Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Hou, Zhiyan, Guo, Haiyun, Ma, Haokai, Sun, Yandu, Yang, Yonghui, Wang, Jinqiao
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2601.13020
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866912832192249856
author Hou, Zhiyan
Guo, Haiyun
Ma, Haokai
Sun, Yandu
Yang, Yonghui
Wang, Jinqiao
author_facet Hou, Zhiyan
Guo, Haiyun
Ma, Haokai
Sun, Yandu
Yang, Yonghui
Wang, Jinqiao
contents 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 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 phenomenon 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, 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 anti-forgetting without adding parameters. Our code will be released upon acceptance.
format Preprint
id arxiv_https___arxiv_org_abs_2601_13020
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle PASs-MoE: Mitigating Misaligned Co-drift among Router and Experts via Pathway Activation Subspaces for Continual Learning
Hou, Zhiyan
Guo, Haiyun
Ma, Haokai
Sun, Yandu
Yang, Yonghui
Wang, Jinqiao
Machine Learning
Artificial Intelligence
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 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 phenomenon 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, 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 anti-forgetting without adding parameters. Our code will be released upon acceptance.
title PASs-MoE: Mitigating Misaligned Co-drift among Router and Experts via Pathway Activation Subspaces for Continual Learning
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2601.13020