Knowledge Editing
121 beliefs (121 IN, 0 OUT)
Knowledge editing addresses the problem of modifying specific factual associations stored in large language models without retraining or degrading unrelated knowledge. This matters because LLMs encode vast amounts of world knowledge parametrically, and errors or outdated facts persist in their outputs unless targeted interventions are applied. The belief set under this topic spans the core editing methods (ROME, MEMIT, MEND, ICE, MeLLo, fine-tuning), the benchmarks designed to evaluate them (RIPPLE EDITS, MQuAKE, COUNTERFACT), the mathematical machinery behind parametric edits, and the fundamental limitations that constrain what can be edited and how reliably knowledge propagates beyond the edited fact.
The central finding across the beliefs is a tension between editing efficacy at the target fact and the model's ability to preserve or propagate logically related knowledge. ROME performs a rank-one weight update to a single mid-layer MLP (rome-rank-one-update-formula, rome-rank-one-weight-update-layer-9), achieving roughly 99 to 100 percent efficacy on the edited fact while maintaining 75 to 79 percent neighborhood specificity (rome-quantitative-benchmark-results). MEMIT extends this to multi-layer batch edits across a range of feedforward layers (memit-multi-layer-batch, memit-targets-mlp-layers-not-attention) and scales to 10,000 simultaneous edits (memit-10k-edit-scores-cOUNTERFACT). Yet both methods show dramatic multi-hop degradation: ROME-edited GPT-J drops from 40.5 percent to 7.4 percent on MQuAKE-CF (rome-gptj-mquake-cf-multi-hop-drop), MEMIT-edited Vicuna-7B drops from 30.2 percent to 4.9 percent (vicuna7b-memit-multi-hop-drop), and fine-tuning is even more catastrophic at just 1.4 percent versus 39.5 percent for the unedited model (ft-catastrophic-mquake-cf-multipath). Paradoxically, a simple prompt-based in-context editing method (ICE) outperforms all dedicated parametric methods on the RIPPLE EDITS benchmark (c2023-in-context-editing-best-ripple-edits, in-context-editing-outperforms-ke), and ICE outperforms ROME by more than 29 points on LLaMA (ripple-edits-ice-outperforms-rome). This suggests the parametric editing machinery introduces more collateral damage than it repairs, consistent with the finding that in at least 68 percent of failure cases the model produces noise or abstention rather than simply retaining the original answer (c2023-ripple-edits-erroneous-change-rate, ripple-edits-error-taxonomy-noise-dominant).
All beliefs in this topic are IN (affirmed); none have been retracted. A key structural distinction separates premises from derived conclusions: the mathematical definitions of ROME's closed-form solution (rome-lambda-closed-form-solution), MEMIT's core update equation (memit-core-update-equation), and the benchmark construction rules (mquake-chain-construction-rule, mquake-cf-conflict-free-subject-relation) serve as axiomatic foundations, while the multi-hop failure results and the 68-percent-noise finding are empirical conclusions derived from applying those methods to those benchmarks. The belief set also draws an important boundary: the linear-subspace assumption underlying ROME and MEMIT limits them to directional entity-relation-object factual triples (memit-limited-directional-relations, rome-scope-limitation), and this same assumption is identified as the shared root cause of why neither can encode non-linear constraints like part-whole relations (linearity-boundary-of-geometric-framework). Additionally, the benchmarks themselves have different scopes: RIPPLE EDITS covers only the 1-to-2-hop close neighbourhood of an edit (c2023-ripple-edits-scope-1-to-2-hop, ripple-edits-2hop-scope), while MQuAKE evaluates multi-hop chains of 2 to 4 triples (mquake-dataset-3000-instances-2-4-hop, c2023-mquake-multi-hop-ke), and white-box gradient-based methods are distinguished from black-box prompting approaches like MeLLo (mquake-whitebox-blackbox-distinction).
The broader implications captured in these beliefs point to a structural problem. Evidence suggests LLMs maintain at least two distinguishable knowledge pathways—a parametric channel amenable to targeted weight editing and a contextual channel in which retrieved passages alter accuracy patterns in ways that invert closed-book trends (parametric-contextual-knowledge-duality-v2)—but the available evidence does not establish full functional independence between them. Knowledge editing side effects are now recognized as one of three causes of intra-memory conflict, alongside training corpus bias and decoding strategy effects (im-three-causes-pipeline-stages), and editing methods like ROME and MEMIT have been shown to create new intra-memory conflicts and increase hallucination as a side effect (ke-creates-im-and-increases-hallucination). The practical consequence is that current parametric editing methods, while effective at the target fact, systematically degrade the model's ability to reason over interconnected knowledge, raising serious questions about whether weight-level modification is the right mechanism for maintaining coherent world knowledge in large language models.
-
IN
c2023-eval-protocol-100-lg-queries
The RIPPLE EDITS evaluation protocol collects model responses to 100 random LG (long-geometry/logical generalization) queries and categorizes each into correct, abstain, or noise -
IN
c2023-ice-prompt-based-correction
ICE (In-Context Editing) uses prompt-based correction rather than parameter-level modification to perform knowledge edits -
IN
c2023-in-context-editing-best-ripple-edits
A simple in-context editing method achieves the best overall results on the RIPPLE EDITS benchmark, outperforming parametric methods ROME and ICE -
IN
c2023-mquake-multi-hop-ke
MQuAKE is a benchmark that evaluates multi-hop knowledge editing, while RIPPLE EDITS covers only single-hop ripple effects -
IN
c2023-ripple-edits-erroneous-change-rate
In ≥ 68% of cases across all settings, factual editing via KE methods introduces erroneous changes (noise or abstention) rather than simply making no change -
IN
c2023-ripple-edits-scope-1-to-2-hop
RIPPLE EDITS benchmark covers only the close neighbourhood of an edit (1–2 hops) and does not test paraphrase robustness, subject specificity, or distantly-related fact retention -
IN
c2023-ripple-edits-wikidata-dependence
The RIPPLE EDITS benchmark relies on Wikidata as the external knowledge base for generating ripple-edit triplets, which may be incomplete or outdated -
IN
c2023-rome-rank-one-weight-edit
ROME performs knowledge editing via rank-one matrix (weight) updates to the model -
IN
c2023-rome-vs-ice-failure-modes
ROME tends to produce more incorrect/noisy changes after editing, while ICE tends to cause the model to generate abstention responses (e.g., 'unknown', 'a mystery') -
IN
ft-catastrophic-mquake-cf-multipath
Fine-tuning achieves only 1.4% multi-hop accuracy on MQuAKE-CF versus 39.5% for the unedited base model, while ROME achieves 37.3% on 2-hop but drops to 10.0% on 3-hop and 7.7% on 4-hop. -
IN
im-three-causes-pipeline-stages
Intra-memory conflict has three causes mapped to pipeline stages: training corpus bias (pre-training), decoding strategy effects (inference), and knowledge editing side effects (post-update). -
IN
in-context-editing-outperforms-ke
A simple in-context editing baseline obtains the best scores on the RIPPLE EDITS benchmark, outperforming all dedicated parameter-level knowledge editing methods -
IN
ke-creates-im-and-increases-hallucination
Knowledge editing (e.g., ROME, MEMIT) can create intra-memory conflicts and increase hallucination as a side effect (Li et al. 2023f; Pinter & Elhadad, 2023). -
IN
ke-entity-relation-object-triplets
Knowledge editing methods operate on entity-relation-object triplets (e, r, o), such as 'Messi → plays for → Inter Miami' -
IN
linearity-boundary-of-geometric-framework
The geometric framework's explanatory boundary is precisely the boundary of linearity: ROME cannot edit non-factual (logical/spatial/numerical) associations because they are not rank-one key-value pairs, and Park's framework cannot encode the 'thing⇒part' relation because it is a non-linear constraint in the unembedding space—both failures share the root cause of the linear-subspace assumption. -
IN
mello-cot-prompt-3-demonstrations
The CoT multi-hop prompt template for MQuAKE evaluation uses 3 in-context demonstrations of 'Question → Thoughts (intermediate reasoning) → Answer' to elicit chain-of-thought decomposition. -
IN
mello-outperforms-rome-multihop-30.9-vs-19.9
MeLLo achieves 30.9% multi-hop accuracy versus ROME's 19.9% on 3,000 MQuAKE instances using GPT-3. -
IN
mello-scale-k-values-cf-t
MeLLo scale evaluation uses k ∈ {1, 100, 1000, 3000} simultaneous edits on MQuAKE-CF and k ∈ {1, 100, 500, 1868} on MQuAKE-T. -
IN
memit-10k-edit-scores-cOUNTERFACT
At 10,000 edits on GPT-J (COUNTERFACT), MEMIT achieves a harmonic-mean Score of ≈85.8, compared to ROME ≈50.3, MEND ≈23.1, and FT-W ≈67.6 (with generation failure). -
IN
memit-causal-tracing-3sigma-noise
Causal tracing for MEMIT layer selection injects Gaussian noise at 3σ (empirical embedding variance) into subject-token hidden states, then restores individual layer outputs to measure causal contribution. -
IN
memit-code-memit-baulab-info
MEMIT code and data are available at memit.baulab.info -
IN
memit-core-update-equation
MEMIT's weight update is computed as Δ = R·K₁ᵀ · (C₀ + K₁K₁ᵀ)⁻¹, where C₀ = λ·E_k[kkᵀ] is an uncentered covariance of pre-existing keys. -
IN
memit-covariance-lambda-gptj
The covariance adjustment factor λ is set to 15,000 for GPT-J and 20,000 for GPT-NeoX-20B, with an optimal range near 10⁴. -
IN
memit-editing-score-harmonic-mean
MEMIT's editing score is defined as the harmonic mean of efficacy, generalization, and specificity (fluency tracked separately) -
IN
memit-ftw-hyperparameters-gptj
FT-W baseline on GPT-J targets layer 21 with weight decay 5×10⁻⁴, max 25 steps, learning rate 5×10⁻⁴, early-stopping at loss ≤ 10⁻², completing 10,000 edits in ~0.48 hours. -
IN
memit-generation-entropy-formula
Generation Entropy (GE) is computed as −(⅔ Σ f₂(k) log₂ f₂(k) + ⁴⁄₃ Σ f₃(k) log₂ f₃(k)), using weighted 2-gram and 3-gram frequency distributions to detect repetitive generation. -
IN
memit-gptj-critical-layer-range
For GPT-J (6B), MEMIT's critical MLP layer range R is {3, 4, 5, 6, 7, 8}, and for GPT-NeoX-20B it is {6, 7, 8, 9, 10}. -
IN
memit-gptj-mquake-cf-multi-hop-drop
MEMIT-edited GPT-J drops from 40.5% to 7.0% multi-hop accuracy on MQuAKE-CF -
IN
memit-hardware-gpu-requirements
GPT-J (6B) runs MEMIT on a single NVIDIA A6000 (48 GB) GPU, while GPT-NeoX-20B requires at least two GPUs (one 48 GB for the model in fp16, one smaller for editing computation). -
IN
memit-iclr-2023-gptj-gptneox
MEMIT was published at ICLR 2023 and demonstrated on GPT-J (6B) and GPT-NeoX (20B) -
IN
memit-lambda-tradeoff-direction
Increasing the covariance adjustment factor λ monotonically increases specificity and fluency while decreasing efficacy and generalization in MEMIT. -
IN
memit-limited-directional-relations
MEMIT is limited to directional (subject, relation, object) factual relations and does not handle symmetric relations, spatial reasoning, temporal knowledge, or procedural knowledge. -
IN
memit-mend-vs-memit-runtime-10k
For 10,000 edits on GPT-J, MEND takes 98.25 seconds, ROME takes ~12.29 hours (~4 s/edit sequential), and MEMIT's z_i computation takes ~6.54 hours (serial) with parallel insertion taking ~0.90 hours. -
IN
memit-multi-layer-batch
MEMIT extends ROME to batch edits by updating feedforward networks across a range of layers to encode many facts simultaneously -
IN
memit-multi-layer-update-layers-5-9
MEMIT performs multi-layer weight updates at layers 5 through 9. -
IN
memit-non-conflicting-edit-constraint
MEMIT's edit set E requires that no two edits share the same (subject, relation) pair with different objects, formally: ¬∃ i,j : (s_i = s_j) ∧ (r_i = r_j) ∧ (o_i ≠ o_j) -
IN
memit-norm-clamp-fractions
MEMIT clamps the L2 norm of δ_i to ¾ ‖hⁱ_L‖ for GPT-J and 3/10 ‖hⁱ_L‖ for GPT-NeoX-20B to prevent destabilizing updates. -
IN
memit-rome-sequential-degradation
ROME processes edits sequentially (~4 sec/edit) and degrades beyond approximately 32 edits, while MEMIT scales to 10,000+ edits via parallel z_i computation and a shared matrix inverse. -
IN
memit-targets-mlp-layers-not-attention
MEMIT targets MLP (feed-forward) layers specifically as key-value memory stores, not attention layers, for factual knowledge editing -
IN
memit-vicuna-layers-5-to-9
MEMIT on Vicuna-7B edits layers {5, 6, 7, 8, 9} with covariance computed from 100K Wikitext samples; ROME on Vicuna-7B edits only layer 9. -
IN
mend-cot-mquake-t-anomaly
MEND achieves 38.2% multi-hop accuracy with CoT prompting on MQuAKE-T, substantially higher than ~4-12% for ROME and MEMIT, attributed to relation-specific editing effectiveness -
IN
mquad-t-extended-golden-labels-text-davinci-003
MQuAKE-T uses extended golden labels generated via GPT-3.5 (text-davinci-003) that broaden the acceptable answer set to include outdated but historically valid answers for smaller models. -
IN
mquake-38-relation-types-templates
MQuAKE covers 38 Wikidata relation types (e.g., P36, P27, P136) with distinct question templates for querying single-hop facts and cloze templates for converting edits into natural-language statements for memory storage. -
IN
mquake-7-heuristic-constraints
Wikidata fact-chain sampling in MQuAKE uses 7 heuristic constraints including no circular chains, no repeated relations, country objects only in last two hops, ≤3 object types, and specific subject requirements for headquarters/capital relations. -
IN
mquake-benchmark-two-variants
MQuAKE benchmark consists of two subsets: MQuAKE-CF (counterfactual edits) and MQuAKE-T (temporal/real knowledge changes), evaluating multi-hop questions whose answers are entailed consequences of edits. -
IN
mquake-cf-3000-instance-no-conflict
The MQuAKE-CF subset of 3000 instances enforces that after all edits, for any subject-relation pair (s, r), at most one object o exists, preventing knowledge conflicts across instances. -
IN
mquake-cf-conflict-free-subject-relation
The MQuAKE-CF subset is conflict-free: for any (subject, relation) pair, at most one object exists across the 3,000 instances. -
IN
mquake-cf-instance-count
MQuAKE-CF contains 9,218 instances constructed from Wikidata fact chains with counterfactual edits -
IN
mquake-chain-construction-rule
In MQuAKE fact chains C = ⟨(s₁,r₁,o₁), …, (sₙ,rₙ,oₙ)⟩, the object of each triple equals the subject of the next (oᵢ = sᵢ₊₁), and the question asks about s₁ with answer oₙ -
IN
mquake-counterfactual-validity-condition
A counterfactual edit in MQuAKE-CF is valid only if the updated chain yields a different answer a* ≠ a from the original answer -
IN
mquake-dataset-3000-instances-2-4-hop
MQuAKE is a benchmark of 3,000 multi-hop questions (2, 3, and 4 hops) constructed from Wikidata fact chains, with variants MQuAKE-CF (counterfactual edits) and MQuAKE-T (real temporal knowledge changes). -
IN
mquake-eval-or-vs-and-logic
MQuAKE multi-hop accuracy uses OR logic (any generated question answered correctly counts as success) while instance-wise accuracy uses AND logic (all associated facts must be recalled). -
IN
mquake-evaluation-edit-sizes
MQuAKE grouped-edit evaluation uses k ∈ {1, 100, 1000, 3000} for MQuAKE-CF and k ∈ {1, 100, 500, 1868} for MQuAKE-T -
IN
mquake-github-repo
MQuAKE code and data are publicly available at https://github.com/princeton-nlp/MQuAKE -
IN
mquake-multi-hop-accuracy-metric
In MQuAKE, multi-hop accuracy counts an instance as correct if any one of the three generated questions is answered correctly -
IN
mquake-question-generation-model
ChatGPT (gpt-3.5-turbo-instruct) generates 3 diverse multi-hop questions per fact chain in MQuAKE -
IN
mquake-questions-generated-by-chatgpt
MQuAKE multi-hop questions are automatically generated by ChatGPT (gpt-3.5-turbo) from Wikidata fact chains of 2-4 triples using a 12-demoneration prompt, not human-authored. -
IN
mquake-recallability-filter-8shot
The MQuAKE construction pipeline discards any chain containing a fact that GPT-J cannot recall via 8-shot in-context learning -
IN
mquake-t-excludes-vicuna7b
MQuAKE-T excludes Vicuna-7B from evaluation due to suspected training data contamination with the temporal knowledge in the benchmark. -
IN
mquake-t-instance-count
MQuAKE-T contains 1,868 instances sourced from Wikidata diffs between 2021-04 and 2023-04, restricted to 6 relations -
IN
mquake-tf-dataset-split-purpose
MQuAKE-CF tests propagation of artificial counterfactual edits (up to 4 per instance) while MQuAKE-T tests real temporal updates (exactly 1 per instance) -
IN
mquake-whitebox-blackbox-distinction
Parameter-editing methods (FT, MEND, ROME, MEMIT) require white-box access to model weights, while MeLLo is designed for black-box LLMs requiring no gradient updates -
IN
mquake-wikidata-construction-params
MQuAKE chains are sampled from a Wikidata subgraph restricted to 37 manually selected relations and top 20% common entities by Wikipedia hyperlink count, with chains of 2-4 triples -
IN
parametric-contextual-knowledge-duality-v2
Evidence suggests LLMs exhibit at least two distinguishable knowledge pathways: a parametric channel amenable to targeted editing via ROME's rank-one weight update (achieving ~99–100% efficacy with ~75–79% neighborhood specificity, in contrast to fine-tuning's 100% efficacy but only ~10–40% specificity), and a contextual channel in which retrieved passages alter the accuracy–rarity relationship (accuracy increases with relevant document count, inverting the closed-book trend). These pathways display distinct behavioral profiles and appear to be amenable to different manipulation strategies, though the available evidence does not establish full functional independence or confirm that they rely on entirely separate internal mechanisms. -
IN
ripple-edits-2hop-scope
RIPPLE EDITS restricts ripple-effect evaluation to facts within a 2-hop distance from the subject or object of the edited triplet. -
IN
ripple-edits-benchmark-size
The RIPPLE EDITS benchmark contains 5,000 factual edits designed to capture various types of ripple effects including team changes, alias propagation, and related entity updates -
IN
ripple-edits-code-repo
The RIPPLE EDITS benchmark code is available at https://github.com/edenbiran/RippleEdits. -
IN
ripple-edits-data-quality-validation
In RIPPLE EDITS data validation, 100% of sampled test queries were semantically sound and 98.5% were grammatically correct, with failures attributed to non-English entity names. -
IN
ripple-edits-error-taxonomy-noise-dominant
In at least 68% of RIPPLE EDITS failures across all model-method settings, the model produces an incorrect or unrelated object (noise) rather than simply retaining the original answer. -
IN
ripple-edits-ice-outperforms-rome
ICE outperforms ROME by more than 10 points on GPT-NeoX and more than 29 points on LLaMA in average accuracy across RECENT, RANDOM, and POPULAR subsets. -
IN
ripple-edits-ice-prompt-format
The In-Context Editing (ICE) baseline prepends the prompt 'Imagine thatwould have been ' to the query, where Pr is a manually written natural-language proposition of the relation r. -
IN
ripple-edits-ke-methods-fail
Prominent parameter-editing methods (ROME, MEMIT, Hernandez et al., Si et al.) fail to introduce consistent changes across logically related facts in the ripple graph -
IN
ripple-edits-models-evaluated
RIPPLE EDITS evaluates four open-source LMs (GPT-2 XL 1.5B, GPT-J 6B, LLaMA 7B, GPT-NeoX 20B) and GPT-3 (text-davinci-003, 175B) across ROME, MEMIT, MEND, and ICE methods. -
IN
ripple-edits-nrel-54-relations
The RIPPLE EDITS benchmark uses Nrel = 54 basic relations for which manually crafted natural-language templates and logical rules are defined. -
IN
ripple-edits-popular-pageviews
POPULAR-subject edits in RIPPLE EDITS have ~17,377 average Wikipedia page views versus ~190 for RECENT-subject edits, a >90× difference. -
IN
ripple-edits-recent-sampling-window
RECENT triplets in RIPPLE EDITS are sampled from Wikidata modifications within 250 days after July 2022. -
IN
ripple-edits-six-criteria
RIPPLE EDITS evaluates six ripple-effect criteria: Logical Generalization (LG), Compositionality I (CI), Compositionality II (CII), Subject Aliasing (SA), Preservation (PV), and Relation Specificity (RS). -
IN
ripple-edits-success-criterion
In RIPPLE EDITS evaluation, a generation is scored as correct if one of the aliases of the target object appears within the first 20 generated tokens. -
IN
ripple-edits-three-subsets-sizes
RIPPLE EDITS is divided into RECENT (2,000 edits, injection), RANDOM (1,000 edits, modification), and POPULAR (1,000 edits, modification) subsets. -
IN
rome-c-gram-matrix-definition
In ROME's derivation, C = KKᵀ is the Gram matrix of the key set, and its inverse C⁻¹ projects the new key into the subspace orthogonal to existing keys to minimize disruption. -
IN
rome-causal-tracing-3sigma-corruption
Causal Tracing requires corrupting subject token embeddings with at least 3σ_t Gaussian noise; noise at σ_t produces effects too small to discern the indirect causal effect. -
IN
rome-causal-tracing-three-run
Causal Tracing uses a three-run protocol (clean, corrupted with subject noise, corrupted-with-restoration) where the Indirect Effect is computed as the difference between runs 2 and 3. -
IN
rome-code-hosted-baulab-info
ROME's code, interactive notebooks, dataset, benchmarks, and visualizations are hosted at https://rome.baulab.info. -
IN
rome-code-repository
ROME's code, dataset, visualizations, and interactive demo notebook are available at https://rome.baulab.info/. -
IN
rome-corruption-magnitude
In the Causal Tracing corruption protocol, subject embeddings are perturbed by ε ~ N(0; ν) where ν = 3× the empirical standard deviation of embeddings. -
IN
rome-corruption-noise-protocol
Causal Tracing corruption adds Gaussian noise ε ~ N(0; 3σ_t) to subject token embeddings (ν = 3 times the observed standard deviation), reducing correct-object probability from ~27% to ~8.47%. -
IN
rome-counterfact-built-on-pararel
COUNTERFACT is a benchmark built on top of the ParaRel dataset (Elazar et al., 2021a) for fine-grained, multi-dimensional measurement of knowledge extraction and editing. -
IN
rome-counterfact-dataset-size
The COUNTERFACT evaluation dataset contains 21,919 counterfactual records including subjects, relations, paraphrases, neighborhood prompts, and generation prompts. -
IN
rome-counterfact-sparql-neighborhood-prompts
COUNTERFACT's neighborhood prompts are derived via SPARQL queries for entities sharing the same predicate, chosen for latent-space proximity to test bleedover with linear editing methods. -
IN
rome-covariance-from-wikipedia
In ROME, the key covariance matrix C = KKᵀ is pre-computed from a Wikipedia text sample and is not recomputed per edit. -
IN
rome-decisive-token-position
The decisive processing step for factual recall in GPT occurs specifically at the last token of the subject name in the prompt, not the first token or an average across tokens. -
IN
rome-directional-edits
ROME edits are directional: 'A is B' and 'B is A' are stored as separate associations and require two independent edits to alter both. -
IN
rome-edit-complexity-o-d2
ROME's rank-one edit is O(D²) per edit (a single outer product Λ(C⁻¹k*)ᵀ) rather than O(params), avoiding full retraining. -
IN
rome-edit-time-2s-a6000
ROME's total edit time is approximately 2 seconds on an NVIDIA A6000 GPU for GPT-2 XL, compared to ~100ms inference for hypernetwork methods that require hours-to-days of training. -
IN
rome-entropy-metrics-miss-fluency-loss
Human raters detected subtle fluency losses in ROME's generated output that an entropy-based automatic metric failed to capture, indicating a gap between statistical and perceptual quality assessment. -
IN
rome-geva-2021-mlp-key-value-memory
Geva et al. (2021) identified MLP layers in masked-LM transformers as key-value memories storing entities and their associated information. -
IN
rome-gpt2xl-clean-27pct-corrupted-8.47pct
In GPT-2 XL Causal Tracing, clean p(o_c) is 27.0%, drops to 8.47% after 3σ_t corruption, and best single hidden-state restoration recovers to 19.5% (layer 15, last subject token). -
IN
rome-gptj-mquake-cf-multi-hop-drop
ROME-edited GPT-J answers only 7.4% of MQuAKE-CF multi-hop questions, down from 40.5% before editing -
IN
rome-human-eval-150-raters-per-axis
The ROME human evaluation study used n=150 raters for factual consistency and n=150 for fluency, with 10 independent test cases per participant, using shuffled AI ordering and forced rankings. -
IN
rome-human-eval-ratios
In human evaluation, ROME is rated 1.8× more likely to be consistent with the inserted fact than FT+L, but 1.3× less likely to be more fluent than FT+L. -
IN
rome-human-ranker-rome-over-ftl-consistency
In the ROME human evaluation, all three sampled raters ranked ROME > FT+L > original GPT in factual consistency with the injected counterfactual. -
IN
rome-key-computation-prefixes
The ROME key k* is computed as the post-nonlinearity activation at the chosen mid-layer, averaged over approximately 50 random 2–10 token prefixes generated by the model. -
IN
rome-kl-divergence-lambda-100
ROME's value optimization uses KL divergence scaling factor λ = 100, Adam optimizer with lr = 0.5, weight decay = 1.5×10⁻³, max 20 steps, and early stopping when L(z) reaches 5×10⁻². -
IN
rome-lambda-closed-form-solution
The ROME edit vector Λ is given by the closed-form solution Λ = (v* − Wk*) / ((C⁻¹k*)ᵀk*), obtained by solving the 2×2 block system from the Lagrangian. -
IN
rome-mlp-key-value-interpretation
In the ROME framework, W_fc in the MLP acts as the key projection and W_proj acts as the value projection, modeling the MLP as a linear associative memory where Wk ≈ v. -
IN
rome-model-layer-counts
GPT-2 XL has 48 layers, GPT-J 6B has 28 layers, and GPT-NeoX 20B has 44 layers; GPT-NeoX uses noise scale ν = 0.03 and GPT-J uses ν = 0.025 for corruption. -
IN
rome-pearl-2001-paired-interventions
ROME's causal analysis of hidden states uses paired interventions (causal indirect effects) rooted in Pearl (2001) to measure the causal contribution of individual hidden state vectors. -
IN
rome-prompt-extraction-learn-vs-recall-risk
Fine-tuning on extraction prompts risks the model learning new facts rather than recalling existing ones (Zhong et al., 2021), motivating paraphrase-based benchmarks like ParaRel and COUNTERFACT. -
IN
rome-publication-venue
The ROME paper (Meng et al.) was published at NeurIPS 2022 (36th Conference on Neural Information Processing Systems) with arXiv ID 2202.05262. -
IN
rome-quantitative-benchmark-results
On COUNTERFACT, ROME achieves ~99–100% efficacy and ~75–79% neighborhood specificity, while fine-tuning achieves 100% efficacy but only ~10–40% specificity. -
IN
rome-rank-one-update-formula
ROME applies a rank-one weight update Ŵ = W + Λ(C⁻¹k*)ᵀ to the W_proj matrix of a single mid-layer MLP, where C = KKᵀ and Λ = (v* − Wk*) / ((C⁻¹k*)ᵀk*). -
IN
rome-rank-one-weight-update-layer-9
ROME performs a rank-one weight update at layer 9 of the language model. -
IN
rome-restoration-run-length-10-layers
Decisive causal effects in Causal Tracing require restoring a run of 10 consecutive layer activations (interval [l*−4, ..., l*+5]); single-vector restorations have negligible effects. -
IN
rome-scale-invariance-334m-to-20b
The two-site causal pattern (early MLP at last subject token + late attention) persists across GPT-2 Medium (334M), GPT-2 Large (774M), GPT-2 XL (1.5B), GPT-J (6B), and GPT-NeoX (20B), though peak layer indices shift. -
IN
rome-scope-limitation
ROME and Causal Tracing address factual (entity-relation-object) associations only; logical, spatial, and numerical knowledge are explicitly out of scope. -
IN
rome-second-moment-100k-all-wikipedia-tokens
ROME's second moment statistics C ∝ E[kk^T] are computed from 100,000 hidden-state vectors sampled from all Wikipedia text tokens (not just subject tokens), stored at float32 precision. -
IN
rome-single-fact-scope
ROME as presented in Meng et al. (2022) edits one factual association at a time; scaling to multiple facts is deferred to a follow-up paper. -
IN
rome-specificity-generalization-advantage
ROME's distinguishing advantage over hyper-network baselines (De Cao et al., Mitchell et al.) is that it updates the target fact and its paraphrases without corrupting other facts, whereas those methods exhibit collateral specificity and generalization failures. -
IN
rome-target-layer-gpt2xl
In the ROME experiments on GPT-2 XL (1.5B parameters), the editing layer was layer 18, and the model tested for main causal tracing results was GPT-2 XL. -
IN
rome-value-optimization-objective
ROME optimizes the value v* by minimizing a two-term objective: maximizing P[o*|prompt] (efficacy) plus a KL-divergence penalty on a 'subject is a ___' prompt to control essence drift. -
IN
vicuna7b-memit-multi-hop-drop
MEMIT-edited Vicuna-7B drops from 30.2% to 4.9% multi-hop accuracy