ON THE EFFECTIVENESS OF LLMS IN UNIT TEST GENERATION FOR STRUCTURED TEXT PROGRAMS

Abstract

The reliability of industrial automation systems heavily depends on the correctness of Programmable Logic Controller (PLC) programs, which are often written in Structured Text (ST). While Large Language Models (LLMs) have shown promise in automating test generation for mainstream programming languages, their effectiveness for the syntactically strict ST language remains underexplored. This thesis presents a systematic empirical evaluation of three state-of-the-art LLMs—GPT-4o, Gemini 2.5 Pro, and Claude Sonnet 4.5—for generating ST unit tests. We examine three prompting strategies: Natural Language (NL), Code Language (CL), and Chain-of-Thought (CoT), across a curated set of 11 ST function blocks. The quality of the generated tests is assessed using Compilation Success Rate (CSR), Statement Coverage (SC), and Branch Coverage (BC). In the zero-shot setting, Claude Sonnet 4.5 achieves the highest CSR, while Gemini 2.5 Pro consistently delivers the best statement and branch coverage, particularly under CL prompts. By incorporating a one-shot CL prompt, all models exhibit substantial improvements—most notably GPT-4o, whose CSR increases from 45.45% to 90.91%, with substantial gains in both SC and BC. To further contextualize these findings, we compare GPT-4o’s one-shot results with PLCAutoTester, a state-ofthe- art ST unit test generation tool, on an additional benchmark dataset. While LLMgenerated tests approach competitive coverage levels, PLCAutoTester maintains significantly higher and more stable coverage across programs. This study provides the first comprehensive benchmark of modern LLMs for ST unit testing, highlighting their strengths, limitations, and improvements through one-shot prompting, and positioning their performance relative to specialized automated testing tools in industrial automation.