THEORY-1031: In GPT-J, inserting 10,000 CounterFact counterfactuals at once by a closed-form least-squares update of the MLP output matrices of layers 3 to 8, spread so that each layer writes a share of each fact's required change to the subject's hidden state, gives an editing score of 85.8 (efficacy 98.9, paraphrase 88.6, neighbourhood 73.7) with fluency kept, against 67.6 for weight-decayed fine-tuning, which breaks generation, 50.3 for ROME applied one fact at a time and 23.1 for MEND; GPT-NeoX scores 82.0, and editing late MLPs or attention instead does worse

Source

Meng, Sen Sharma, Andonian, Belinkov and Bau (2022), ICLR 2023, arXiv:2210.07229, LIT-1582, read in full in NOTE-1197.

What was actually shown

Method. For each fact, a change δᵢ to the hidden state at the last subject token at layer 8 (GPT-J, layers numbered from 0) is optimized to make the new object likely. Then, layer by layer from 3 to 8, each MLP output matrix gets the update Δ = R Kᵀ (C + K Kᵀ)⁻¹, where K holds the facts’ keys at that layer, R a 1/(L − l + 1) share of each remaining δᵢ, and C = λ·E[kkᵀ] a covariance of the layer’s inputs from Wikitext (λ = 15,000).

Results. CounterFact, 10,000 edits, GPT-J: MEMIT S 85.8 (ES 98.9, PS 88.6, NS 73.7, generation entropy 619.9 against 622.4 unedited, reference score 40.1); FT-W 67.6 (NS 46.9, generation entropy 293.9); ROME sequential 50.3 (reference score 3.3); MEND 23.1; unedited 22.4. GPT-NeoX (layers 6–10): 82.0 against 23.7 unedited. zsRE, 10,000 true facts, GPT-J: 50.7 against FT-W 42.1, MEND 20.0, ROME 2.6. Scaling: ROME holds to about 10 edits and degrades from 32; MEMIT is best at large n and slightly behind ROME on paraphrase at small n. Ablations: fewer layers, a late range of MLPs, or attention modules do worse; λ trades specificity against efficacy, best near 10⁴.

What could have come out otherwise. The batch update could have failed to generalize (high efficacy, low paraphrase), or damaged neighbours as fine-tuning does, or broken generation; spreading over layers could have done no better than one layer; a late range could have done as well as the traced one.

What this does not say

• Not where facts are stored. The layer range is chosen from causal traces averaged over facts. Per fact, tracing does not predict where an edit works (THEORY-1024), and an edit at a layer can produce the output whether or not the layer holds the fact (THEORY-282). The storage reading is THEORY-1011, Proposed.
• Not that the edits are integrated knowledge. The tests are the edit prompt, paraphrases, neighbouring subjects and generations about the subject. Reverse relations are not edited, as the authors say.
• Not that it is better at every scale. At a few edits ROME generalizes slightly better.
• Not beyond the baselines. SERAC was not compared, FT-W was tuned only at 10,000 edits, and MEND’s hypernetwork was trained for single edits.
• Not with uncertainty across runs for the main tables, which report one configuration with confidence intervals over cases.