?
ReText.Ai Team at PAN 2025: Applying a Multiple Classification Heads to a Transformer Model for Human-AI Collaborative Text Classification
This paper presents the ReText.Ai team’s solution to the Human-AI Collaborative Text Classification subtask of the PAN-2025 Generative AI Authorship Verification Task. Our approach involves fine-tuning transformer models, such as RoBERTa-base and Gemma-2 2B, with a custom multi-head classifier that includes a main multiclass head and auxiliary binary heads to better distinguish closely related labels. Through utilizing a transformer-based model augmented with multiple classification heads and a confidence-based override mechanism, our method outperforms the baseline, achieving macro Recall scores of 80.36% and 83.00% for RoBERTa-base and Gemma-2 2B, respectively, compared to 68.67% and 75.70% for the baseline models. In the competition, our team’s fine-tuned Gemma-2-2B model achieved seventh place in the automated evaluation on the test set with a score of 56.11%.