NetLLM: Adapting Large Language Models for Networking **Authors:** Duo Wu, Xianda Wang, Yaqi Qiao, Zhi Wang, Junchen Jiang, Shuguang Cui, Fangxin Wang **Institutional Affiliations:** The Chinese University of Hong Kong, Tsinghua University, The University of Chicago **Abstract:** This paper addresses the challenges of adapting large language models (LLMs) to networking tasks, aiming to achieve a more sustainable design philosophy. Current deep learning (DL) algorithms for networking suffer from high engineering overhead and poor generalization performance on unseen data. Inspired by the success of LLMs in natural language processing, NetLLM proposes a framework that leverages pre-trained LLMs to solve various networking problems with minimal engineering effort. NetLLM includes a multimodal encoder to process multimodal data, a networking head to generate task-specific answers efficiently, and a data-driven low-rank networking adaptation (DD-LRNA) scheme to reduce fine-tuning costs. Extensive evaluations across three networking tasks—viewport prediction, adaptive bitrate streaming, and cluster job scheduling—show that NetLLM significantly outperforms state-of-the-art algorithms, demonstrating its effectiveness and strong generalization capabilities. **Keywords:** Deep Learning, Network Optimization, Video Streaming, Job Scheduling, Large Language Model Adaptation **Contributions:** - Identifies key challenges of LLM adaptation for networking. - Designs NetLLM, the first framework for LLM adaptation in networking. - Extensively evaluates NetLLM across three networking tasks, showing superior performance and generalization. **Design and Contributions:** - **Multimodal Encoder:** Encodes multimodal input data into token-like embeddings for efficient processing. - **Networking Head:** Generates task-specific answers directly, ensuring reliability and reducing latency. - **DD-LRNA:** Reduces fine-tuning costs by incorporating a data-driven adaptation pipeline and low-rank matrices. **Evaluation:** - Extensive experiments across three networking tasks (viewport prediction, adaptive bitrate streaming, cluster job scheduling) demonstrate the effectiveness of NetLLM. - NetLLM significantly outperforms state-of-the-art algorithms, achieving performance improvements of 10.1-36.6% for VP, 14.5-36.6% for ABR, and 6.8-41.3% for CJS. **Conclusion:** NetLLM addresses the limitations of existing learning-based algorithms in networking by leveraging the capabilities of LLMs. It provides a coherent design to efficiently adapt LLMs to various networking tasks, reducing engineering overhead and improving generalization.NetLLM: Adapting Large Language Models for Networking **Authors:** Duo Wu, Xianda Wang, Yaqi Qiao, Zhi Wang, Junchen Jiang, Shuguang Cui, Fangxin Wang **Institutional Affiliations:** The Chinese University of Hong Kong, Tsinghua University, The University of Chicago **Abstract:** This paper addresses the challenges of adapting large language models (LLMs) to networking tasks, aiming to achieve a more sustainable design philosophy. Current deep learning (DL) algorithms for networking suffer from high engineering overhead and poor generalization performance on unseen data. Inspired by the success of LLMs in natural language processing, NetLLM proposes a framework that leverages pre-trained LLMs to solve various networking problems with minimal engineering effort. NetLLM includes a multimodal encoder to process multimodal data, a networking head to generate task-specific answers efficiently, and a data-driven low-rank networking adaptation (DD-LRNA) scheme to reduce fine-tuning costs. Extensive evaluations across three networking tasks—viewport prediction, adaptive bitrate streaming, and cluster job scheduling—show that NetLLM significantly outperforms state-of-the-art algorithms, demonstrating its effectiveness and strong generalization capabilities. **Keywords:** Deep Learning, Network Optimization, Video Streaming, Job Scheduling, Large Language Model Adaptation **Contributions:** - Identifies key challenges of LLM adaptation for networking. - Designs NetLLM, the first framework for LLM adaptation in networking. - Extensively evaluates NetLLM across three networking tasks, showing superior performance and generalization. **Design and Contributions:** - **Multimodal Encoder:** Encodes multimodal input data into token-like embeddings for efficient processing. - **Networking Head:** Generates task-specific answers directly, ensuring reliability and reducing latency. - **DD-LRNA:** Reduces fine-tuning costs by incorporating a data-driven adaptation pipeline and low-rank matrices. **Evaluation:** - Extensive experiments across three networking tasks (viewport prediction, adaptive bitrate streaming, cluster job scheduling) demonstrate the effectiveness of NetLLM. - NetLLM significantly outperforms state-of-the-art algorithms, achieving performance improvements of 10.1-36.6% for VP, 14.5-36.6% for ABR, and 6.8-41.3% for CJS. **Conclusion:** NetLLM addresses the limitations of existing learning-based algorithms in networking by leveraging the capabilities of LLMs. It provides a coherent design to efficiently adapt LLMs to various networking tasks, reducing engineering overhead and improving generalization.