PAST-TIDE: Prototype-Anchored Statement Tuning with Topic-Invariant Normalization for Stance Detection

PAST-TIDE is a stance detection system submitted to both subtasks of the StanceNakba Shared Task at NakbaNLP@LREC-COLING 2026. It reframes stance prediction as cloze-style masked language modeling, using a verbalizer and the pretrained MLM head instead of a randomly initialized classifier. The system also adds learnable class prototypes for batch-size-independent prototypical contrastive learning and topic-conditional layer normalization for cross-topic Arabic stance detection. According to the abstract, it reached macro-F1 scores of 0.75 and 0.74 on the official leaderboard for Subtasks A and B.