Development of a multi-agent, LLM-driven system to enhance human-machine interaction: integrating DSPy with modular agentic strategies and logical reasoning layers for the autonomous generation of smart contracts
Year: 2024
Type: dissertation
Abstract: Recent advances in hardware capabilities have significantly accelerated the development of neural networks, leading to the emergence of Large Language Models (LLMs). Originally designed for text generation, these models have evolved to perform an increasingly broad spectrum of tasks, demonstrating capabilities that extend beyond early predictions based on their sole auto-regressive nature. During my internship at Sony’s R&D Center Brussels Laboratory, I studied the relevance of these models for inclusion in a broader project focused on fan engagement in Web3, integral to the Loyalty 3.0 initiative, operating on the Ethereum Virtual Machine (EVM) to streamline the deployment of decentralized protocols. The module developed in this thesis autonomously generates these protocols by interpreting natural language inputs from non-expert users, showcasing state-of-the-art language models’ capabilities in translating early propositions into advanced source code. The system incorporates a Human-Machine Interaction Module, syntax generation with specialized layers, and a novel Survey Augmented Generation for Validation (SuAV) technique for alignment assessment. This system integrates DSPy for declarative syntax prompting, custom modular strategies for enhanced flexibility, and LangChain sequence patterns to ensure robust and accurate protocol generation. A comprehensive study comparing various techniques revealed the effectiveness of a hybrid model with conversational agents, particularly for smaller language models. Yet, the research highlights the complexities of optimizing LLMs for specialized tasks, emphasizing the need for nuanced approaches that balance model capabilities with task-specific requirements. To gain a comprehensive understanding of the overall implications and contextual relevance of the material presented in this work, it is advisable to prioritize reading the conclusions, followed by the epilogue. These sections provide valuable insights beyond the numerical results of the thesis, offering a broader perspective on the study’s significance and impact.
Author: Mendoza Juan, Yago
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7046654788