{"paper":{"title":"Locating and Editing Factual Associations in GPT","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Factual associations in GPT models are stored in localized mid-layer feed-forward computations that can be directly edited via rank-one weight updates.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Alex Andonian, David Bau, Kevin Meng, Yonatan Belinkov","submitted_at":"2022-02-10T18:59:54Z","abstract_excerpt":"We analyze the storage and recall of factual associations in autoregressive transformer language models, finding evidence that these associations correspond to localized, directly-editable computations. We first develop a causal intervention for identifying neuron activations that are decisive in a model's factual predictions. This reveals a distinct set of steps in middle-layer feed-forward modules that mediate factual predictions while processing subject tokens. To test our hypothesis that these computations correspond to factual association recall, we modify feed-forward weights to update s"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We find that ROME is effective on a standard zero-shot relation extraction (zsRE) model-editing task, comparable to existing methods. To perform a more sensitive evaluation, we also evaluate ROME on a new dataset of counterfactual assertions, on which it simultaneously maintains both specificity and generalization, whereas other methods sacrifice one or another.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that the causal intervention correctly isolates the decisive feed-forward computations for factual recall, and that a rank-one update to those weights changes the association without creating unmeasured side effects on the broader distribution of model behavior.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Factual associations in autoregressive transformers are localized to mid-layer feed-forward modules and can be edited via rank-one model editing while preserving both specificity and generalization on counterfactual tests.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Factual associations in GPT models are stored in localized mid-layer feed-forward computations that can be directly edited via rank-one weight updates.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b2abb54c7d4c21749bfc35a611cee2fafeabc9c1546669f3b09b475630eed4ea"},"source":{"id":"2202.05262","kind":"arxiv","version":5},"verdict":{"id":"596fd162-bcc3-417a-9bd0-541a17654282","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T09:55:25.774121Z","strongest_claim":"We find that ROME is effective on a standard zero-shot relation extraction (zsRE) model-editing task, comparable to existing methods. To perform a more sensitive evaluation, we also evaluate ROME on a new dataset of counterfactual assertions, on which it simultaneously maintains both specificity and generalization, whereas other methods sacrifice one or another.","one_line_summary":"Factual associations in autoregressive transformers are localized to mid-layer feed-forward modules and can be edited via rank-one model editing while preserving both specificity and generalization on counterfactual tests.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that the causal intervention correctly isolates the decisive feed-forward computations for factual recall, and that a rank-one update to those weights changes the association without creating unmeasured side effects on the broader distribution of model behavior.","pith_extraction_headline":"Factual associations in GPT models are stored in localized mid-layer feed-forward computations that can be directly edited via rank-one weight updates."},"references":{"count":43,"sample":[{"doi":"","year":2017,"title":"Fine-grained analysis of sentence embeddings using auxiliary prediction tasks","work_id":"b7bd2b01-1ceb-409d-9ec3-e841491e27b1","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1972,"title":"Anderson, J. A. A simple neural network generating an interactive memory. Mathematical biosciences, 14 0 (3-4): 0 197--220, 1972","work_id":"e6955ef9-c3c3-47d3-8a55-a361da59cac2","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Rewriting a deep generative model","work_id":"30b22db5-f3a2-4cd6-83dc-bb134d36c8b7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1162/coli_a_00422","year":2021,"title":"Probing Classifiers: Promises, Shortcomings, and Advances","work_id":"3eabab74-ac71-4292-86ce-b0469cd4e6cf","ref_index":4,"cited_arxiv_id":"2102.12452","is_internal_anchor":true},{"doi":"10.1162/tacl_a_00254","year":2019,"title":"Analysis methods in neural language processing: A survey","work_id":"62f1a2a2-d87f-4f14-90cf-b4b293a4780c","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":43,"snapshot_sha256":"73260a8da562501fd2f1f3f7a4d7f9b70137a4a1f7ce0d835bd03497c761cf52","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"32953a508cd941293af0a1776cfdcfae878e118b70e9d4b2121fff6aef4acbd5"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}