- Rating
- 4(+1)
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies regional weather forecasting, focusing on cross-region generalization and global information filtering. The authors propose ARGO, which combines Geo-aware Position Encoding, random regional sampling, and an auxiliary alignment loss to support cross-region generalization. ARGO also incorporates a Global Token Router to select relevant global information. The authors evaluate the approach on ERA5 data against Pangu-Weather, Stormer, OneForecast, and STCast, reporting strong performance in both fixed-region forecasting and cross-region generalization.
Critical strengths
- ARGO combines Geo-aware Position Encoding, random regional sampling, and an auxiliary alignment loss to support cross-region generalization.
- The Global Token Router selects global information based on geographical proximity and the current weather state. Experiments on ERA5 data demonstrate improved regional forecasting accuracy over the evaluated baselines and promising generalization 3. to a held-out region.
Critical weaknesses
- GaPE incorporates geographical information effectively, but the explanation of GaPE could be more detailed, particularly regarding how it encodes geographical relationships and supports transfer across regions. Further clarification of these mechanisms would make the design easier to understand.
- The auxiliary alignment loss is well motivated, but an ablation removing auxiliary alignment loss, together with an analysis of its effect on regional–global representations, would help substantiate its proposed alignment benefit.
- GTR provides a flexible way to select global information. Discussing related approaches such as BiFormer[1], Native Sparse Attention [2], and FlexPrefill [3] would help clarify its distinctive design.
- Figure 2 provides a useful overview, but clearer annotations would make the data flow easier to follow.
[1] BiFormer: Vision Transformer with Bi-Level Routing Attention
[2] Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
[3] FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference
Questions for authors
- Could you explain in more detail how GaPE encodes geographical relationships and how this design supports cross-region generalization?
- Could you provide an ablation removing the auxiliary alignment loss, along with an analysis of whether it improves regional–global representation alignment?
- Could you expand the discussion of related work on token routing and clarify how GTR differs from existing token selection and routing approaches?
5 comments
Thank you for your detailed responses and revisions, which have addressed my previous questions and concerns. I particularly appreciate the additional experimental results, which provide convincing evidence that the auxiliary alignment loss is effective.
I have one further question about ARGO’s potential as an efficient surrogate for a global forecasting model at 0.25° resolution. Can ARGO produce accurate regional weather forecasts while adapting to arbitrary geographic locations and grid sizes without region-specific retraining?
Thank you again for acknowledging our additional results. Your comments are very professional and constructive.
Our experiments demonstrate that a single ARGO checkpoint can generate 0.25° regional forecasts across multiple geographical locations and grid sizes without region-specific retraining or fine-tuning. By combining global background information at 1.5° resolution with high-resolution regional inputs, ARGO offers a promising alternative to 0.25° global forecasting models for regional prediction while avoiding high-resolution computation over the entire globe.
Responses to Reviewer ADih
We are truly grateful to Reviewer ADih for the encouraging assessment and helpful comments. Further explanations of GaPE and additional results on the auxiliary alignment loss are provided in the updated version.
W1, Q1: Clarification of GaPE
We have expanded Appendix B.2, “Geo-aware Position Encoding (GaPE),” (lines 752–779) in the updated version with the derivation from the periodic similarity kernel 1+cos(α−β), its corresponding rotation matrix, and its extension to longitude–latitude coordinates. We further explain how applying these rotations to queries and keys encodes relative geographical relationships while preserving longitude periodicity. Since GaPE uses geographical coordinates rather than domain-specific grid indices, it provides a consistent positional reference across different regional windows, supporting cross-region transfer.
W2, Q2: Auxiliary Alignment Loss
We compare ARGO with and without the auxiliary alignment loss, keeping the forecasting loss and remaining training settings matched. Both variants are evaluated without multi-step fine-tuning.
At the 72-hour lead time, the auxiliary alignment loss reduces RMSE across all four variables: by 5.2% for T2m, 6.2% for U10, 3.6% for MSLP, and 2.7% for Z500. ACC increases by 0.01 for T2m and MSLP and by 0.02 for U10, while Z500 ACC remains at 0.96 at the reported precision. These results show that auxiliary loss improves forecast accuracy.
W3, Q3: Related Token Selection Methods
We have expanded Section 2.2 (Related Work) (lines 141–146) in the updated version to discuss BiFormer, Native Sparse Attention, and FlexPrefill. We should clarify that GTR addresses regional–global fusion across grids of different resolutions by combining geographical proximity, dynamic content-based remote selection, and pooled global background information. The selected candidates are then processed by GaPE-equipped cross-attention, allowing each regional token to independently weight the shared candidates. This discussion clarifies GTR’s task-specific design in relation to existing sparse-attention approaches.
W4: Clarity of Figure 2
We have revised the annotations in the updated version's Figure 2 (lines 162–183) to clarify the data flow through the global and regional branches, GTR-based token selection, and GaPE-equipped cross-attention. The updated labels and arrows make the connections between these components easier to follow.