acceptodds
Under review as a conference paper at ICLR 2027

ARGO: A Regional-Global Joint Method for Region-Flexible Weather Forecasting

Abstract

Regional weather forecasting is crucial for many domains such as transportation, energy and disaster prevention. Existing data-driven methods typically perform global forecasting directly or build weather forecasting systems tailored to specific regions. However these paradigms present two key limitations: (1) generalization across regions with different geographical locations and spatial extents, and (2) filtering relevant global information to guide regional prediction. To address these challenges, we propose ARGO, a regional–global joint method for region-flexible weather forecasting. To support regions with varying locations and spatial extents, ARGO combines Geo-aware Position Encoding (GaPE), random regional sampling, and an auxiliary alignment loss to learn transferable forecasting patterns and compatible geographical representations. To filter global information, a Global Token Router (GTR) selects relevant nearby and remote tokens while retaining broad global context, followed by latent-space cross-attention for regional–global fusion. Experiments at 0.25 resolution show that ARGO outperforms state-of-the-art baselines in regional forecasting and generalizes across different regions without retraining from scratch.

Primary area:
learning on time series and dynamical systems
Keywords:
Deep Learning, Spatiotemporal Analysis, Weather Forecasting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ADihOpenReview ↗
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
  1. ARGO combines Geo-aware Position Encoding, random regional sampling, and an auxiliary alignment loss to support cross-region generalization.
  2. 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
  1. 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.
  2. 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.
  3. 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.
  4. 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
  1. Could you explain in more detail how GaPE encodes geographical relationships and how this design supports cross-region generalization?
  2. Could you provide an ablation removing the auxiliary alignment loss, along with an analysis of whether it improves regional–global representation alignment?
  3. 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
AuthorsOpenReview ↗

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)1+\cos(\alpha-\beta), 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.

T2m (72 h)U10 (72 h)MSLP (72 h)Z500 (72 h)
ConfigurationRMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑
w/o auxiliary alignment loss1.730.841.950.79184.860.94180.030.96
w auxiliary alignment loss1.640.851.830.81178.200.95175.120.96

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.

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.

Reviewer_8otTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ARGO, a regional-global joint framework for region-flexible weather forecasting. It aims to address cross-region generalization and relevant global information filtering. The method introduces Geo-aware Position Encoding (GaPE), random regional sampling, an auxiliary alignment loss, and a Global Token Router (GTR) with selective cross-attention. Experiments on ERA5 data at 0.25° resolution demonstrate state-of-the-art regional forecasting performance and zero-shot generalization to unseen regions.

Critical strengths
  1. Novel Formulation for Region-Flexible Forecasting: The combination of GaPE and random regional sampling elegantly addresses the longitude periodicity and cross-region generalization issues, which are often overlooked in fixed-domain regional models. This provides a solid foundation for arbitrary-region deployment.

  2. Efficient Global-Regional Fusion Mechanism: The Global Token Router (GTR) provides a computationally efficient and conceptually interesting way to filter remote atmospheric forcings (teleconnections) without dense cross-attention over the entire global grid, effectively balancing computational cost and global context retention.

Critical weaknesses

1. Insufficient Ablation Studies: Only Module-Level, Lacking Hyperparameter and Method-Level Analysis

  • The ablation experiments in Table 3 are strictly module-level (w/o GTR, w SAA, w/o two-step fine-tuning). No direct quantitative comparison between GaPE and standard RoPE (Su et al., 2024) on forecasting metrics (RMSE/ACC) is provided. Although Figure 4 visually illustrates the longitude periodicity advantage of GaPE, this remains a qualitative observation without empirical validation on downstream prediction performance.

  • Critical hyperparameters of GTR, such as the number of nearby tokens KnK_n (currently 64), remote tokens KfK_f (currently 128), background tokens ∣B∣|B| (currently 32), and the number of subregions MM (currently 16, i.e., 4×44\times4 layout), are stated in Appendix E.2 (p. 14) but never varied in any experiment. No sensitivity analysis is provided to demonstrate robustness.

  • The two-step fine-tuning strategy is ablated only in a binary manner (with vs. without). No comparison with longer autoregressive fine-tuning horizons (e.g., 4-step, 8-step) is conducted.

2. Non-Standard Experimental Protocol: 9-Year Training Split Deviates from Community Standards

The paper uses only 2010–2018 for training, 2019 for validation, and 2020 for testing, totaling merely 9 years of training data. In contrast, all cited baselines, Pangu-Weather (Bi et al., 2023), Stormer (Nguyen et al., 2024), and the WeatherBench 2 benchmark (Rasp et al., 2024), use approximately 40 years of ERA5 data (typically 1979–2015/2017) for training. The degraded baseline performance is clearly visible: Pangu-Weather's Z500 4-day RMSE in Table 1 (p. 7) is 658.60, which is substantially worse than the metrics reported in its original publication (Bi et al., 2023) using 40 years of training data. Similarly, Stormer's aggregate performance falls well below its published results (Nguyen et al., 2024).

3. Absence of Computational Complexity and Efficiency Comparison with Baselines

The paper provides only an asymptotic complexity comparison for selective attention versus O(NR⋅NG⋅d)O(N_R \cdot N_G \cdot d) for dense attention. No empirical wall-clock training time, inference latency, peak GPU memory, or FLOPs are reported for ARGO or any baseline. No comparison of computational cost is made against the baselines (Pangu-Weather, Stormer, OneForecast, STCast) under identical hardware conditions.

4. Motivation: Ultra-High-Resolution Global Models Challenge the Necessity of Regional-Global Coupling

The paper motivates regional forecasting by arguing that "increasing the global resolution will incur substantial computational and memory costs over the entire globe". However, recent works such as FengWu-GHR and CorrDiff (Mardani et al., 2025) have demonstrated the feasibility of ultra-high-resolution (≤0.25° or finer) global or near-global forecasting. The paper does not discuss or compare against these emerging approaches.

5. Subregion Pooling Destroys Fine-Grained Regional Details; GTR Lacks Physical Interpretability

GTR first mean-pools regional tokens within each 4×44\times4 subregion to compute routing scores (xˉR,r=1∣Ir∣∑i∈IrxR,i\bar{x}_{R,r} = \frac{1}{|I_r|}\sum_{i \in I_r} x_{R,i}). This pooling inevitably destroys fine-grained spatial details within each subregion, which is particularly detrimental for localized extreme weather events (e.g., convective storms, orographic precipitation). The global branch operates at 1.5° resolution, which is already extremely coarse relative to the 0.25° regional grid (approximately 36× fewer grid points). The additional area-weighted pooling into 32 background tokens further compresses this already sparse information.

The "content relevance" scoring in GTR is computed via a learned bilinear form between subregion features and global token features, but no physical interpretability analysis is provided. There is no visualization or case study showing whether GTR actually selects meteorologically meaningful remote tokens versus spurious statistical correlations.

Questions for authors

See Weakness

9 comments
AuthorsOpenReview ↗

Responses to Reviewer 8otT

We sincerely appreciate Reviewer 8otT’s recognition of our contributions and thoughtful suggestions. Your concerns about ablation studies, standard evaluation and computational efficiency are provided in the following contents.

W1: Insufficient Ablations

We compare GaPE with standard 2D RoPE and no rotary encoding, vary GTR’s nearby, remote, and background token counts and subregion layout, and compare two-step and four-step autoregressive fine-tuning. Except for the two-step and four-step fine-tuning variants, all results are obtained after single-step regional training without multi-step fine-tuning. Both fine-tuning variants start independently from the same Stage 2: Regional Training (one-step) checkpoint. The table below reports 72-hour RMSE and ACC.

T2m (72 h)U10 (72 h)MSLP (72 h)Z500 (72 h)
VariantsRMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑
ARGO1.640.851.830.81178.200.95175.120.96
w standard RoPE1.690.851.890.80183.680.94178.070.96
w/o rotary encoding1.670.851.860.81180.780.95176.160.96
GTR (32, 64, 16, 2×2)1.740.841.960.79185.510.94181.790.96
GTR (128, 256, 64, 8×8)1.760.841.980.78187.270.94181.970.96
T2m (72 h)U10 (72 h)MSLP (72 h)Z500 (72 h)
VariantsRMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑
ARGO1.640.851.830.81178.200.95175.120.96
w two-step fine-tuning1.610.861.810.82174.700.95169.180.96
w four-step fine-tuning1.590.861.800.81175.490.95169.410.96

Positional encoding. GaPE achieves the lowest 72-hour RMSE across all four variables. RoPE’s weaker performance reflect its lack of longitude periodicity. Without rotary encoding, GTR still provides geographical guidance through distance-based nearby-token selection, potentially explaining the competitive results.

GTR configuration. The default configuration (64,128,32,4×4) performs best among the tested settings. Coarser pooling obscure local information needs, while smaller token budgets limit global context. Finer pooling provides less routing context, and larger budgets admit less relevant candidates. These jointly varied settings assess the combined effect of routing granularity and token budget.

Multi-step fine-tuning. Both two-step and four-step fine-tuning improve RMSE across all four variables. Four steps yield slightly lower T2m and U10 RMSE but higher MSLP and Z500 RMSE and lower U10 ACC. The four-step model starts directly from the one-step checkpoint, requiring adaptation to a longer prediction chain without an intermediate two-step fine-tuning stage.

W2: Training and Evaluation Protocol

Both the shorter training period and the restriction to regional inputs lead to the gap from the originally reported global forecasting results. As described in Section 4.1 (lines 374–377), Pangu-Weather and Stormer are trained directly on regional inputs and targets because they do not include a dedicated regional–global coupling module. Their regional forecasts therefore lack the evolving global background available in their original global forecasting settings, which can contribute to weaker performance, particularly at longer lead times.[1] Their reported results here are consequently not directly comparable with those in the original publications. All retrained models use the same training, validation, and test years, ensuring comparison fairness.

To evaluate Pangu-Weather in its original global forecasting setting, we downloaded the standard global forecasts from WeatherBench 2 and cropped them to the same East Asian domain used to evaluate ARGO. The Pangu-weather's results provide forecasts initialized at 00 and 12 UTC. We used temporal interpolation to obtain forecasts for the intermediate 06 and 18 UTC cycles, matching the four daily evaluation times used for ARGO. The comparison of RMSE and ACC between these East Asian Pangu-Weather forecasts and ARGO is shown below.

Pangu (WB2)ARGO
VariableLead timeRMSE ↓ACC ↑RMSE ↓ACC ↑
U101 day1.230.9151.030.94
Z5001 day69.780.99360.661.00
U104 days1.850.8062.070.76
T8504 days1.490.9061.710.88

ARGO performs better on 1-day U10 and Z500. Pangu-Weather, trained on fine-resolution global data that capture atmospheric interactions across the entire globe, performs better on 4-day U10 and T850. Given the different training data and evaluation preprocessing, this additional comparison serves as a complementary reference rather than a matched benchmark.


W3: Computational Efficiency

We benchmarked ARGO and the two baselines using a single NVIDIA RTX 4090 with batch size 1 and FP32 precision. The table reports 6h step latency, latency and peak GPU memory for a 72h forecast, and a lower bound on 6h-step FLOPs.

Model6h step latency (ms) ↓72h forecast latency (s) ↓Peak GPU memory for 72h (MiB) ↓FLOPs per 6h step (GFLOPs, lower bound) ↓
ARGO67.870.8712,315583.12
ARGO (w/o GTR)54.080.7062,312580.56
OneForecast137.571.6451,3241,165.70
STCast101.951.2243,502321.08

ARGO produces a 72-hour forecast in 0.871 s, faster than STCast (1.224 s) and OneForecast (1.645 s). Removing GTR reduces latency to 0.706 s with nearly unchanged peak memory.

GTR reduces cross-attention pairs but adds routing, top-k selection, and token-gathering costs. At the tested grid sizes, these overheads and irregular memory access outweigh the savings compared with regular batched attention over all global tokens.

W4: Motivation

We should clarify that Fengwu is a high-resolution global forecasting model, while CorrDiff is a downscaling approach. Compared with ARGO’s regional–global joint forecasting, they represent different approaches to obtaining high-resolution regional weather predictions.

FengWu-GHR operates globally at 0.09° resolution with over 4b parameters, retaining computation over the entire globe even when only a specific region is of interest. For hardware context, the original FengWu reports training on 32 NVIDIA A100 GPUs with 80GB memory each, whereas our experiments use 8 RTX 4090 GPUs with 24 GB memory each. ARGO concentrates fine-resolution computation on the target region while maintaining a coarse global background. This provides a clear motivation for regional modeling even as high-resolution global forecasting advances. CorrDiff generates high-resolution regional fields by downscaling coarse global inputs. ARGO instead uses both coarse global fields and high-resolution regional initial states to forecast regional weather evolution, whitch is different from CorrDiff. Therefore, our method is a promising approach for region weather forecasting.


W5: GTR Pooling and Interpretability

The pooled regional representation is used to construct a shared candidate set and routing bias; it does not replace the individual regional features used in cross-attention. Each regional token retains its own query and independently weights the candidates, allowing different locations to use the same candidate set differently. Moreover, nearby tokens are retained independently of content scores, while background tokens provide broad global context. These complementary paths reduce reliance on pooled content alone. Pooling therefore affects the granularity of candidate selection, while the subsequent cross-attention preserves token-specific interactions.

To examine whether the selected remote information is meteorologically meaningful, we added a GTR Case Studies subsection in the updated version's Section 4.4 (lines 493–527) and a new visualization in the updated version's Figure 4 (lines 500–515), which overlays the selected tokens and their final cross-attention weights on ERA5 circulation fields. For Typhoon Hagupit, the selected remote tokens include locations along the southern flank of the western North Pacific subtropical ridge, where the 500 hPa height gradient is consistent with midlevel easterlies that form part of the storm’s steering environment. During the East Asian cold surge, the remote selections also include locations near the upstream surface high over western Siberia and adjacent Central Asia. Northerly flow on the eastern flank of this high can carry cold air toward East Asia. These two cases provide qualitative evidence that GTR selects remote regions associated with circulation features relevant to each event.

[1] Hao Chen, Tao Han, Jie Zhang, Song Guo, and Lei Bai. Stcast: Adaptive boundary alignment for global and regional weather forecasting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2026.

Reviewer_muBuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ARGO combines a coarse-resolution global forecaster with a high-resolution regional branch for region-flexible weather forecasting. Its main components are geographical position encoding (GaPE), global token routing (GTR), and a training procedure incorporating random regional sampling. Experiments on ERA5 report improved regional forecasting performance, including zero-shot evaluation on a held-out East Asian domain.

Critical strengths

The paper addresses the useful problem of reusing a regional forecasting model across geographical domains. The reported improvements over the selected baselines and the held-out-region experiment provide promising initial evidence for the proposed framework.

Critical weaknesses
  1. Region-flexibility is insufficiently validated. Sections 3.2.2 and 4.1 explicitly state that all three training settings are evaluated on the same fixed East Asian window. The zero-shot experiment demonstrates transfer to one held-out domain, but there is no reported evaluation across multiple target locations or spatial extents. Random training crops and GTR’s internal subregions do not establish this capability.

  2. The GaPE ablation does not isolate the contribution of positional encoding. Section 4.3 replaces the entire GaPE-based fusion mechanism with SAA, changing both positional treatment and feature interaction. Consequently, Table 3 cannot attribute the performance difference specifically to GaPE. A direct comparison against RoPE within the same cross-attention architecture, using consistent geographical coordinates and otherwise matched configurations, is needed.

  3. Typographical and grammatical issues. Several editorial errors should be corrected: “Eq. equation 13” appears on pp. 13–14 and in Algorithm 1, step 24, and should read “Eq. (13)”; the Appendix E heading “Experiments Details” should be “Experimental Details”; “for arbitrary region” in the Figure 1 caption should be “for an arbitrary region”; and “india” in the IndiaWeatherBench reference should be capitalized as “India.”

Questions for authors
  1. Can the same trained checkpoint be evaluated on additional held-out regions and different window sizes without target-specific fine-tuning, to directly test geographical and spatial-extent generalization?

  2. Can you provide GaPE, geographically aligned RoPE, and no-rotary-encoding comparisons while keeping GTR, attention dimensions, rotary-channel proportions where applicable, and the training protocol unchanged?

3 comments
AuthorsOpenReview ↗

Responses to Reviewer muBu

We would like to sincerely thank Reviewer muBu for recognizing the motivation and contributions of our work. Your questions about regional flexibility and the controlled comparison of positional encodings are addressed in the updated version.

W1, Q1: Regional Flexibility

We evaluate the few-shot checkpoint across four geographically diverse regions with different grid sizes, without target-specific fine-tuning:

  • Europe (30°–65°N, 0°–45°E);
  • Australia (44.75°S–12.75°S, 120.25°–174.75°E);
  • Tropical Africa (15°S–20°N, 10°–55°E);
  • North Atlantic (35.25°–74.75°N, 12.25°W–42.25°E).

The table below reports the aggregated RMSE and ACC for each region.

6-hour1-day3-day7-day
RegionGrid sizeRMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑
Europe141 × 1818.930.9716.670.9444.000.88127.130.59
Australia129 × 2198.730.9713.560.9429.670.8880.920.65
Tropical Africa141 × 1818.540.9211.230.8516.120.7525.470.58
North Atlantic159 × 2199.150.9718.150.9550.710.88151.150.58
East Asia139 × 1899.580.9717.010.9440.160.86102.180.61

Across all regions, ACC ranges from 0.75 to 0.88 at 3 days and remains between 0.58 and 0.65 at 7 days. These results provide quantitative evidence of regional flexibility beyond the original East Asia evaluation. Absolute RMSE varies across regions and should be interpreted alongside ACC because regional variability differs.


W2, Q2: Positional Encoding Ablation

We compare GaPE, geographically aligned RoPE, and no rotary encoding within the same cross-attention architecture. GTR, attention dimensions, and the training protocol remain unchanged. All three variants were evaluated after Stage 2: Regional Training (one-step), without two-step fine-tuning.

T2m (72 h)U10 (72 h)MSLP (72 h)Z500 (72 h)
VariantsRMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑RMSE ↓ACC ↑
No rotary encoding1.670.851.860.81180.780.95176.160.96
Geographically aligned RoPE1.690.851.890.80183.680.94178.070.96
GaPE (ours)1.640.851.830.81178.200.95175.120.96

GaPE achieves the lowest 72-hour RMSE across all four variables. RoPE’s weaker performance reflect its lack of longitude periodicity, which can distort spatial relationships across the longitude seam. Without rotary encoding, GTR still retains nearby tokens by spherical distance and selects remote tokens by weather features, potentially explaining the competitive performance.


W3: Editorial Corrections

The typographical and grammatical errors you mentioned are corrected in the updated version.

Dear Reviewer muBu,

Thank you for your valuable and constructive feedback, which has inspired further improvements to our paper. We have made substantial efforts to try to successfully address your concerns by evaluating regional flexibility across multiple locations and grid sizes and conducting controlled comparisons of positional encodings. We would be grateful for your feedback on these additional results. If you have any further questions or concerns, please do not hesitate to let us know, and we will respond promptly.

Best regards,

Authors

Dear Reviewer muBu,

As the rebuttal period is coming to an end, we would welcome your final feedback on our responses. If they have resolved your concerns, we kindly ask you to reconsider your score.

Thanks again for your time and consideration.

Best regards,

The Authors

Comments on the paper

1 comment

Dear reviewers,

The authors have submitted their response to your review comments. Please kindly take a moment to read their response and assess whether your concerns have been adequately addressed. Further discussion with the authors to reach a deeper mutual understanding would be greatly appreciated.

Thanks, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.