Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Amith Kamath Belman

Second Advisor

Ethel Tshukudu

Third Advisor

Yingjie Liu

Keywords

agentic AI, generative AI, LLM benchmark

Abstract

LaTeX allows authors to have full control of their scholarly work; however, learning LaTex can be difficult and using LaTex to write an article or dissertation can take time. Spartan Write is a desktop authoring tool that enables authors to edit LaTeX-based documents using a natural-language interface instead of having to manually type all of the commands. Authors can compose text, construct tables, insert images, as well as repair format-related problems in a chat-type environment. The author will also maintain access to the original source file as well as the compiled view of their document. The system has a React/Tauri front-end, a fast API server-side component for handling file operations and compilation, and a LangGraph agent for processing updates to the document content. In addition to presenting the design and implementation of the system's architecture, this report also provides an evaluation of Spartan Write based on a reproducible benchmark executed against several different LLM (Large Language Model) service endpoints. The benchmark measured several aspects of task execution including completion rate, total time required for task execution, tokens used during execution, costs associated with executing the task, and behaviors exhibited by the tool. Out of a total of 119 test cases, the system successfully completed 98.3% of them. Results indicated significant variability in terms of cost and speed between services. Variability was largely due to differences in usage patterns exhibited by the models being employed by each service. Gemini 3 Flash is recommended as the default choice for most users in this benchmark. Gemini 3 Flash demonstrated perfect success in completing every single one of the tasks in the test suite. It also demonstrated the shortest average runtime of approximately 12.9 seconds. Additionally, Gemini 3 Flash showed very low mean cost per run ($0.0094).

Download

Available for download on Saturday, May 22, 2027

Share

COinS