A New AI Building Block Targets the Cost of Reading Long Medical, Genomic, and Physics Data

Training an AI on a 3D MRI volume, a full human genome, or a turbulence simulation costs more than training the same model on a stack of books. The reason is the math. Today's dominant design, the transformer, scales with the square of the input length: double the sequence, roughly four times the compute. A new preprint argues there is a way out, but only by teaming up with the design it was supposed to replace.

The paper is called HyenaND, submitted to arXiv on 1 July 2026 by D R Wessels and collaborators, according to the arXiv author listing. It introduces a subquadratic operator, a building block that processes long, multi-dimensional data without the quadratic blowup that limits transformers. "Subquadratic" means its compute grows slower than the square of input size; in practice, the authors report a path that scales closer to linearithmic (proportional to length times log length), because it uses FFT-based convolution under the hood.

What changes is what kind of data the operator can read natively. Most large language models work on flat sequences of tokens, so they can stack the same operator over and over. But a 3D MRI scan is a volume, a genome is a long string with structure at many scales, and a physics simulation is a grid that evolves in time. Naively flattening any of those into a single sequence loses the geometry. HyenaND instead applies implicitly parametrized, input-dependent kernels directly on the multi-dimensional structure, the authors write in the full paper.

The cost story is the practical hook. The authors implemented the operator as a CUDA kernel called nSubQ, which fuses the FFT-convolution path so the favorable scaling shows up as wall-clock speed, not just theory. They evaluate it across four domains the field cares about: long-context genomics, computer vision, medical imaging, and PDE modeling, the partial differential equations used in fluid dynamics, weather forecasting, and turbulence research. On those tasks, pure stacks of the new operator reportedly match strong attention baselines. Hybrid stacks that interleave the new operator with attention reportedly beat both pure attention and strong recurrence-based hybrids.

The honest read on that result: "beats transformers" is a phrase with a mixed track record in this subfield, and the strongest evidence here is hybrid, not replacement. The paper does not claim to retire the transformer; it claims the new operator is a useful building block when mixed in.

Cheaper training on volumetric and grid-structured data does more than save money. It changes which scientific questions are tractable. A radiology model that fits a full chest CT in memory at training time, rather than tiling slices, learns different features. A PDE surrogate that can swallow a long simulation trajectory becomes a usable design tool, not a curiosity.

The caveats are visible. The work is a preprint, not a peer-reviewed paper. The wall-clock numbers are reported on a specific GPU and CUDA configuration, and the abstract's "strong attention baselines" are not enumerated, so any specific head-to-head claim should be hedged. The third-party adoption signal is thin: the MONAI medical-imaging project's master-branch changelog carries a HyenaND-adjacent entry, which is a discovery hint, not a deployment.

What the paper does add is a clean read on where the cost of multi-dimensional AI actually comes from, and a building block that attacks it without pretending the rest of the stack goes away. For genomics labs, radiology teams running on volumetric data, and physics groups modeling PDEs, the question is whether the hybrid result holds up under independent benchmarking. The paper is the first public reference; the next move is replication.