PaperlayerView exact revision on arXiv

Source-linked AI summary

ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall

Jiayu Yang, Yuxuan Fan, Songning Lai, Shengen Wu, Jiaqi Tang, Chun Kang, Zhijiang Guo, Yutao Yue

arXiv:2510.07896v3cs.CL

TL;DR

Existing knowledge-editing methods can lose performance on multi-hop recall, especially when an edit changes an implicit subject in the reasoning chain. ACE uses neuron-level attribution to identify and edit query-value pathways; it outperforms PMET in multi-hop accuracy on GPT-J and Qwen3-8B.

  • Problem

    Existing knowledge-editing methods show reduced performance on multi-hop factual recall, particularly when edits involve implicit subjects in reasoning chains.

    Source: §1 · p. 2

  • Method

    ACE uses neuron-level attribution to identify and edit query-value pathways involved in multi-hop reasoning.

    Source: §1 · p. 2

  • Results

    ACE outperforms PMET in multi-hop accuracy by 9.44% on GPT-J and 37.46% on Qwen3-8B.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Query neurons orchestrate sequential activation of semantically interpretable value neurons across the reasoning chain.

    Source: §7 · p. 10

Abstract

from arXiv · show

Large Language Models (LLMs) require efficient knowledge editing (KE) to update factual information, yet existing methods exhibit significant performance decay in multi-hop factual recall. This failure is particularly acute when edits involve intermediate implicit subjects within reasoning chains. Through causal analysis, we reveal that this limitation stems from an oversight of how chained knowledge is dynamically represented and utilized at the neuron level. We discover that during multi hop reasoning, implicit subjects function as query neurons, which sequentially activate corresponding value neurons across transformer layers to accumulate information toward the final answer, a dynamic prior KE work has overlooked. Guided by this insight, we propose ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall, a framework that leverages neuron-level attribution to identify and edit these critical query-value (Q-V) pathways. ACE provides a mechanistically grounded solution for multi-hop KE, empirically outperforming state-of-the-art methods by 9.44% on GPT-J and 37.46% on Qwen3-8B. Our analysis further reveals more fine-grained activation patterns in Qwen3 and demonstrates that the semantic interpretability of value neurons is orchestrated by query-driven accumulation. These findings establish a new pathway for advancing KE capabilities based on the principled understanding of internal reasoning mechanisms.

1 INTRODUCTION

Efficient knowledge editing is needed because LLM facts can become outdated, while existing methods lose effectiveness on multi-hop recall, especially when edits involve implicit subjects. ACE uses neuron-level attribution to edit query-value pathways and reports higher multi-hop accuracy than PMET on GPT-J and Qwen3-8B.

1 INTRODUCTION

  • Knowledge editing targets specific factual associations cost-effectively because full retraining is computationally prohibitive.Edits aim to minimize effects on unrelated knowledge.Source: §1 · p. 2
  • Locate-then-edit methods such as ROME and MEMIT lose efficacy on multi-hop recall, particularly when the edited knowledge involves an implicit subject.Prior work improved multi-hop performance with deeper FFN edits, but the mechanism behind their importance remained unexplained.Source: §1 · p. 2
  • The paper finds that implicit subjects act as query neurons, sequentially activating value neurons to accumulate information across reasoning hops.It also reports structurally similar localization patterns for semantically analogous knowledge across transformer components.Source: §1 · p. 2
  • 9.44% on GPT-J and 37.46% on Qwen3-8B: ACE outperforms PMET in multi-hop accuracy.The framework identifies and edits critical query-value pathways through neuron-level attribution.Source: §1 · p. 2

2 RELATED WORK

Prior work edits model weights to update facts, but such methods face difficulties preserving reasoning chains in multi-hop settings. ACE builds on neuron-level knowledge localization and query-value interactions to target activation pathways for multi-hop edits.

Knowledge Editing and Multi-hop Reasoning in LLMs

  • Locate-then-edit methods such as ROME and MEMIT update FFN weights, while PMET distinguishes MHSA for general patterns from FFNs for factual content.Other approaches tune small parameter subsets or use hypernetworks.Source: §2 · p. 3
  • Editing intermediate facts can disrupt multi-hop reasoning chains, a challenge highlighted by the MQuAKE benchmark.IFMET advances multi-hop performance through layer-level interventions.Source: §2 · p. 3

Mechanistic Interpretability and Knowledge Localization

  • Prior work characterizes FFNs as key-value stores and identifies interactions between query and value neurons in prediction.ACE traces and modulates these activation pathways for targeted, interpretable editing in multi-hop reasoning.Source: §2 · p. 3

3 PRELIMINARIES

The preliminaries define factual recall and knowledge edits as changes to subject-relation-object facts that models must use across reasoning chains. They then introduce attribution measures for value neurons and query neurons in transformer layers.

3.1 MODELING AND TASK

  • The model processes token representations through transformer layers, where attention outputs combine across heads and FFNs apply nonlinear transformations.The text represents FFN output as a weighted sum of FFN neurons.Source: §3.1 · p. 3
  • A knowledge edit replaces an object in a subject-relation-object fact with a new object.The task represents knowledge as entity-relation-entity triplets.Source: §3.1 · p. 4
  • Factual recall evaluates whether a model answers single-hop and multi-hop questions, with multi-hop chains proceeding from an explicit subject through intermediate facts to a final answer.The editing task tests whether the model uses an updated fact to answer multi-hop questions.Source: §3.1 · p. 4

3.2 ATTRIBUTION METRICS

  • Value-neuron importance is measured by the increase in target-token log probability attributable to a neuron vector.For FFN layers, the measure replaces the previous hidden state with that state plus the attention output.Source: §3.2 · p. 4
  • Query-neuron importance uses the inner product between a query neuron's subkey and itself to estimate its ability to activate value neurons.Larger inner products indicate stronger activation of value neurons.Source: §3.2 · p. 4

4 MECHANISM OF MULTI-HOP REASONING

This section examines how LLMs store semantic knowledge and accumulate information during multi-hop reasoning. It reports structurally similar storage for related semantics and sequential query–value activation through reasoning chains.

4.1 HOW LLMS STORE THE SEMANTICS KNOWLEDGE?

  • Semantically similar knowledge tends to occupy structurally similar components, with MHSA positions shared across categories and FFN layers varying by semantic group.For example, f24, f27, and f16 rank highly for similar semantics, while author, sports-team, and company-founder knowledge appears in distinct layers.Source: §4.1 · p. 5
  • Intervening on 1% of critical semantic-related neurons caused over 90% accuracy decrease in semantic-related subsets.The reported decrease was 9.47% in GPT-J and 8.19% in Qwen3-8B for 1% random intervention.Source: §4.1 · p. 5

4.2 HOW IS THE INFORMATION ACCUMULATED?

  • In GPT-J, value neurons peak in middle-to-deeper FFNs, deplete in deepest layers, and query activation precedes value activation by 1–2 layers.Figure 3 plots query-layer log increase and value-neuron counts by layer.Source: §4.2 · p. 6, Figure 3 · p. 6
  • Ablating the top 100 query neurons in fq16 and fq18 reduced model capability by 46.2% and 61.9%, respectively.Activated value neurons in subsequent layers f17, f18, and f19 fell from (28,16,33) to (6,4,7).Source: §4.2 · p. 6
  • Final-answer information accumulates through implicit query neurons that sequentially activate corresponding value neurons across the reasoning chain.The section frames this as a query–value activation process that progressively encodes answer information.Source: §4.2 · pp. 6–7

5 ACE: ATTRIBUTION-CONTROLLED KNOWLEDGE EDITING

ACE extends locate-then-edit with attribution-guided identification and editing of query–value pathways. On MQuAKE-3K, Table 2 compares multi-hop accuracy across methods in a few-shot setting.

5 ACE: ATTRIBUTION-CONTROLLED KNOWLEDGE EDITING

  • ACE identifies critical query and value neurons, ranks their layers, and edits selected components in three sequential operations.Identification uses importance scores from multi-hop questions; the subsequent operations edit explicit-subject value components and implicit-reasoning query mechanisms.Source: §5 · p. 7
  • ACE applies locate-then-edit to FFNs while leaving attention heads unchanged, using PMET as its editing backbone.The method modifies FFN value components to introduce new factual values.Source: §5 · p. 7
  • Complementary edits to query layers help updated knowledge be activated and traversed during multi-step reasoning.ACE also edits deeper value components and middle-to-shallow query mechanisms.Source: §5 · pp. 7–8
  • Table 2 reports that ACE outperforms other knowledge-editing methods on MQuAKE-3K multi-hop accuracy in a few-shot setting.The table caption does not provide the comparison values.Source: Table 2 · p. 8

6 EXPERIMENTS

Experiments evaluate ACE on a curated MQuAKE-3K subset using GPT-J and Qwen3-8B, with multi-hop answering accuracy as the primary metric. ACE outperforms PMET, and ablations and analyses examine the roles and activation patterns of query and value layers.

6.1 EXPERIMENTAL SETUP

  • The evaluation uses a filtered MQuAKE-3K subset where the base model reasons accurately, with GPT-J-6B and Qwen3-8B assessed by multi-hop answering accuracy.PMET serves as ACE’s primary editing backbone.Source: §6.1 · p. 8

6.2 MAIN RESULTS

  • On the MQuAKE-3K subset across different data settings, ACE demonstrates superiority over existing methods, with reported performance gains of 9.44% on GPT-J and 37.46% on Qwen3-8B.Source: §6.2 · p. 8
  • ACE is reported to outperform state-of-the-art methods across efficacy, paraphrase, and specificity metrics.Efficacy tests single-hop recall, paraphrase tests alternate wording, and specificity assesses effects on other stored facts.Source: §6.2 · p. 8, Table 3 · p. 9

6.3 ABLATION STUDY

  • Skipping the three most important query layers reduces performance by 16.51%, while skipping the two most important value layers reduces it by 40.45%.Both drops are measured relative to ACE; the reported ablation compares skipped-layer edits for GPT-J-6B and Qwen3-8B.Source: §6.3 · p. 9, Table 4 · p. 9
  • With reduced in-context information, ACE’s performance degradation ranges from 0.4% to 9.5% across the reported prompt and model settings.For OOD few-shots, degradation is 0.4% on GPT-J and 0.6% on Qwen3; the evaluation includes zero-shot, one-shot, and OOD few-shot prompts.Source: §6.3 · p. 9, Table 5 · p. 9

6.4 ANALYSIS

  • Analysis attributes prior editing failures to missed query layers and deeper value layers, and finds that Qwen3’s query and value layer positions shift by knowledge domain.GPT-J’s query and value layers instead show consistent separation and invariant positions across domains.Source: §6.4 · pp. 9–10
  • Removing 27 interpretable neurons associated with the correct target token reduces accuracy to 3.2%, versus 59.4% after removing high-importance neurons without that interpretability.The case study uses 1,000 sampling trials and reports the contrast for the single-hop prompt “Tim Duncan plays the sport of”.Source: §6.4 · p. 10
  • The analysis links semantically convergent tokens and shared neurons to recent RL research on token entropy as a future research focus.The authors describe alternating semantic divergence and convergence during token prediction as a fine-grained activation behavior modulated by knowledge domain.Source: §6.4 · p. 10

7 CONCLUSION

The conclusion presents query-value coordination as a mechanism underlying multi-hop factual recall in language models.

7 CONCLUSION

  • Query neurons, including those representing implicit subjects or final-answer components, sequentially activate semantically interpretable value neurons across the reasoning chain.The authors position this mechanism as a contribution to understanding how LLMs organize and process knowledge, and as a direction for more interpretable models.Source: §7 · p. 10

REPRODUCIBILITY STATEMENT

The reproducibility statement describes publicly available implementation materials and experimental details.

REPRODUCIBILITY STATEMENT

  • The supplementary materials provide the code, experimental pipeline, and hyperparameter choices to support independent reproduction.The pipeline covers dataset preparation, model configurations, prompts, and training procedures.Source: REPRODUCIBILITY STATEMENT · p. 11

ETHICS STATEMENT

The supplied material covers ethics and language-model use, related work, dataset preparation, ACE optimization, and the knowledge-bottleneck analysis.

ETHICS STATEMENT

  • The study uses publicly available benchmarks, reports no private or sensitive personal data, and states that experimental code will be made publicly available.The authors also report efforts to mitigate evaluation bias and follow the ICLR Code of Ethics.Source: ETHICS STATEMENT · p. 11

A USE OF LARGE LANGUAGE MODELS

  • An LLM was used occasionally to refine language, but not to formulate questions, design methods, conduct experiments, analyze results, or draft substantive scientific content.The authors state that they reviewed and selectively incorporated linguistic suggestions.Source: A USE OF LARGE LANGUAGE MODELS · p. 14

B RELATED WORK IN DETAILS

  • Prior locate-then-edit methods perform well on single-hop edits but often decline on multi-hop tasks; ACE targets neuron-level query-value pathways, building on work on interpretable value neurons.The cited background describes FFNs as key-value memories and prior query neurons as activators of semantically interpretable value neurons.Source: B RELATED WORK IN DETAILS · pp. 14–15

C.1 SUBSET CONSTRUCTION

  • The subset construction selects MQuAKE knowledge triples answerable by GPT-J-6B in zero-shot settings to reduce response ambiguity.The protocol also establishes a baseline for recall under constrained conditions.Source: §C.1 · p. 15

C.2 DATASET LABELING

  • GPT-4o labels dataset knowledge by categories including geography, organizations, personal attributes, sports, and entertainment.The passage also lists language and culture, education, religion, literature, and events.Source: §C.2 · p. 15

D ACE OPTIMIZATION

  • ACE performs a value-neuron edit followed by a query-neuron edit, optimizing target weights for knowledge preservation and editing.The optimization passage identifies KL prompts and editing prompts as inputs to the process.Source: D ACE OPTIMIZATION · pp. 15–16

E KNOWLEDGE BOTTLENECK

  • ACE avoids the value-only editing bottleneck observed in other editors as the number of edited layers increases.The analysis describes query-neuron editing as efficient and necessary, while excessive layer editing is infeasible and can encounter a knowledge bottleneck.Source: E KNOWLEDGE BOTTLENECK · p. 16

F DETAILS OF INTERNAL FINE-GRAINED REASONING Q-V PAIRS FLOW IN QWEN3-8B

The section links token-position activation patterns to query–value interactions and contrasts the interpretability of query and value neurons.

F DETAILS OF INTERNAL FINE-GRAINED REASONING Q-V PAIRS FLOW IN QWEN3-8B

  • At “plays” and “of,” importance rises as semantic flow converges toward a prediction; other token positions favor exploration and lose interpretability.The authors hypothesize that post-training reinforcement learning confers this intra-sentence pattern.Source: F DETAILS OF INTERNAL FINE-GRAINED… · pp. 16–17
  • Query neurons show little vocabulary interpretability, while value neurons are more interpretable; query layers appear to help by amplifying value-neuron activations through Q-V pairs.The authors argue that query layers enhance editing without directly modifying knowledge representations or token embeddings.Source: F DETAILS OF INTERNAL FINE-GRAINED… · p. 17, Table 8 · p. 17

G MORE DETAILS OF THE ANALYSIS

The Qwen3-8B analysis finds deeper-layer activation and distinguishes dense reasoning propagation from the sparse interpretable neurons associated with correct outputs.

G MORE DETAILS OF THE ANALYSIS

  • In Qwen3-8B, activated layers concentrate in deeper regions, while final layers remain relatively low-activation and are described as primarily generating outputs rather than storing knowledge.Figure 5 presents layer-level log increase for the case study.Source: G MORE DETAILS OF THE ANALYSIS · pp. 17–18, Figure 5 · p. 18
  • Attention log increase drops around layer 25, followed by an FFN drop around layer 27; the authors interpret the pattern as attention activating progressively deeper FFN layers.Figure 6’s darker heatmap regions indicate higher importance scores.Source: G MORE DETAILS OF THE ANALYSIS · p. 18
  • Unlike GPT-J’s shallower query-value activation, Qwen3-8B sustains the process in deeper layers, with active query layers immediately preceding corresponding value layers.The authors relate this pattern to Qwen3-8B’s finer-grained intra-sentence activation behavior.Source: G MORE DETAILS OF THE ANALYSIS · p. 18
  • Removing 27 target-token-associated neurons across 1,000 trials reduced accuracy to 3.2%, while removing 27 high-importance but less interpretable neurons left accuracy at 59.4%.The authors infer that correct output generation depends on a sparse set of specialized, interpretable neurons despite broader information propagation through the network.Source: G MORE DETAILS OF THE ANALYSIS · p. 18

H EXPERIMENTAL SETTINGS

These sections specify ACE’s evaluation and implementation settings, then report its AlphaEdit extension, latency, labeling alignment, and locality results.

H EXPERIMENTAL SETTINGS

  • The reported critical query/value layers are GPT-J-6B Rq={3,4,5,6,7,8}, Rv={26,27,28}, and Qwen3-8B Rq={25,26,27}, Rv={28,29,30,31,32}; ACE mainly updates their FFNs.PMET uses its specified settings, including a 30-step optimization limit and learning rate 0.2; experiments use MQuAKE and few-shot CoT evaluation.Source: H EXPERIMENTAL SETTINGS · pp. 18–19

I PROMPTS

  • The prompt appendix covers multi-hop evaluation, single-hop knowledge screening, and GPT-4o question-category annotation.Evaluation prompts demonstrate reasoning with “Thoughts”; screening probes base-model knowledge, while annotation requests a main and subcategory.Source: I PROMPTS · p. 20, §I.1 · p. 20, §I.2 · p. 20

I.1 FEW-SHOT AND CHAIN-OF-THOUGHT EVALUATION PROMPTS

  • The few-shot CoT examples trace multi-hop facts through intermediate entities before producing the final answer.Examples include questions about a town hall’s country and a company’s creator and location.Source: §I.1 · p. 20

I.2 PROMPTS TO RECALL SINGLE-HOP FACT

  • Single-hop factual questions probe base-model knowledge so experiments can select MQuAKE chains whose ground truth the model already knows.This screening is intended to avoid interference from missing foundational knowledge.Source: §I.2 · p. 20

I.3 DATASET ANNOTATION PROMPT

  • The GPT-4o annotation prompt enforces a consistent two-category output format for dataset question classification.The requested fields are main category and subcategory.Source: §I.3 · pp. 20–21

J ACE LATENCY

  • With knowledge values precomputed and cached, ACE takes 3.10 seconds per GPT-J update and 3.14 seconds per Qwen3-8B update; computing all knowledge vectors takes 26,474.35 and 27,195.26 seconds, respectively.The authors identify inversion of a large square matrix as the most computationally expensive step and say the computation can be cached for incoming edits.Source: J ACE LATENCY · p. 21

K ACE EXTENSION

  • ACE’s Identify-Locate-then-Edit extension outperformed AlphaEdit by 9.97% and 15.95%, and the authors state that the framework extends to later KE methods.The extension replaces AlphaEdit as the backbone and also adds it as a baseline.Source: K ACE EXTENSION · p. 21

L HUMAN UNDERSTANDING ALIGNMENT

  • Most human participants showed acceptable understanding overlap with GPT-4o labeling on random batches of 80 multi-hop questions.Participants included students from literature, social sciences, and STEM backgrounds.Source: L HUMAN UNDERSTANDING ALIGNMENT · p. 21

M.2 ATTRIBUTION ROBUSTNESS

The sections report attribution stability, test whether edits propagate through implicit subjects, assess counterfactual editing, and derive the attribution measure and its relationship to PMET updates.

M.2 ATTRIBUTION ROBUSTNESS

  • Critical-layer importance scores were stable across Pass@k trials, and top-layer rankings changed little across prompting settings, supporting identification with a standard prompt setup.The reported example f27 varies by only about 0.47 points across Pass@k settings; rank changes across prompt templates mostly occur in the lower half of the Top-9 list.Source: §M.2 · p. 22

M.3 VERIFYING INTERMEDIATE REASONING PROCESS CHANGES

  • After masking the final value edit, intermediate “plays” activations remained stable and interpretable, while the final “from” activation dropped by over 40%; the authors interpret this as propagation through implicit subjects rather than late answer injection.The final token’s vocabulary projection also shifted from country names such as “USA” to generic terms such as “country.”Source: §M.3 · pp. 22–23

M.4 COUNTER-FACTUAL EDIT

  • ACE edited counterfactual knowledge chains and retained locality across CSQA, BBH, MMLU, and GSM8k with minimal negative impact on unrelated capabilities.The authors interpret this as evidence that ACE modifies internal representations rather than merely amplifying correct reasoning patterns.Source: §M.4 · pp. 23–24

N ACE ATTRIBUTION DERIVATION

  • The derivation builds ACE attribution metrics from residual-stream algebra, logit additivity, a first-order Taylor approximation, and the PMET update mechanism.The authors state that the additivity assumption is mathematically isomorphic to PMET’s update mechanism.Source: N ACE ATTRIBUTION DERIVATION · p. 25

N.1 ALGEBRAIC STRUCTURE OF THE TRANSFORMER

  • The residual stream accumulates layer outputs, while FFNs are represented as key-value memories whose output is a weighted sum of value vectors.The section defines hidden-state propagation and describes the FFN parameter matrices as keys and values.Source: §N.1 · p. 25

N.2 THEOREM OF LOGIT LINEARITY

  • A target-token logit is the inner product of the final hidden state and its token embedding, and each FFN neuron’s contribution is strictly additive in logit space.The derivation states that interaction terms between active neurons are zero in logit space.Source: §N.2 · pp. 25–26

N.3 DERIVATION OF IMPORTANCE SCORE

  • ACE’s importance score uses change in log probability as a first-order approximation of a neuron’s causal effect on the output distribution.The derivation treats neuron activation as a perturbation to logits and projects its effect through the log-softmax gradient.Source: §N.3 · p. 26

N.4 IMPORTANCE ADDITIVITY

  • Under the assumption that local log-likelihood curvature is negligible, the importance score of a neuron set is the sum of individual scores.The authors restrict the validity of layer importance to the local neighborhood defined by the Taylor expansion.Source: §N.4 · pp. 26–27
Loading 2510.07896v3…