J, chosen where severing MLPs cuts the causal effect. Solve for each layer’s update with a least-squares objective that also preserves a covariance of existing keys
  • zsRE, 10,000 real facts on GPT-J (Table 1): efficacy 96.7, paraphrase 89.7, specificity 26.6 (unedited 27.0). ROME collapses to 2.6 overall. Plain fine-tuning (42.1) beats both ROME and MEND at this scale
  • CounterFact, 10,000 counterfactuals (Table 2): MEMIT 85.8 on GPT-J (ES 98.9, PS 88.6, NS 73.7, against 83.5 unedited) and 82.0 on GPT-NeoX-20B. ROME 50.3. Fine-tuning scores 67.6, with fluency collapsing (generation entropy 294 against 622)
  • Scaling curve (Figure 5): ROME is fine to about 10 edits and degrades from 32. MEND loses efficacy by 1,000. At small n, ROME generalizes slightly better than MEMIT
  • Cost: 7.4 hours for 10,000 edits on GPT-J, done serially in the released code

Standing in the anthology

Filed from #163 (LARQL). It extends ROME and sources SOTA-347, which RippleEdits (LIT-577) contests.

What it takes from ROME and what it changes. From ROME (LIT-578) it keeps causal tracing to choose where to edit and the view of an MLP as a linear associative memory whose update preserves existing keys. ROME writes one fact as a rank-one update to one layer, and degrades after about 10 edits. MEMIT solves for many facts at once in a batched least-squares update spread over a range of critical layers (3–8 in GPT-J). At 10,000 CounterFact edits that is 85.8 against ROME’s 50.3.

What “10,000 memories” means here. The test is recall of the edited triple, its paraphrases and unrelated neighbors. Whether the edited facts propagate to their implications is not tested, and when it later was, MEMIT did not (LIT-577).

Read — NOTE-316.

Backlinks