Gradient Descenders at DeSegMa-IT: Leveraging Monolingual Transformer for LLM-Generated Text Detection and Boundary Identification
Tran Phuoc Thanh Nhan, Bui Hong Son, and Dang Van Thin
In International Workshop on Evaluation of Natural Language and Speech Tools for Italian, Feb 2026
In this paper, we present the Gradient Descenders systems developed for the DeSegMa-IT shared task at EVALITA 2026. Our approach yielded highly competitive results: we achieved the 1st rank in Subtask A with an accuracy of 94.58% and secured the 3rd rank in Subtask B with a Mean Absolute Error (MAE) of 62.66. In the MachineGenerated Text (MGT) detection task, we demonstrated the effectiveness of UmBERTo, a RoBERTa-based model pre-trained on Italian text, validating the advantage of monolingual baselines over multilingual counterparts. In the Boundary Detection task, we approached the problem as a fine-grained token classification task using DeBERTa-base-Italian as the backbone encoder. Looking ahead, we aim to address the identified limitations regarding domain generalization and sequence dependency. Future work will focus on integrating DomainAdversarial Neural Networks (DANN) to handle domain shifts and exploring more complex classification layers, such as Bi-LSTM, to further refine boundary localization precision.