Analyzing Hallucination Layers in LLM-Based Automated Program Repair

Analysis of 832 Defects4J bugs reveals that LLM hallucinations in automated program repair occur not just in final patches, but also in intermediate artifacts like testcase identification and line coverage prediction. This multi-layered study characterizes how ungrounded intermediate steps compromise repair reliability.