Evolving AI-Driven Workflow Management, Part B: Non-Unique Engineering Workflows and Scalable Open-weight Agents

This is the latest version of this item.

Abstract

Workflow engines play an important role for modern engineering, especially when the complexity of the subject is high. Large Language Models could potentially provide a powerful user interface, if they can be set up to reliably transform natural language inputs into correct workflow instructions. Previous works investigated this possibility with a single proprietary Large Language Model as the only involved AI system. The accompanying part A of the current work enhances that method by implementing a multi-agent architecture of proprietary models and Open Weight Models resulting in reduced context window sizes, and by also incorporating a data provenance system. The present part B addresses the real-world problem of ambiguity of workflows and how to solve this problem in a scalable manner. The main contributions are twofold. First, the non-uniqueness of results of queries to knowledge graphs of realistic workflows for computational engineering is classified as either "by multi-fidelity" or "by redundancy". Second, it is shown that LLMs with large context window can be capable of resolving such non-uniqueness whereas fine-tuned Small Language Models contribute in other ways to the scalability of the multi-agent system of part A.

Available Versions of this Item

• Evolving AI-Driven Workflow Management, Part B: Non-Unique Engineering Workflows and Scalable Open-weight Agents. (deposited 09 Jan 2025 16:12) [Currently Displayed]