Two AI Models Just Merged Into One That Sees Better

Two researchers, Badri N. Patro and Vijay S. Agneeswaran, have published a method for stapling together two kinds of image-recognition AI that normally don't mix. One is a Vision Transformer, which studies a photo in square patches, like reading a page tile by tile. The other is a state-space model, a leaner design that scans an image step by step. Their problem: the two models store what they've learned in completely different shapes, so you can't just average them like two batches of cookie dough. The researchers solved it by translating both models' internal settings into a shared curved mathematical space — think of converting two recipes into the same units before mixing the batter — then encoding the result back into each branch.

Why should you care? Training a large image AI from scratch can cost millions of dollars and enormous amounts of electricity. If you can instead combine models that are already trained, you get better results for a fraction of the cost and carbon. That matters for the photo apps, medical scanners, security cameras, and shopping-search tools that increasingly depend on AI recognising what's in a picture.

The results are real but modest. On CIFAR-10, a small image test set, accuracy hit 82.37%, up from 76.54% for the best of the two alone. On a pet-breed test set it reached 75.04% versus 71.42%. On ImageNet-1K, the big benchmark, it hit 78.58% top-1 versus 76.42%. The honest catch is that this is not a free lunch: the paper itself notes the final model was fine-tuned (given extra training) and the combining gate was learned, so it isn't a purely training-free merge. It's also a preprint — not yet reviewed by other scientists — and tested only on standard benchmarks, not real products.

So don't expect this in your phone next month. What it signals is a direction: instead of building ever-bigger AI from scratch, researchers are getting better at recycling and combining what already exists. That's cheaper, greener, and faster — and it's how a lot of AI progress may arrive in the next few years.