arXiv:2607.01128 2026 Architecture 2 ideas extracted · analyzed Aug 30, 2026
GAIA contributes a transferable pattern for learning operators whose input and output fields live on different, irregular spatial domains: represent geometry, boundary information, and sampled fields as tokens, then condition local integral transforms through cross-attention. The key asset is not a PDE-specific discretization, but the ability to make the kernel depend on local geometry and domain membership without constructing a graph or retraining for each shape. This can be adapted to point-cloud transformers, irregular scientific data, mesh-free neural operators, and amortized inverse problems. The most direct experiments are geometry-conditioned attention layers and a one-pass encoder-decoder that maps sparse measurements plus a domain description to a field on a separate query domain.
Use separate measurement-domain and target-domain token sets so a network can infer a field on one spatial domain from sparse observations on another in one forward pass. The same decoder can answer arbitrary target query points, avoiding an optimization loop for each inverse instance.
Replace a geometry-agnostic kernel or standard point-cloud attention layer with an integral transform whose weights are conditioned on geometry tokens. The layer receives field values at arbitrary points, domain indicators, boundary features, and local geometric descriptors, and uses cross-attention to produce location-dependent kernels without graph construction.