作者位次 #4/6ACL 2026Poster· Natural Language Processing
PAS s- M o E : Mitigating Misaligned Co-drift among Router and Experts via Pathway Activation Subspaces for Continual Learning
ZhiYan Hou · Haiyun Guo · Haokai Ma · Yandu Sun · Yonghui Yang · Jinqiao Wang
摘要
Continual instruction tuning (CIT) requires multimodal large language models (MLLMs) to adapt to a stream of tasks without forgetting prior capabilities. A common strategy is to isolate updates by routing inputs to different LoRA experts. However, existing LoRA-based Mixture-of-Experts (MoE) meth...