nucleation

LIT-1536: Locating and Editing Factual Associations in GPT

1 min read

Properties6
StatusActive — read 2026-10-26 (NOTE-1166) in full, from the arXiv copy (v5, the NeurIPS 2022 version); worth reading for causal tracing, a corrupt-and-restore intervention that locates where a fact's recall depends on hidden states, for the finding that in GPT-2 XL that dependence concentrates at the last subject token in middle-layer MLPs, and for the rank-one edit that tests it (THEORY-998). Read it knowing that the storage hypothesis it draws is wider than what it shows (THEORY-1011), that LIT-968 proves a working rank-one edit does not locate a fact, and that several numbers and figure references in v5 do not match.
Version2
Tagsrepresentation-learning, causality, anthology-candidate
Cited as “extends” byLIT-1582 — Mass-Editing Memory in a Transformer
Cited as “paper” byNOTE-1166 — Meng, Bau, Andonian and Belinkov — Locating and editing factual associations in GPT
Cited as “source” byTHEORY-1011 — In autoregressive transformer language models, factual associations are stored in middle-layer MLP modules as linear key–value memories: the MLP takes a key encoding the subject at its last token and returns a value encoding its attributes, which late attention moves to the prediction, and any middle layer could store a given fact equally, as Meng, Bau, Andonian and Belinkov propose, THEORY-1039 — In autoregressive transformer language models, factual associations are recalled through middle-layer MLP modules acting as key–value memories, and any fact could be stored equally in any one of the middle MLP layers, as Meng, Bau, Andonian and Belinkov conjecture, THEORY-998 — In GPT-2 XL, for 1000 facts the model knows, restoring single clean hidden states into a run whose subject tokens were noised recovers the most of the correct object's probability at two sites, middle layers at the last subject token, where MLP outputs carry most of the effect, and late layers at the last token, where attention does; and a rank-one edit of one middle MLP output matrix keyed on the last subject token inserts counterfactual facts that generalize to paraphrases and spare neighbouring subjects better than the fine-tuning, knowledge-neuron and hypernetwork editors tested, working best at that token and in middle layers

Graph View

Meng, K., Bau, D., Andonian, A. and Belinkov, Y. (2022). Locating and editing factual associations in GPT. NeurIPS 2022. arXiv:2202.05262.

Read in full on 2026-10-26 (NOTE-1166). Two accounts were drawn from it: THEORY-998 and THEORY-1011.

Why this is here

Registered and read at the owner’s request of 2026-10-26, after an answer to “how’s the record on neural circuit theory?” found that the record held almost none of the core circuits literature. It is the founding paper of locating knowledge by intervention in language models, and the record already held its critique without it: Makelov et al. (LIT-968, THEORY-282) show the interpretability illusion on GPT2-XL fact patches and prove that any one-dimensional fact patch has a rank-one edit with the same output, so that a working ROME edit does not show where a fact is stored. With this paper held, that argument has its target in the record. It is also the MLP counterpart of the attention-head circuits of LIT-1534 and LIT-1540.

Backlinks