Selective Target Geometry for Test-Time Multimodal Sentiment Adaptation

Adapting a multimodal sentiment model to an unlabeled target corpus is difficult when language, recording conditions, and annotation conventions change together. Source-free adaptation methods that align target features with a reconstructed source distribution can then reinforce a mismatched geometry, while target-only pseudo-labeling can corrupt prototypes when initial predictions are unreliable. We present Selective Target Geometry (STG), an offline test-time adaptation method that initializes class-conditional prototypes from statistics saved with the source model and selectively updates them using unlabeled target batches. A frozen text teacher supplies semantic anchors, lightweight audio-visual evidence moderates uncertain pseudo-labels, and a router reuses, creates, or defers prototype updates according to batch compatibility. We evaluate ten directed transfers among four English and Chinese multimodal sentiment datasets. In the manuscript’s reported results, STG reduces mean absolute error relative to the frozen source model in all ten transfers and achieves the lowest mean absolute error among the compared methods in each transfer. An ablation on MOSI-to-SIMS shows that removing the router largely eliminates the gain. These findings suggest that selective updates to target geometry can reduce negative transfer when source-alignment assumptions are unreliable.