TransformerTrio at SemEval-2026 Task 13: Navigating Domain Shift and Representation Instability in Machine-Generated Code Detection

AbstractDetecting machine-generated code is increasingly challenging due to advances in code generation models and domain variation across programming tasks. We present our submissions to SemEval-2026 Task 13, evaluating detection in three settings: binary human vs. machine classification, multi-class generator attribution, and four-way authorship classification including hybrid and adversarial cases. We compare feature-based, transformer-based, and hybrid approaches under domain shift and limited supervision. Results show that domain-specific signals often dominate model decisions, degrading generalization when training and test distributions diverge. Increasing model complexity does not consistently improve performance in low-resource or cross-domain settings and may amplify spurious correlations. These findings emphasize robustness and feature alignment over model sophistication for reliable detection.