CoSPlay: Cooperative Self-Play for Test-Time Code Generation

There's a new framework called CoSPlay, which is explained in the arXiv paper 2605.23491. It addresses the issue of needing ground-truth unit tests for code generation in large language models (LLMs). Existing techniques, like reinforcement learning with verifiable rewards (RLVR) and test-time scaling (TTS), rely on costly ground-truth tests for proper training, which can be a disadvantage. CoSPlay changes the game by removing the need for these tests altogether, using cooperative self-play to enhance both the code and the unit tests. It starts by generating different solution concepts and spotting potential failure modes to devise solid unit test ideas. Then it uses bidirectional pass-count signals to refine the code and tests, reducing noise and false connections in self-generated tests.