2026/01/01 by Zhiyan Hou, Haiyun Guo, Haokai Ma +3
Computer Science · #Domain Adaptation and Few-Shot Learning #Advanced Graph Neural Networks #Multimodal Machine Learning Applications
paper · pdf · doi:10.18653/v1/2026.acl-long.1474
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 LoRAbased Mixture-of-Experts (MoE) methods often jointly update the router and experts in an indiscriminate way, causing the router's preferences to co-drift with experts' adaptation pathways and gradually deviate from early-stage input-expert specialization.We term this as Misaligned Co-drift, which blurs expert responsibilities and exacerbates forgetting.To address this, we introduce the pathway activation subspace (PASs), a LoRA-induced subspace that reflects which low-rank pathway directions an input activates in each expert, providing a capability-aligned coordinate system for routing and preservation.Based on PASs, we propose a fixed-capacity PASsbased MoE-LoRA method with two components: PAS-guided Reweighting, which calibrates routing using each expert's pathway activation signals, and PAS-aware Rank Stabilization, which selectively stabilizes rank directions important to previous tasks.Experiments on a CIT benchmark show that our approach consistently outperforms a range of conventional continual learning baselines and MoE-LoRA variants in both accuracy and antiforgetting without adding parameters.Our code is publicly available at