Source-linked AI summary
Uncovering Entity Identity Confusion in Multimodal Knowledge Editing
Shu Wu, Xiaotian Ye, Xinyu Mou, Dongsheng Liu, Xiaohan Wang, Mengqi Zhang
TL;DR
Multimodal knowledge editing lacks sufficient analysis of post-edit behavior, motivating investigation of Entity Identity Confusion, in which text-only queries about an original entity return the new entity’s name. The paper introduces EC-Bench to diagnose this failure, attributes it to confusion between image-entity and entity-entity knowledge, and identifies early-stage editing constraints as a promising mitigation direction.
Problem
Post-edit behavior in multimodal knowledge editing is underexplored, while existing benchmarks largely fail to detect Entity Identity Confusion.
Method
The paper constructs EC-Bench and analyzes image-entity versus entity-entity knowledge, including editing constraints targeted at early-stage representations.
Results
Existing methods overfit entity-entity associations instead of updating image-entity bindings, producing systemic EIC; FT and MEND on LLaVA approach a 99% confusion rate.
Takeaways & Limitations
Faithful multimodal editing should update image-entity binding while preserving the structural integrity of entity-entity relational knowledge.
Takeaways & Limitations
Old Binding Persistence and New Binding Generalization remain challenging, and faithful multimodal editing warrants further exploration because text-only multi-hop reasoning remains open.
Abstract
from arXiv · showhide
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.
1 Introduction
Multimodal knowledge editing remains underexplored despite its deployment relevance, and this paper identifies Entity Identity Confusion as a systemic failure mode. It introduces EC-Bench, diagnoses the mechanism behind EIC, and proposes targeting image-entity processing to mitigate it.
- Motivation: Multimodal knowledge editing extends post-deployment knowledge revision from text-only systems to large vision-language models, where systematic post-edit behavior remains insufficiently studied.MKE aims to modify image-entity associations, such as correcting an image misidentified as Biden so it is recognized as Trump.
- Failure mode: Entity Identity Confusion occurs when text-only identity questions about an original entity unexpectedly receive the new entity’s name after editing.For example, after correcting an image association from Biden to Trump, asking for Biden’s full name may produce Trump.
- Benchmark: EC-Bench diagnoses EIC and evaluates how image bindings to original and new entities evolve through Old Binding Persistence and New Binding Generalization tasks.The benchmark is designed to expose behavior that standard metrics may miss.
- Mechanistic diagnosis: Existing MKE methods often leave the image bound internally to the original entity while using the new entity’s label as a spurious identity description.This shortcut can preserve the original entity’s relational facts, such as its alma mater, on more complex questions.
- Mitigation: The paper advocates separating image-entity binding from entity-entity relations and restricting edits to the processing region responsible for image-entity binding.The authors report this as a promising mitigation direction validated across multiple baseline methods.
2 Preliminaries
The paper formalizes multimodal knowledge editing as rebinding an image to a target entity while preserving unrelated behavior. It distinguishes image-entity bindings from entity-entity relations and defines evaluation around post-edit answers.
- LVLM background: A large vision-language model combines a vision encoder, projector, and language-model backbone to generate answers from image and text inputs.The vision encoder produces visual tokens, the projector maps them into the language model’s embedding space, and the backbone generates autoregressively.
- Knowledge types: Image-entity binding knowledge captures which entity an image refers to, whereas entity-entity knowledge captures semantic facts and attributes relating entities.These knowledge types may be handled by different model components and layers.
- Problem formulation: MKE takes an image originally bound to entity e and seeks to rebind it to a target entity e∗ using an edit set and updated model parameters.The edited model should produce the target answer while preserving unrelated behavior.
- Problem formulation: The standard editing objective requires the edited model to output the target answer for the image and query while preserving unrelated model behavior.The formulation uses θ′ = M(θ, Dedit) for the updated parameters.
3 Observing Entity Identity Confusion: A Preliminary Experiment
The preliminary experiment evaluates EIC by testing whether edits contaminate text-only identity knowledge. It finds that high edit efficacy can coexist with widespread confusion across editing paradigms.
- Experimental setup: The preliminary setup extends VLKEB with EIC-focused evaluation tasks to observe post-edit behavior under varied editing methods.The editing objective rebinds an image from an original entity to a target entity, while evaluation includes text-only identity queries.
- Experimental setup: EIC measures the proportion of text-only queries about the original entity that incorrectly return the new entity’s label.For example, after editing Biden to Trump, asking for Biden’s full name may produce Donald Trump.
- Characteristics of EIC: High edit success rates coexist with severe identity confusion across editing methods, showing that single-prompt efficacy is insufficient as the sole quality indicator.Models can succeed on original edit queries while exhibiting anomalous identity behavior elsewhere.
- Characteristics of EIC: EIC recurs across parameter-modifying, external-memory-based, and prompt-based methods, indicating a structural issue in the current MKE formulation.Severity varies by method, but the pattern is not confined to one editing paradigm.
- Characteristics of EIC: Text-only anomalies show that editing intended for image-conditioned I-E behavior contaminates textual knowledge representations instead of acting precisely on the I-E relationship.The resulting identity conflation can occur without visual input.
4 Analyzing Post-Edit Binding Behavior with EC-Bench
EC-Bench combines conventional editing measures with binding diagnostics to distinguish successful rebindings from hidden identity confusion. Results show severe EIC, persistent old bindings, weak new-binding generalization, and an apparent shortcut through entity-level associations.
- EC-Bench: EC-Bench combines Efficacy, Generality, and Locality with diagnostic probes for EIC, OBP, and NBG.The benchmark examines both basic editing competency and how image-entity associations change after editing.
- EC-Bench: Efficacy measures whether the edited model returns the target entity on the original edit query.This is the minimal criterion for a successful intervention.
- EC-Bench: The binding diagnostics separately test spurious E-E association, persistence of the old I-E binding, and formation of the new I-E binding.EIC probes the first, OBP the second, and NBG the third through identity or multi-hop reasoning tasks.
- Experiments and findings: Nearly all editing methods produce severe EIC, with FT and MEND on LLaVA reaching confusion rates approaching 99%.The result indicates substantial contamination of textual-modal knowledge during I-E editing.
- Experiments and findings: Post-edit models retain high OBP scores while most NBG scores remain low, showing that old bindings persist and injected new bindings are difficult to use for complex reasoning.NBG is consistently and substantially lower than OBP, suggesting recognition of the original entity before downstream reasoning.
- Experiments and findings: FT-Vis achieves the best EIC scores and approaches the unedited base model, but still performs poorly on OBP and NBG and has locality deficiencies.Editing the visual side leaves the language decoder, where E-E knowledge is encoded, unmodified.
- Conclusion: EC-Bench reveals an inconsistent internal structure in which the original I-E pathway remains active, the new pathway is weak, and an unintended E-E shortcut produces the new label.The model can therefore appear to succeed while still perceiving the image as the original entity.
5 Mitigating Entity Identity Confusion: A Preliminary Exploration
The section argues that editing location can mitigate EIC by targeting early I-E binding representations rather than downstream E-E relation decoding. Experiments with FT and MEND show reduced EIC for shallow or vision-side edits, although OBP and NBG remain challenging.
- Editing should decouple I-E binding, which must change, from E-E relational knowledge, which should remain intact.
- 5.1 Background and Rationale: Knowledge Recall in LLMs: A two-stage recall framework motivates editing before entity representations fully consolidate, because deeper edits may perturb relation decoding while leaving upstream binding intact.
- 5.2 Mitigating EIC via Editing-Location Control: FT experiments on LLaVA vary editing locations to test whether shallow-layer interventions reduce EIC.
- 5.2 Mitigating EIC via Editing-Location Control: EIC severity remains relatively flat in early layers, rises in middle layers, and becomes steep in deeper layers, supporting the entity-representation solidification hypothesis.
- 5.2 Mitigating EIC via Editing-Location Control: MEND shows a similarly significant EIC reduction when edits are confined to shallow or vision-side layers instead of the default deep-layer configuration.
- 5.3 Discussion & Implications for Future Research: Shallow-layer editing improves OBP and NBG less than EIC, leaving multimodal multi-hop reasoning as an unresolved challenge.
- 5.3 Discussion & Implications for Future Research: Editing location is presented as a useful design principle, alongside calls for diagnostic evaluations and mechanisms that localize edits to appropriate representational stages.
6 Related Work
The related-work section situates the paper within knowledge editing for language models and its extension to large vision-language models. It highlights parameter-modifying methods and multimodal benchmarks organized around efficacy, generalization, and locality.
- Knowledge Editing in Large Language Models: Knowledge editing updates model knowledge while aiming to preserve unrelated knowledge, using parameter-modifying and other method categories.
- Multimodal Knowledge Editing: Multimodal knowledge editing extends this paradigm to LVLMs through migrated editing methods and benchmarks including MMEdit, MIKE, VLKEB, and MC-MKE.
7 Conclusion
The conclusion identifies EIC as an overlooked failure mode caused by conflating I-E and E-E knowledge, and presents EC-Bench for diagnosing it. It proposes early-stage editing as a mitigation direction and frames the findings as a basis for more faithful MKE.
- EIC is a systemic, previously overlooked failure mode that existing benchmarks largely fail to detect.
- Current MKE methods can overfit E-E associations as a shortcut instead of modifying the underlying I-E binding.
- EC-Bench introduces challenging diagnostic tasks that expose EIC beyond standard evaluations.
- Constraining edits to early-stage representations is identified as a promising mitigation direction for more faithful and robust multimodal knowledge editing.
A Limitations
The paper’s scope is largely confined to image–entity binding behavior in classical MKE, while broader multimodal knowledge and misuse risks remain open concerns.
- Scope: EC-Bench primarily analyzes edited-model behavior for image–entity bindings in the classical MKE setting.The authors identify broader entity categories, image types, and knowledge such as image style as directions for future extension.
- Benchmark: EC-Bench is designed to detect EIC and evaluate how bindings to original and new entities evolve after editing.It includes EIC, OBP, and NBG diagnostics alongside conventional MKE tasks.
- Benchmark: The benchmark extends existing MKE data and baselines with restructured data and novel binding-diagnostic tasks.Part of the dataset is sourced from open-source benchmarks, including VLKEB.
- Diagnostic construction: EIC construction rewrites image-conditioned identity questions into text-only questions about the original entity, using the target entity as the reference answer.Answering the original-entity question with the target entity indicates entity confusion.
- Diagnostic construction: NBG tests whether the new image–entity binding supports factual reasoning beyond the edited prompt, while OBP tests persistence of the old binding.These diagnostics distinguish memorized output changes from broader binding behavior.
C.3 Examples of Dataset Entries
The dataset examples pair counterfactual edits with EIC, locality, NBG, and OBP queries, and the benchmark computes task scores using averaged token-level evaluation.
- Dataset entries: EC-Bench examples specify an original entity, target entity, rewritten EIC question, locality query, and NBG question for each edit.Examples cover actors, films, counties, bands, and national flags.
- Metrics: EC-Bench reports task-level scores as averages over valid evaluation samples using a common token-level scoring rule.The model is evaluated before and after editing, with predicted-token and probability-based computations.
- Metrics: The probability score is an average token probability rather than the product probability of the complete answer sequence.Main result tables normally report post-edit scores for edited rows and pre-edit scores for base rows.
- Metrics: The EIC reference answer is the target entity, while OBP uses an original-entity answer and NBG uses a target-associated fact.Reference answers therefore differ by diagnostic task.
- Metrics: Locality scores measure consistency between pre-edit and post-edit predictions for unrelated text and image-conditioned inputs.The same consistency principle applies to both locality modalities.
D.2.1 Training
Training configurations specify data, optimization, checkpointing, and test-time settings for trained editors, while editing-location controls compare language-side and vision-side interventions.
- Training: MEND, SERAC, and KE are trained before evaluation on 5,000 edit cases with a held-out validation set for checkpoint selection.Validation monitors generalization during training.
- Training: Table 4 groups shared training settings, KE-specific optimization settings, and model-specific settings for trained editors.Shared settings include batch size, optimizer, gradient clipping, and loss weights.
- Training: MEND and SERAC use validation-based early stopping, whereas KE uses a fixed update budget.Validation is run every 1k steps for the trained editors using checkpoint performance.
- Test-time configuration: Table 5 records test-time configurations, including edit locations, update steps, learning rates, and the distinct in-context setup used by IKE.MEND, SERAC, and KE use trained editors without additional test-time gradient steps.
- Editing-location controls: Editing-location experiments vary FT and MEND between selected LLM layers and vision-side components.Language-side rows change edited LLM MLP layers, while VIS rows move editing to vision-side components.
E.1 Results on Owl-2
On mPLUG-Owl2, EIC persists across methods, with accompanying OBP and NBG anomalies; FT-Vis achieves the best comprehensive performance across these metrics.
- Results: EIC persists on mPLUG-Owl2, where most methods show increased EIC and anomalies on OBP and NBG.This is consistent with findings reported for the broader EC-Bench analysis.
- Results: FT-Vis achieves the best comprehensive performance across EIC, OBP, and NBG on mPLUG-Owl2.The comparison is reported in the main EC-Bench results context.
E.2 Probability Metric Results on Ec-Bench
This section supplements the main EC-Bench experiments with probability-based metrics and reports results for both the main experiment and editing-location control.
- Probability-based metrics supplement the accuracy metrics reported in the main text.These continuous metrics provide finer granularity and are included for reference.
- Table 8 reports probability-metric results for the main EC-Bench experiment.
- Probability computation requires specifying the target answer token sequence, whereas the locality metric has no target.
- Table 9 reports the editing-location control experiment results for FT and MEND under the probability metric.All other experimental settings are identical to those in the main text.