Academic

CoMoL: Efficient Mixture of LoRA Experts via Dynamic Core Space Merging

arXiv:2603.00573v1 Announce Type: new Abstract: Large language models (LLMs) achieve remarkable performance on diverse downstream and domain-specific tasks via parameter-efficient fine-tuning (PEFT). However, existing PEFT methods, particularly MoE-LoRA architectures, suffer from limited parameter efficiency and coarse-grained adaptation due to the proliferation of LoRA experts and instance-level routing. To address these issues, we propose Core Space Mixture of LoRA (\textbf{CoMoL}), a novel MoE-LoRA framework that incorporates expert diversity, parameter efficiency, and fine-grained adaptation. Specifically, CoMoL introduces two key components: core space experts and core space routing. Core space experts store each expert in a compact core matrix, preserving diversity while controlling parameter growth. Core space routing dynamically selects and activates the appropriate core experts for each token, enabling fine-grained, input-adaptive routing. Activated core experts are then merg

J
Jie Cao, Zhenxuan Fan, Zhuonan Wang, Tianwei Lin, Ziyuan Zhao, Rolan Yan, Wenqiao Zhang, Feifei Shao, Hongwei Wang, Jun Xiao, Siliang Tang · · 1 min read · 11 views
#cs.CL

arXiv:2603.00573v1 Announce Type: new Abstract: Large language models (LLMs) achieve remarkable performance on diverse downstream and domain-specific tasks via parameter-efficient fine-tuning (PEFT). However, existing PEFT methods, particularly MoE-LoRA architectures, suffer from limited parameter efficiency and coarse-grained adaptation due to the proliferation of LoRA experts and instance-level routing. To address these issues, we propose Core Space Mixture of LoRA (\textbf{CoMoL}), a novel MoE-LoRA framework that incorporates expert diversity, parameter efficiency, and fine-grained adaptation. Specifically, CoMoL introduces two key components: core space experts and core space routing. Core space experts store each expert in a compact core matrix, preserving diversity while controlling parameter growth. Core space routing dynamically selects and activates the appropriate core experts for each token, enabling fine-grained, input-adaptive routing. Activated core experts are then merged via a soft-merging strategy into a single core expert, which is combined with a shared LoRA to form a specialized LoRA module. Besides, the routing network is projected into the same low-rank space as the LoRA matrices, further reducing parameter overhead without compromising expressiveness. Extensive experiments demonstrate that CoMoL retains the adaptability of MoE-LoRA architectures while achieving parameter efficiency comparable to standard LoRA, consistently outperforming existing methods across multiple tasks.

Executive Summary

The proposed CoMoL framework addresses the limitations of existing MoE-LoRA architectures by introducing core space experts and routing, enabling fine-grained adaptation and parameter efficiency. CoMoL achieves comparable performance to standard LoRA while retaining the adaptability of MoE-LoRA architectures, outperforming existing methods across multiple tasks. This novel approach has significant implications for large language models and parameter-efficient fine-tuning.

Key Points

  • ▸ Introduction of core space experts and routing for fine-grained adaptation
  • ▸ Compact core matrix storage for preserving diversity and controlling parameter growth
  • ▸ Soft-merging strategy for combining activated core experts with shared LoRA

Merits

Improved Parameter Efficiency

CoMoL achieves parameter efficiency comparable to standard LoRA while retaining adaptability

Fine-Grained Adaptation

Core space routing enables dynamic selection and activation of core experts for each token

Demerits

Increased Complexity

CoMoL's novel components may add complexity to the overall architecture

Expert Commentary

The CoMoL framework represents a significant advancement in MoE-LoRA architectures, addressing long-standing limitations and achieving impressive results. The introduction of core space experts and routing enables fine-grained adaptation, while the compact core matrix storage preserves diversity and controls parameter growth. As the field of large language models continues to evolve, CoMoL's improvements will likely have a profound impact on the development of more efficient and adaptable models.

Recommendations

  • ✓ Further research on the applications of CoMoL in various NLP tasks
  • ✓ Exploration of potential extensions and modifications to the CoMoL framework

Sources

Related Articles