This paper presents the system for SemEval-2026 Task 13, addressing both binary detection (Subtask A) and multi-class attribution (Subtask B). For Subtask A, we propose a robust multi-view training framework using UniXcoder-base, incorporating domain-specific structural prefixes, delexicalization with symmetric KL consistency loss, and token dropout. Our system achieves a high macro F1 of 0.845 on the out-of-distribution test set, demonstrating strong generalization across five unseen languages and two unseen domains. For Subtask B, we provide a rigorous diagnostic analysis of majority-class bias in transformer-based detectors. We reveal a significant performance gap where an 88.4% accuracy masks a near-complete failure in minority-class attribution (0.086 Macro F1), highlighting that standard fine-tuning is insufficient for fine-grained generator identification. Our results expose distinct regimes in code detection and motivate the need for imbalance-aware, structure-focused modeling in future work.

We present a production-ready automated item generation pipeline integrated into a large-scale assessment program’s workflows. A one-shot approach prompts LLMs from an operational source item and existing guidelines, then populates metadata, screens quality, and uploads items for human review. Generated items passed expert review, and difficulty prompting reliably shifted difficulty.