Elamur Research⌕
☰
home / papers / 2026 / Stochastic Gradient Descenders at DeSegMa-IT: Instruction-Tuned LLM and Token Classification for MGT Detection and Boundary Localization

Stochastic Gradient Descenders at DeSegMa-IT: Instruction-Tuned LLM and Token Classification for MGT Detection and Boundary Localization

Paper · 2026 · Huynh Nguyen Phu, Bui Hong Son and Dang Van Thin

Summary

This paper describes the Stochastic Gradient Descenders system for DeSegMa-IT (EVALITA 2026), a shared task on detecting machine-generated text (MGT) in Italian. The task has two subtasks: deciding whether a whole document is human- or machine-written (Subtask A), and locating the character index where a human-written opening gives way to an LLM continuation (Subtask B).

The authors built a separate pipeline for each subtask:

  • Subtask A: Qwen2.5-0.5B-Instruct, adapted with LoRA, is prompted in ChatML format to output a single label digit. If generation fails to produce a valid digit, a logit-based fallback picks the label.
  • Subtask B: Instead of regressing the offset directly, they recast it as per-token human/machine tagging with a fully fine-tuned UmBERTo encoder. The predicted boundary is the start offset of the first token tagged as machine.

The token-level approach ranked first in Subtask B (MAE 52.54). The classifier placed fifth in Subtask A (accuracy 0.92). The results suggest that dense token supervision and a monolingual encoder work well for fine-grained segmentation. Error analysis shows that formal human prose is often mistaken for machine output.

Key points

  • Data: Subtask A has 33,138 balanced training documents, plus 5,059 dev and 15,135 test documents. Subtask B has 19,945 / 4,579 / 26,015 mixed documents. In Subtask B, the boundary falls at about 265 characters on average, roughly 20% of the document length.
  • Subtask A setup: LoRA with r=16 and α=32 on all linear layers, trained on 2× T4 GPUs. The 0.5B model beat the 3B variant in validation, which the authors attribute to overfitting in the larger model.
  • Subtask B setup: UmBERTo (umberto-commoncrawl-cased-v1), 512-token inputs, 10 epochs of full fine-tuning. It reached 94.1% token-level accuracy on validation.
  • Centering bias: Boundaries are over-predicted for short prefixes (up to +73 characters) and under-predicted for long ones (about −53 characters).
  • Main error sources: smooth stylistic transitions (38% of high-error cases), formal human text that resembles LLM output (27%), and oscillating predictions near the boundary (21%).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

Cite

Cite as Huynh Nguyen Phu, Bui Hong Son and Dang Van Thin. 2026. Stochastic Gradient Descenders at DeSegMa-IT: Instruction-Tuned LLM and Token Classification for MGT Detection and Boundary Localization. Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{phu2026stochastic,
  title = {{Stochastic Gradient Descenders at DeSegMa-IT: Instruction-Tuned LLM and Token Classification for MGT Detection and Boundary Localization}},
  author = {Huynh Nguyen Phu, Bui Hong Son and Dang {Van Thin}},
  booktitle = {Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026)},
  year = {2026},
  url = {https://aclanthology.org/2026.evalita-1.9/}
}

References

14 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. G. Puccetti, A. Pedrotti, A. Esuli, DeSegMa-IT at EVALITA 2026: Overview of the Detection and Segmentation of Machine Generated Text in Italian Task, in: Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026), CEUR.org,… find
  2. F. Cutugno, A. Miaschi, A. P. Aprosio, G. Rambelli, L. Siciliani, M. A. Stranisci, EVALITA 2026: Overview of the 9th Evaluation Campaign of Natural Language Processing and Speech Tools for Italian, in: Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for I… find
  3. Qwen Team, Qwen2.5 Technical Report, arXiv preprint arXiv:2412.15115 (2024). find
  4. E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, W. Chen, LoRA: Low-Rank Adaptation of Large Language Models, arXiv preprint arXiv:2106.09685 (2021).
  5. Musixmatch Research, UmBERTo: an Italian Language Model trained with Whole Word Masking, https://huggingface.co/Musixmatch/umberto-commoncrawl-cased-v1, 2020. Accessed: 2025-12-31. ↗
  6. UnslothAI, Unsloth: Efficient Fine-tuning and Reinforcement Learning for LLMs, https://github.com/unslothai/unsloth, 2024. Accessed: 2025-12-31. ↗
  7. I. Loshchilov, F. Hutter, Decoupled Weight Decay Regularization, in: International Conference on Learning Representations (ICLR), 2019, pp. 1–18. find
  8. T. Dettmers, M. Lewis, S. Shleifer, L. Zettlemoyer, 8-bit Optimizers via Block-wise Quantization, arXiv preprint arXiv:2110.02861 (2021).
  9. Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V. Stoyanov, RoBERTa: A Robustly Optimized BERT Pretraining Approach, arXiv preprint arXiv:1907.11692 (2019).
  10. A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, PyTorch: An Imperative Style, High-Performance Deep Learning Libr… — PyTorch: An Imperative Style, High-Performance Deep Learning Library
  11. T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Brew, Transformers: State-of-the-Art Natural Language Processing, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 2020, pp.…
  12. P. He, X. Liu, J. Gao, W. Chen, DeBERTa: Decoding-enhanced BERT with Disentangled Attention, arXiv preprint arXiv:2006.03654 (2020).
  13. J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, in: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2019, pp. 4171–4186.
  14. DeSegMa-IT Organizers, DeSegMa-IT Leaderboard (Results), https://desegma.github.io/results/, 2025. Accessed: 2025-12-31. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/phu-2026-stochastic-gradient-descenders-desegma-it/