PaperlayerView exact revision on arXiv

Source-linked AI summary

EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory

Hongru Cai, Ran Wei, Wenjie Wang, Chengfa Wu, Ning Song, Yongqi Li, Wenjie Li

arXiv:2610.10533v1cs.CL

TL;DR

Conditional memory could separate factual storage from general-purpose computation, but expression-dependent lookups and shared embeddings complicate selective updates. EngramEdit computes target representations across expressions and jointly updates shared embeddings with reuse-based penalties. It achieves near-perfect editing, supports generalization and multi-hop use, and largely preserves unrelated knowledge and capabilities.

  • Problem

    Conditional memory offers a way to separate factual storage from general-purpose computation, but different expressions can activate different n-grams and shared-embedding updates can affect unrelated factual predictions.

    Source: §1 · p. 2

  • Method

    EngramEdit computes target memory representations across multiple expressions, then jointly updates the shared n-gram embeddings to match them while penalizing frequently reused embeddings more strongly.

    Source: §1 · p. 2

  • Results

    EngramEdit achieves near-perfect editing success on CounterFact and ZsRE, generalizes across expressions, and reaches nearly 3× the strongest baseline’s MQuAKE multi-hop accuracy with CoT prompting.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The findings support conditional memory as an editable knowledge interface for decoupled updates, with unrelated knowledge and general capabilities largely preserved as edits accumulate.

    Source: §1 · p. 3, §6 · p. 11

    Evaluation uses counterfactual edit targets from existing benchmarks, which should not be treated as verified real-world facts; other architectures, scales, and higher-frequency stable updates remain future work.

    Source: Ethics statement · p. 12, §6 · p. 12

Abstract

from arXiv · show

Conditional memory architectures such as DeepSeek Engram use input n-grams to look up learned embeddings, expanding the capacity of large language models (LLMs) with limited additional computation. Beyond model scaling, this architecture has demonstrated the potential to decouple factual knowledge storage from general-purpose computation, offering a promising route to updating factual knowledge while keeping the Transformer backbone fixed. Realizing this potential is challenging because different expressions of a fact may activate different n-gram embeddings, while updating shared embeddings can unintentionally change the model's predictions about other facts. We propose EngramEdit for decoupled knowledge updates through conditional memory. EngramEdit first computes target memory representations that make the model predict the updated fact across multiple expressions. It then jointly updates the shared n-gram embeddings to match these targets across expressions and edits, penalizing updates to frequently reused embeddings more strongly to preserve unrelated knowledge. Experiments show that EngramEdit enables independent factual knowledge updates through conditional memory, achieving near-perfect editing success. Revised knowledge is usable across unseen expressions and in multi-hop reasoning, with nearly three times the strongest baseline's accuracy under chain-of-thought (CoT) prompting. Unrelated knowledge and general capabilities are largely preserved even as factual updates accumulate. These findings show that EngramEdit turns conditional memory into an editable knowledge interface, extending its role beyond model scaling to support decoupled knowledge updates.

1 Introduction

Conditional memory separates reusable factual representations from the Transformer’s contextual processing, suggesting facts could be updated without changing the backbone. EngramEdit addresses expression and embedding conflicts through target representations and jointly regularized embedding updates, with strong editing and preservation results.

1 Introduction

  • Conditional memory uses input n-grams to retrieve learned embeddings for Transformer computation, while the backbone dynamically integrates and processes those reusable representations in context.Disabling DeepSeek Engram memory degrades factual-knowledge performance more than in-context understanding and reasoning, according to the cited ablation.Source: §1 · pp. 1–2
  • Prior work focused mainly on model scaling, personalization, and domain adaptation, leaving decoupled factual updates through conditional memory largely unexplored.This separation suggests updating factual memory while keeping the general-purpose backbone fixed.Source: §1 · p. 2
  • Different expressions can activate different n-grams, edits can demand conflicting changes to shared embeddings, and modifying short or frequent n-grams risks changing unrelated factual predictions.These challenges concern editing efficacy, generalization across expressions, and specificity.Source: §1 · p. 2
  • EngramEdit generates multiple expressions per fact, optimizes a shared temporary perturbation to produce target memory representations, then jointly updates the activated n-gram embeddings to match those targets.Frequently reused embeddings receive stronger penalties to help preserve unrelated knowledge.Source: §1 · pp. 2–3
  • EngramEdit achieves near-perfect editing success on CounterFact and ZsRE, nearly 3× the strongest baseline’s MQuAKE multi-hop accuracy with CoT prompting, and largely preserves unrelated knowledge and general capabilities.Updated facts generalize to unseen expressions, while unrelated-task performance remains largely unchanged during extended sequential editing.Source: §1 · p. 3

2 Preliminaries

The preliminaries introduce factual knowledge editing and conditional memory, then connect the paper’s approach to locate-then-edit methods. They frame editing as changing selected knowledge while preserving unrelated knowledge.

2 Preliminaries

  • Factual knowledge editing updates specific facts in a pretrained model while aiming to preserve unrelated knowledge.An edit request specifies a subject, relation, and desired replacement object.Source: §2 · p. 3
  • EngramEdit follows locate-then-edit methods by finding a representation that elicits the desired object and updating the parameters that produce it.The preliminaries illustrate this framing with FFN editing at a subject-token position, using a temporary perturbation optimized against the desired-object prediction loss.Source: §2 · p. 3
  • Conditional memory associates each n-gram with a learned embedding used when that n-gram is activated, with the same n-gram sharing its embedding across inputs.The memory representation combines activated embeddings through architecture-specific processing and interaction with the backbone state.Source: §2 · pp. 3–4

3 Method

EngramEdit computes memory targets across expressions, maps expressions to shared n-gram embeddings, and jointly solves for updates with reuse-based regularization. The method also establishes cost and stability properties for its update allocation.

3 Method

  • The method computes target memory representations across expressions, maps those expressions to activated n-grams, and jointly solves for one regularized update per shared embedding.Reuse-based regularization penalizes large changes to frequently reused embeddings.Source: §3 · p. 4, §3.3 · p. 5

3.1 Conditional Memory Target Computation

  • For each edit, EngramEdit generates semantically equivalent expressions and optimizes a shared temporary perturbation across their memory representations to predict the updated fact.The target computation holds model parameters fixed; the optimized perturbation defines target representations, with a norm clamp and penalties on perturbation size and subject-description prediction changes.Source: §3.1 · pp. 4–5

3.2 Multi-Expression Memory Mapping

  • EngramEdit maps each expression’s target to its activated n-grams and records which expressions use each distinct embedding in a batch-level mapping matrix.Expressions that activate the same n-gram share its embedding, which is retained once in the batch mapping.Source: §3.2 · p. 5

3.3 Joint Memory Update Allocation

  • Under sum aggregation, embedding updates are chosen to match target memory changes, while shorter or more frequent n-grams receive larger penalties because they are more likely to affect unrelated inputs.The objective jointly balances potentially conflicting expression targets and has a unique solution when all regularization weights are positive.Source: §3.3 · p. 6
  • Weighting update penalties by activation probability bounds the average squared change in memory representations on unrelated inputs under sum aggregation.The regularization combines a reuse-based penalty with an optional uniform ridge penalty.Source: §3.3 · p. 6
  • The joint solution minimizes weighted embedding-update cost among updates producing the same achieved memory changes, and regularization bounds how target variations affect solved updates.The stability result concerns fixed mapping and regularization.Source: §3.3 · pp. 6–7

4 Experiments

Experiments evaluate factual editing and preservation, general capabilities during sequential edits, and the roles of EngramEdit’s design choices. EngramEdit performs strongly on cross-expression and multi-hop use of revised facts while largely preserving unrelated knowledge and general capabilities.

4.1 Experimental Setup

  • The evaluation uses sequential editing on CounterFact and ZsRE, measures editing and preservation metrics, and tests multi-hop reasoning on MQuAKE.The experiments compare MoE-applicable editors and direct memory fine-tuning methods using LongCat-Flash-Lite.Source: §4 · p. 7, §4.1 · p. 7

4.2 Knowledge Update and Preservation (RQ1)

  • With CoT prompting, EngramEdit achieves nearly 3× the strongest baseline’s MQuAKE accuracy and leads across all hop groups.On CounterFact and ZsRE, it also achieves the highest Utility, with Specificity close to pre-edit levels, and leads in Fluency and Consistency.Source: Figure 3 · p. 8, §4.2 · p. 8

4.3 General Capabilities (RQ2)

  • After 5,000 sequential edits, EngramEdit retains over 96% of pre-edit mean F1, with small changes on most tasks but a larger decline on MRPC.The authors relate the MRPC decline to local embedding updates affecting sentence representations differently in equivalence judgments.Source: §4.3 · p. 9, Figure 4 · p. 9

4.4 In-Depth Analysis

  • Removing joint allocation lowers Efficacy and Generalization, while removing reuse-based regularization lowers Efficacy and Specificity; frequency and length weighting also support Utility.Generated expressions improve Generalization in the unregularized variant, and over 92% of new unrelated-query errors involve updated embeddings.Source: §4.4 · p. 9, Table 2 · p. 10
  • Using generated expressions for memory mapping increases held-out paraphrase Coverage and yields most of the Generalization gain; target computation adds further gains at unchanged Coverage.With the same generated expressions, EngramEdit outperforms MFT-S in Generalization at similar Efficacy and Specificity.Source: §4.4 · pp. 9–10, Figure 5 · p. 10
  • Disabling fact-related updates sharply lowers Efficacy and Generalization, whereas disabling a matched set of random updates leaves both unchanged.Fact-related disabling nearly matches disabling all updates and makes the model favor original over revised facts on average.Source: §4.4 · p. 10, Figure 5 · p. 10

5 Related Work

Related work covers conditional memory and n-gram embeddings for model scaling and adaptation, as well as knowledge-editing methods that preserve or modify model parameters. EngramEdit applies pretrained conditional memory to decoupled factual updates.

Conditional Memory and n-gram Embeddings

  • Conditional-memory and n-gram approaches use reusable local representations for model scaling, memory construction, personalization, and domain adaptation [6] [7] [8] [23].EngramEdit develops pretrained conditional memory as an editable interface while aiming to preserve unrelated knowledge and general capabilities.Source: §5 · p. 11

Knowledge Editing

  • Knowledge-editing methods either preserve original parameters with external information or added components, or modify existing parameters through constrained updates, locate-then-edit, or other mechanisms [15].EngramEdit focuses on decoupled updates through pretrained conditional memory.Source: §5 · p. 11

6 Conclusion and Future Work

The paper concludes that EngramEdit supports decoupled factual updates through conditional memory, with high editing success and broad preservation. It identifies other architectures, model scales, and efficient stable updates at higher frequencies as directions for future work.

6 Conclusion and Future Work

  • EngramEdit jointly updates shared n-gram embeddings with reuse-based regularization and supports factual editing, generalization, multi-hop use, and largely preserved unrelated knowledge and capabilities.The authors frame conditional memory as an editable knowledge interface beyond its model-scaling role.Source: §6 · p. 11
  • Future work can explore other conditional-memory architectures and model scales, and efficient, stable updates to more facts at higher frequencies.The conclusion presents these as future directions.Source: §6 · p. 12

AI use statement

  • The authors report using generative AI for implementation, proof drafting, writing and revision, figure creation, and discussion of result interpretation, while taking responsibility for the final work.They state that they reviewed the implementation and proofs and tested the code.Source: AI use statement · p. 12

Ethics statement

  • The evaluation uses benchmark counterfactual targets, which the authors caution should not be treated as verified real-world facts; they recommend verification, access restrictions, and auditing.The ethics statement notes that knowledge revision could also be misused to introduce false or misleading content.Source: Ethics statement · p. 12

Reproducibility statement

  • The authors state that results are reproducible and that the code is available in their repository.They point to Section 3 and Appendix A for the method and proofs, and Section 4.1 and Appendix B for datasets, protocols, and implementation settings.Source: Reproducibility statement · p. 12

A Method Details and Proofs

The appendix details EngramEdit’s target-computation and memory-update stages, including expression coverage, aggregation variants, reuse-based regularization, and sequential editing. It also gives the joint procedure’s computational cost.

A.1 Locate-Then-Edit Formulation

  • Causal tracing uses restoration of a clean hidden state to identify or motivate the FFN layer and subject-token position used for editing.A large indirect effect indicates that the restored state contributes strongly to recalling the fact.Source: §A.1 · p. 15
  • Target computation temporarily perturbs the selected FFN output to predict the desired object while a KL term limits changes on a preservation prompt; the perturbation is removed afterward.The optimized perturbation defines the target output rather than becoming a persistent model update.Source: §A.1 · p. 15
  • The parameter update jointly matches target outputs while regularizing changes on sampled preservation activations, with a positive ridge term ensuring invertibility.The formulation updates the FFN output projection for a batch of edits.Source: §A.1 · p. 16

A.2.1 Expression Generation

  • Expression generation varies wording immediately before the subject to cover different activated n-grams, using prompts for CounterFact and atomic MQuAKE edits and a same-meaning, same-answer-type instruction for ZsRE.Generated candidates are filtered for duplicates, malformed outputs, and answer strings before accepted expressions are used.Source: §A.2.1 · p. 16, Figure 6 · p. 17

A.2.2 Target Computation Objective

  • Target computation shares a perturbation across each edit’s expression set and combines preservation loss with a norm penalty and a clamp on perturbation size.The preservation prompt is “s_i is a,” and the norm penalty uses the original prompt’s memory representation for normalization.Source: §A.2.2 · p. 17

A.3.1 Sum Aggregation

  • Under sum aggregation, stacking expression-level target relations yields AU ≈ B, so each shared embedding update is constrained by every expression that uses it.Fixed components cancel when current memory representations are subtracted from their targets.Source: §A.3.1 · p. 18

A.3.2 Context-Aware Gated Aggregation

  • For context-aware gating, target matching accounts for changes to projected values and gates; a local Jacobian approximation yields a linear system, while reuse-based regularization is retained.The displayed derivation initially holds the upstream query fixed; if updates affect upstream computation, the query is recomputed during optimization.Source: §A.3.2 · pp. 18–19

A.4 Complete Editing Procedure

  • EngramEdit computes targets, solves joint embedding updates, and can repeat the procedure sequentially using the current memory; its solver has stated costs in batch size, unique n-grams, and embedding dimension.When selected n-grams are not shared between expression groups, the system can be split without changing the joint solution.Source: §A.4 · pp. 19–20

A.5 Proofs for Joint Memory Update Allocation

The appendix characterizes when joint target matching is feasible and how the objective behaves when targets conflict. Exact matching depends on whether each target column lies in the mapping matrix’s column space.

A.5.1 Feasibility of Exact Matching

  • Expressions sharing the same selected n-grams receive the same memory-representation change, so differing targets can make exact matching impossible.When exact matching is infeasible, squared matching error permits a joint compromise across conflicting targets.Source: §A.5.1 · p. 21
  • Exact matching AU = B is possible if and only if every column of B lies in the column space of A.The condition applies to the matching target matrix B and mapping matrix A.Source: §A.5.1 · p. 21
  • Positive regularization can favor an inexact match even when exact matching is feasible.Source: §A.5.1 · p. 21

A.5.2 Reuse-Based Regularization

Reuse-based penalties connect embedding update size to expected memory-representation changes on unrelated inputs. The appendix also establishes properties of the joint solution, including uniqueness, minimum weighted cost, sensitivity bounds, and exact decomposition over disconnected components.

A.5.2 Reuse-Based Regularization

  • More frequent activation on unrelated inputs gives the same embedding update a larger cost, motivating stronger penalties for frequently reused embeddings.For fixed targets, increasing a penalty tightens the bound on its embedding update; the bounds concern memory representations, not downstream predictions.Source: §A.5.2 · pp. 22–23

A.5.3 Closed-Form Solution and Uniqueness

  • When every regularization coefficient is positive, the joint objective has a unique solution, even if overlapping activations make the mapping matrix rank deficient.Positive regularization makes the coefficient matrix positive definite and the objective strictly convex.Source: §A.5.3 · p. 23

A.5.4 Minimum-Cost Update Allocation

  • Among updates producing the same achieved memory-representation changes, the joint solution uniquely minimizes weighted update cost.This comparison remains valid when exact target matching is infeasible; it guarantees cost minimality, not downstream prediction behavior.Source: §A.5.4 · p. 24

A.5.5 Sensitivity to Target Changes

  • For fixed mapping and positive-definite regularization, the solution-change bound scales with target changes and is controlled by the smallest regularization coefficient.The bound applies even if the mapping matrix is rank deficient, but does not guarantee amplification is below one.Source: §A.5.5 · p. 24

A.5.6 Component Decomposition

  • The joint update decomposes exactly over disconnected components of the graph connecting expression rows to selected n-gram columns.Groups sharing no selected embeddings can be solved separately without changing the joint solution.Source: §A.5.6 · p. 25

B Experimental Setup

  • The experimental-setup section describes the datasets, evaluation, baselines, and implementation details for the experiments.Source: B Experimental Setup · p. 25

B.1 Datasets

The evaluation uses factual-editing benchmarks, a multi-hop reasoning benchmark, and separate general-capability tasks. The datasets vary in the requested edits and the held-out evaluations used to test them.

CounterFact

  • CounterFact tests counterfactual factual-association edits using edit requests, held-out paraphrases, neighborhood prompts, and generation prompts.Generated expressions are used during editing, while held-out paraphrases are reserved for evaluation.Source: §B.1 · p. 25

ZsRE

  • ZsRE tests question-answer factual associations by editing to a supplied reference answer and evaluating a held-out paraphrase plus an unrelated question-answer pair.The edit requests the supplied answer rather than a counterfactual replacement.Source: §B.1 · p. 25

MQuAKE

  • MQuAKE tests whether revised facts can be combined to answer multi-hop questions across 3,000 counterfactual cases covering two- to four-hop reasoning.Cases include one or more fact-level edits, question variants, and answers implied by the revised facts.Source: §B.1 · p. 26

General-Capability Tasks

  • Separate general-capability evaluations cover sentiment, paraphrase recognition, linguistic acceptability, textual entailment, and knowledge across subjects.The tasks use 100 examples each, none included in the edit requests.Source: §B.1 · p. 26

B.2 Evaluation

The evaluation defines editing, generalization, specificity, and utility, then compares EngramEdit with model-editing and memory fine-tuning baselines. Implementation details describe sequential editing, multi-expression target computation, and joint memory updates; extended experiments examine editing through 5,000 facts and multi-hop reasoning.

B.2 Evaluation

  • Efficacy measures success on edit prompts, Generalization on held-out paraphrases, Specificity on unrelated queries, and Utility is their arithmetic mean.MQuAKE reports the percentage of cases answered correctly; general-capability evaluation uses mean task-level F1 across six tasks.Source: §B.2 · pp. 26–27
  • Evaluation intervals vary by benchmark, and reported editing confidence intervals quantify uncertainty across cases rather than random seeds.CounterFact is evaluated every 100 edits, ZsRE at common 200-edit intervals, and general capabilities every 500 edits.Source: §B.2 · p. 28, §B.4.1 · p. 29

B.3 Baselines

  • Baselines include gradient-based expert editors, MoEEdit, and Memory-FT variants that update subject-position or all-token memory embeddings.FT, FT-L, AdaLoRA, UnKE, and MoEEdit edit routed experts; the Memory-FT variants keep the backbone and other specified parameters fixed.Source: §B.3 · p. 28

B.4.1 Model and Editing Setup

  • Experiments use LongCat-Flash-Lite, a 68.5B-parameter MoE model with 31.4B parameters in n-gram embeddings, and apply sequential batches of edits.The main CounterFact and ZsRE runs apply 2,000 edits in batches of 100; longer runs extend to 5,000 facts.Source: §B.4.1 · p. 29

B.4.2 Target Computation and Memory Mapping

  • For each edit, four generated expressions and the original prompt support target computation; a shared memory perturbation is optimized while decoder parameters remain fixed.Prediction loss is evaluated across expressions alone and after five generated prefix contexts.Source: §B.4.2 · pp. 29–30

B.4.3 Memory Updating

  • EngramEdit maps expressions to activated suffix n-grams, then jointly solves one update per edited embedding, with reuse-based regularization to limit changes to frequently reused embeddings.Separate cumulative update vectors are indexed by exact n-gram sequence, while pretrained hashed tables remain fixed.Source: §B.4.2 · p. 30, §B.4.3 · p. 30

C.1.1 Sequential Editing with 5,000 Facts

  • EngramEdit maintains high Efficacy and a Generalization lead on CounterFact and ZsRE through 5,000 sequential edits, while preserving higher Specificity than several expert-editing baselines on CounterFact.Evaluations include all edits applied so far; the trajectories are shown in Figure 7.Source: §C.1.1 · pp. 30–31, Figure 7 · p. 31

C.1.2 Multi-Hop Reasoning

  • EngramEdit gains substantially from CoT at all three MQuAKE hop depths, and leads every hop group with CoT despite lower accuracy and smaller margins at greater depths.FT-L, AdaLoRA, and MoEEdit lose accuracy in every group under CoT.Source: §C.1.2 · p. 31

C.1.3 Analysis of ZsRE Specificity

This analysis compares unrelated-query predictions before and after editing on ZsRE, separating retained correctness from newly correct or incorrect answers. EngramEdit retains most initially correct predictions, while other methods’ higher post-edit Specificity includes corrections of previously wrong answers.

C.1.3 Analysis of ZsRE Specificity

  • The transition analysis compares seven methods on the same 11,379 neighborhood prompts and evaluates EngramEdit using the first 2,000 edits of its separate 5,000-edit run.Correctness transitions are averaged equally across cases.Source: §C.1.3 · pp. 31–32
  • 93.83% of correctness states remain unchanged for EngramEdit, and about 87.4% of initially correct predictions are retained after 2,000 edits.Its W→C rate is 0.92%, so post-edit Specificity predominantly reflects retained correct predictions.Source: Table 5 · p. 32, §C.1.3 · p. 33
  • FT-L, AdaLoRA, UnKE, and MoEEdit exceed pre-edit Specificity because their W→C rates exceed their C→W rates.Post-edit Specificity counts both retained correct answers and newly correct answers, so it does not measure retention alone.Source: §C.1.3 · p. 32

C.1.4 Task-Level General Capabilities

The additional analyses examine capability preservation, expression and n-gram choices, regularization sensitivity, and how memory activation and sharing relate to editing outcomes. Across these tests, broader expression coverage supports generalization, while updates to reused or shared embeddings are linked to preservation trade-offs and editing failures.

C.1.4 Task-Level General Capabilities

  • Across 5,000 edits, EngramEdit preserves pre-edit F1 on SST and CoLA, while RTE, MMLU, and NLI stay within four F1 points of their pre-edit scores.Larger changes concentrate on MRPC; MFT-S shows a similar pattern.Source: §C.1.4 · p. 33

C.2.1 Multi-Expression Editing

  • Adding four generated expressions improves Generalization for MFT-S, MFT-A, and EngramEdit on both datasets, whereas MoEEdit’s Efficacy and Generalization decrease.EngramEdit leads in Efficacy, Generalization, and Utility on both CounterFact and ZsRE with either expression count.Source: §C.2.1 · p. 34, Table 6 · p. 34

C.2.2 Number of Generated Expressions

  • Two generated expressions provide most of the Generalization gain on both datasets, with further improvements at four; beyond four, gains are slight and Utility does not consistently increase.Four expressions yield the highest Utility on ZsRE and nearly the highest on CounterFact.Source: §C.2.2 · p. 35, Table 7 · p. 35

C.2.3 Editable n-gram Lengths

  • Using 2-, 3-, and 4-grams together achieves the highest Efficacy, Generalization, and Utility, with higher Specificity than using 2-grams alone.Excluding single-token updates raises Specificity without reducing Efficacy or Generalization; individual tokens can affect many unrelated inputs.Source: §C.2.3 · pp. 35–36, Figure 9 · p. 35

C.2.4 Regularization and Clamp Factor

  • All tested nonzero reuse coefficients outperform disabling reuse-based regularization on Efficacy, Specificity, and Utility, while results remain close across ridge coefficients and clamp factors.Increasing the clamp factor from 8 to 64 yields no consistent Utility gain.Source: §C.2.4 · pp. 36–37, Table 8 · p. 36, Figure 10 · p. 37

C.3.1 Disabling Memory Updates

  • Disabling fact-related updates turns 5,228 of 5,850 previously successful prompts into failures, while matched-random disabling causes none to fail.Disabling an average of 4.679 fact-related updates nearly matches disabling all 9,253 updates in mean margin and success-to-failure rate.Source: §C.3.1 · pp. 37–38

C.3.2 Memory Activation and Editing Outcomes

  • Current-only and Both groups cover 98.05% of held-out paraphrases and have success rates of 96.97–98.97%; Other-only and None account for 50.36% of failures.These groups are defined by which updated embeddings prompts activate.Source: §C.3.2 · p. 38, Figure 11 · p. 38

C.3.3 Cross-Edit Sharing

  • At 5,000 edits, shared n-grams account for 1.25% of updated embeddings and 8.66% of facts; the Sharing group contains 48 of 49 Efficacy failures.Facts without sharing achieve 99.98% Efficacy and 96.7% Generalization.Source: §C.3.3 · p. 39, Figure 12 · p. 39
Loading 2610.10533v1…