Large Language Models (LLMs) such at OpenAI’s ChatGPT-4b and Google’s Gemini have shown extensive capability for general-purpose language understanding. Included within their training data is extensive information pertaining to Tactical Combat Casualty Care (TCCC) medical procedures, including: proper steps to perform  skills, rationale for those steps, and most importantly what those steps look like in an image.  Specifically, this last capability means that the user can present the LLM with a sequential series of TCCC images and the LLM can describe what is happening relative to the known procedure.  

This paper explores how these capabilities can be utilized by the simulation and training community.  We have identified several areas and show preliminary work and implementation that utilizes these concepts.  Our exploration includes 1) Using LLM’s to assess a student’s performance on a TCCC procedure; 2) Using LLM’s to answer questions for a TCCC student as a learning aid; 3) ability to use the LLM’s to label training data for traditional special purpose object recognition models.

Until recently each of the above applications employed a “traditional” approach, such as the use of YOLO models for computer vision. This paper compares our traditional methods with new methods that LLMs offer, including performance metrics on benchmark example cases. We will focus on the capabilities of ChatGPT-4b and Google Gemini Advanced.  We report lessons learned and provide insight into how this technology can be used in the future for the concept of operations described and other uses in simulation and training.

Keywords
AI;ASSESSMENT;CHARACTERIZING SYSTEM PERFORMANCE ;COMBAT CASUALTY CARE;MACHINE LEARNING

Additional Keywords
LARGE LANGUAGE MODELS