Post by Evren Ayberk Munis

I’m excited to share that our paper with MALTO, “Detecting Human, AI, and Hybrid Code via Hard Negative Mining and Curriculum-Driven Ensembles,” is now available in the proceedings of SemEval (International Workshop on Semantic Evaluation) 2026. 📄 You can find the paper through the link below: https://lnkd.in/e7wKpdJ6 In this work, we tackled two challenging problems in machine-generated code detection: • Subtask B: attributing code to either a human author or one of ten different LLM families • Subtask C: distinguishing between human-written, machine-generated, hybrid, and adversarial code Our proposed framework combines UniXcoder and CodeT5 through a weighted soft-voting ensemble. It also integrates: • Tree-sitter-based code canonicalization • Data augmentation for underrepresented classes • Hard negative mining to focus on the most ambiguous samples • A three-phase curriculum training strategy • Custom mean and max pooling for richer code representations The final ensemble achieved Macro-F1 scores of 0.405 on Subtask B and 0.616 on Subtask C, outperforming the individual models and the official baseline. I want to thank you all, my co-authors Hüseyin Arda Arslan, Timofei Khudonogov, Mert Akgun , Murat Beşli , Ayhan Meherrem, and our supervisors Claudio Savelli and Flavio Giobergia.