Many physical systems are one-way coupled: an exogenous field—such as medium density or obstacle geometry—shapes the dynamics of interest without being shaped back. A new preprint on arXiv (2610.14687) asks whether transformer-based surrogate models benefit from representing these fields explicitly rather than leaving them implicit.
The authors propose patchwise adaptive normalization as a mechanism for injecting exogenous field information into the transformer architecture. The approach conditions the surrogate on the field at a local patch level, which is a more direct way to encode spatial structure than a global or learned embedding.
Because the source is a single short abstract, no experimental results or comparisons are available yet. The contribution at this stage is the conceptual framing and the proposed method; there are no conflicting findings to weigh.