Back to feed
arXiv cs.LG
arXiv cs.LG
6/17/2026
Correct When Paired, Wrong When Split: Decoupling and Editing Modality-Specific Neurons in MLLMs

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

Short summary

arXiv researchers found that knowledge edits in multimodal AI models work for paired text-image inputs but revert to outdated facts for single-modality queries. Entity knowledge is stored across separate modality-specific neural pathways, not unified. The DECODE method explicitly targets these pathways to ensure consistent knowledge updates across all input modalities.

  • •Knowledge editing in MLLMs fails when multimodal inputs are split into text or image alone
  • •Entity knowledge is distributed across decoupled modality-specific neural pathways
  • •DECODE method disentangles and targets modality-specific neurons for consistent edits

Generated with AI, which can make mistakes.

#research-breakthrough#ai-tools#ai-agents
Read full article at arXiv cs.LG

Is this a good recommendation for you?

Explore more

MODE: Modality-Decomposed Expert-Level Mixed-Precision Quantization for MoE Multimodal LLMs
arXiv cs.LG
Breaking the Illusion: When Positive Meets Negative in Multimodal Decoding
arXiv cs.LG
Where Reliability Lives in Vision-Language Models: A Mechanistic Study of Attention, Hidden States, and Causal Circuits
arXiv CS.AI
Self-Generated Error Training for Token Editing in Diffusion Language Models
arXiv cs.CL