DiffCodeGen: Coverage-Guided Test-Time Scaling for Code Generation
A new method for improving code generation at inference time, DiffCodeGen, uses coverage-guided differential analysis to select the best solution without relying on public test cases or extensive LLM inference. It generates diverse code candidates via varied sampling and prompting, synthesizes inputs through coverage-guided fuzzing, clusters candidates by behavioral similarity, and outputs the medoid of the largest cluster. This approach reduces token consumption and time overhead compared to prior test-time scaling methods.
Key facts
- DiffCodeGen is a novel test-time scaling method for code generation.
- It uses coverage-guided differential analysis.
- It generates diverse code candidates using various sampling and prompting strategies.
- It applies coverage-guided fuzzing to synthesize inputs without existing tests or LLMs.
- Candidates are clustered based on behavioral similarity from execution on synthesized inputs.
- The medoid of the largest cluster is selected as the final output.
- The method does not rely on public test cases.
- It reduces token consumption and time overhead compared to prior methods.
Entities
Institutions
- arXiv