This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Yihua Shao, Minxi Yan, Yang Liu, Siyu Chen, Wenjie Chen, Xinwei Long, Ziyang Yan, Lei Li, Chenyu Zhang, Nicu Sebe, Hao Tang, Yan Wang, Hao Zhao, Mengzhu Wang, Jingcai Guo
Outline
In this paper, we propose a novel approach, In-Context Meta LoRA (ICM-LoRA), based on in-context meta-learning to address the inefficiency of fine-tuning multi-task specialized large-scale language models (LLMs) using low-dimensional adaptation (LoRA). ICM-LoRA takes task descriptions as input and generates task-specific LoRA weights using a conditional variational autoencoder (CVAE). The generated weights are then integrated into the LLM to generate task-specific models without additional fine-tuning. In particular, we utilize in-context meta-learning to identify and map relationships between tasks, enabling more accurate LoRA parameter generation. As a result, ICM-LoRA achieves more accurate parameter reconstruction than conventional LoRA methods, and occupies only 1% (283 MB) of the storage space compared to conventional LoRA.
Takeaways, Limitations
•
Takeaways:
◦
A novel method for efficient fine-tuning of LLM in a multi-task environment
◦
Create task-specific models that are more accurate and storage-efficient than existing LoRA methods
◦
Identifying and mapping correlations between tasks using meta-learning in context
◦
Efficient LoRA parameter generation using CVAE
•
Limitations:
◦
The performance of ICM-LoRA can be significantly affected by the performance of CVAE.
◦
Additional validation of generalization performance across different task types is needed.
◦
Further studies are needed to investigate the scalability of the proposed method and its applicability to various LLMs.
