The Anatomy of an Edit: Mechanism-Guided Activation Steering for Knowledge Editing

Yuan Cao; Mingyang Wang; Hinrich Schütze

cs.CLMar 21, 2026

The Anatomy of an Edit: Mechanism-Guided Activation Steering for Knowledge Editing

Authors: Yuan Cao, Mingyang Wang, Hinrich Schütze

Organizations: Technical University of Munich · LMU Munich · Munich Center for Machine Learning (MCML)

Abstract

Large language models (LLMs) are increasingly used as knowledge bases, but keeping them up to date requires targeted knowledge editing (KE). However, it remains unclear how edits are implemented inside the model once applied. In this work, we take a mechanistic view of KE using neuron-level knowledge attribution (NLKA). Unlike prior work that focuses on pre-edit causal tracing and localization, we use post-edit attribution -- contrasting successful and failed edits -- to isolate the computations that shift when an edit succeeds. Across representative KE methods, we find a consistent pattern: mid-to-late attention predominantly promotes the new target, while attention and FFN modules cooperate to suppress the original fact. Motivated by these findings, we propose MEGA, a MEchanism-Guided Activation steering method that performs attention-residual interventions in attribution-aligned regions without modifying model weights. On CounterFact and Popular, MEGA achieves strong editing performance across KE metrics on GPT2-XL and LLaMA2-7B. Overall, our results elevate post-edit attribution from analysis to engineering signal: by pinpointing where and how edits take hold, it powers MEGA to deliver reliable, architecture-agnostic knowledge edits.

Figures & tables

Figure 1: Overview of MEGA. (Left): Post-edit NLKA identifies mid-to-late attention layers with high leverage on the target vs. original token. (Middle): MEGA uses this signal to select an edit zone and (Right): applies PCA-stabilized attention–residual steering at inference time, promoting the edited fact without changing weights.

Panel of Figure 2 — Figure 2: MEMIT success vs. failure. Mean contribution differences across multiple cases; positive promotes new target, negative suppresses original.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 3 — Figure 3: IKE success vs. failure on CounterFact (GPT2-XL).

Panel of Figure 4 — Figure 4: ROME success vs. failure on CounterFact (GPT2-XL).

Panel of Figure 5 — Figure 5: Fine-tuning success vs. failure on CounterFact (GPT2-XL).

Panel of Figure 6 — Figure 6: ROME success vs. failure on CounterFact (LLaMA2-7B).

Panel of Figure 7 — Figure 7: IKE success vs. failure on CounterFact (LLaMA2-7B).

Panel of Figure 8 — Figure 8: MEMIT success vs. failure on CounterFact (LLaMA2-7B).

Panel of Figure 9 — Figure 9: Fine-tuning success vs. failure on CounterFact (LLaMA2-7B).

Panel of Figure 10 — Figure 10: ROME success vs. failure on Popular (GPT2-XL).

Panel of Figure 11 — Figure 11: IKE success vs. failure on Popular (GPT2-XL).

Panel of Figure 12 — Figure 12: MEMIT success vs. failure on Popular (GPT2-XL).

Panel of Figure 13 — Figure 13: Fine-tuning success vs. failure on Popular (GPT2-XL).

Panel of Figure 14 — Figure 14: ROME success vs. failure on Popular (LLaMA2-7B).

Panel of Figure 15 — Figure 15: IKE success vs. failure on Popular (LLaMA2-7B).

Panel of Figure 16 — Figure 16: MEMIT success vs. failure on Popular (LLaMA2-7B).

Panel of Figure 17 — Figure 17: Fine-tuning success vs. failure on Popular (LLaMA2-7B).

Setting	Range	Acc	Gen	Spec	DI	DII	Avg
26-layer	0–25	0.16	0.08	0.99	1.00	0.47	0.54
26-layer	15–40	0.98	0.84	0.87	1.00	0.98	0.93
26-layer	22–47	0.88	0.71	0.87	0.99	0.96	0.88
17-layer	0–16	0.01	0.02	1.00	0.22	0.02	0.25
17-layer	20–36	0.86	0.73	0.88	0.98	0.93	0.88
17-layer	31–47	0.21	0.25	0.94	0.82	0.78	0.60

Appendix

Table 3: Attention steering on CounterFact for GPT2-XL: moving a fixed-size window. Attribution-guided mid layers (15–40, 20–36) clearly outperform early or very-late windows.

Comp	PCA	Acc	Gen	Spec	DI	DII	Avg
Range 0–25
ATTN	128	0.158	0.079	0.992	1.000	0.467	0.539
ATTN	512	0.183	0.075	0.987	0.925	0.533	0.541
ATTN-res	128	0.092	0.075	0.915	0.708	0.217	0.401
ATTN-res	512	0.217	0.233	0.897	0.858	0.783	0.598
FFN	128	0.008	0.004	0.998	0.008	0.017	0.207

Appendix

Table 4: Multi-layer steering on CounterFact (GPT2-XL), 26-layer windows.

Comp	PCA	Acc	Gen	Spec	DI	DII	Avg
Range 0–16
ATTN	128	0.008	0.017	1.000	0.217	0.017	0.252
ATTN	512	0.008	0.013	1.000	0.167	0.008	0.239
ATTN-res	128	0.025	0.008	0.969	0.533	0.017	0.310
ATTN-res	512	0.017	0.017	0.978	0.283	0.092	0.277
FFN	128	0.008	0.013	1.000	0.000	0.033	0.211

Appendix

Table 5: Multi-layer steering on CounterFact (GPT2-XL), 17-layer windows.

Comp	PCA	Acc	Gen	Spec	DI	DII	Avg
Layer 0
ATTN	1600	0.008	0.008	1.000	0.067	0.000	0.217
ATTN	128	0.008	0.008	1.000	0.067	0.000	0.217
ATTN	512	0.008	0.008	1.000	0.067	0.000	0.217
ATTN-res	1600	0.008	0.008	1.000	0.067	0.000	0.217
ATTN-res	128	0.008	0.008	0.998	0.000	0.000	0.203

Appendix

Table 6: Single-layer steering on CounterFact (GPT2-XL), layers 0–15.

Comp	PCA	Acc	Gen	Spec	DI	DII	Avg
Layer 20
ATTN	1600	0.008	0.013	1.000	0.125	0.000	0.229
ATTN	128	0.008	0.013	1.000	0.142	0.000	0.233
ATTN	512	0.008	0.013	1.000	0.133	0.000	0.231
ATTN-res	1600	0.025	0.029	1.000	0.467	0.067	0.318
ATTN-res	128	0.025	0.038	0.999	0.825	0.042	0.386

Appendix

Table 7: Single-layer steering on CounterFact (GPT2-XL), layers 20–30.

Comp	PCA	Acc	Gen	Spec	DI	DII	Avg
Range 13–31
ATTN	512	0.942	0.825	0.888	0.917	0.950	0.904
ATTN-res	512	0.992	0.913	0.888	1.000	0.967	0.952
FFN	512	0.000	0.008	0.888	0.250	0.433	0.316
Res	512	0.908	0.779	0.888	0.767	0.875	0.844
Range 15–27

Appendix

Table 8: Multi-layer steering on CounterFact (LLaMA2-7B), layer windows.

Comp	PCA	Acc	CI	CII	RS	SA	Avg
Range 15–40
ATTN	128	0.667	0.275	0.400	0.382	0.703	0.485
ATTN	512	0.392	0.250	0.400	0.441	0.658	0.428
ATTN-res	128	0.642	0.225	0.400	0.324	0.679	0.454
ATTN-res	512	0.817	0.200	0.450	0.324	0.648	0.488
FFN	128	0.050	0.410	0.000	0.660	1.000	0.424

Appendix

Table 9: Multi-layer steering on Popular (GPT2-XL), layer windows.

Comp	PCA	Acc	CI	CII	RS	SA	Avg
Range 13–31
ATTN	512	0.517	0.082	0.284	0.354	0.637	0.375
ATTN-res	512	0.817	0.074	0.247	0.329	0.637	0.421
Res	512	0.692	0.081	0.148	0.399	0.651	0.394
Range 15–31
ATTN	512	0.500	0.082	0.284	0.367	0.630	0.372

Appendix

Table 10: Multi-layer steering on Popular (LLaMA2-7B), layer windows.

Language Model Editing

Language model editing changes a model’s knowledge or behavior, or revises its outputs, while aiming to preserve unrelated capabilities.

See all 76 papers

Exposing the Illusion of Erasure in Knowledge Editing for LLMs
Jun 22, 2026Advik Raj Basani, Anshuman ChhabraKnowledge Editing Knowledge Conflicts in Language Models
Birla Institute of Technology and Science, Goa · University of South Florida
Can Fine-Tuning Erase Your Edits? On the Fragile Coexistence of Knowledge Editing and Adaptation
Nov 8, 2025Yinjie Cheng, Paul Youssef, Christin Seifert +2Knowledge Editing LLM Fine-Tuning
School of Computer Science University of Sheffield Sheffield, UK · Mathematics and Computer Science Marburg University Marburg, Germany
Scalable Knowledge Editing for Mixture-of-Experts LLMs via Tensor-Structured Updates
May 15, 2026Roman Maksimov, Vladimir Aletov, Dmitry Bylinkin +3Knowledge Editing Mixture-of-Experts Language Models
1Moscow Independent Research Institute of Artificial Intelligence · 2Basic Research of Artificial Intelligence Laboratory · 3Innopolis University

Jun 22, 2026J/K move · Enter open · S save

Exposing the Illusion of Erasure in Knowledge Editing for LLMs

Advik Raj Basani, Anshuman Chhabra

Birla Institute of Technology and Science, Goa · University of South Florida

Knowledge Editing Knowledge Conflicts in Language Models LLM Interpretability Factual Knowledge in Language Models

Open paper AskPDF arXiv

Knowledge Editing (KE) has emerged as a frontier for updating specific facts in LLMs without costly retraining, but its reliability and underlying mechanisms remain poorly understood. In this work, we examine KE from an adversarial elicitation perspective, revealing that edited knowledge is often not fully erased and continues to surface, with consistent failures observed across diverse model architectures. To explain this behavior, we conduct a mechanistic analysis of popular KE methods. We show that low-rank updates do not overwrite existing knowledge but instead redistribute it within the model's representation space. Furthermore, we find that these methods act as targeted suppression mechanisms that reduce the likelihood of expressing original facts, rather than removing them from the model. Analysis of the loss landscape reveals that edited knowledge lies in narrow, anisotropic regions that are highly sensitive to perturbations, making them highly vulnerable to indirect prompting and adversarial attacks. By exposing these profound architectural vulnerabilities, our work proves that KE algorithms are inherently bypassable and motivates a fundamental reevaluation of how we deploy post-hoc updates in several LLM applications.

Abstract

Language Model Editing

Explore similar work

Exposing the Illusion of Erasure in Knowledge Editing for LLMs

Can Fine-Tuning Erase Your Edits? On the Fragile Coexistence of Knowledge Editing and Adaptation

Scalable Knowledge Editing for Mixture-of-Experts LLMs via Tensor-Structured Updates