acceptodds
Under review as a conference paper at ICLR 2027

DriveGuard-WM: Dynamic-Only Cross-Region Meteorological Downscaling

Abstract

Supervised meteorological downscaling relies on paired low-resolution (LR) and high-resolution (HR) data, yet observationally supported HR data remain geographically limited. This motivates cross-region downscaling from paired source-region data to target regions with only LR inputs. However, regional differences in spatiotemporal dynamics limit the transferability of learned reconstruction relationships. Our central premise is that transfer should account for the predictive contributions of source-learned relationships and adapt their use to current LR conditions. To address this challenge, we propose DriveGuard-WM, a cross-region downscaling framework that reconstructs HR fields from dynamic LR inputs alone, without target HR supervision or fine-tuning. The framework translates contribution assessments from a world model into LR-conditioned guidance for adaptive spatiotemporal fusion. Across six directed transfers among Australia, the contiguous United States, and Europe, DriveGuard-WM reduces macro-averaged, source-standardized mean absolute error (MAE) by 38.6% for temperature and 46.0% for precipitation relative to the best evaluated baseline on this metric under the reported reference-field protocol. Our code is available at https://anonymous.4open.science/r/WeatherDownscaling-7732/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cross-region downscaling, Spatiotemporal modeling, Transfer learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_CLaROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the cross regional meteorological downscaling without HR monitoring in the target area. The author proposes a driveguard WM framework. Firstly, wmca uses the recursive potential state structure inspired by the world model to represent the global background, cross scale messages and local evolution. Secondly, LGL uses the HR of the source region to assist in the contribution supervision of the structural relationship, and learns the drive/guard guidance that only relies on LR information. Finally, ASR performs guided spatial reconstruction and adaptive fusion with temporal branch prediction. The paper carries out migration experiments in six directions among Australia, the United States and Europe. The results reported that under its reference field evaluation protocol, compared with the baseline stvd with the optimal average Mae, the average Mae of temperature and precipitation decreased by 38.6% and 46.0% respectively, and the model behavior was analyzed through state replacement, guidance and fusion.

Critical strengths
  1. The method design is innovative. The potential state modeling inspired by the world model is combined with the relationship contribution assessment, and the drive/guard is used to guide the cross regional meteorological downscaling.
  2. The temperature and precipitation were evaluated in six migration directions in three regions, and compared with five baselines in multiple indicators, and the consistent performance advantages were reported.
Critical weaknesses
  1. The introduction does not clearly explain the shortcomings of the existing cross regional downscaling methods. At the same time, the challenges proposed in the introduction and the subsequent solutions do not seem to have a good logical correspondence.
  2. The context of meteorological downscaling and cross domain generalization methods studied in this paper is not clearly sorted out in the relevant work, and the analysis in the appendix can be advanced to the text.
  3. The expression of the world model in the method part needs to be considered. The general performance of the world model is to input the state of the past time to predict the state of the future time, while the current method is to input the low-scale map of the past time to predict the high-scale map of the current time.
  4. Experiments have compared a variety of downscaling methods, but there is a lack of comparison with the recent methods directly for geographical generalization.
Questions for authors
  1. Please supplement the quantitative comparative experiment with the recent downscaling method specifically for cross regional generalization.
  2. Please further explain why the idea of world model is used to solve the problem of cross regional meteorological downscaling, and the main advantages of this idea over the existing downscaling methods.
  3. Please explain in more detail how the proposed contribution responds to the challenges raised in the introduction.
Reviewer_3ojPOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DriveGuard-WM, a framework for the cross-region meteorological downscaling problem in which a downscaling model must be transferred from a source region to a new region without access to the HR data in the target region at inference time. The method consistes of 3 main components, (1) WMCA which learns global and local temporal representations so that the relationship between representations can be used for downscaling, (2) LGL which learns to predict "Drive" and "Guard" scores to identify when to use such relationships in the downscaling, and (3) ASR which does a purely spatial reconstruction that can be fused with the former paths. The main experiments evaluate cross-region transfer across 6 scenarios for temperature and precipitation, and the authors report improvements over multiple downscaling baselines.

Critical strengths
  1. The problem setting is relevant and challenging. (The paper considers geographic transfer of meteorological downscaling models without target-region HR supervision, static geographic features, or test-time adaptation.)
  2. The cross-region transfer results are encouraging, provided the baseline implementations and evaluation protocol can be adequately verified.
Critical weaknesses
  1. Several important aspects of the architecture and experimental implementation are unclear.
    • Sections 4 and 5 do not adequately explain how the HR encoder, gated correction mechanism, and HR-informed posterior states are implemented. Some experimental controls in the plots, including the "free fusion" and "unconstrained", are also insufficiently specified.
    • Appendix C.2 identifies differences between the architecture described in Section 4 and the implementation used to produce the reported results. These include whole-state versus single-message corrections, fixed versus sample-dependent guidance targets, and top-2 versus dense expert routing. It is therefore unclear how closely the reported experiments validate the proposed formulation in the main paper, and it does not make sense to me why the implementation used in the experiments was not explicitly laid out in the main paper methods sections.
  2. The ablations do not convincingly justify the architectural complexity or establish the contribution of the proposed mechanisms. This is especially important since the paper proposes such as architecturally complex mechanism.
    • The main ablations are conducted on separate held-out source-region experiments, rather than under the cross-region benchmark supporting the paper's main claims. Important components, including the G/M/L design, mixture-of-experts spatial reconstruction, and windowed state-space propagation / SSMs in general, also don’t have ablations.
    • Section 5.4 does not compare the learned Drive scores with any simple alternative, making it difficult to determine whether input-dependent guidance is actually beneficial. For example, if one set of scores can be estimated from the training data and still perform well, it calls into question the benefit of the input guidance component.
    • The evidence for the Guard part is lacking. Although the full fusion gate performs well in Table A8, Appendix Table A9 reports AUROC values of only 0.137–0.313 for identifying harmful spatial corrections, which falls below the random baseline under the paper’s stated score interpretation.
  3. The evaluation is incomplete, and the fairness of the baseline comparisons is difficult to assess.
    • Some baselines, particularly GeoFAR and SGD, originally use geographic or satellite information excluded by the proposed protocol. The paper does not sufficiently describe how these models were adapted, nor the training budgets, pretraining, or sampling procedures used for the baselines, some of which are noted as missing in the appendix. Simple baselines such as interpolated ERA5 and a residual CNN, which are typically included in downscaling papers, are also absent.
    • While MAE, RMSE, and SSIM are common metrics, the evaluation does not go further and evaluate precipitation extremes or fine-scale spatial structure, both are central to the purpose of meteorological downscaling, arguably more so for the region specific evaluation in this paper. Related work evaluates these properties, including upper-percentile precipitation errors [1], Fractions Skill Score, and spatial power spectra [2].
    • The results do not include variation across random seeds / multiple initialization so cannot assess reliability of results.
    • Numerous experiments in Appendices B.2–B.4 and C.3 are marked as "pending," including some key ablations, so it reads like an evaluation plan and not completed worth with empirical evidence to back up the claims made in the main paper.

Overall, both the methodology and experiments sections are not well written and difficult to parse. Specific language is often used before things are defined, experimental setups are not described in detail or motivated, and discussion of the implications of results is incomplete.

References

[1] Harder, P., et al. (2026). Benchmarking the geographic generalization of deep learning models for precipitation downscaling. Scientific Reports, 16, 3733. https://doi.org/10.1038/s41598-025-34557-4

[2] Singh, S., et al. (2026). Comprehensive inter-comparison of generative AI models for super-resolution precipitation downscaling across hydroclimatic regimes. Geoscientific Model Development, 19, 7545–7567. https://doi.org/10.5194/gmd-19-7545-2026

Questions for authors
  1. Which exact architecture was used to produce the main results in Table 1? Appendix C.2 describes several differences between the proposed method and the reported implementation. Could the authors clarify which components were actually evaluated, and why the experimental implementation is not the one described in Section 4?
  2. Can the authors clarify the baseline implementations and evaluation protocol behind Table 1? Specifically, how were GeoFAR and SGD adapted to the LR-only setting, and were training budgets, pretraining, and sampling procedures comparable? I would also like to see results across multiple random seeds (mean ± standard deviation) to assess the robustness of the reported improvements.
  3. Can the authors do an ablation study under the main cross-region benchmark? A cumulative ablation table, progressively adding components (e.g., WMCA only, WMCA + unguided ASR, adding Drive guidance, and finally Drive/Guard fusion), would help establish which components contribute to the reported gains.
Reviewer_F6jJOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces DriveGuard-WM, a weather downscaling method conceived to be trained on one region and being transferred to another. The authors claim 3 contributions: a (misclassified) world model for contribution assessment, a low-resolution conditioned guidance learning framework and an adaptive spatiotemporal reconstruction framework. The full system is tested and compared in terms of cross-region generalisation capabilities. The core contributions of the paper are also individually tested.

Critical strengths

The paper attempts to solve an important problem. Conceptually, it makes sense to create a framework aimed at identifying and using source relations that can reduce reconstruction errors.

Critical weaknesses

Unfortunately, the paper is poorly written, and, in its current state, it doesn’t meet publication standards at this or any other venue. Additionally, I would argue that the authors have not even read their appendix, as either human- or LLM-generated comments explicitly highlight missing information or poor experimental design.

Questions for authors

Since I do not think the paper can be easily elevated to a publishable state within the review timeframe, I will use this section to provide recommendations for a future submission and to substantiate the critical weaknesses rather than asking questions. Note that the list of comments reported below is not comprehensive.

  1. The model proposed in this paper is repeatedly categorised as a world model. While I agree that the term world model is often abused, WMCA does not fit any of the formal definitions (e.g., [i, ii, iii, iv]).

  2. Many variables are not properly defined, or they appear as mathematical symbols which are explained only in later sections. For instance, at line 156, what are G, M, L? At that point, not only does the reader not know whether we are talking about scalars, vectors, tensors, functions, etc., but it's not even 100% clear if they are the factors for regional context, background-local interactions, and local evolutions. What are terms like QGQ^G or QLQ^L? Where do they come from? Are WQW_Q and similar symbols matrices or functions? Also, Drive/Guards are mentioned at line 215 but not explained until line 234. Things like GRU are not even ever disambiguated (so I assumed they were Gated Recurrent Units, yet couldn’t be 100% sure). My suggestion is: be rigorous and explain what variables are when they first appear.

  3. Methodological sections don’t explain the rationale for architectural choices and go straight into the math (which is also incomplete and uses variables that are never introduced).

  4. Some terms or sentences are most likely made up by LLMs, for instance: what is offline scoring (line 135), what is a two-stage encoder (line 160), what about a distributional head (line 190), etc.? Similarly, there are instances in which terms like “bilinear interaction” (line 193) are taken for granted, but what is happening here? Also, sentences like “Drive weights positive gain by variance attenuation and sign consistency” (line 234) are hard to grasp. Please stick to standard terminology and use equations where they help disambiguate.

  5. From Fig. 2 it looks like 3 sets of features are extracted from the “two-stage encoder”, what is correct? The figure or the text and math?

  6. It is unclear to me what makes features local vs global. From Fig. 2, it looks like the local features are extracted from the deeper layers of the encoder, in which case I would argue they are global rather than local.

  7. Is the cross-attention described at line 190 performed in space or time? I suppose time, but it is never explained.

  8. What do you mean by “Target HR is used only for scoring” (line 323)? By looking at Section 3, it looks like y is HR and thus used as GT in the loss function.

  9. If the methodology is aimed at proving transfer capabilities, ablations should be performed in this experimental setting.

  10. I am struggling to understand what happened with the resolutions and their aspect ratios. By looking at the appendix, I infer that the downscaling is from 0.25 to approximately 0.04 degrees (even if it has never been officially stated). However, it looks like the downscaling differs depending on the axis. Applying the 5.5 magnification factor between the vertical resolutions of CONUS to the horizontal dimension, we should get 237x5.5~1304 and not 1041. Most numbers in Fig A1 are inconsistent.

  11. The qualitative results reported in Fig. A4 and A5 are very pixelated and surely aren’t at a 0.04 degree resolution. I am not even sure they represent the LR inputs. I would suggest showing close-ups. Evaluating downscaling outputs on a whole country/continent is impossible.

  12. As mentioned in the weaknesses, it looks like some comments in the appendix are highlighting the incompleteness of the paper, for instance: • There are “pending” paragraphs (line 970 and 977) • Lines 899 to 903 seem a reminder to specify some hparams as well as details regarding baselines (which I agree, are also needed). • If I understand correctly, according to line 1381, it looks like the reported results are not even those of the final model.


References

[i] LeCun, Y. (2022). A path towards autonomous machine intelligence.

[ii] Maes, L., Lidec, Q. L., Scieur, D., LeCun, Y., & Balestriero, R. (2026). Leworldmodel: Stable end-to-end joint-embedding predictive architecture from pixels.

[iii] Chen, X., Guo, H., Guo, S., Jiang, B., Shen, C., Shen, X., ... & Zhou, M. (2026). A Definition and Roadmap for World Models.

[iv] Ha, D., & Schmidhuber, J. (2018). Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.