Code-A1: Adversarial Co-Evolution of Code LLM and Test LLM via Reinforcement Learning
arXiv 2603.15611·high signal
Addresses the bootstrapping problem in RL-for-code: high-quality test suites are scarce but RL requires verifiable rewards. Code-A1 trains a code model and a test model adversarially via RL — each model's failures become the other's training signal, eliminating dependence on human-curated test suites. This self-improving adversarial loop could dramatically reduce the benchmark dependency bottleneck for training next-generation coding agents.