[Paper Review] Multilevel Large Language Models for Everyone

Yuanhao Gong|arXiv (Cornell University)|Jul 25, 2023

Topic Modeling4 citations

TL;DR

This paper proposes Multilevel Large Language Models (MLLM), a hierarchical framework unifying global, field-specific, and personal LLMs to enhance efficiency, privacy, and performance. By distilling large cloud-based models into specialized and locally run personal models, MLLM reduces redundancy, improves response speed, and protects user data—while integrating a blockchain-based economic model to incentivize high-quality user input and decentralize development for broader accessibility.

ABSTRACT

Large language models have made significant progress in the past few years. However, they are either generic {\it or} field specific, splitting the community into different groups. In this paper, we unify these large language models into a larger map, where the generic {\it and} specific models are linked together and can improve each other, based on the user personal input and information from the internet. The idea of linking several large language models together is inspired by the functionality of human brain. The specific regions on the brain cortex are specific for certain low level functionality. And these regions can jointly work together to achieve more complex high level functionality. Such behavior on human brain cortex sheds the light to design the multilevel large language models that contain global level, field level and user level models. The user level models run on local machines to achieve efficient response and protect the user's privacy. Such multilevel models reduce some redundancy and perform better than the single level models. The proposed multilevel idea can be applied in various applications, such as natural language processing, computer vision tasks, professional assistant, business and healthcare.

Motivation & Objective

• To unify generic and field-specific large language models into a cohesive, hierarchical system that improves performance and efficiency.
• To address privacy and computational inefficiency in large language models by deploying user-level models locally on personal devices.
• To reduce model redundancy and enhance response speed through multi-level distillation and specialization.
• To develop a decentralized, blockchain-integrated economic model that incentivizes high-quality user input and lowers barriers to LLM development.
• To enable broader access to LLM technology by decentralizing training and inference across users and devices.

Proposed method

• Proposes a three-tiered architecture: global LLM (cloud-based, pre-trained on broad data), field-level LLM (fine-tuned and distilled from global model for domains like healthcare or finance), and user-level LLM (further distilled for local deployment on personal devices).
• Employs knowledge distillation to transfer knowledge from larger models to smaller, faster, and privacy-preserving local models.
• Introduces a blockchain-based system where users act as decentralized nodes, contributing computation and receiving services in return, creating a mutual economic incentive.
• Treats user input as a valuable resource that improves LLMs at all levels, embedding user feedback into model refinement across the hierarchy.
• Designs a decentralized, parallel computation model inspired by blockchain architecture to support scalable and secure model training and inference.
• Uses the brain’s cortical functional specialization as a biological analogy: low-level specialized regions collaborate to enable high-level cognition, mirroring model interactions across levels.

Experimental results

Research questions

• RQ1 How can global, field-specific, and personal large language models be hierarchically linked to improve performance and efficiency?
• RQ2 Can local deployment of distilled models enhance response speed and user privacy while maintaining high accuracy?
• RQ3 How can user input be systematically leveraged as a valuable resource to improve LLMs across all levels?
• RQ4 What economic and technical mechanisms can decentralize LLM development and incentivize high-quality contributions?
• RQ5 Can blockchain-based infrastructure support scalable, secure, and efficient multilevel LLM training and inference?

Key findings

• The multilevel framework reduces redundancy and improves efficiency by distributing model specialization across global, field-specific, and personal tiers.
• Local deployment of user-level LLMs enables faster inference and stronger privacy protection compared to cloud-only models.
• The proposed system supports efficient knowledge distillation from large global models to smaller, application-specific models at each level.
• User input is treated as a valuable, incentive-driven resource that contributes to model improvement across all levels, enhancing system-wide performance.
• The integration of blockchain enables a decentralized, self-sustaining economic model where users and developers mutually benefit from computation and service exchange.
• The framework is extensible to diverse applications, including NLP, computer vision, healthcare, and business intelligence, demonstrating broad applicability.