An intern discovered that directly prompting LLMs to rewrite legacy Java code is unreliable due to hallucinations, lack of stopping criteria, and difficulty in verifying functional equivalence. To address this, they developed a sequential agent system where an APIDesigner agent generates a JSON contract, followed by a CodeGenerator and a TestGenerator. The core innovation is using JUnit tests derived from extracted rules as oracles, ensuring the generated code passes specific criteria rather than relying on the LLM's output alone. This approach revealed ambiguities in the initial rule extraction, highlighting that the LLM's failures often stemmed from misinterpretations of these rules. AI
IMPACT This approach could improve the reliability of LLM-driven code generation by introducing rigorous testing and verification steps.
RANK_REASON The cluster describes the development of a specialized tool/system for code generation and testing, rather than a new model release or core research.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
