Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·21h ago
AI Research
arXiv CS.AI·May 25·8 min read
CoSPlay: Cooperative Self-Play at Test-Time with Self-Generated Code and Unit Test
TL;DR
LLMs struggle to generate correct code without ground-truth unit tests (reference tests), which are expensive to create and limit scalability. CoSPlay, a training-free framework, jointly improves code and self-generated unit tests through cooperative iteration, using execution feedback to prune weak solutions.
✦ Why It Matters
Engineers can now generate reliable code without expensive ground-truth test data, scaling inference-time computation instead of training.
Key Takeaways
- Utilize CoSPlay for GT-free code generation.
- Leverage self-generated unit tests to improve code quality.
- Explore diverse solutions to identify potential code failures.
- Achieve scalable performance improvements in code generation tasks.
How It Works
CoSPlay operates by generating its own unit tests and using them to evaluate and refine code candidates. It employs a cooperative self-play mechanism where both code and tests evolve together.
The framework identifies diverse solution ideas and their failure modes, producing discriminative unit test ideas. By analyzing the execution results, it iteratively prunes weak code and updates unreliable tests, ensuring that the best-performing code is selected based on consensus among outputs.