ACE: Self-Evolving LLM Code Generation via Adversarial Testing

Researchers have introduced ACE, a framework for self-evolving code generation tailored for large language models (LLMs), which employs a solver-adversary structure. In this innovative setup, a single LLM alternates between crafting candidate programs and generating adversarial unit test inputs aimed at triggering execution failures, such as runtime errors or non-termination. This method overcomes the shortcomings of conventional solver-verifier systems, where tests produced by verifiers become less effective as solvers advance. ACE operates exclusively on execution-derived supervision, removing the necessity for extensive annotated solutions. The framework emphasizes active failure discovery to facilitate ongoing self-improvement in code generation. This research is documented in arXiv preprint 2605.16299.