ReDyGait: Representation Disentanglement with Gated Attention for Invariant-Contextual Transfer in Stance Detection

Cross-topic stance detection degrades when encoders entangle stance signals with topic-specific vocabulary, causing representations that fail to transfer to unseen targets. Existing methods commit to either topic-invariant or topic-aware representations and apply the same strategy uniformly to every input, sacrificing complementary information. We propose ReDyGait, a three-stage framework that disentangles these two types of signals through dedicated contrastive pre-training and recombines them adaptively at inference time. Stage 1 trains a topic-invariant encoder with supervised contrastive loss over cross-topic positives. Stage 2 trains a topic-contextual encoder with bidirectional pair contrastive loss over within-topic positives; both stages employ topic-aware hard negative mining to prevent shortcut learning. Stage 3 freezes the two contrastive encoders and learns a gating network that produces per-instance weights over invariant, contextual, and base-encoder pathways. On VAST, ReDyGait achieves a macro-averaged F1 of 0.782 in the zero-shot setting and 0.752 in the few-shot setting, improving over the strongest baseline by 1.1 points in both; on SEM16t6 in a leave-one-target-out setup, ReDyGait reaches an average F1 of 0.612. Analysis of the learned gate weights shows that the model shifts toward the invariant pathway for unfamiliar topics and toward the contextual pathway when topic-specific patterns are available, confirming that the disentanglement operates as intended.