A new technique fuses two very different AI architectures into one model - and it beats both of its parents on standard vision benchmarks.
Researchers built a method called Riemannian-Lorentz Parameter Fusion (RLPF) that combines a Vision Transformer and a state-space model, two architectures that process images in fundamentally different ways and don't normally share weights. Instead of averaging parameters directly, RLPF matches components by their role in the network, then projects them into hyperbolic space - a coordinate system suited to hierarchical relationships - to compute a blended midpoint before decoding it back into both branches. A trained gate then decides how much each branch contributes to a given prediction. After fine-tuning, the combined system hit 82.37% accuracy on CIFAR-10, 75.04% on Oxford-IIIT Pet, and 78.58% top-1 on ImageNet-1K, each beating the better of its two source models, which scored 76.54%, 71.42%, and 76.42% respectively.
Model merging usually promises a cheap alternative to retraining: no gradient descent, no new data, just arithmetic on existing checkpoints. This work stretches that idea to architectures that were never merge-compatible in the first place, which matters as vision AI increasingly mixes transformer and state-space designs to balance accuracy against compute cost. But the savings are smaller than they sound - the gate and final model still need training, so this isn't the free, instant merge the model-merging pitch usually implies.
Call it a supervised hybrid with extra geometry, not the training-free shortcut model merging usually promises.