Determine the source of late-fusion performance advantages
Determine whether the reported advantage of late-fusion architectures in automated pain assessment arises from the fusion principle itself or from the more specialised and extensively engineered modality-specific encoders used by those architectures.
References
However, it remains unclear whether their advantage arises from the fusion principle itself or from their more specialised and extensively engineered modality-specific encoders.
— Efficient Architecture Search under Leave-One-Subject-Out Evaluation
(2609.21457 - Hihn et al., 18 Sep 2026) in Section 1, Introduction
We conjecture this is because of how attention mechanisms are already powerful modality fusion algorithms~\citep{attn_strong,mm_transformer_survey}. Adding to that is the fact that the CLIP and BERT encoders for both modalities are state-of-the-art.
— Mutual Equilibrium: Multimodal Representation Learning through Reciprocal Feedback
(2609.39456 - Park et al., 30 Sep 2026) in Section 4.1, Accuracy Results