anthology-of-the-sota

SOTA-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

2 min read

Properties6
StatusProposed
Consensuscontested
Tagsadaptation-and-tuning, analysis-and-evaluation
SourceLIT-576 — Mass-Editing Memory in a Transformer, LIT-574 — Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models
Introduced byLIT-576 — Mass-Editing Memory in a Transformer
Contested byLIT-577 — Evaluating the Ripple Effects of Knowledge Editing in Language Models

Graph View

Source

Meng et al. (2023), LIT-576, read as NOTE-316. Hase et al. (2023), LIT-574, for how to choose layers.

The practice

If you must edit facts into weights, for example without control over the prompt at inference:

  • Batch the edits and spread them over several layers. ROME-style single-layer rank-one edits interfere past about 10–30 facts. MEMIT’s joint solve held up to 10,000 on GPT-J and GPT-NeoX-20B
  • Choose the layer range by editing results. Sweep edit success, paraphrase and neighborhood scores over candidate ranges on held-out edits. Do not use a per-fact tracing peak
  • Measure neighborhood damage at the edit count you use. It grows with the number of edits, and at 10,000 it cost about 10 points of neighborhood success (83.5 to 73.7)
  • Test implications, not just recall (SOTA-348)

Conditions

  • Contested as a way to update knowledge. On RippleEdits, weight editors average 38–66, and in-context editing beats them
  • Directional facts, on GPT-J and GPT-NeoX-class models with ungated MLPs

The dispute is LIT-577 (RippleEdits), and it is not about whether MEMIT writes the edited triple — it does — but about whether that amounts to updating knowledge. Across 5K edits, MEMIT and the other weight editors leave the edit’s implications (logical consequences, multi-hop compositions) mostly unchanged, with logical generalization as low as 5.5–7.0 on popular subjects, and simply stating the new fact in the prompt scores better overall.

Backlinks