Pixel Phantoms at SemEval-2026 Task 13: Exploring Classical and Neural Approaches for AI-Generated Code Detection

AbstractThis paper describes our system for SemEval-2026 Task 13, Subtask A: detecting whether a given code snippet is AI-generated or human-written. We explored a range of approaches from classical machine learning baselines using TF-IDF representations to fine-tuned transformer models pre-trained on code, specifically CodeBERT and GraphCodeBERT. Our experiments revealed a notable degradation in model performance when CodeBERT was trained beyond an optimal number of steps, indicating that continued training within an epoch leads to overfitting or representation drift. GraphCodeBERT, by contrast, yielded our best submission with a macro F1 score of 0.36866. Our findings highlight the sensitivity of code-specific transformers to training duration and suggest that early checkpoint selection is critical for this task.