매일 아침, 어제의 AI를 한 통으로 정리해 보내드립니다메일로 받아보기

METAL LAB

NeuPAT: Neuron-aware Plasticity Allocation Tuning for Language-Preserving MLLMs

arXiv:2608.081072026-08-11

arXiv:2608.08107v1 Announce Type: new Abstract: Multimodal expansion of large language models (LLMs) enables new perceptual capabilities but often compromises the language intelligence acquired during pretraining. In this work, we investigate this phenomenon from the perspective of internal adaptation dynamics and discover that neurons in pretrained LLMs exhibit heterogeneous plasticity during multimodal learning: some neurons are critical for preserving language capabilities, while others are m

저자 · Jiayue Jin, Jingwei Zhang, Chen Wang, Jing Liu, Longteng Guo

arXiv에서 원문 보기