Uncovering Entity Identity Confusion in Multimodal Knowledge Editing

# Uncovering Entity Identity Confusion in Multimodal Knowledge Editing
Source: [https://arxiv.org/html/2605.06096](https://arxiv.org/html/2605.06096)
Shu Wu1, Xiaotian Ye2∗, Xinyu Mou1,3∗, Dongsheng Liu1,4∗, Xiaohan Wang5, Mengqi Zhang6

1New Laboratory of Pattern Recognition \(NLPR\) State Key Laboratory of Multimodal Artificial Intelligence Systems \(MAIS\) Institute of Automation, Chinese Academy of Sciences 2Beijing University of Posts and Telecommunications 3School of Artificial Intelligence, University of Chinese Academy of Sciences 4School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences 5Huazhong University of Science and Technology 6Shandong University shu\.wu@nlpr\.ia\.ac\.cn, yexiaotian@bupt\.edu\.cn \{mouxinyu2025, liudongsheng2025\}@ia\.ac\.cn shawn\_wang@hust\.edu\.cn, mengqi\.zhang@sdu\.edu\.cn

###### Abstract

Multimodal knowledge editing \(MKE\) aims to correct the internal knowledge of large vision\-language models after deployment, yet the behavioral patterns of post\-edit models remain underexplored\. In this paper, we identify a systemic failure mode in edited models, termed Entity Identity Confusion \(EIC\): edited models exhibit an absurd behavior where text\-only queries about the original entity’s identity unexpectedly return information about the new entity\. To rigorously investigate EIC, we construct EC\-Bench, a diagnostic benchmark that directly probes how image\-entity bindings shift before and after editing\. Our analysis reveals that EIC stems from existing methods failing to distinguish between Image\-Entity \(I\-E\) binding and Entity\-Entity \(E\-E\) relational knowledge in the model, causing models to overfit E\-E associations as a shortcut: the image is still perceived as the original entity, with the new entity’s name serving only as a spurious identity label\. We further explore potential mitigation strategies, showing that constraining edits to the model’s I\-E processing stage encourages edits to act more faithfully on I\-E binding, thereby substantially reducing EIC\. Based on these findings, we discuss principled desiderata for faithful MKE and provide methodological guidance for future research\.

## 1Introduction

Today’s knowledge editing \(KE\)\(Zhanget al\.,[2024b](https://arxiv.org/html/2605.06096#bib.bib83)\)has established itself as a key research area in the large language model \(LLM\)\(Zhaoet al\.,[2025](https://arxiv.org/html/2605.06096#bib.bib74)\)field\. In real\-world deployments, maintaining LLMs often requires revising their encoded knowledge to address outdated facts or to meet safety, policy, and privacy requirements\. Knowledge editing focuses on targeted modifications to the internal knowledge of LLMs, thereby enabling more practical and auditable post\-deployment maintenance\. With the growing adoption of large vision\-language models \(LVLMs\)\(Liuet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib123); Zhuet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib140); Baiet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib134)\)in real\-world applications, these needs have naturally extended from purely textual systems toMultimodal Knowledge Editing \(MKE\)\(Chenget al\.,[2023a](https://arxiv.org/html/2605.06096#bib.bib130)\)\.

Unlike text\-based knowledge editing\(Menget al\.,[2022](https://arxiv.org/html/2605.06096#bib.bib61); Zhanget al\.,[2026](https://arxiv.org/html/2605.06096#bib.bib110)\), which typically targets relationships between real\-world entities \(e\.g\., modifying that “Trump, graduate from, UPenn”\), mainstream multimodal KE settings focus on binding the content depicted in a specific image to a different entity\. As shown in Figure[1](https://arxiv.org/html/2605.06096#S1.F1)\(a\), for an imageAAof Trump that the pre\-edit model erroneously recognizes as Biden, the post\-MKE model correctly identifies the content in the image as the true entity Trump\. Despite this natural motivation, multimodal KE remains considerably less mature than its text\-only counterpart, and systematic analysis of post\-edit model behavior is largely absent from the literature\.

In this work, we observe a previously undiscovered failure mode during our analysis of post\-edit model behavior, which we termEntity Identity Confusion \(EIC\): after the entity bound to imageiiis modified fromeetoe∗e^\{\*\}, when asked identity\-related questions aboutee, the model surprisingly responds with the name ofe∗e^\{\*\}\. To illustrate this issue, consider the aforementioned case of rectifying the image\-entity association for Trump: as illustrated in Figure[1](https://arxiv.org/html/2605.06096#S1.F1)\(b\), when prompted with identity queries such as “Who is this?”, the edited model may indeed output “Trump,” and its performance might appear normal under existing benchmark metrics\. However, deeper probing reveals a behavior that even non\-experts would find absurd: when the model is asked text\-only questions aboutBiden\(the entity previously associated with the image before editing\), such as “What is the full name of Biden?”, the model unexpectedly answers “Trump” This is clearly highly anomalous\. We conducted a pilot study and consistently observed this pattern across various editing methods, indicating that such an issue is a systemic phenomenon rather than an isolated error\.

![Refer to caption](https://arxiv.org/html/2605.06096v1/x1.png)Figure 1:Overview of Entity Identity Confusion \(EIC\) in multimodal knowledge editing\.We further perform an in\-depth analysis of the characteristics of EIC\. Given that EIC is difficult to detect using standard metrics in traditional benchmarks, we construct a more comprehensive benchmark,EC\-Bench\. In addition to tasks specifically designed to examine EIC, EC\-Bench introduces two generalization tasks: Old Binding Persistence \(OBP\), and New Binding Generalization \(NBG\), to evaluate how the bindings between images and the original/new entities evolve after editing\. This allows us to analyze more characteristics of EIC and explore its underlying mechanisms\. Ideally, MKE should decouple imageiifrom the original entityeeand establish a new binding with entitye∗e^\{\*\}\. Our experimental analysis, however, reveals that existing MKE methods largely fail to affect the image\-entity binding; instead, the edited model still perceivesiias the original entityee\(e\.g\., Biden\) but uses thee∗e^\{\*\}label “Trump” to describeee’s identity, which explains the phenomena we observed\. Consequently, on more complex tasks such as asking “Which university did the person in the image graduate from?”, the model still provides the alma mater of Biden\. This suggests that even when the internal mechanism is fundamentally flawed, the model can still exhibit seemingly ideal behavior on simple tasks, thereby “deceiving” many existing benchmarks\.

What causes EIC? We posit that EIC stems from the fact that existing MKE methods fail to explicitly account for the complexity of different knowledge types in multimodal settings\. As shown in Figure[1](https://arxiv.org/html/2605.06096#S1.F1)\(c\), the objectives of current MKE methods typically only require the model to produce the correct string on given samples\(Huanget al\.,[2024](https://arxiv.org/html/2605.06096#bib.bib147)\), a superficial behavioral constraint: they achieve this through parameter updates and similar mechanisms, without any constraint of how it is internally realized\. However, knowledge in LVLMs involves two distinct categories\(Zhanget al\.,[2025a](https://arxiv.org/html/2605.06096#bib.bib152)\): Image\-Entity \(I\-E\) binding\(i,e\)\(i,e\)and Entity\-Entity \(E\-E\) relations\(e1,r,e2\)\(e\_\{1\},r,e\_\{2\}\), which may rely on different retrieval mechanisms at the model’s architecture levels\.This discrepancy means the model may in practice satisfy the editing objective through incorrect underlying mechanisms\.For instance, the model may implicitly force a spurious association between Biden and Trump – which yields correct answers on simple questions but is fundamentally incorrect at the underlying level, exposing issues like EIC under complex tests\.

We therefore advocate that a principled editing strategy should decouple two types of knowledge, ensuring that editing interventions precisely target I\-E binding representations while preserving the structural integrity of E\-E relational knowledge\. To provide methodological guidance for future research, we further we further explored and proposed a potential mitigation strategy for EIC: we propose that, since I\-E recall and E\-E recall occur at different locations during model inference,restricting the editing target to the region responsible for I\-E binding may help direct the editing effect toward the correct type of knowledge, thereby mitigating EIC and enabling more accurate knowledge editing\. We validate this hypothesis across multiple baseline methods by varying the editing location, and confirm that this constitutes a promising and robust direction for future research\. Furthermore, we discuss future directions for correct multimodal knowledge editing, thereby providing principled guidance for future MKE research\.

The core contributions of this paper are summarized as follows:

- •We identify and define Entity Identity Confusion \(EIC\) as an overlooked systematic failure mode in multimodal knowledge editing\.
- •We construct a diagnostic benchmark EC\-Bench and introduce more demanding generalization tasks to thoroughly assess the internal knowledge structure of the edited model, facilitating future in\-depth analysis of this issue\.
- •We conduct mechanistic diagnosis and analysis of MKE based on the benchmark, and propose a preliminary mitigation strategy, thereby providing methodological guidance for future multimodal editing research\.

## 2Preliminaries

This section provides definitions of key concepts and necessary backgrounds relevant to our work\.

### 2\.1Architecture of Large Vision\-Language Models

A typical large vision\-language model \(LVLM\)\(Liuet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib123); Zhuet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib140); Liet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib139)\)consists of three components: avision encoder, aprojector, and anLLM backbone\.

Given an input imageii, the vision encoder \(e\.g\., a Vision Transformer\) extracts a sequence of visual token embeddings𝐯=\[v1,…,vn\]\\mathbf\{v\}=\[v\_\{1\},\\dots,v\_\{n\}\]\. The projector \(e\.g\., a linear layer or MLP\) maps these tokens into the LLM’s embedding space, yielding𝐡=Proj​\(𝐯\)\\mathbf\{h\}=\\mathrm\{Proj\}\(\\mathbf\{v\}\)\. The LLM backbone then takes the concatenation of𝐡\\mathbf\{h\}and the text token embeddings as input and performs autoregressive generation to produce the output\.

### 2\.2Problem Formulation

Knowledge in LVLMscan be decomposed into two distinct types\(Zhanget al\.,[2025a](https://arxiv.org/html/2605.06096#bib.bib152)\)\.Image\-entity \(I\-E\) binding knowledge\(i,e\)\(i,e\)captures the correspondence between visual evidence and entity identity, answers “who or what does this image refer to?”Entity\-entity \(E\-E\) relational knowledge\(e1,r,e2\)\(e\_\{1\},r,e\_\{2\}\)captures facts and attributes connected to an entity through semantic relations, such as birthplace, occupation, or affiliation\. These two types may be handled by different components and layers of the model, a premise that motivates our analysis in later sections\.

Multimodal Knowledge Editing \(MKE\)aims to modify I\-E bindings: given an imageiioriginally bound to entityee, the goal is to rebind it to a target entitye∗e^\{\*\}\. Formally, letf​\(⋅;θ\)f\(\\cdot;\\theta\)denote a pretrained LVLM with parametersθ\\theta\. Given an imageiiand a textual queryxx, the model outputs an answery=f​\(i,x;θ\)y=f\(i,x;\\theta\)\. We are given an edit set

𝒟edit=\{\(i,x,y,y′\)\},\\mathcal\{D\}\_\{\\text\{edit\}\}=\\\{\(i,x,y,y^\{\\prime\}\)\\\},\(1\)wherexxis a query about the identity of the entity depicted inii,yyis the model\-consistent pre\-edit answer, andy′y^\{\\prime\}is the target answer expected after editing\.

An editing methodℳ\\mathcal\{M\}produces updated parametersθ′=ℳ​\(θ,𝒟edit\)\\theta^\{\\prime\}=\\mathcal\{M\}\(\\theta,\\mathcal\{D\}\_\{\\text\{edit\}\}\)\. The standard objective is

f​\(i,x;θ′\)=y′,f\(i,x;\\theta^\{\\prime\}\)=y^\{\\prime\},\(2\)while preserving unrelated model behavior\.

## 3Observing Entity Identity Confusion: A Preliminary Experiment

To empirically validate Entity Identity Confusion \(EIC\), we conduct a preliminary experiment\. In this section, We first detail the experimental setup, including the evaluation tasks we adopt\. Subsequently, based on the experimental results, we elaborate on the performance of EIC in downstream tasks and verify its prevalence across different basemodels and MKE methods\.

### 3\.1Preliminary Experiments Settings

Our preliminary experiments are based on a representative MKE Benchmark, VLKEB\(Huanget al\.,[2024](https://arxiv.org/html/2605.06096#bib.bib147)\), and extend its pipeline with additional evaluation tasks targeting EIC to observe the post\-edit behavior of models under various editing methods\. Descriptions of the baselines are provided in Appendix[D\.1](https://arxiv.org/html/2605.06096#A4.SS1)\.

Editing Task\.The editing objective of MKE is to modify an image\-entity binding within the model, i\.e\.,\(i,e\)→\(i,e′\)\(i,e\)\\rightarrow\(i,e^\{\\prime\}\)\. In practice, it provides a set of training samples containing images paired with questions querying the identity of the entity depicted; for example,\[Image of Biden\] What’s the full name of the person in this image?; and requires performing acounterfactualedit such that the model responds withDonald Trump\.

Evaluation Task\.To evaluate EIC, we query the identity of the original entityeein a pure text modality that contains no images, and examine the proportion of cases where the model erroneously predicts the label of the new entity,e∗e^\{\*\}, as the answer\. For example, we askWhat’s the full name of Biden?Models exhibiting EIC will anomalously respond withDonald Trump\. We also provide the efficacy metric, which is the classic edit success rate metric\.

### 3\.2Characteristics of EIC

We observe three recurring characteristics of EIC from the preliminary experiment\.

![Refer to caption](https://arxiv.org/html/2605.06096v1/x2.png)Figure 2:Performance of LLaVA edited with various MKE methods\.Characteristic 1: High Efficacy Coexists with High Confusion\.Across all editing methods, models achieve high edit success rates on the original edit queries while simultaneously exhibiting severe identity confusion\. This implies that single\-prompt efficacy is insufficient as a sole indicator of edit quality in LVLMs\.

Characteristic 2: Universality Across Editing Paradigms\.EIC is not confined to any single class of editing methods\. It manifests in parameter\-modifying approaches \(e\.g\., FT, MEND\), external\-memory\-based methods \(e\.g\., SERAC\), and prompt\-based strategies \(e\.g\., IKE\) alike\. While the severity differs across methods, the recurrence of this pattern across fundamentally different editing paradigms indicates that EIC is a structural issue inherent to the current MKE formulation\.

Characteristic 3: Text\-side Knowledge Contamination\.MKE targets the model’s I\-E binding, which should be image\-conditioned behavior that only manifests when image input is provided; however, we observe that the model also exhibits clearly anomalous behavioral patterns under text\-only queries, indicating that the editing has contaminated the model’s textual knowledge representations rather than acting precisely on the I\-E relationship\.

Conclusion\.Based on these observations, we provide a formal definition of the EIC phenomenon\. Given an editing instance that rebinds imageiifrom entityeeto target entitye∗e^\{\*\}, we define EIC as the phenomenon where the post\-edit modelf​\(⋅;θ′\)f\(\\cdot;\\theta^\{\\prime\}\), when queried about the identity ofeethrough a text\-only promptxtextx\_\{\\text\{text\}\}\(i\.e\., without any image input\), erroneously outputse∗e^\{\*\}:

EIC:f\(xtext\(e\);θ′\)=e∗,wheref\(xtext\(e\);θ\)=e\.\\text\{EIC\}:\\quad f\(x\_\{\\text\{text\}\}^\{\(e\)\};\\theta^\{\\prime\}\)=e^\{\*\},\\quad\\text\{where \}f\(x\_\{\\text\{text\}\}^\{\(e\)\};\\theta\)=e\.\(3\)In other words, the editing procedure intended to modify only the correspondence between images and entities, which is visual\-conditioned behavior, but causes the model to conflate the identities ofeeande∗e^\{\*\}even in the absence of any visual input\.

## 4Analyzing Post\-Edit Binding Behavior with EC\-Bench

To provide a more detailed analysis of how EIC manifests across different model architectures and editing methods, we introduceEC\-Bench\(Entity Confusion Benchmark\), an evaluation framework that extends standard MKE protocols\(Huanget al\.,[2024](https://arxiv.org/html/2605.06096#bib.bib147); Chenget al\.,[2023a](https://arxiv.org/html/2605.06096#bib.bib130)\)with dedicated diagnostics for identity corruption and binding inconsistency\. In this section, we first describe the tasks introduced by EC\-Bench, and then assess the performance of editing methods, accompanied by a diagnostic analysis of how internal knowledge associations are altered in post\-edit models\.

### 4\.1EC\-Bench

EC\-Bench consists of threefundamental tasksand threebinding diagnostic tasks\. The fundamental tasks align with conventional MKE benchmark settings and measure each method’s basic editing competency, coveringEfficacy, Generality, and Locality\. The binding diagnostic tasks are specifically designed to detect the EIC phenomenon and to analyze how internal knowledge associations are formed in edited models; to this end, we introduce three dedicated probes:Entity Identity Confusion \(EIC\), Old Binding Persistence \(OBP\), andNew Binding Generalization \(NBG\)\.

Fundamental Tasks\.Specifically, we introduce the following three fundamental tasks\.

- •Efficacymeasures whether the edited model returns target entitye∗e^\{\*\}on the original edit query\. This is the minimal criterion for successful intervention\.
- •Generalityevaluates whether edited behavior transfers to semantically equivalent variants\.*T\-Gen*uses paraphrased text prompts with the same image;*I\-Gen*uses alternative images of the same entity with the same query intent\. High generality indicates that the edit is not merely a string\-level patch to one prompt template\.
- •Localitymeasures whether unrelated knowledge remains stable\.*T\-Loc*compares pre\-/post\-edit answers on unrelated text\-only queries;*I\-Loc*compares pre\-/post\-edit behavior on visually similar but non\-target entities\.

Binding Diagnostic Tasks\.Consider the running example where an imageiiofBiden\(ee\) is edited to be rebound toTrump\(e∗e^\{\*\}\)\. If we use a multimodal knowledge graph\(Liuet al\.,[2019](https://arxiv.org/html/2605.06096#bib.bib112)\)to represent the underlying knowledge structure of the model, MKE is primarily concerned with three edges: \(1\) avoid introducing a spurious E\-E edge\(Biden,Trump\)\(\\text\{Biden\},\\text\{Trump\}\), \(2\) erase the old I\-E edge\(i,Biden\)\(i,\\text\{Biden\}\), and \(3\) establish the new I\-E edge\(i,Trump\)\(i,\\text\{Trump\}\)\. We introduce three binding diagnostic tasks to probe these three edges respectively, thereby characterizing how editing alters entity binding at a finer granularity\.

- •Entity Identity Confusion \(EIC\)probes edge \(1\): whether a spurious E\-E association\(e,e∗\)\(e,e^\{\*\}\)has been created\. After editing, we ask identity questions abouteewithout image input \(e\.g\.,What is the full name of Biden?\)\. If the model responds withe∗e^\{\*\}\(Trump\), we count it as confusion\.
- •Old Binding Persistence \(OBP\)probes edge \(2\): whether the old I\-E binding\(i,e\)\(i,e\)still survives after editing\. Note that directly asking “Who is in this image?” cannot reliably test this, because the spurious E\-E edge from EIC may redirect the answer toe∗e^\{\*\}even when the model still internally perceivesiiasee\. We therefore test the old binding*indirectly*via multi\-hop reasoning\(i→e,r,e1\)\(i\\rightarrow e,r,e\_\{1\}\): we present imageiiand ask relational facts unique toee\(e\.g\., “Which university did the person in this image graduate from?”\)\. Correct answers foreeindicate the old binding remains active\.
- •New Binding Generalization \(NBG\)probes edge \(3\): whether the new binding\(i,e∗\)\(i,e^\{\*\}\)supports factual reasoning beyond the edited prompt\. This task takes the form of a multi\-hop reasoning task consistent with OBP, but probes relations involving the new entity\(i→e∗,r,e2\)\(i\\rightarrow e^\{\*\},r,e\_\{2\}\): we present imageiiand query facts unique toe∗e^\{\*\}\(e\.g\., “In which city was the person in this image born?”\)\. Correct answers fore∗e^\{\*\}indicate that the model has formed a functional new grounding rather than merely memorizing one output string\.

### 4\.2Experiments and Findings

To conduct a thorough analysis of EIC, we employ six editing methods: FT\-Vis, FT\-LLM, KE, MEND, IKE, and SERAC \(Details in Appendix\.[D\.1](https://arxiv.org/html/2605.06096#A4.SS1)\), to edit LLaVA\-1\.5\(Liuet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib123)\), MiniGPT\-4\(Zhuet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib140)\), mPLUG\-Owl2\(Yeet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib122)\), and Qwen\-VL\(Baiet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib134)\), evaluating performance on EC\-Bench\. Detailed results are presented in Table[1](https://arxiv.org/html/2605.06096#S4.T1), while results for Owl2 are presented in Appendix[E\.1](https://arxiv.org/html/2605.06096#A5.SS1)\. Based on these results, we summarize our findings as follows:

Table 1:Main EC\-Bench results on inherited and diagnostic metrics\.Finding 1\. Nearly all editing methods exhibit severe EIC\.As shown in Table[1](https://arxiv.org/html/2605.06096#S4.T1), every method produces a significant and anomalous increase in EIC scores relative to the base model\. FT and MEND on LLaVA even reach a confusion rate approaching 99%, and the phenomenon is pervasive across different LLM backbones\. Such high rates reveal that existing methods cause severe contamination of textual\-modal knowledge when editing I\-E bindings: even under purely text\-based queries, the post\-edit model produces highly erroneous outputs with extremely high probability\. This clearly violates the expectations for knowledge editing in real\-world deployment\.

Finding 2\. Results on challenging tasks reveal that existing editing methods fail to achieve their underlying editing objectives\.A successful MKE intervention should dissolve the binding\(i,e\)\(i,e\)and establish a new\(i,e∗\)\(i,e^\{\*\}\)\. These two core objectives are measured by the OBP and NBG tasks, respectively\. However, as shown in Table[1](https://arxiv.org/html/2605.06096#S4.T1), performance on both metrics remains far from satisfactory: post\-edit models still retain very high OBP scores, with methods such as MEND and SERAC yielding values that remain close to those of the pre\-edit baseline; on the NBG task, the majority of models still score very low, indicating that it is extremely difficult for models to leverage the I\-E binding injected during editing for complex reasoning\. Overall, NBG scores are consistently and substantially lower than OBP scores, suggesting that the model’s internal processing pipeline still tends to first recognize the image as the original entity before performing downstream reasoning\.

Finding 3\. Methods that edit the visual side of models exhibit less EIC, though they still fall short on OBP and NBG\.Among the baseline methods compared in the main experiment, there is a category of approaches that perform editing on the visual side: FT\-Vis, which targets the vision encoder or projector module of LVLMs\. As shown in Table[1](https://arxiv.org/html/2605.06096#S4.T1), FT\-Vis achieves the best EIC scores among all compared methods, approaching the performance of the unedited base model, indicating that it barely contaminates the model’s purely text\-modal knowledge during the editing process\. We attribute this to the fact that E\-E type knowledge is necessarily encoded within the decoder of the LLM backbone; consequently, leaving this component unmodified naturally prevents overfitting to the editing objective through the contamination of E\-E knowledge\. Nevertheless, FT\-Vis still fails to achieve satisfactory performance on tasks such as OBP and NBG, and continues to exhibit deficiencies on basic metrics such as locality\.

Conclusion\.Taken together, EC\-Bench reveals that the apparent success of current MKE methods often conceals a inconsistent internal knowledge structure: \(1\) the original image\-to\-entity pathway\(i,e\)\(i,e\)remains active, \(2\) the new image\-to\-entity pathway\(i,e∗\)\(i,e^\{\*\}\)is weak and difficult to be leveraged for complex reasoning, and \(3\) an unintended entity\-level shortcut betweeneeande∗e^\{\*\}is introduced in the language space\. When querying the model, it still perceives the imageiias the original entityee, and then exploits the shortcut\(e,e∗\)\(e,e^\{\*\}\)to output the label ofe∗e^\{\*\}, thereby creating the illusion of a successful edit\.

## 5Mitigating Entity Identity Confusion: A Preliminary Exploration

The above analysis suggests that the lack of explicit distinction between I\-E and E\-E type knowledge in existing editing strategies likely leads models to incorrectly fit editing targets by forcibly altering E\-E associations, rather than modifying the intended I\-E binding relationships\. We therefore argue thata principled editing strategy should decouple these two types of knowledge, ensuring that editing interventions precisely target I\-E binding representations while preserving the structural integrity of E\-E associative knowledge\.

To address this issue, inspired by the observation that methods targeting visual modules exhibit significantly less severe EIC phenomenon, we hypothesize that controlling the location of the editing target module may serve as a minimalist yet effective mitigation strategy\. In this section, we aim to conduct a preliminary exploratory analysis of EIC mitigation strategies, thereby providing methodological guidance for future research\. We first introduce the theoretical foundations underlying the proposed mitigation strategy, then present empirical evidence of its effectiveness, and finally discuss the broader implications for future research directions\.

### 5\.1Background and Rationale: Knowledge Recall in LLMs

Two\-Stage Knowledge Recall in LLMs\.Recent interpretability research on both LLMs and LVLMs\(Gevaet al\.,[2021](https://arxiv.org/html/2605.06096#bib.bib90),[2023](https://arxiv.org/html/2605.06096#bib.bib78); Venhoffet al\.,[2025](https://arxiv.org/html/2605.06096#bib.bib148)\)has outlined a common two\-stage pipeline for knowledge recall\. As individual tokens carry only partial, locally\-scoped semantic content, attention modules in shallow layers first aggregate scattered token representations into a unified*entity representation*that encodes the entity identity referred to by the input; mid\-layer MLPs then inject relevant factual knowledge based on this representation, which is subsequently extracted in deeper layers for downstream reasoning\(Menget al\.,[2022](https://arxiv.org/html/2605.06096#bib.bib61); Gevaet al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib78); Yeet al\.,[2025](https://arxiv.org/html/2605.06096#bib.bib102)\)\. Specifically for LVLMs, visual tokens are first aggregated into a coherent entity representation—a process that corresponds precisely to the I\-E binding most central to MKE, in the shallow layers of the LVLM, before any relational knowledge can be retrieved\.

Implications for MKE\.This two\-stage structure has direct implications for knowledge editing\. Based on this, we posit thatif editing interventions are applied at layers*before*the entity representation is fully consolidated, the edit is more likely to target the I\-E binding pathway rather than disrupting downstream E\-E relational knowledge decoding\.Conversely, editing deeper layers – as most existing MKE methods do, likely perturbs relation decoding while leaving upstream binding intact, which is precisely the failure pattern we observe in EIC\. We therefore propose that controlling the editing location may be a potentially effective strategy for multimodal knowledge editing\.

### 5\.2Mitigating EIC via Editing\-Location Control

To validate this hypothesis, we use FT to edit different layers of LLaVA\-1\.5 and examine the EIC performance of the resulting edited models\. The specific results are shown in Figure[3](https://arxiv.org/html/2605.06096#S5.F3)\. We summarize our observations as follows:

![Refer to caption](https://arxiv.org/html/2605.06096v1/x3.png)Figure 3:Results for FT on LLaVA with different editing locations\.Obs1\. Editing Shallow LLM Layers Reduces EIC\.As shown in Figure[3](https://arxiv.org/html/2605.06096#S5.F3), the model’s EIC performance exhibits a strong correlation with the editing location: the severity of EIC increases monotonically as the edited layer moves deeper\. We observe that editing shallow LLM layers does not produce severe EIC, yielding levels close to those of FT\-Vis and the original model, suggesting that the shallow layers of the LLM backbone can still preserve textual entity identity\. In contrast, at deeper layers, EIC reaches extremely high levels approaching 100%, implying that the model may have completely overfit and lost its normal capacity for processing entity knowledge\. To be noted, editing layer 0 results in a slight drop in edit success rate, which may be attributed to the inevitable negative perturbation that fine\-tuning introduces to that layer’s parameters; and layer 0 is particularly critical as it directly processes the input\.

Table 2:Editing\-location comparison for FT and MEND\.Obs2\. The Shape of the Curve Corroborates the Entity Representation Solidification Hypothesis\.Another aspect of the experiment lies in the shape of the curve, which provides supporting evidence for our theoretical framework\. The severity of EIC does not increase linearly with editing depth: the curve remains relatively flat across the first few layers, its slope rises markedly upon entering the middle layers of the model, and becomes very steep in the deeper layers\. We posit that this abrupt transition point likely corresponds to the layer at which entity representations solidify: before this point, edits primarily act on the image\-to\-entity \(I\-E\) binding pathway; after this point, edits primarily disrupt downstream relation decoding, giving rise to the characteristic identity confusion of EIC\. Interestingly, the layers implicated by EIC closely align with those identified in prior mechanistic interpretability work\(Venhoffet al\.,[2025](https://arxiv.org/html/2605.06096#bib.bib148)\), further corroborating the consistency between our EIC framework and established mechanistic understanding of entity representation formation in transformer models\.

Generalization to Other Methods\.A natural question is whether this finding is specific to FT or generalizes across editing paradigms\. Since methods such as IKE and SERAC rely on external prompts and modules and do not involve editing specific layers, we select MEND – a representative of another parameter\-modification paradigm, and apply it to vision\-side modules as well as shallow LLM layers\. We observe the same mitigation effect: as shown in Table[2](https://arxiv.org/html/2605.06096#S5.T2), MEND with edits confined to shallow or vision\-side layers achieves a similarly significant reduction in EIC compared to the default deep\-layer configuration\. This suggests that shallow\-layer editing is a generalizable principle and can serve as a design reference for parameter\-modification\-based knowledge editing methods\. We also note that the improvements brought by shallow\-layer editing on the OBP and NBG tasks remain less pronounced than on EIC itself, which further reflects that multi\-hop reasoning in multimodal settings may be equally challenging as in text\-only settings and warrants further exploration in future work\.

### 5\.3Discussion & Implications for Future Research

In conclusion, our analysis suggests that faithful should distinguish between I\-E binding, which must be updated, and E\-E relational knowledge, which should remain intact\. Meanwhile, OBP and NBG tasks remain equally challenging to resolve; given that multi\-hop reasoning in the text\-only modality is still an open problem, how to achieve truly faithful multimodal editing warrants further exploration in future work\.

We frame our analysis of editing location as an exploration in this direction, and the robust reduction of EIC it yields indicates that editing location can still serve as a useful design principle for future MKE frameworks\. More broadly, we believe that effective MKE requires further attention to diagnostic evaluations beyond surface efficacy, and calls for better mechanisms that localize edits to the appropriate representational stages\.

## 6Related Work

Knowledge Editing in Large Language Models\.Knowledge editing aims to update model knowledge precisely and efficiently while preserving unrelated knowledge intact\(Zhanget al\.,[2024b](https://arxiv.org/html/2605.06096#bib.bib83); Wanget al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib137)\)\. Knowledge editing methods can be broadly categorized into two types\(Zhanget al\.,[2024b](https://arxiv.org/html/2605.06096#bib.bib83),[2025b](https://arxiv.org/html/2605.06096#bib.bib106),[2025c](https://arxiv.org/html/2605.06096#bib.bib96); Zhouet al\.,[2026](https://arxiv.org/html/2605.06096#bib.bib153)\):parameter\-modifyingmethods directly modify internal weights to enforce the injection of target factsl for example, FT directly fine\-tunes model parameters; KE\(De Caoet al\.,[2021](https://arxiv.org/html/2605.06096#bib.bib120)\)and MEND\(Mitchellet al\.,[2022b](https://arxiv.org/html/2605.06096#bib.bib116)\)train a hypernetwork to generate parameter updates; ROME\(Menget al\.,[2022](https://arxiv.org/html/2605.06096#bib.bib61)\), MEMIT\(Menget al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib79)\), and GLAME\(Zhanget al\.,[2024a](https://arxiv.org/html/2605.06096#bib.bib95)\)first locate knowledge storage positions before performing targeted updates\.Parameter\-preservingmethods rewrite model behavior through retrieval or external memory; IKE\(Zhenget al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib118)\)alters model outputs via in\-context learning, while memory\-based methods such as SERAC\(Mitchellet al\.,[2022a](https://arxiv.org/html/2605.06096#bib.bib117)\)modify model behavior through an additional memory module\.

Multimodal Knowledge Editing\.Recent research on multimodal editing has extended the knowledge editing paradigm to LVLMs, migrating a series of editing methods to LVLMs\(Huanget al\.,[2024](https://arxiv.org/html/2605.06096#bib.bib147); Panet al\.,[2024](https://arxiv.org/html/2605.06096#bib.bib149); Zenget al\.,[2025](https://arxiv.org/html/2605.06096#bib.bib150)\)and producing a range of benchmark works, such as the representative datasets MMEdit\(Chenget al\.,[2023b](https://arxiv.org/html/2605.06096#bib.bib119)\), MIKE\(Liet al\.,[2024](https://arxiv.org/html/2605.06096#bib.bib151)\), VLKEB\(Huanget al\.,[2024](https://arxiv.org/html/2605.06096#bib.bib147)\), and MC\-MKE\(Zhanget al\.,[2025a](https://arxiv.org/html/2605.06096#bib.bib152)\), forming an evaluation framework centered on efficacy, generalization, and locality\. However, current evaluations of MKE remain dominated by surface\-level efficacy on simple questions, with insufficient analysis of the specific behavioral patterns of edited models\. This allows many methods with underlying issues to still achieve favorable results\. Our work provides a valuable complement to this line of research and reveals that high efficacy scores may conceal severe internal knowledge inconsistency\.

## 7Conclusion

In this work, we identified and characterized Entity Identity Confusion \(EIC\), a systemic yet previously overlooked failure mode in multimodal knowledge editing that existing benchmarks largely fail to detect\. We demonstrated that EIC stems from the failure of current MKE methods to distinguish between I\-E and E\-E knowledge, leading models to overfit E\-E associations as a shortcut rather than the underlying I\-E binding\. To rigorously diagnose this phenomenon, we introduced EC\-Bench, a benchmark featuring challenging tasks that expose EIC where standard evaluations cannot\.

Building on our mechanistic analysis, we identified constraining edits to early\-stage representations as a promising mitigation direction, and discussed the principled desiderata that a faithful MKE method should satisfy\. We hope the problem formulation, benchmark, and insights presented here provide a useful foundation for future research toward more faithful and robust multimodal knowledge editing\.

## References

- J\. Bai, S\. Bai, S\. Yang, S\. Wang, S\. Tan, P\. Wang, J\. Lin, C\. Zhou, and J\. Zhou \(2023\)Qwen\-vl: a versatile vision\-language model for understanding, localization, text reading, and beyond\.arXiv preprint arXiv:2308\.12966\.Cited by:[§1](https://arxiv.org/html/2605.06096#S1.p1.1),[§4\.2](https://arxiv.org/html/2605.06096#S4.SS2.p1.1)\.
- S\. Cheng, B\. Tian, Q\. Liu, X\. Chen, Y\. Wang, H\. Chen, and N\. Zhang \(2023a\)Can we edit multimodal large language models?\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 13877–13888\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.856),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.856)Cited by:[§1](https://arxiv.org/html/2605.06096#S1.p1.1),[§4](https://arxiv.org/html/2605.06096#S4.p1.1)\.
- S\. Cheng, B\. Tian, Q\. Liu, X\. Chen, Y\. Wang, H\. Chen, and N\. Zhang \(2023b\)Can we edit multimodal large language models?\.arXiv preprint arXiv:2310\.08475\.Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p2.1)\.
- N\. De Cao, W\. Aziz, and I\. Titov \(2021\)Editing factual knowledge in language models\.InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,M\. Moens, X\. Huang, L\. Specia, and S\. W\. Yih \(Eds\.\),Online and Punta Cana, Dominican Republic,pp\. 6491–6506\.External Links:[Link](https://aclanthology.org/2021.emnlp-main.522),[Document](https://dx.doi.org/10.18653/v1/2021.emnlp-main.522)Cited by:[2nd item](https://arxiv.org/html/2605.06096#A4.I1.i2.p1.1),[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- M\. Geva, J\. Bastings, K\. Filippova, and A\. Globerson \(2023\)Dissecting recall of factual associations in auto\-regressive language models\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 12216–12235\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.751/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.751)Cited by:[§5\.1](https://arxiv.org/html/2605.06096#S5.SS1.p1.1)\.
- M\. Geva, R\. Schuster, J\. Berant, and O\. Levy \(2021\)Transformer feed\-forward layers are key\-value memories\.InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,M\. Moens, X\. Huang, L\. Specia, and S\. W\. Yih \(Eds\.\),Online and Punta Cana, Dominican Republic,pp\. 5484–5495\.External Links:[Link](https://aclanthology.org/2021.emnlp-main.446/),[Document](https://dx.doi.org/10.18653/v1/2021.emnlp-main.446)Cited by:[§5\.1](https://arxiv.org/html/2605.06096#S5.SS1.p1.1)\.
- H\. Huang, H\. Zhong, T\. Yu, Q\. Liu, S\. Wu, L\. Wang, and T\. Tan \(2024\)VLKEB: a large vision\-language model knowledge editing benchmark\.InAdvances in Neural Information Processing Systems,A\. Globerson, L\. Mackey, D\. Belgrave, A\. Fan, U\. Paquet, J\. Tomczak, and C\. Zhang \(Eds\.\),Vol\.37,pp\. 9257–9280\.External Links:[Document](https://dx.doi.org/10.52202/079017-0294),[Link](https://proceedings.neurips.cc/paper_files/paper/2024/file/1198b53fa686831d5f0c0860d7ec4f34-Paper-Datasets_and_Benchmarks_Track.pdf)Cited by:[§C\.2](https://arxiv.org/html/2605.06096#A3.SS2.p1.1),[Appendix C](https://arxiv.org/html/2605.06096#A3.p2.1),[Appendix D](https://arxiv.org/html/2605.06096#A4.p1.1),[§1](https://arxiv.org/html/2605.06096#S1.p5.2),[§3\.1](https://arxiv.org/html/2605.06096#S3.SS1.p1.1),[§4](https://arxiv.org/html/2605.06096#S4.p1.1),[§6](https://arxiv.org/html/2605.06096#S6.p2.1)\.
- J\. Li, M\. Du, C\. Zhang, Y\. Chen, N\. Hu, G\. Qi, H\. Jiang, S\. Cheng, and B\. Tian \(2024\)MIKE: a new benchmark for fine\-grained multimodal entity knowledge editing\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 5018–5029\.External Links:[Link](https://aclanthology.org/2024.findings-acl.298/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.298)Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p2.1)\.
- J\. Li, D\. Li, S\. Savarese, and S\. Hoi \(2023\)BLIP\-2: bootstrapping language\-image pre\-training with frozen image encoders and large language models\.InICML,Cited by:[§2\.1](https://arxiv.org/html/2605.06096#S2.SS1.p1.1)\.
- H\. Liu, C\. Li, Q\. Wu, and Y\. J\. Lee \(2023\)Visual instruction tuning\.arXiv preprint arXiv:2304\.08485\.External Links:[Link](https://arxiv.org/abs/2304.08485)Cited by:[§1](https://arxiv.org/html/2605.06096#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.06096#S2.SS1.p1.1),[§4\.2](https://arxiv.org/html/2605.06096#S4.SS2.p1.1)\.
- Y\. Liu, H\. Li, A\. Garcia\-Duran, M\. Niepert, D\. Onoro\-Rubio, and D\. S\. Rosenblum \(2019\)MMKG: multi\-modal knowledge graphs\.InEuropean Semantic Web Conference,pp\. 459–474\.Cited by:[§4\.1](https://arxiv.org/html/2605.06096#S4.SS1.p2.6)\.
- K\. Meng, D\. Bau, A\. Andonian, and Y\. Belinkov \(2022\)Locating and editing factual associations in gpt\.Advances in Neural Information Processing Systems35,pp\. 17359–17372\.Cited by:[§1](https://arxiv.org/html/2605.06096#S1.p2.1),[§5\.1](https://arxiv.org/html/2605.06096#S5.SS1.p1.1),[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- K\. Meng, A\. S\. Sharma, A\. J\. Andonian, Y\. Belinkov, and D\. Bau \(2023\)Mass\-editing memory in a transformer\.InThe Eleventh International Conference on Learning Representations,Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- E\. Mitchell, C\. Lin, A\. Bosselut, C\. Finn, and C\. D\. Manning \(2022a\)Memory\-based model editing at scale\.InInternational Conference on Machine Learning,External Links:[Link](https://arxiv.org/pdf/2206.06520.pdf)Cited by:[5th item](https://arxiv.org/html/2605.06096#A4.I1.i5.p1.1),[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- E\. Mitchell, C\. Lin, A\. Bosselut, C\. Finn, and C\. D\. Manning \(2022b\)Fast model editing at scale\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/pdf?id=0DcZxeWfOPt)Cited by:[3rd item](https://arxiv.org/html/2605.06096#A4.I1.i3.p1.1),[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- K\. Pan, Z\. Fan, J\. Li, Q\. Yu, H\. Fei, S\. Tang, R\. Hong, H\. Zhang, and Q\. Sun \(2024\)Towards unified multimodal editing with enhanced knowledge collaboration\.InAdvances in Neural Information Processing Systems,A\. Globerson, L\. Mackey, D\. Belgrave, A\. Fan, U\. Paquet, J\. Tomczak, and C\. Zhang \(Eds\.\),Vol\.37,pp\. 110290–110314\.External Links:[Document](https://dx.doi.org/10.52202/079017-3500),[Link](https://proceedings.neurips.cc/paper_files/paper/2024/file/c705ba25f183b875c9359ef83fa262e8-Paper-Conference.pdf)Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p2.1)\.
- C\. Venhoff, A\. Khakzar, S\. Joseph, P\. Torr, and N\. Nanda \(2025\)Too late to recall: explaining the two\-hop problem in multimodal knowledge retrieval\.External Links:2512\.03276,[Link](https://arxiv.org/abs/2512.03276)Cited by:[§5\.1](https://arxiv.org/html/2605.06096#S5.SS1.p1.1),[§5\.2](https://arxiv.org/html/2605.06096#S5.SS2.p3.1)\.
- P\. Wang, N\. Zhang, X\. Xie, Y\. Yao, B\. Tian, M\. Wang, Z\. Xi, S\. Cheng, K\. Liu, G\. Zheng,et al\.\(2023\)Easyedit: an easy\-to\-use knowledge editing framework for large language models\.arXiv preprint arXiv:2308\.07269\.Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- Q\. Ye, H\. Xu, J\. Ye, M\. Yan, A\. Hu, H\. Liu, Q\. Qian, J\. Zhang, F\. Huang, and J\. Zhou \(2023\)MPLUG\-owl2: revolutionizing multi\-modal large language model with modality collaboration\.External Links:2311\.04257Cited by:[§4\.2](https://arxiv.org/html/2605.06096#S4.SS2.p1.1)\.
- X\. Ye, M\. Zhang, and S\. Wu \(2025\)LLM unlearning should be form\-independent\.External Links:2506\.07795,[Link](https://arxiv.org/abs/2506.07795)Cited by:[§5\.1](https://arxiv.org/html/2605.06096#S5.SS1.p1.1)\.
- Z\. Zeng, L\. Gu, X\. Yang, Z\. Duan, Z\. Shi, and M\. Wang \(2025\)Visual\-oriented fine\-grained knowledge editing for multimodal large language models\.InProceedings of the IEEE/CVF International Conference on Computer Vision \(ICCV\),pp\. 2491–2500\.Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p2.1)\.
- C\. Zhang, M\. Zhang, X\. Ye, R\. Cheng, Z\. Zhou, Y\. Zhou, P\. Ren, and Z\. Chen \(2026\)Spectral characterization and mitigation of sequential knowledge editing collapse\.External Links:2601\.11042,[Link](https://arxiv.org/abs/2601.11042)Cited by:[§1](https://arxiv.org/html/2605.06096#S1.p2.1)\.
- J\. Zhang, H\. Zhang, X\. Yin, B\. Huang, X\. Zhang, X\. Hu, and X\. Wan \(2025a\)MC\-MKE: a fine\-grained multimodal knowledge editing benchmark emphasizing modality consistency\.InFindings of the Association for Computational Linguistics: ACL 2025,W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 17430–17445\.External Links:[Link](https://aclanthology.org/2025.findings-acl.896/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-acl.896),ISBN 979\-8\-89176\-256\-5Cited by:[§1](https://arxiv.org/html/2605.06096#S1.p5.2),[§2\.2](https://arxiv.org/html/2605.06096#S2.SS2.p1.2),[§6](https://arxiv.org/html/2605.06096#S6.p2.1)\.
- M\. Zhang, B\. Fang, Q\. Liu, X\. Ye, S\. Wu, P\. Ren, Z\. Chen, and L\. Wang \(2025b\)KELE: residual knowledge erasure for enhanced multi\-hop reasoning in knowledge editing\.InFindings of the Association for Computational Linguistics: EMNLP 2025,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 24537–24552\.External Links:[Link](https://aclanthology.org/2025.findings-emnlp.1334/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.1334),ISBN 979\-8\-89176\-335\-7Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- M\. Zhang, X\. Ye, Q\. Liu, P\. Ren, S\. Wu, and Z\. Chen \(2024a\)Knowledge graph enhanced large language model editing\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 22647–22662\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.1261/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.1261)Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- M\. Zhang, X\. Ye, Q\. Liu, S\. Wu, P\. Ren, and Z\. Chen \(2025c\)Uncovering overfitting in large language model editing\.InThe Thirteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=t8qcGXaepr)Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- N\. Zhang, Y\. Yao, B\. Tian, P\. Wang, S\. Deng, M\. Wang, Z\. Xi, S\. Mao, J\. Zhang, Y\. Ni, S\. Cheng, Z\. Xu, X\. Xu, J\. Gu, Y\. Jiang, P\. Xie, F\. Huang, L\. Liang, Z\. Zhang, X\. Zhu, J\. Zhou, and H\. Chen \(2024b\)A comprehensive study of knowledge editing for large language models\.External Links:2401\.01286,[Link](https://arxiv.org/abs/2401.01286)Cited by:[§1](https://arxiv.org/html/2605.06096#S1.p1.1),[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- W\. X\. Zhao, K\. Zhou, J\. Li, T\. Tang, X\. Wang, Y\. Hou, Y\. Min, B\. Zhang,et al\.\(2025\)A survey of large language models\.External Links:2303\.18223,[Link](https://arxiv.org/abs/2303.18223)Cited by:[§1](https://arxiv.org/html/2605.06096#S1.p1.1)\.
- C\. Zheng, L\. Li, Q\. Dong, Y\. Fan, Z\. Wu, J\. Xu, and B\. Chang \(2023\)Can we edit factual knowledge by in\-context learning?\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 4862–4876\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.296),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.296)Cited by:[4th item](https://arxiv.org/html/2605.06096#A4.I1.i4.p1.1),[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- Z\. Zhou, M\. Zhang, S\. Wu, X\. Ye, C\. Zhang, Z\. Chen, and P\. Ren \(2026\)Uncovering context reliance in unstructured knowledge editing\.External Links:2602\.19043,[Link](https://arxiv.org/abs/2602.19043)Cited by:[§6](https://arxiv.org/html/2605.06096#S6.p1.1)\.
- D\. Zhu, J\. Chen, X\. Shen, X\. Li, and M\. Elhoseiny \(2023\)MiniGPT\-4: enhancing vision\-language understanding with advanced large language models\.arXiv preprint arXiv:2304\.10592\.Cited by:[§1](https://arxiv.org/html/2605.06096#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.06096#S2.SS1.p1.1),[§4\.2](https://arxiv.org/html/2605.06096#S4.SS2.p1.1)\.

## Appendix ALimitations

EC\-Bench follows the classical MKE setting and primarily focuses on the binding relationship between images and entities\. Accordingly, the scope of this paper is largely confined to analyzing the behavioral patterns of edited models under this type of multimodal editing\. However, real\-world multimodal scenarios may involve more complex tasks, including a richer variety of entity categories, image types, and knowledge beyond entity\-level understanding, such as knowledge of image style\. Future work may extend EC\-Bench to broader real\-world image distributions, more diverse entity types, and explore knowledge categories beyond entity knowledge\.

## Appendix BImpact Statement

This paper presents work whose goal is to advance the field of multimodal knowledge editing for large vision\-language models\. By identifying and formalizing Entity Identity Confusion \(EIC\) as a systemic failure mode, introducing EC\-Bench as a diagnostic benchmark, and proposing editing\-location control as a principled mitigation strategy, our work improves the transparency and reliability of multimodal knowledge editing\. These contributions help ensure that knowledge updates in deployed LVLMs are faithful and consistent, rather than producing superficially correct yet internally corrupted behavior\.

While multimodal knowledge editing has broad positive applications, including correcting outdated information, enforcing safety policies, and enabling continual model maintenance, we acknowledge potential ethical considerations\. The ability to alter a model’s internal knowledge bindings could be misused to inject biased or misleading associations between visual content and entity identities\. We encourage future work to develop safeguards against such misuse and to ensure that multimodal knowledge editing techniques are deployed in alignment with ethical AI principles\.

## Appendix CDetails on EC\-Bench Benchmark

EC\-Bench is a benchmark designed for evaluate the effectiveness of multimodal knowledge editing methods for LVLMs\. Its goal is to detect EIC and evaluate how the bindings between images and the original/new entities evolve after editing, thereby allows us to analyze the characteristics of EIC and analyze its underlying mechanisms\.

To be noted, part of the EC\-Bench dataset are sourced from existing open\-source MKE benchmarks\. We have specifically restructured and extended the VLKEB\[Huanget al\.,[2024](https://arxiv.org/html/2605.06096#bib.bib147)\]dataset, constructing new data and added novel tasks, while building upon its original baselines and hyperparameter settings to enable a more comprehensive and in\-depth evaluation of the EIC issue we investigate\.

### C\.1Dataset Composition

At the benchmark level, the base data unit is a counterfactual image–entity edit tuple

\(i,e\)→\(i,e∗\),\(i,e\)\\rightarrow\(i,e^\{\*\}\),\(4\)whereiiis an image originally associated with entityee, and the edited model is expected to recognize the same image under the target entitye∗e^\{\*\}\.

For each edit tuple, EC\-Bench instantiates a set of fundamental tasks, including the original edit query, a text\-side rephrasing, an image\-side rephrasing, and locality examples\. In addition to these standard evaluation dimensions, EC\-Bench includes three Binding Diagnostic Tasks\. EIC is the central diagnostic in EC\-Bench and measures whether the edit creates language\-side entity confusion betweeneeande∗e^\{\*\}; OBP measures whether the original image–entity binding remains active after editing; and NBG evaluates whether the edited image supports target\-side relational reasoning aboute∗e^\{\*\}\. The scale of the current evaluation split used in our experiments is summarized in Table[3](https://arxiv.org/html/2605.06096#A3.T3)\.

Table 3:Scale summary of the current EC\-Bench evaluation split used in our experiments\.
### C\.2Dataset Construction Details

The edit tuples and fundamental task data in EC\-Bench are sourced from the established benchmark VLKEB\[Huanget al\.,[2024](https://arxiv.org/html/2605.06096#bib.bib147)\], specifically leveraging a subset of their test splits\. We further constructing new data and added novel binding diagnostic tasks for assess post\-edit multi\-modal binding behavior\.

As a concrete running example, consider an edit case in which the original image is associated with*Jack Webb*and the target entity is*Joseph Cotten*\. The corresponding edit tuple is

\(i,Jack Webb\)→\(i,Joseph Cotten\)\.\(i,\\text\{Jack Webb\}\)\\rightarrow\(i,\\text\{Joseph Cotten\}\)\.\(5\)The examples below instantiate EIC, OBP, and NBG around this same edit tuple before presenting additional real evaluation cases\.

##### EIC construction\.

The goal of EIC is to test whether the edited model incorrectly connects the original entityeeto the target entitye∗e^\{\*\}on the language side\. The construction starts from the image\-conditioned entity\-identification question already present in Efficacy Task\. Let this source question be denoted byqimg​\(i\)q\_\{\\mathrm\{img\}\}\(i\), whose answer in the original data is the entityee\. We convert it into a text\-only question by replacing the image referent in the question with the exact entity name, while keeping the remaining semantics unchanged\. The resulting EIC question is a text\-only question aboutee:

qEIC​\(e\)=ℛ​\(e,qimg\),q\_\{\\mathrm\{EIC\}\}\(e\)=\\mathcal\{R\}\(e,q\_\{\\mathrm\{img\}\}\),\(6\)whereℛ\\mathcal\{R\}denotes the rewrite operation performed by an external LLM\. In the Jack Webb example, the source image\-conditioned question is

qimg​\(i\)=“Who is the actor featured in this image?”q\_\{\\mathrm\{img\}\}\(i\)=\\text\{\`\`Who is the actor featured in this image?''\}\(7\)and the rewritten EIC question becomes

qEIC​\(e\)=“Who is the actor Jack Webb?”q\_\{\\mathrm\{EIC\}\}\(e\)=\\text\{\`\`Who is the actor Jack Webb?''\}\(8\)with target answer*Joseph Cotten*\. If the edited model answers the question about Jack Webb with Joseph Cotten, the edit has introduced entity confusion\.

We useDeepSeek\-Chatas the rewriting model\. The prompt is designed to preserve the semantic structure of the original image\-conditioned question while removing image dependence and directly inserting the original entity name\. The full prompt used in construction is shown below\.

Prompt Block A1: EIC question rewritingModel\.DeepSeek\-ChatSystem prompt\.You are a powerful question\-and\-answer generator\. The user provides a question about a specific entity \(person, location, sign, scene, poster, logo, sight, occupation, etc\.\) in an image, where the answer to the question is the entity itself\. Your task is to rewrite the question by replacing any reference to “the image” with the exact entity provided, while keeping the rest of the question unchanged and ensuring linguistic and logical coherence\.Output format: Directly output the rewritten question only\. Do not include any additional information\.Examples:User Input: Entity: Denton, Texas∣\\midQuestion: What city of Texas is depicted in the image?Output: What city of Texas is Denton, Texas?User Input: Entity: Lily Tomlin∣\\midQuestion: Who is the actor featured in this image?Output: Who is the actor Lily Tomlin?User Input: Entity: Carl Stalling∣\\midQuestion: What is the full name of the musician depicted in the image?Output: What is the full name of the musician Carl Stalling?User Input: Entity: A\.C\. Milan∣\\midQuestion: Which football club is represented in the image?Output: Which football club is A\.C\. Milan?User template\.Entity: \{entity\}∣\\midQuestion: \{question\}

##### OBP construction\.

The goal of OBP is to test whether the old I\-E binding\(i,e\)\(i,e\)still survives after editing\. To make this test discriminative, we construct a relation\-controlled A/B question from an entity pair in which the original entity and the target entity correspond to different answers under the same relation:

\(e,r,o\),\(e∗,r,o∗\),o≠o∗\.\(e,r,o\),\\qquad\(e^\{\*\},r,o^\{\*\}\),\\qquad o\\neq o^\{\*\}\.\(9\)This condition ensures that the question can distinguish whether the model is still following the original image\-entity pathway or has shifted away from it\. In this construction, the option associated with the original entity is always assigned toA, and the option associated with the target entity is assigned toB\. The OBP question is then obtained by instantiating the resulting template with an image referent phrase, yielding an image\-conditioned multiple\-choice question whose old\-binding answer remainsAif the original binding is still active\.

In the Jack Webb example, the generated relation\-controlled question tests a role associated with the two entities\. The resulting OBP question is:

> the man in this image is best known for playing the role of A\. joe friday B\. holly martins Answer with one letter only \(A or B\):

The old\-binding answer isA\. If the edited model still prefersAafter editing, the original image–entity binding remains active\.

We again useDeepSeek\-Chatto construct these questions\. The model receives the original entity and the target entity and is instructed to produce a question template, one answer option for the original entity, one answer option for the target entity, and a strong one\-letter answer cue\. The prompt explicitly constrains the generated question to be simple, factual, short, and non\-ambiguous\. The full construction prompt is shown below\.

Prompt Block A2: OBP question generationModel\.DeepSeek\-ChatGeneration setting\.Temperature = 0\.25System prompt\.You are a precision multiple\-choice question designer for multimodal knowledge\-editing datasets\. Given a pred entity and an alt entity, design a very simple question whose answer is a single letter: A or B\.Goals:1\.Choose a very simple property, relation, or description of the pred entity that makes the correct option highly predictable\.2\.The question must strongly guide the model to output exactly one letter: A or B\.3\.The question must end with a strong answer cue that explicitly requests one letter only\.4\.Option A must be the correct completion for the pred entity\.5\.Option B must be the correct completion for the alt entity\.6\.Keep both options short, factual, and easy to read\.7\.Use the exact four labels below\.8\.Output exactly four lines:•question\_template: <sentence with \_\_SUBJECT\_\_ placeholder and no options\>•answer\_a: <pred answer option\>•answer\_b: <alt answer option\>•question\_suffix: <strong final cue, for example ‘‘Answer with one letter only \(A or B\):’’\>Rules:1\.Do not use yes/no wording\.2\.Do not output explanations\.3\.Do not make the question too hard or too specific\.4\.Do not use punctuation that makes the answer ambiguous\.5\.Keep option text lowercase if possible\.6\.Do not include the options inside the template line\.Example 1:Input:pred entity: Monte Carloalt entity: MarrakeshOutput:question\_template: \_\_SUBJECT\_\_ is a district of the principality ofanswer\_a: monacoanswer\_b: moroccoquestion\_suffix: Answer with one letter only \(A or B\):Example 2:Input:pred entity: Shine \(film\)alt entity: Please GiveOutput:question\_template: \_\_SUBJECT\_\_ is directed byanswer\_a: scott hicksanswer\_b: kirsten johnsonquestion\_suffix: Answer with one letter only \(A or B\):User template\.pred entity: \{pred\_entity\}alt entity: \{alt\_entity\}

The output of this prompt is a template with a subject placeholder\_\_SUBJECT\_\_, together with two short answer options\. The OBP question is constructed by replacing the subject placeholder with an image referent phrase such as “the man in this image” or “the film in this image”\.

##### NBG construction\.

The goal of NBG is to test whether the new binding\(i,e∗\)\(i,e^\{\*\}\)supports factual reasoning beyond the edited prompt\. In practice, we use the portability tasks from the original VLKEB dataset as the data source for this task after filtering and simple processing, as both tasks involve multi\-hop reasoning and essentially probe the same content\. Each NBG example is an image\-conditioned open\-ended question whose answer is a factual attribute or relation associated withe∗e^\{\*\}\.

### C\.3Examples of Dataset Entries

The following examples are drawn from the EC\-Bench Dataset\.

Example 1Edit queryWho is the actor featured in this image?Text generalization queryWhich actor is shown in this picture?Original entityJack WebbTarget entityJoseph CottenEIC questionWho is the actor Jack Webb?EIC target answerJoseph CottenImage locality questionWhat educational institution is represented in the image?Image locality answerUniversity of FloridaNBG question1\-hop: Where did the person associated with the individual in the picture pass away?NBG target answerLos AngelesOBP questionthe man in this image is best known for playing the role ofA\. joe fridayB\. holly martinsAnswer with one letter only \(A or B\):OBP old\-binding answerA

Example 2Edit queryWhat is the title of the movie in this scene?Text generalization queryWhich film does this image represent?Original entityThe Full MontyTarget entityHustle & FlowEIC questionWhat is the title of the movie The Full Monty?EIC target answerHustle & FlowImage locality questionWhat is the title of the film represented in the image?Image locality answerThe Ladykillers \(2004 film\)NBG question1\-hop: Which company distributed the film represented in the picture?NBG target answerParamount VantageOBP questionthe film in this image is a film about a group ofA\. unemployed steelworkersB\. aspiring rappersAnswer with one letter only \(A or B\):OBP old\-binding answerA

Example 3Edit queryWhat Irish county is depicted in the image?Text generalization queryWhich Irish county is seen in the picture?Original entityCounty TipperaryTarget entityCounty WaterfordEIC questionWhat Irish county is County Tipperary?EIC target answerCounty WaterfordImage locality questionWhat city in Ireland is depicted in the image?Image locality answerWaterfordNBG question1\-hop: What types of political leadership can be found in the area neighboring the location shown in the picture?NBG target answerFine Gael, Independent politician, Labour Party \(Ireland\), Social democracyOBP questionthe place in this image is located in the province ofA\. munsterB\. leinsterAnswer with one letter only \(A or B\):OBP old\-binding answerA

Example 4Edit queryWhat band is represented in this image?Text generalization queryWhich musical group is depicted in the picture?Original entityBee GeesTarget entityThe ClashEIC questionWhat band is Bee Gees?EIC target answerThe ClashImage locality questionWho is the person present in this image?Image locality answerAlan ParsonsNBG question1\-hop: What music genre is associated with the group shown in the picture?NBG target answerPunk rockOBP questionthe group in this image is a music group fromA\. australiaB\. englandAnswer with one letter only \(A or B\):OBP old\-binding answerA

Example 5Edit queryWhat country’s flag is seen in the image?Text generalization queryWhich Asian country is represented in the picture?Original entityBelgiumTarget entityJapanEIC questionWhat country’s flag is Belgium?EIC target answerJapanImage locality questionWhich noble house is depicted in the image?Image locality answerHouse of OldenburgNBG question1\-hop: What is the capital city of the country associated with the item in the picture?NBG target answerTokyoOBP questionthe country in this image is a country inA\. europeB\. asiaAnswer with one letter only \(A or B\):OBP old\-binding answerA

### C\.4Details on Metrics

In this subsection, we detail the computation rules for EC\-Bench metrics\.

We compute EC\-Bench metrics with a common token\-level scoring rule whenever a task has a prompt and a reference answer\. Letτ\\taudenote an evaluation task and let thejj\-th evaluation sample in EC\-Bench be:

ujτ=\(xjτ,vjτ\),𝐲jτ=\(yj,1τ,…,yj,Ljτ\),u\_\{j\}^\{\\tau\}=\(x\_\{j\}^\{\\tau\},v\_\{j\}^\{\\tau\}\),\\qquad\\mathbf\{y\}\_\{j\}^\{\\tau\}=\(y\_\{j,1\}^\{\\tau\},\\ldots,y\_\{j,L\_\{j\}\}^\{\\tau\}\),\(10\)wherexjτx\_\{j\}^\{\\tau\}is the text prompt,vjτv\_\{j\}^\{\\tau\}is the optional image input, and𝐲jτ\\mathbf\{y\}\_\{j\}^\{\\tau\}is the non\-padding token sequence of the reference answer\. We uses∈\{pre,post\}s\\in\\\{\\mathrm\{pre\},\\mathrm\{post\}\\\}to denote the model state before and after editing\. Given statess, the modelf\(s\)f^\{\(s\)\}generates a next token distribution

pj,t\(s\)​\(w\)=pf\(s\)​\(w∣ujτ,yj,<tτ\)\.p\_\{j,t\}^\{\(s\)\}\(w\)=p\_\{f^\{\(s\)\}\}\\\!\\left\(w\\mid u\_\{j\}^\{\\tau\},y\_\{j,<t\}^\{\\tau\}\\right\)\.\(11\)The corresponding predicted token is

y^j,t\(s\)=arg⁡maxw⁡pj,t\(s\)​\(w\)\.\\hat\{y\}\_\{j,t\}^\{\(s\)\}=\\arg\\max\_\{w\}p\_\{j,t\}^\{\(s\)\}\(w\)\.\(12\)The accuracy for thejj\-th sample and its probability score are then computed as

accj\(s\)=1Lj​∑t=1Lj𝟏​\[y^j,t\(s\)=yj,tτ\],probj\(s\)=1Lj​∑t=1Ljpj,t\(s\)​\(yj,tτ\)\.\\operatorname\{acc\}\_\{j\}^\{\(s\)\}=\\frac\{1\}\{L\_\{j\}\}\\sum\_\{t=1\}^\{L\_\{j\}\}\\mathbf\{1\}\\\!\\left\[\\hat\{y\}\_\{j,t\}^\{\(s\)\}=y\_\{j,t\}^\{\\tau\}\\right\],\\qquad\\operatorname\{prob\}\_\{j\}^\{\(s\)\}=\\frac\{1\}\{L\_\{j\}\}\\sum\_\{t=1\}^\{L\_\{j\}\}p\_\{j,t\}^\{\(s\)\}\\\!\\left\(y\_\{j,t\}^\{\\tau\}\\right\)\.\(13\)The reported task\-level scores are averages over the valid evaluation samples:

Accτ\(s\)=1Nτ​∑j=1Nτaccj\(s\),Probτ\(s\)=1Nτ​∑j=1Nτprobj\(s\)\.\\operatorname\{Acc\}\_\{\\tau\}^\{\(s\)\}=\\frac\{1\}\{N\_\{\\tau\}\}\\sum\_\{j=1\}^\{N\_\{\\tau\}\}\\operatorname\{acc\}\_\{j\}^\{\(s\)\},\\qquad\\operatorname\{Prob\}\_\{\\tau\}^\{\(s\)\}=\\frac\{1\}\{N\_\{\\tau\}\}\\sum\_\{j=1\}^\{N\_\{\\tau\}\}\\operatorname\{prob\}\_\{j\}^\{\(s\)\}\.\(14\)Thus, the probability score is an average token probability, rather than the product probability of the whole answer sequence\. In the main result tables, edited rows report post\-edit scores unless explicitly stated otherwise, while thebase \(unedited\)rows report pre\-edit scores\.

In practice, the reference answer tokens depend on the task\. For efficacy, text generalization, and image generalization,𝐲jτ\\mathbf\{y\}\_\{j\}^\{\\tau\}is the target\-side answer associated with the edit, typically the target entitye∗e^\{\*\}\. For EIC, the reference answer ise∗e^\{\*\}; For OBP, the reference answer is the option or answer related to original\-entity in the question; For NBG, the reference answer is the target\-side associated fact in the open\-ended image\-conditioned query\.

To be noted, in the main text we use the Acc metric for analysis\. We additionally provide calculation results based on the probability metric in Appendix\.[E\.2](https://arxiv.org/html/2605.06096#A5.SS2)as supplementary reference\.

##### Locality Metrics\.

Locality metrics use a different rule because their goal is not to reward a new target answer, but to measure whether unrelated behavior is preserved\. For a locality sample, let𝐲^j,loc\(pre\)\\hat\{\\mathbf\{y\}\}\_\{j,\\mathrm\{loc\}\}^\{\(\\mathrm\{pre\}\)\}and𝐲^j,loc\(post\)\\hat\{\\mathbf\{y\}\}\_\{j,\\mathrm\{loc\}\}^\{\(\\mathrm\{post\}\)\}denote the predicted tokens, or the selected prediction identifiers, produced by the pre\- and post\-edit models on the same locality input\. The locality score is computed as a consistency rate:

Locτ=1Nτ​∑j=1Nτ1Mj​∑t=1Mj𝟏​\[y^j,loc,t\(pre\)=y^j,loc,t\(post\)\]\.\\operatorname\{Loc\}\_\{\\tau\}=\\frac\{1\}\{N\_\{\\tau\}\}\\sum\_\{j=1\}^\{N\_\{\\tau\}\}\\frac\{1\}\{M\_\{j\}\}\\sum\_\{t=1\}^\{M\_\{j\}\}\\mathbf\{1\}\\\!\\left\[\\hat\{y\}\_\{j,\\mathrm\{loc\},t\}^\{\(\\mathrm\{pre\}\)\}=\\hat\{y\}\_\{j,\\mathrm\{loc\},t\}^\{\(\\mathrm\{post\}\)\}\\right\]\.\(15\)This rule is used for both text locality and multimodal locality, with the latter applying the same consistency principle to image\-conditioned locality inputs\.

## Appendix DExperiment Setup Details

Our experiments build on the codebase implemented byHuanget al\.\[[2024](https://arxiv.org/html/2605.06096#bib.bib147)\]\. All the baseline implementations, including hyperparameters, remain consistent with the setup ofHuanget al\.\[[2024](https://arxiv.org/html/2605.06096#bib.bib147)\]\.

### D\.1Baselines

We focus on six representative editing methods: FT\-LLM, FT\-Vis, KE, MEND, IKE, and SERAC, spanning four broad paradigms\.

- •FTdirectly fine\-tunes different components of the LVLM\. It contains two variants:FT\-LLMfine\-tunes the LLM backbone, whileFT\-Visfine\-tunes the vision encoder module or projector\.
- •KE\[De Caoet al\.,[2021](https://arxiv.org/html/2605.06096#bib.bib120)\]is a hypernetwork\-based editing method that trains a bidirectional LSTM hypernetwork to predict weight updates to specific layers of the LLM directly based on gradients\.
- •MEND\[Mitchellet al\.,[2022b](https://arxiv.org/html/2605.06096#bib.bib116)\]likewise trains a hypernetwork, but predicts low\-rank weight updates to specific LLM layers given the gradient information of an edit pair\.
- •IKE\[Zhenget al\.,[2023](https://arxiv.org/html/2605.06096#bib.bib118)\]directly leverages in\-context learning to achieve the editing effect, which prepends retrieved demonstration examples to the query context without any parameter modification\.
- •SERAC\[Mitchellet al\.,[2022a](https://arxiv.org/html/2605.06096#bib.bib117)\]performs editing via an external memory module, stores edit tuples in an external memory and routes queries through a scope classifier at inference time, leaving base model parameters unchanged\.

### D\.2Details on EC\-Bench Evaluation

In this section, we briefly introduces the configurations used to obtain the EC\-Bench results\.

#### D\.2\.1Training

MEND,SERAC, andKErequire training before evaluation\. The trainable editors are trained on 5000 edit cases, and a held\-out validation set is used to monitor generalization and select the final checkpoint\.

Table[4](https://arxiv.org/html/2605.06096#A4.T4)groups the shared training settings,KE\-specific optimization settings, and model\-specific training settings\.

Table 4:Training configuration for trained editors\.Shared settings\.

KE\-specific settings\.

Model\-specific trained\-editor settings\.

The shared panel lists the batch size, optimizer, gradient clipping, and loss weights used by the trained editors\. TheKE\-specific panel records additional objective and optimization parameters used only byKE\. In the lower panel, iterations is the training budget, early stop is the patience window used for checkpoint selection,l​rlris the optimizer learning rate, and edit lr is the update scale used by the editor\-specific update mechanism\.

The lower panel reports only the method\- and LVLM\-specific settings that differ across trained editors, while shared values are kept in the upper panels\. A dash denotes a parameter that is not used by the corresponding method\.

MENDandSERACuse validation\-based early stopping, whereasKEuses a fixed update budget\. Validation is run every 1k steps for these trained editors, and the selected checkpoint is the one with the best validation performance\. ForMEND, the learned learning\-rate parameters usel​rl​r=1​e​\-​4lr\_\{lr\}=1e\\text\{\-\}4\.

#### D\.2\.2Evaluation

Table[5](https://arxiv.org/html/2605.06096#A4.T5)summarizes the test\-time configuration used by each method\.

Table 5:Test\-time configuration for EC\-Bench methods\.MethodLVLMEdit location\# edit stepsedit lrFTLLaVAlayer 31 MLP,down\_proj/up\_proj101×10−41\\times 10^\{\-4\}MiniGPT\-4layer 31 MLP,down\_proj/up\_proj31×10−41\\times 10^\{\-4\}Qwen\-VLlayer 31 MLP,w1/w2/c\_proj21×10−41\\times 10^\{\-4\}Owl\-2layer 31 MLP,gate\_proj/down\_proj/up\_proj201×10−41\\times 10^\{\-4\}FT\-VISLLaVAmultimodal projector,mm\_projector101×10−41\\times 10^\{\-4\}MiniGPT\-4Q\-Former151×10−41\\times 10^\{\-4\}Qwen\-VLfinal visual\-transformer MLP, resblock 47252×10−32\\times 10^\{\-3\}Owl\-2vision model251×10−31\\times 10^\{\-3\}MENDLLaVAlayers 29–31 MLP,down\_proj/up\_proj——MiniGPT\-4layers 29–31 MLP,down\_proj/up\_proj——Qwen\-VLlayers 29–31,w1/w2/c\_proj——Owl\-2layers 29–31,down\_proj/up\_proj——SERACLLaVAlayers 29–31 MLP,down\_proj/up\_proj——MiniGPT\-4layers 29–31 MLP,down\_proj/up\_proj——Qwen\-VLlayers 29–31,w1/w2/c\_proj——Owl\-2layers 29–31,down\_proj/up\_proj——KELLaVAlayers 29–31,down\_proj/up\_proj——MiniGPT\-4layers 29–31,down\_proj/up\_proj——Qwen\-VLlayers 29–31,w1/w2/c\_proj——Owl\-2layers 29–31,down\_proj/up\_proj——IKEallretrieved demonstrations \(k=32k=32\) withall\-MiniLM\-L6\-v2——The edit location specifies the model component or parameter group to which an edit is applied\. For per\-case update methods, \# edit steps is the number of gradient\-update steps and edit lr is the corresponding learning rate\.MEND,SERAC, andKEapply trained editors without additional test\-time gradient steps; andIKEis shown separately because it uses retrieved in\-context demonstrations rather than an edited parameter location\. All methods are evaluated as single\-sample edits, where each edit case is handled independently\.

### D\.3Details on Editing\-Location Control Experiments

The editing\-location control experiments vary the editing location forFTandMEND\. Table[6](https://arxiv.org/html/2605.06096#A4.T6)lists the language\-side locations included in the comparison and the Vis setting used as the vision\-side reference\.

Table 6:Configuration of the editing\-location control experiments\.MethodLVLMEdit locationParameter group\# edit stepsedit lrFTLLaVAlayers 05, 31 MLPdown\_proj/up\_proj101e\-4multimodal projectormm\_projectorMiniGPT\-4layers 10, 31 MLPdown\_proj/up\_proj31e\-4Q\-Former—15MENDLLaVAlayers 15–17 MLPdown\_proj/up\_proj——layers 29–31 MLPdown\_proj/up\_proj——multimodal projectormm\_projector——MiniGPT\-4layers 1–3 MLPdown\_proj/up\_proj——layers 29–31 MLPdown\_proj/up\_proj——Q\-Former layer 11intermediate/output\_query——In this table, edit location specifies either the selected LLM layers or the vision\-side setting, parameter group specifies the edited module within that location, and the last two columns use the same per\-case update notation as Table[5](https://arxiv.org/html/2605.06096#A4.T5)\. ForMEND, dashes in the last two columns indicate that the trained editor is applied without additional test\-time gradient steps\. The language\-side rows vary the edited LLM MLP layers, while the VIS rows move the editing location to the vision\-side component for the corresponding method\. In addition, the editing locations of FT\-shallow selected in Table[2](https://arxiv.org/html/2605.06096#S5.T2)are layer 5 and layer 10 for LLaVA and MiniGPT\-4, respectively;

## Appendix ESupplementary Experimental Results

### E\.1Results on Owl\-2

In this section, we extend our experiments to the mPLUG\-Owl2 model using the EC\-Bench dataset, with results presented in Tables[7](https://arxiv.org/html/2605.06096#A5.T7)\. Consistent with findings in Section[4\.2](https://arxiv.org/html/2605.06096#S4.SS2), EIC persists on this model, where most methods also exhibit an increase in EIC, along with anomalies on the OBP and NBG tasks\. The conclusions on FT\-Vis also apply to this model, where it achieves the best comprehensive performance across these three metrics\.

Table 7:Main EC\-Bench results on inherited and diagnostic metrics\.
### E\.2Probability Metric Results on Ec\-Bench

In addition to the accuracy metrics reported in the main text, we additionally provide metric results based on probability computation \(See Appendix\.[C\.4](https://arxiv.org/html/2605.06096#A3.SS4)\) as a supplement\. The probability\-based metrics, being continuous metrics, offer finer granularity and are provided for reference\. Table[8](https://arxiv.org/html/2605.06096#A5.T8)reports the EC\-Bench main experiment results under the probability metric\. Note that, since probability computation requires specifying the target answer token sequence, and the locality metric has no target answer as it only evaluates the consistency of outputs before and after editing, the locality results are not reported here\.

Table 8:Probability results for the main EC\-Bench results\.ModelMethodEfficacy↑\\uparrowT\-Gen↑\\uparrowI\-Gen↑\\uparrowEIC↓\\downarrowOBP↓\\downarrowNBG↑\\uparrowLLaVAbase \(unedited\)26\.129\.125\.824\.088\.132\.9FT99\.498\.999\.498\.868\.629\.3FT\-VIS98\.591\.787\.924\.050\.945\.0MEND98\.598\.298\.496\.388\.036\.7SERAC98\.896\.998\.874\.487\.644\.7IKE99\.697\.999\.665\.947\.151\.0KE97\.896\.997\.692\.188\.035\.9MiniGPT\-4base \(unedited\)22\.925\.722\.727\.349\.831\.5FT98\.697\.697\.666\.441\.831\.9FT\-VIS99\.998\.799\.627\.350\.935\.7MEND99\.098\.698\.891\.951\.135\.6SERAC97\.393\.597\.375\.949\.646\.2IKE99\.097\.199\.067\.242\.646\.9KE97\.196\.896\.980\.425\.334\.5Qwen\-VLbase \(unedited\)20\.824\.020\.719\.156\.625\.9FT99\.896\.999\.482\.738\.725\.4FT\-VIS100\.093\.199\.319\.129\.927\.7MEND99\.398\.197\.463\.457\.828\.8SERAC66\.762\.666\.850\.445\.025\.4IKE99\.297\.899\.255\.734\.146\.9KE98\.895\.098\.288\.337\.929\.1Owl\-2base \(unedited\)28\.132\.028\.124\.482\.137\.9FT100\.099\.5100\.097\.080\.933\.1FT\-VIS99\.795\.499\.024\.430\.546\.4MEND99\.198\.097\.975\.982\.038\.9SERAC97\.794\.497\.676\.581\.249\.9IKE99\.898\.799\.863\.628\.154\.7KE63\.662\.562\.147\.883\.740\.2

Table[9](https://arxiv.org/html/2605.06096#A5.T9)provides the results of the editing\-location control experiment under the probability metric\. All other experimental settings are identical to those in the main text\.

Table 9:Editing\-location comparison for FT and MEND\.