acceptodds
Withdrawn from ICLR 2027

Weather Text as Observation for Meteorological Field Generation via a Unified Multimodal Model

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Natural-language weather descriptions have a long history as meteorological records but have rarely been exploited as an information source for atmospheric state inference, particularly for reconstructing meteorological fields. Directly conditioning field generators on weather text, however, provides weak constraints because descriptions are ambiguous and are sparsely aligned with gridded meteorological observation. We address this problem by converting weather descriptions into quantitative meteorological pseudo-observations before field generation. To support this formulation while providing dense supervision, we construct MeteoText, a large-scale dataset that aligns daily city-level weather descriptions with meteorological variables and fields, and define five tasks spanning weather-text understanding and field generation. We further propose MeteoOmni, a unified multimodal model with a Value Probe for text-to-value mapping and Text-Derived Observation Guidance for injecting the resulting psedo-observations as local constrains during generation inference. Experiments show that MeteoOmni consistently outperforms strong baselines across all five tasks and generalizes zero-shot to real-world meteorological warnings. More importantly, we demonstrate that weather texts can serve as pseudo-observations to improve local generation, with their benefits extending to the full meteorological field.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
weather and climate, generative modeling, multimodal learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EqMDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MeteoText, a dataset aligning weather descriptions with meteorological variables and fields, and proposes MeteoOmni, a unified multimodal model for weather understanding and field generation. The method converts weather descriptions into numerical pseudo-observations using a Value Probe and incorporates them into field generation through TDOG.

Critical strengths
  1. The idea of using weather descriptions as observational constraints for meteorological field generation is interesting and has potential applications in historical climate reconstruction.
  2. MeteoText provides a relatively large dataset covering different geographic regions and meteorological conditions, supporting five understanding and generation tasks.
  3. The paper includes ablation studies examining the contributions of joint training, the Value Probe, and TDOG.
Critical weaknesses
  1. Since the model is trained on synthetic data, it remains unclear whether it can handle the inherent ambiguity and complexity found in natural language descriptions of weather.
  2. The "Value Probe" and TDOG methods proposed in the paper primarily combine numerical regression supervision with particle-based resampling; however, the paper fails to provide a sufficiently compelling argument regarding the novelty of these approaches compared to existing methods.
  3. The evaluation in real-world scenarios is limited to 739 weather warnings across 22 cities in China and covers only two tasks—a scope insufficient to support the authors' claims regarding the model's broader generalization capabilities.
  4. As shown in Table 4, the improvements gained from using real-world weather texts are marginal (the root mean square error, or RMSE, for temperature decreased from 1.759 to 1.747, and for precipitation from 7.737 to 7.693), and the paper lacks the statistical analysis necessary to verify the reliability of these differences.
  5. The "field generation" baseline model employed in the study differs significantly from MeteoOmni in terms of both architecture and capacity; furthermore, the paper fails to compare the proposed model against other methods that operate under the same constraints—namely, the use of identical textual or numerical observational data.
  6. Although the paper cites "historical climate reconstruction" as a key research motivation, the experimental section does not incorporate actual historical weather records; this weakens the evidence supporting the claim that the model holds practical, real-world utility.
Questions for authors

See above

5 comments
AuthorsOpenReview ↗

Answer of W1 To Reviewer EqMD

Dear Reviewer EqMD,

Thank you for your valuable time and feedback. We hope our clarifications address your concerns.

[W1]: Ambiguity and Complexity in Natural Weather Language

[A1]: We sincerely thank the reviewer for raising this concern. We address ambiguity and complexity from two aspects.

Ambiguity. We account for ambiguity during synthetic data construction. Specifically, we define variable-specific numerical ranges based on national meteorological standards to guide qualitative text generation. Thus, descriptions correspond to numerical intervals rather than exact values. The grading schemes are provided below.

Table. ERA5 wind grading based on the Beaufort scale.

LevelWind speed range (m/s)Label
0[0, 0.2]Calm
1(0.2, 1.5]Light Air
2(1.5, 3.3]Light Breeze
3(3.3, 5.4]Gentle Breeze
4(5.4, 7.9]Moderate Breeze
5(7.9, 10.7]Fresh Breeze
6(10.7, 13.8]Strong Breeze
7(13.8, 17.1]Near Gale
8(17.1, 20.7]Gale
9(20.7, 24.4]Strong Gale
10(24.4, 28.4]Storm
11(28.4, 32.6]Violent Storm
12> 32.6Hurricane

Table. ERA5 precipitation grading based on 24-hour accumulated precipitation.

LevelPrecipitation range (mm/day)Label
0< 0.1Trace Rain
1[0.1, 10.0)Light Rain
2[10.0, 25.0)Moderate Rain
3[25.0, 50.0)Heavy Rain
4[50.0, 100.0)Torrential Rain
5[100.0, 250.0)Severe Torrential Rain
6≥ 250.0Extreme Torrential Rain

Table. Standardized-anomaly grading for ERA5 temperature and CMIP6 variables.

LevelStandardized anomaly rangeLabel
0z < −2.0Extremely / Anomalously Low
1[−2.0, −1.5)Significantly Low
2[−1.5, −0.5)Slightly Low
3[−0.5, 0.5]Normal
4(0.5, 1.5]Slightly High
5(1.5, 2.0]Significantly High
6z > 2.0Extremely / Anomalously High

Complexity. We evaluate MeteoOmni using real-world weather texts and different text generators to examine its adaptability beyond synthetic descriptions.

  • Real-world text transfer. We evaluate MeteoOmni on CMA weather warnings without additional training or fine-tuning. We conduct zero-shot T5 evaluation using real-world warnings paired with MIROC6 fields under SSP2-4.5 from 2021–2023.
SettingCity-t2m RMSE (1×1) ↓City-tp RMSE (1×1) ↓City-t2m RMSE (5×5) ↓City-tp RMSE (5×5) ↓
w/o real text2.9379.5233.1449.475
with real text2.2448.0782.6208.828
RMSE Reduction (%)23.61%15.17%16.66%6.83%

Real-world text reduces temperature and precipitation RMSE by 23.61% and 15.17% at 1×1, and 16.66% and 6.83% at 5×5, respectively.

  • Text generator sensitivity. For T1, we replace GPT-5.1-generated ground truth (GT) with GPT-5.5-generated GT while keeping predictions unchanged. For T2 and T3, we replace GPT-5.1-generated inputs with GPT-5.5-generated texts while keeping numerical targets and model parameters unchanged.

T1: City Weather QA

MeteoOmni maintains broadly comparable performance across both GT settings and consistently outperforms GPT-5.5 across all three metrics.

ModelTraining GT GeneratorEvaluation GT GeneratorROUGE-L ↑METEOR ↑BERTScore ↑
MeteoOmniGPT-5.1GPT-5.10.5310.5330.926
MeteoOmniGPT-5.1GPT-5.50.4880.5360.905

T2: CMIP6 Text Quantification

MeteoOmni maintains comparable quantification performance under both input generators.

ModelInput Text GeneratorSettingA-T ↑A-W ↑A-P ↑A-O ↑
MeteoOmniGPT-5.1In-domain0.8690.8490.7130.810
MeteoOmniGPT-5.5Zero-shot0.8890.8470.6920.809

T3: ERA5 Text Quantification

MeteoOmni also maintains comparable performance on T3 under both generators.

ModelInput Text GeneratorSettingA-T ↑A-U ↑A-V ↑A-P ↑A-O ↑
MeteoOmniGPT-5.1In-domain0.8720.7010.7020.3790.664
MeteoOmniGPT-5.5Zero-shot0.8840.6930.6950.3600.655

Overall, MeteoOmni shows good robustness to test-time text generator changes. Its performance remains broadly stable across T1, T2, and T3, demonstrating adaptability to different weather text expressions.

Together, these experiments demonstrate MeteoOmni's adaptability to real-world descriptions and variations in text generation styles.

AuthorsOpenReview ↗

Answer of W2&3&4 To Reviewer EqMD

[W2]: Novelty of Value Probe and TDOG

[A2]: We sincerely thank the reviewer for raising this important concern regarding the novelty of our proposed methods.

We acknowledge that numerical regression and particle-based resampling are established techniques.

Our main contribution is a new framework that transforms natural language weather descriptions into spatially localized, station-like pseudo-observations and uses them to constrain meteorological field generation.

To realize this framework, we introduce Value Probe and TDOG as complementary technical components. The key innovation of TDOG lies in treating text-derived information as station-like observations and using local observation consistency to guide particle selection for whole-field generation. This design allows sparse, spatially localized textual observations to influence the generation of complete meteorological fields, rather than merely correcting individual grid cells.

To the best of our knowledge, this is the first attempt to use text-derived, station-like pseudo-observations as local physical constraints to guide the generation of complete meteorological fields through particle sampling.

[W3]: Real-World Evaluation and Generalization

[A3]: We sincerely thank the reviewer for highlighting the limited scope of our real-world evaluation.

We acknowledge that the real-world evaluation is limited to 739 weather warnings from 22 cities in China and two tasks.

To further extend the evaluation, we conduct an additional real-world text experiment on T5 using a different CMIP6 model, MIROC6, under the SSP2-4.5 scenario for 2021–2023. We pair the daily CMIP6 fields with contemporaneous real-world weather warnings for zero-shot evaluation.

SettingCity-t2m (1×1) ↓City-tp (1×1) ↓City-t2m (5×5) ↓City-tp (5×5) ↓
w/o real text2.93749.52253.14429.4751
with real text2.24408.07762.62038.8280
RMSE Reduction (%)23.61%15.17%16.66%6.83%

As shown in the table, real-world text reduces the RMSE of both temperature and precipitation across all spatial ranges. The improvements on MIROC6 under SSP2-4.5 demonstrate that text-guided correction remains effective under zero-shot cross-model and cross-scenario transfer.

[W4]: Marginal Improvements and Statistical Reliability

[A4]: We sincerely thank the reviewer for pointing out the need for further statistical validation of the observed improvements.

The gains from real-world text depend on the background field. For T4, the unmasked ERA5 regions already provide reliable spatial context, so text mainly serves as supplementary local observational constraints, resulting in relatively modest improvements.

To assess the reliability of these improvements, we repeat the experiments using 5 different random seeds. We report the mean and standard deviation and paired statistical tests.

SettingTemperature RMSE ↓Precipitation RMSE ↓
w/o real text1.7330 ± 0.01647.6830 ± 0.0402
with real text1.7179 ± 0.01737.6457 ± 0.0385
p-value0.00050.0482

The results show statistically significant improvements in both temperature (p = 0.0005) and precipitation (p = 0.0482), supporting the effectiveness of real-world text.

AuthorsOpenReview ↗

Answer of W5&6 To Reviewer EqMD

[W5] Baseline Selection and Comparability

[A5]: We sincerely thank the reviewer for raising this important concern regarding baseline selection and comparability.

We adopt meteorology-specific models as field generation baselines, following WeatherGFM [1]. A similar evaluation practice is also used in OmniWeather [2], which compares BAGEL-based unified model with task-specific models.

Our baselines use the same textual information and observations under matched input conditions, compared to our MeteoOmni.

We acknowledge the differences in architecture and model capacity. Therefore, these comparisons aim to evaluate the task performance of unified models against specialized models, rather than isolate the effects of architecture or model size.

[1] WeatherGFM (ICLR-25)

[2] OmniWeather (ICLR-26)

[W6] Historical Climate Reconstruction and Practical Applications

[A6]: We sincerely thank the reviewer for highlighting the gap between our research motivation and the current experimental validation.

We agree that our current experiments do not directly validate meteorological field reconstruction from historical weather archives. Historical climate reconstruction is a key motivation and potential application of our framework.

However, historical archives often lack explicit correspondence between weather descriptions and quantitative observations. Moreover, qualitative descriptions need to be interpreted against the climate conditions of their respective historical periods, requiring both paired text–observation data and period-specific climatological references.

Our real-world text experiments demonstrate two essential capabilities: extracting quantitative meteorological information from natural language and using text-derived observations to constrain field reconstruction. These capabilities provide a foundation for extending our framework to historical climate reconstruction. We will clarify this distinction in the revised manuscript and identify validation on historical archives as an important direction for future work.

Reviewer_bXiTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MeteoOmni, a unified multimodal model that converts natural-language weather descriptions into quantitative pseudo-observations to guide meteorological field generation. To train MeteoOmni, the paper introduces MeteoText, a large dataset that aligns weather text with meteorological values and fields. The experimental results show that MeteoOmni consistently outperforms strong baselines across five tasks and generalizes zero-shot to real-world meteorological warnings.

Critical strengths

The paper is well-presented and easy to follow.

The paper's framing of treating weather text as quantitative pseudo-observations for field generation seems novel and introduces a promising data generation pipeline given the scarcity of weather observations.

The proposed MeteoOmni model unifies weather-text understanding and field generation, the experimental results seem promising and generalizes zero-shot to real-world warnings.

Critical weaknesses

The main concern is the grid size not matching the city scale setup of the dataset. As each city is reduced to a single cell of the ~1.4° (128 × 256) grid, yet the paper never addresses the representativeness error between city-level descriptions and grid-cell values. Many cities are much smaller than a single cell, so coastal, mountainous and urban effects are smoothed out, and nearby cities may share a cell. The “city-level” CMIP6 Correction metric is also computed over 5 × 5 cells (~700 × 700 km). This limits the claimed benefit for local constraints, and in particular for reconstruction from historical local weather documents.

To my knowledge, there are many CMIP6 models from various institutions, and they have different systematic biases. Studies typically use multi-model ensembles for calibration and bias correction. Only using CNRM-CM6-1 as CMIP6 Correction may therefore learn model-specific biases, and it is unclear whether the gains carry over to other CMIP6 models.

Questions for authors

Please refer to the strengths and weaknesses. Additionally, how sensitive is MeteoOmni to the choice of text generator? For example, could GPT-5.5 performance improve if the ground truth text generator is also GPT-5.5, especially on text-only tasks such as City Weather QA?

4 comments
AuthorsOpenReview ↗

Answer of W1 To Reviewer bXiT

[W1]: Grid–City Spatial Mismatch

[A1]: We sincerely thank the reviewer for highlighting the spatial mismatch between grid cells and city-scale weather descriptions. We acknowledge that the 1.4° grid cannot fully resolve local meteorological variations caused by coastal, mountainous, and urban effects. Our goal is to use city-associated weather descriptions to constrain and correct meteorological fields at the available grid resolution, rather than to reconstruct fine-scale urban weather patterns.

Regarding the spatial overlap, we verify that each of the 314 selected cities is mapped to a unique grid cell, with no shared cells.

To further evaluate the effectiveness of local constraints, we conduct additional T5 experiments using the 1×1 center grid. Compared with the no-text setting, synthetic text reduces RMSE by 24.10%, 39.07%, 35.89%, and 17.24% for t2m, U10, V10, and tp, respectively. These improvements are larger than those obtained in the 5×5 evaluation for temperature and wind.

The detailed results are as follows:

RegionSettingt2m RMSE ↓U10 RMSE ↓V10 RMSE ↓tp RMSE ↓
1×1w/o synthetic text3.4913.3193.35028.036
1×1with synthetic text2.6492.0222.14823.203
1×1RMSE Reduction (%)24.10%39.07%35.89%17.24%
5×5w/o synthetic text3.3243.5603.39813.074
5×5with synthetic text2.6752.6402.53410.691
5×5RMSE Reduction (%)19.52%25.84%25.43%18.23%

We will clarify that the 5×5 metric measures reconstruction quality over a city-centered region (~700 × 700 km), rather than the city itself. We will also replace "city-level RMSE" with "city-centered regional RMSE" to better reflect its spatial scope.

AuthorsOpenReview ↗

Answer of W2 To Reviewer bXiT

[W2]: Cross-Model Generalization

[A2]: We sincerely thank the reviewer for raising this important concern.

We agree that different CMIP6 models exhibit distinct systematic biases and that multi-model ensembles are valuable for bias correction. To evaluate cross-model generalization, we test our model, originally trained on CNRM-CM6-1 historical data, on daily data from another CMIP6 model, MIROC6, without additional training. The evaluation includes historical data from 2012 and SSP2-4.5 scenario data from 2021–2023.

Synthetic-text evaluation. We first evaluate our model on the 2012 synthetic-text test set. The results show that text-guided correction improves both global and city-centered reconstruction. The improvements are particularly substantial at the city-centered scales. These results demonstrate that the learned text-guided correction is not limited to CNRM-CM6-1.

RegionSettingt2m RMSE ↓U10 RMSE ↓V10 RMSE ↓tp RMSE ↓
Globalw/o text3.9424.7354.7656.038
Globalwith text3.8354.6844.7165.930
GlobalReduction (%)2.71%1.07%1.02%1.79%
1×1w/o text3.0632.3922.10724.977
1×1with text2.3461.6291.50520.526
1×1Reduction (%)23.42%31.90%28.59%17.82%
5×5w/o text3.0453.0032.65411.390
5×5with text2.3442.1812.10510.191
5×5Reduction (%)23.02%27.37%20.68%10.53%

Real-world text evaluation. We further evaluate the model using daily MIROC6 SSP2-4.5 fields from 2021–2023, paired with contemporaneous real-world weather warnings. No additional training or fine-tuning is performed. The results show consistent RMSE reductions across all three spatial scales.

RegionSettingt2m RMSE ↓tp RMSE ↓
1×1w/o text2.9379.523
1×1with text2.2448.078
1×1Reduction (%)23.61%15.17%
5×5w/o text3.1449.475
5×5with text2.6208.828
5×5Reduction (%)16.66%6.83%
10×10w/o text3.3249.478
10×10with text2.9969.219
10×10Reduction (%)9.89%2.74%

Overall, these experiments demonstrate that text-guided correction generalizes to MIROC6 without additional training. The improvements are observed with both synthetic and real-world weather texts, supporting the cross-model generalization capability of our framework.

AuthorsOpenReview ↗

Answer of Q1 To Reviewer bXiT

[Q1]: Sensitivity to Text Generator

[A1]: We sincerely thank the reviewer for raising this important question.

The original synthetic weather texts are generated by GPT-5.1. To evaluate sensitivity to text generators, we additionally use GPT-5.5 to regenerate test texts for zero-shot evaluation on T1, T2, and T3. We examine both the effect of changing the evaluation references and the effect of changing the input text generator.

T1: City Weather QA. We replace the original GPT-5.1-generated ground truth (GT) with GPT-5.5-generated GT while keeping model predictions unchanged. GPT-5.5 achieves higher ROUGE-L and METEOR scores with its own GT, but a lower BERTScore. MeteoOmni shows a similar metric-dependent pattern. Its METEOR score remains comparable, while ROUGE-L and BERTScore decrease under GPT-5.5-generated GT. Nevertheless, MeteoOmni consistently outperforms GPT-5.5 across all three metrics.

ModelTraining GT GeneratorEvaluation GT GeneratorROUGE-L ↑METEOR ↑BERTScore ↑
GPT-5.5N/AGPT-5.10.2970.2730.879
GPT-5.5N/AGPT-5.50.3050.2950.871
MeteoOmniGPT-5.1GPT-5.10.5310.5330.926
MeteoOmniGPT-5.1GPT-5.50.4880.5360.905

T2: CMIP6 Text Quantification. We replace GPT-5.1-generated input texts with GPT-5.5-generated texts while keeping the meteorological conditions unchanged. MeteoOmni maintains comparable quantification performance under both generators. It also consistently outperforms GPT-5.5.

ModelInput GeneratorSettingA-T ↑A-W ↑A-P ↑A-O ↑
GPT-5.5GPT-5.1—0.7410.5540.5670.621
GPT-5.5GPT-5.5—0.7840.5680.5540.669
MeteoOmniGPT-5.1In-domain0.8690.8490.7130.810
MeteoOmniGPT-5.5Zero-shot0.8890.8470.6920.809

T3: ERA5 Text Quantification. We conduct the same evaluation on T3 using the two sets of input texts. MeteoOmni also maintains comparable quantification scores under both generators.

ModelInput GeneratorSettingA-T ↑A-U ↑A-V ↑A-P ↑A-O ↑
GPT-5.5GPT-5.1—0.8090.0200.1050.3280.315
GPT-5.5GPT-5.5—0.8140.0000.0000.3680.295
MeteoOmniGPT-5.1In-domain0.8720.7010.7020.3790.664
MeteoOmniGPT-5.5Zero-shot0.8840.6930.6950.3600.655

Overall, MeteoOmni shows good robustness to test-time text generator changes. GPT-5.5 shows only modest improvements on some metrics with its own generated texts, with no consistent gains across all metrics.

AuthorsOpenReview ↗

Kindly Requesting for Review of Our Response

Dear Reviewer bXiT,

We are eager to ensure that we have adequately addressed your concerns and are prepared to offer further clarifications or address any additional questions you may have.

Should you find that our revisions have satisfactorily addressed your previous concerns, we would be most grateful if you would reconsider the evaluation of our paper with a view to enhancing its standing.

We would like to express our heartfelt gratitude for the time and effort you have dedicated to reviewing our work. It has been a pleasure to engage with you throughout this process.

Best regards,

Authors

Reviewer_Z7hBOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is quite interesting. It asks whether weather text can provide local observational constraints for meteorological field generation. It proposes to ground city-level weather descriptions into quantitative meteorological variables as pseudo-observations, and then use these pseudo-observations as explicit numerical constraints during generation. To support this, the authors construct MeteoText, a spatiotemporally indexed dataset that jointly aligns daily city-level weather descriptions, corresponding meteorological values, and meteorological fields, using the ERA5 and CMIP6, and define five understanding and generation tasks. They also propose MeteoOmni, a unified multimodal model with Value Probe for text-to-value grounding and Text-Derived Observation Guidance (TDOG) for injecting pseudo-observations during inference. The paper claims that text-derived pseudo-observations can provide local constraints that improve masked ERA5 reconstruction and CMIP6 correction at city level.

Critical strengths
  1. The paper addresses a novel and interesting problem. Instead of treating weather text merely as a description or evaluation target, it proposes to ground city-level weather descriptions into quantitative meteorological pseudo-observations and then use these as explicit local numerical constraints for meteorological field reconstruction and correction. This “text →\rightarrow quantified pseudo-observation →\rightarrow soft structual constraint for field generation” formulation is conceptually clear and more interpretable than direct text-conditioned generation.

  2. The authors also construct MeteoText, a dataset that aligns weather text, meteorological values, and ERA5/CMIP6 fields, and define five understanding and generation tasks, providing a reusable task framework and evaluation basis. The inclusion of real-world CMA warning text for zero-shot OOD evaluation further strengthens the application relevance of the work.

Critical weaknesses
  1. The actual improvements obtained using real-world weather texts seems marginal. In Table 4, incorporating real text reduces the temperature RMSE from 1.759 to 1.747 and precipitation RMSE from 7.737 to 7.693. Furthermore, Table 3 reports relatively low clamped APE scores for real-text wind quantification, with scores of only 0.1093 and 0.1040 for the two wind components.

  2. The real-world evaluation is also restricted to T3 and T4, without corresponding validation of the T5.

  3. The temporal splitting strategy raises a concern regarding the independence of training and evaluation data across different tasks. The T1 uses ERA5 data from 1979 to 2014 for training, whereas T3, T4, and T5 use 2012 as their test period. Therefore, although each task has a chronological train/validation/test split, the evaluation year for T3–T5 falls within the training period of T1. The authors should clarify whether identical ERA5 fields or overlapping city-date samples occur across task-specific training and test sets.

Questions for authors

See weakness. I will raise the score if my concerns are clarified.

3 comments
AuthorsOpenReview ↗

Answer of W1 To Reviewer Z7hB

Dear Reviewer Z7hB,

Thank you for the valuable time and feedback you have invested in this review. We sincerely hope our further clarification could address your concerns:

[W1-1]: Limited Gains from Real-World Text

[A1-1]: We thank the reviewer for pointing out the modest improvements in the original Table 4. The gains from real-world text depend strongly on the background field.

For T4, the unmasked ERA5 regions already provide reliable spatial context, so text mainly serves as supplementary local observational constraints, resulting in relatively modest improvements. In contrast, the CMIP6 field in T5 exhibits biases relative to ERA5, and text-derived observations provide effective local correction signals, leading to more substantial gains.

With real-world text, T5 also achieves substantial RMSE reductions, as shown below:

SettingCity-t2m (1×1) ↓City-tp (1×1) ↓City-t2m (5×5) ↓City-tp (5×5) ↓
w/o real text2.9379.5233.1449.475
with real text2.2448.0782.6208.828
RMSE Reduction (%)23.61%15.17%16.66%6.83%

[W1-2]: Real-World Wind Quantification

[A1-2]: We thank the reviewer for highlighting the relatively low quantification scores for the two wind components. The main reason is the lack of wind-direction information in real-world CMA texts.

Specifically, CMA real-world texts typically report wind intensity but omit wind direction. However, recovering U10 and V10, the two components of the horizontal wind vector, requires both wind speed and direction.

Therefore, we additionally report wind speed magnitude (W=U102+V102W=\sqrt{U10^2+V10^2}), which achieves an APE-W of 0.427, about 4× the values for U10 (0.109) and V10 (0.104) as follows:

ModelAPE-W ↑
GPT-5.6-Sol0.052
GPT-6-Astra0.109
Ours0.427
AuthorsOpenReview ↗

Answer of W2&3 To Reviewer Z7hB

[W2]: Zero-Shot T5 Evaluation with Real-World Text

[A2]: We sincerely thank the reviewer for pointing out the lack of real-world evaluation on T5. The original T5 dataset uses CNRM-CM6-1 historical fields from 1850 to 2014, which have no temporal overlap with real-world CMA weather warnings from 2021 to 2023. Therefore, direct evaluation was not possible.

To address this limitation, we use daily CMIP6 data from another model, MIROC6, under the SSP2-4.5 scenario for 2021–2023 as input fields. We pair them with contemporaneous real-world weather text for zero-shot evaluation.

SettingCity-t2m RMSE (1×1) ↓City-tp RMSE (1×1) ↓City-t2m RMSE (5×5) ↓City-tp RMSE (5×5) ↓
w/o real text2.9379.5233.1449.475
with real text2.2448.0782.6208.828
RMSE Reduction (%)23.61%15.17%16.66%6.83%

As shown in the table, real-world text reduces temperature and precipitation RMSE across both spatial scales. These results demonstrate the effectiveness of text-guided correction under zero-shot cross-model and cross-scenario transfer.

[W3]: Independence of Training and Evaluation Data

[A3]: We sincerely thank the reviewer for raising this important concern.

To prevent potential data leakage, we removed all T1 training samples overlapping with the 2012 test sets of T3, T4, and T5. All results reported in the paper are obtained using this filtered T1 training set. This ensures that no identical ERA5 fields or overlapping city-date samples are shared between T1 training and the test sets of T3–T5. We will clarify this data filtering procedure in the revised manuscript.

AuthorsOpenReview ↗

Kindly Requesting for Review of Our Response

Dear Reviewer Z7hB,

We are eager to ensure that we have adequately addressed your concerns and are prepared to offer further clarifications or address any additional questions you may have.

Should you find that our revisions have satisfactorily addressed your previous concerns, we would be most grateful if you would reconsider the evaluation of our paper with a view to enhancing its standing.

We would like to express our heartfelt gratitude for the time and effort you have dedicated to reviewing our work. It has been a pleasure to engage with you throughout this process.

Best regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comments and Revision Summary

We sincerely thank all the reviewers for their insightful reviews and helpful suggestions, as well as for acknowledging the importance, comprehensiveness, and contributions of our work.


We would like to re-emphasize the main contribution of our work. MeteoOmni establishes a new framework that transforms natural language weather descriptions into spatially localized, station-like pseudo-observations and uses them as local physical constraints to guide meteorological field generation. Unlike direct text-conditioned generation, our framework connects quantitative information extracted from text with the iterative generation process through Value Probe and TDOG. Together with MeteoText, which aligns weather descriptions, meteorological variables, and ERA5/CMIP6 fields across five tasks, our work provides a new pathway for text-guided meteorological field reconstruction and correction.


We have revised our paper and provide a summary of the additional experiments and clarifications below:

  1. In response to Reviewer Z7hB and Reviewer EqMD, we have extended real-world text evaluation to T5. We use MIROC6 data under SSP2-4.5 for 2021–2023, paired with contemporaneous CMA weather warnings, for zero-shot evaluation. Real-world text reduces temperature and precipitation RMSE across both 1×1 and 5×5 spatial scales.

  2. Following Reviewer Z7hB's suggestion, we have clarified the relatively low wind-component quantification scores. Real-world CMA texts typically report wind speed but omit wind direction. We therefore additionally evaluate wind speed magnitude, achieving an APE-W of 0.427, approximately four times the scores of U10 and V10.

  3. In response to Reviewer Z7hB's concern about data leakage, we have removed all T1 training samples overlapping with the 2012 test sets of T3, T4, and T5. All reported results use the filtered training set, ensuring independence between training and evaluation samples.

  4. Following Reviewer bXiT's feedback, we have clarified the spatial representativeness limitations of the 1.4° grid. We verified that all 314 selected cities map to unique grid cells. We also added 1×1 center-grid evaluations and clarified that the 5×5 metric represents a city-centered regional evaluation rather than fine-scale city reconstruction.

  5. As suggested by Reviewer bXiT, we have conducted additional cross-model generalization experiments using MIROC6, including historical data from 2012 and SSP2-4.5 data from 2021–2023, without additional training. Both synthetic-text and real-world-text experiments demonstrate zero-shot transfer beyond the original CNRM-CM6-1 model.

  6. To address Reviewer bXiT's concern about text generator sensitivity, we have additionally evaluated MeteoOmni using GPT-5.5-generated texts on T1, T2, and T3. The results show broadly comparable performance across different text generators, demonstrating robustness to variations in text generation.

  7. In response to Reviewer EqMD's concerns about ambiguity and novelty, we have clarified the variable-specific numerical grading schemes based on meteorological standards. We have also further explained the role of Value Probe and TDOG. In particular, TDOG treats text-derived information as station-like observations and uses local observation consistency to guide particle selection for whole-field generation.

  8. Following Reviewer EqMD's suggestion, we have conducted additional experiments using five random seeds and reported the mean and standard deviation, together with paired statistical tests. The improvements in temperature and precipitation RMSE achieve p-values of 0.0005 and 0.0482, respectively, providing statistical evidence for the effectiveness of real-world text.

  9. Regarding baseline selection, we have clarified our use of meteorology-specific field generation models, following WeatherGFM and OmniWeather. We also explain that our comparisons use matched textual information and observations, while acknowledging differences in model architecture and capacity.

  10. Regarding historical climate reconstruction, we have clarified the distinction between our current experimental validation and this broader application. We further discuss the challenges of historical archives, including the lack of paired text–observation data and the need for period-specific climatological references. We identify validation on historical archives as an important direction for future work.


Lastly, we would like to express our sincere gratitude to all reviewers for their constructive feedback. Their comments have helped us improve the experimental evaluation, clarify the methodological contributions, and strengthen the presentation of our work. We sincerely hope that our revisions and additional experiments adequately address their concerns.

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

Related papers

Discussion (0)

Sign in to comment.