Barista: Synthesizing Typestate Specifications with LLM Agents (catarinagamboa.github.io)

At the upcoming ASE 2026 conference in Munich, researchers introduced Barista, an innovative system that leverages large language model (LLM) agents to synthesize typestate specifications from Java documentation. By interpreting implicit rules from Javadoc and transforming them into LiquidJava specifications, Barista enhances the compiler's ability to detect violations before code execution. The system classifies the various states of Java classes, such as `java.net.Socket`, and provides a method to ensure that developers follow the proper ordering of operations like connecting or closing a socket. The significance of Barista for the AI and machine learning community lies in its ability to automate complex code validation processes that traditionally require expert intervention. Barista demonstrates impressive precision (87%) in identifying protocol violations and shows an efficient execution time, completing tasks in just 7.3 minutes compared to 31 minutes taken by human experts. While its effectiveness decreases with more complex protocols, Barista presents a robust solution for enhancing software reliability, showcasing the promising intersection of LLMs and automated software engineering.