Beyond Binary Edits Robust Multimodal Knowledge Editing with Adversarial Subspace Alignment

2026-05-22

AI Summary

This work addresses the limited robustness of multimodal large language models in knowledge editing, particularly their poor generalization to semantically equivalent visual and linguistic variants. To tackle this issue, the authors propose ASAM, a novel approach that first formalizes editing robustness through semantically equivalent multimodal knowledge units. It then employs Latent Adversarial Robustification (LAR) to generate semantically consistent adversarial variants that expose model vulnerabilities. Furthermore, Rank-Constrained Subspace Learning (RCSL), based on singular value decomposition, is introduced to align low-rank subspaces within the editing layer. Experimental results demonstrate that ASAM substantially enhances the generalization, reliability, and locality of knowledge edits, overcoming the limitations of conventional methods that rely on fixed sample anchors and constrained editing scopes.

Abstract

Multimodal large language models (MLLMs) need efficient mechanisms to update knowledge without degrading existing capabilities. While intrinsic multimodal knowledge editing achieves strong reliability and locality, it often exhibits limited generality, failing to propagate edits across semantically equivalent visual and linguistic variations. This issue arises from the lack of explicit semantic supervision, rigid editing scopes, and biased anchoring to individual samples in high-dimensional multimodal spaces. We address robust intrinsic multimodal knowledge editing by explicitly targeting generalization. We formalize robustness through knowledge units that group semantically equivalent multimodal inputs and define generality as consistent predictions within each unit. To expose fragile semantic regions, we introduce Latent Adversarial Robustification (LAR), which generates adversarial yet semantically coherent variants in the joint latent space. We further propose Rank-Constrained Subspace Learning (RCSL), enforcing low-rank alignment of adversarial representations at the edit layer via a singular value-based objective. Extensive analysis demonstrates the effectiveness of ASAM empirically.

Problem

• multimodal knowledge editing
• semantic generalization
• adversarial robustness
• knowledge updating
• multimodal large language models

Innovation

• Multimodal Knowledge Editing
• Adversarial Subspace Alignment
• Latent Adversarial Robustification
• Rank-Constrained Subspace Learning
• Semantic Generalization