Abstract
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) methods often jointly update the router and experts indiscriminately, 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 PASs-based MoE-LoRA method with two components:
- PAS-guided Reweighting: calibrates routing using each expert's pathway activation signals.
- PAS-aware Rank Stabilization: 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 resistance to forgetting, without increasing model parameters. Our code is publicly available at GitHub Repository.
Blogger's Review: The pathway activation subspace method proposed in this paper offers a novel solution to the expert routing issue in continual learning. By accurately calibrating routes using pathway activation signals, it significantly enhances the model's stability and memory retention when facing new tasks, showcasing high practical value and application prospects.