Controlled Comparison of KAN and MLP Fusion Heads in Multimodal Sentiment Analysis

This study tests whether Kolmogorov–Arnold Networks (KAN) outperform the Multilayer Perceptron (MLP) as a cross-modal fusion head. Sixteen configurations (BERT/DistilBERT text encoders, ViT/DeiT image encoders, concatenation or gated fusion, and MLP or KAN heads) were compared on MASAD (38,066 binary-labeled text-image pairs) and MVSA-Multiple (17,027 three-class pairs) under identical splits and protocols (ten seeds for the confirmatory BERT+ViT pair, five for the others). The primary metrics were domain-averaged F1 (MASAD) and weighted F1 (MVSA-Multiple). On MASAD, all configurations fell within a 95.3–96.0% weighted F1 band. Among the sixteen configurations, the best mean, 96.01 ± 0.19%, came from BERT+DeiT (concat)+MLP, and the best KAN configuration reached 95.74 ± 0.12%. For the confirmatory BERT+ViT pair, the head effect (KAN − MLP) in domain-averaged F1 was +0.19 points under concatenation (90% CI [−0.11; +0.49]) and −0.08 under gated fusion ([−0.36; +0.21]), within the pre-specified ±0.5-point equivalence margin; no other encoder pair showed a consistent head advantage. On MVSA-Multiple, no head was consistently superior, and wide intervals precluded an equivalence conclusion; a class-macro-F1 gain for gated KAN in the BERT+ViT pair (+1.77 points, 90% CI [0.00; +3.53]) was not reproduced for the other encoder pairs and remains exploratory. The findings do not support the assumption that KAN automatically improves multimodal fusion.