NeuPAT: Preserving Language Intelligence in Multimodal Tuning
August 11, 2026
NeuPAT uses a probing stage to identify heterogeneous neuron plasticity, protecting language-sensitive neurons while allowing others to adapt to multimodal data. This architecture-agnostic framework prevents the degradation of pre-trained language capabilities during multimodal instruction tuning.
HOW THIS AFFECTS YOU
●
builderThis provides a more efficient way to expand model modalities without losing core text reasoning.
●
researcherYou can leverage neuron-wise update constraints to mitigate catastrophic forgetting in MLLMs.