Weather foundation models rely on transformers that process dense reanalysis fields—gridded data that blends observations and model output. But not all parts of those fields are equally trustworthy: reliability varies by variable, by location, and by weather regime. Standard token merging cuts computation by grouping similar tokens, yet it ignores this uneven data quality.
UtoMe, short for Observation-Uncertainty-Guided Token Merging, addresses that gap. Instead of merging tokens based purely on feature similarity, it factors in observation uncertainty, guiding the model to preserve information where confidence is low and merge more aggressively where data is reliable. The result is a more principled trade-off between speed and fidelity.
The work is described in a new arXiv preprint. The abstract notes that token merging can reduce computation, but that feature similarity alone does not account for the reliability structure of reanalysis fields—UtoMe is proposed as a correction to that limitation. Further details on experimental results are not included in the available abstract.