Type: url
Original: https://arxiv.org/pdf/2403.17134
Status: generated
Summary
The paper introducing RepairAgent, which its authors present as the first autonomous LLM-based agent for automated program repair: the LLM freely interleaves gathering information about a bug, gathering repair ingredients and validating fixes by invoking 14 repair-related tools, guided by a dynamically updated prompt and a finite state machine. On Defects4J it correctly fixes 164 bugs, including 39 not fixed by prior techniques, at a median cost of about 270,000 tokens (14 US cents) per bug.