Rathnasuriya & Yang's Input Adaptation Cuts Code Model Errors Without Retraining

Code language models (CLMs) power many software engineering tasks, from generation to classification, yet they still suffer from notable mispredictions in real-world use—even when trained on up-to-date data. Existing fixes like retraining, modifying model architecture, or re-engineering prompts are computationally expensive, require extensive data labeling, and often fail to generalize across tasks or models. This new work from Ravishka Rathnasuriya and Wei Yang (University of Texas at Dallas) flips the script by adapting the input instead of the model.

Their approach has two stages: input validation, which identifies inputs that are likely to trigger mispredictions, and input adaptation, which rewrites those inputs using syntax- and semantics-preserving operations so they better align with the model's learned behavior. The method requires no retraining, no parameter changes, and no additional supervision—only lightweight transformations at inference time. Evaluations across diverse code understanding tasks show significant reductions in error rates. This is particularly promising for high-stakes applications (e.g., safety-critical systems) where model reliability is paramount and retraining overhead is prohibitive.