CoSPlay: Cooperative Self-Play at Test-Time with Self-Generated Code and Unit Test

Researchers introduced CoSPlay, a training-free test-time framework for code generation that jointly refines candidate solutions and self-generated unit tests.

By iteratively filtering, repairing, and selecting code with cooperative self-play, it improves both code quality and unit-test accuracy without ground-truth tests.

On multiple benchmarks, CoSPlay matched or sometimes beat a supervised RLVR baseline while generalizing across models and token budgets.

The result points to a scalable, inference-only path to stronger code generation without expensive ground-truth unit tests.