Automated program repair is the field of automatically finding and fixing bugs in software with little or no human intervention. Classical approaches, dating to systems like GenProg, treated repair as a search problem: mutate the buggy code and run the test suite until a variant passes. These methods worked but were slow and often produced patches that passed the tests yet did not really fix the underlying problem.
The arrival of large code models changed the field. The paper “Practical Program Repair in the Era of Large Pre-trained Language Models” (arXiv, October 2022, by Chunqiu Steven Xia, Yuxiang Wei, and Lingming Zhang, published at ICSE 2023) evaluated nine pretrained models from 125 million to 20 billion parameters and found that directly applying them already substantially outperformed all existing repair techniques across the datasets tested. Larger models generally fixed more bugs, and giving the model the code after the buggy line, not just before it, improved both the number of patches and how often they compiled.
Automated program repair matters because fixing bugs consumes a large share of software maintenance effort and cost. The shift to language-model-based repair is part of the broader move toward AI agents that can diagnose and patch real software, the same capability measured by benchmarks like SWE-bench. For businesses, reliable automated repair promises lower maintenance costs and faster response to defects.