Large Language Models in Automated Repair of Haskell Type Errors
Sofia Leal Santos, João Saraiva, Francisco Ribeiro · 2024
This paper introduces a new method of Automated Program Repair that relies on a combination of the GPT-4 Large Language Model and automatic type checking of Haskell programs. This method identifies the source of a type error and asks GPT-4 to fix that specific portion of the program. Then, QuickCheck is used to automatically generate a large set of test cases to validate whether the generated repair behaves as the correct solution. Our publicly available experiments revealed a success rate of 88.5% in normal conditions. However, more detailed testing should be performed to more accurately evaluate this form of APR.