CodeImprove: Program Adaptation for Deep Code Models

Ravishka Rathnasuriya, Zijie Zhao, Wei Yang

Summary

CodeImprove improves deep code models by adapting their input programs rather than retraining the models. A validity score identifies inputs outside a model’s handling capability, and a genetic search applies semantics-preserving program transformations to adapt them. Evaluation across three models and two software engineering tasks shows improvements in accuracy and promising detection of out-of-scope inputs.