Semantic Transformations vs. Iterative Repair: Handling Bugs with LLMs

Two recent papers on applying large language models to software bugs. The first explored using semantic code transformations to improve defect detection, but found they were often unreliable and did not boost accuracy. The second focused on automatic program repair, showing that even minimal fine-tuning and an iterative 'generate-and-test' approach make LLMs much more effective at fixing bugs.