Towards Siloed LLM-based Systems for Mission-critical Planning

Tyler D Comisky, Leslie N Smith, Mak Roberts, Joshua D Lovejoy, Avni Garg, Luke Nam, Adrian R Li, Leora B Samuels

Keywords: large language models, automated planning, mission critical systems

TL;DR: This paper proposes an agent-based framework to deploy large language models in mission-critical environments for planning purposes, using travel planning and aerial refuel scheduling as example applications.

Abstract: Adoption of Large Language Models (LLMs) is occurring at an accelerated pace. One area where LLMs currently fall short is in producing reliable and verifiable outputs. This makes it challenging to deploy LLMs to mission-critical environments, where both reliability and keeping the data private are required. Additional challenges occur when the environment must remain entirely disconnected from the internet. We propose a framework for deploying siloed LLMs to these mission-critical environments using tools and verifiers. First, we demonstrate this framework's feasibility using commercial LLMs for travel planning. Then, we use this framework for the mission-critical application of generating aerial refueling schedules using disconnected, open-source LLMs.

Submission Number: 8