Improving the Robustness of Large Language Models for Code Tasks via Fine-tuning with Perturbed Data

Researchers have developed a method to improve the robustness of code-generating LLMs by fine-tuning them with 'perturbed' data—intentionally altered code at character, word, and sentence levels. This approach significantly improves model resilience against adversarial inputs, reducing robustness degradation (RD) by 4-6%. The trade-off is a minor 1-3% drop in raw performance (pass@1), highlighting a crucial balance between reliability and capability for AI coding assistants.