Correct When Paired, Wrong When Split: Decoupling and Editing Modality-Specific Neurons in MLLMs

• URL: http://arxiv.org/abs/2606.17057v1
• Date: Mon, 20 Apr 2026 05:43:19 GMT
• Title: Correct When Paired, Wrong When Split: Decoupling and Editing Modality-Specific Neurons in MLLMs
• Abstract summary: We propose DECODE, which explicitly disentangles and localizes modality-specific neuron groups for targeted knowledge.<n>Our in-depth empirical analysis reveals that the entity knowledge in MLLMs is not stored as a unified representation, but is instead distributed across disentangled modality-specific pathways.
• Score: 9.32596777876781
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Although Knowledge Editing provides an efficient mechanism for updating the knowledge of Multimodal Large Language Models (MLLMs), we find that current paradigms still suffer from an important yet remain underexplored issue : editing decoupling failure, where entity-related knowledge can be updated when the model is triggered by multimodal inputs (text--image query pairs), however, it often reverts to outdated pre-edit facts when the paired inputs are split into unimodal ones. Our in-depth empirical analysis reveals that the entity knowledge in MLLMs is not stored as a unified representation, but is instead distributed across disentangled modality-specific pathways. As a result, updates biased toward multimodal queries fail to propagate effectively to unimodal circuits. To bridge this gap, we propose DECODE, which explicitly disentangles and localizes modality-specific neuron groups for targeted knowledge. Extensive experiments demonstrate that DECODE consistently achieves effective knowledge updates under different modality triggers, thereby mitigating editing decoupling failures.

Related papers

• Towards Localized and Disentangled Knowledge Editing for Multimodal Large Language Models [12.330598426331091]
  Localized and Disentangled Knowledge Editing achieves precise and generalized editing by localizing fact-specific model layers and disentangling target-relevant inputs from irrelevant ones.<n>Experiments demonstrate LDKE achieves superior performance in propagating edits to related contexts while maintaining high locality.
  arXiv  Detail & Related papers  (2026-05-28T12:06:39Z)
• Are We Evaluating the Edit Locality of LLM Model Editing Properly? [68.441768731381]
  We find that existing specificity evaluation protocols are inadequate for this purpose.<n>Existing specificity metrics are weakly correlated with the strength of specificity regularizers.<n>We also find that current metrics lack sufficient sensitivity, rendering them ineffective at distinguishing the specificity performance of different methods.
  arXiv  Detail & Related papers  (2026-01-24T07:07:21Z)
• Consistency-Aware Editing for Entity-level Unlearning in Language Models [53.522931419965424]
  We introduce a novel consistency-aware editing (CAE) framework for entity-level unlearning.<n>CAE aggregates a diverse set of prompts related to a target entity, including its attributes, relations, and adversarial paraphrases.<n>It then jointly learns a low-rank update guided by a consistency regularizer that aligns the editing directions across prompts.
  arXiv  Detail & Related papers  (2025-12-19T15:18:07Z)
• PropMEND: Hypernetworks for Knowledge Propagation in LLMs [82.99849359892112]
  We present a hypernetwork-based approach for knowledge propagation, named PropMEND.<n>We show almost 2x accuracy on challenging multi-hop questions whose answers are not explicitly stated in the injected fact.<n>We also introduce a new dataset, Controlled RippleEdit, to evaluate the generalization of our hypernetwork.
  arXiv  Detail & Related papers  (2025-06-10T15:44:19Z)
• Disentangling Knowledge Representations for Large Language Model Editing [38.244171146682206]
  We propose DiKE, a novel approach that Disentangles Knowledge representations for LLM Editing.<n>DiKE consists of two key components: a Knowledge Representation Disentanglement (KRD) module that decomposes the subject representation into target-knowledgerelated and -unrelated components, and a Knowledge Edit (DKE) module that updates only the target-related component while explicitly preserving the unrelated one.<n>To rigorously evaluate fine-grained irrelevant knowledge preservation, we construct FINE-KED, a new benchmark comprising fine-grained irrelevant knowledge at different levels of relational similarity to the edited knowledge.
  arXiv  Detail & Related papers  (2025-05-24T16:24:04Z)
• MindBridge: Scalable and Cross-Model Knowledge Editing via Memory-Augmented Modality [55.01380617388064]
  Most existing methods overfit to specific models, causing edited knowledge to be discarded during each update.<n>We introduce MindBridge, a scalable solution inspired by the low coupling between modality processing and LLMs in multi-modal models.<n>MindBridge achieves superior performance even in editing tens of thousands of knowledge entries and can flexibly adapt to different LLMs.
  arXiv  Detail & Related papers  (2025-03-04T15:17:57Z)
• LLMs as Repositories of Factual Knowledge: Limitations and Solutions [1.7764955091415962]
  We study the appropriateness of Large Language Models (LLMs) as repositories of factual knowledge.<n>We evaluate their reliability in responding to time-sensitive factual questions.<n>We propose "ENtity-Aware Fine-tuning" (ENAF) to improve the model's performance.
  arXiv  Detail & Related papers  (2025-01-22T10:16:53Z)
• Robust and Scalable Model Editing for Large Language Models [75.95623066605259]
  We propose EREN (Edit models by REading Notes) to improve the scalability and robustness of LLM editing. Unlike existing techniques, it can integrate knowledge from multiple edits, and correctly respond to syntactically similar but semantically unrelated inputs.
  arXiv  Detail & Related papers  (2024-03-26T06:57:23Z)
• Multimodal Representation Learning by Alternating Unimodal Adaptation [73.15829571740866]
  We propose MLA (Multimodal Learning with Alternating Unimodal Adaptation) to overcome challenges where some modalities appear more dominant than others during multimodal learning. MLA reframes the conventional joint multimodal learning process by transforming it into an alternating unimodal learning process. It captures cross-modal interactions through a shared head, which undergoes continuous optimization across different modalities. Experiments are conducted on five diverse datasets, encompassing scenarios with complete modalities and scenarios with missing modalities.
  arXiv  Detail & Related papers  (2023-11-17T18:57:40Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.