Segmentation Fault at SemEval-2026 Task 13: A Regularization-First Approach with Generator-Based Out-of-Distribution Splits for Detecting AI-Generated Code
Year: 2026
Type: conference-paper
Abstract: This paper describes the system for SemEval-2026 Task 13 Subtask A on detecting AIgenerated code.The proposed approach finetunes CodeBERT-base with a generator-based out-of-distribution validation split, holding out some model families to better match the test conditions.Heavy regularization like data augmentation, dropout, weight decay, and label smoothing is applied, training on a balanced 100K sample subset with early stopping based on validation trend monitoring.Four approaches are experimented with -logistic regression, UniXcoder, vanilla CodeBERT, and the final regularized system.The results suggest that data quality and evaluation design play an important role in OOD generalization, and the final system achieves 0.439 macro F1 score on the test dataset, which is a 62% improvement over the initial approaches.Further analysis of the validation-test gap that is still present attributes it to human code source homogeneity, which is still a challenge and requires future work.
Institution: Sri Sivasubramaniya Nadar College of Engineering
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166881845