IBM's ASTER uses LLM prompting guided by lightweight program analysis to generate high-coverage, natural unit tests for Python and Java
ASTER incorporates software mocking methods to manage external dependencies like database transactions and web service calls
Smaller models like Granite-34B and Llama-3-8B perform nearly as well as larger models, offering cost and privacy advantages for enterprise environments
Over 70% of surveyed developers at IBM preferred ASTER-generated tests, finding them more usable and comprehensible
ASTER received the Distinguished Article Award at the 2025 International Conference on Software Engineering (ICSE) for its innovative approach
Plans include refining testing models to reduce LLM costs, expanding ASTER to more programming languages and testing levels
ASTER addresses the issues with a four-step process: static analysis, LLM-guided test generation, postprocessing and refinement, and coverage augmentation