Most long-horizon forecasting models produce the entire future sequence in a single forward pass. A new arXiv preprint asks whether a smaller, shared operator can instead be applied repeatedly to generate the forecast in contiguous blocks, one after another.
This design shifts the cost from stored parameters to sequential computation. The authors describe it as a trade-off: the model stays compact, but generating the full horizon requires more compute steps as the operator is reused recursively.
The abstract is brief and does not yet include experimental results or architectural details. It only introduces the concept and frames the central question of whether such a parameter-shared, blockwise approach can be competitive with one-shot forecasters.