Curricular Adversarial Training for Robust Code Generation via Hierarchical Reinforcement Learning

ICLR 2026 Conference Submission25575 Authors

Keywords: Robust Code Generation

Abstract: In this paper, we propose a novel approach to boost the robustness of code generation models by curricular adversarial training driven by hierarchical reinforcement learning. Existing code generation systems are prone to breaks by adversarial perturbations, so we propose a two-tiered approach in which a high-level curriculum policy is used to adaptivelyChange complexity of adversarial challenges dynamically while a low-level perturbation policy will be used to generate specific input modifications. The high-level policy goes from simple to sophisticated perturbation based on model performance, which will ensure the gradient of adapting without overwhelming the generator too much.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 25575