acceptodds
Under review as a conference paper at ICLR 2027

Textual Environmental Context and Spatial Graphs for LLM-Based Regional SST Forecasting

Abstract

Sea surface temperature (SST) forecasting depends on local temporal persistence, regional spatial dependence, and environmental conditions that evolve with the forecast date. We study how these heterogeneous conditions can be presented to a large language model (LLM) for regional multi-step forecasting without serializing the full SST grid as text. We formulate forecasting as conditional numerical generation: historical SST and anomaly sequences, date-aligned environmental records, and static ocean knowledge form a textual context, while regional spatial state is supplied through continuous graph-derived prefixes. A static graph encodes persistent geographic–climatological relations, and a dynamic graph encodes recent SST correlations and localized tropical-cyclone influence. Two graph neural networks produce a target-node representation that is mapped by a spatial-prefix fusion and injected into the LLM input. On SST forecasting in the South China Sea, the complete configuration achieves the best MAE and among the compared methods over ten forecast steps. Alongside the numerical forecast, a rule-based module matches predicted trends and environmental-factor directions with knowledge entries to return source-linked, post-hoc contextual explanations.

Primary area:
learning on time series and dynamical systems
Keywords:
SST prediction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_MiLiOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an LLM-based framework for sea surface temperature forecasting. It combines historical SST and environmental information as textual inputs with spatial features extracted from static and dynamic graphs. The graph representations are converted into prefix tokens and fed into the LLM for multi-step forecasting. Experiments on the South China Sea show improvements over several baselines.

Critical strengths
  1. The paper's motivation is clear, and it presents a reasonable method to combining environmental text and spatial information for SST forecasting.

  2. The proposed method achieves competitive performance against several baselines.

Critical weaknesses
  1. The authors evaluate the effect of removing the entire GNN module, but they did not consider the static and dynamic GNNs separately. It is unclear how much each graph actually contributes to the performance.

  2. Although the proposed method outperforms SST forecasting baselines, it is still unclear whether the LLM is necessary. In addition, the experiments are conducted on the South China Sea dataset, it would be helpful to evaluate the method on other ocean datasets.

Questions for authors
  1. Could the authors provide more details about the dataset, including training/validation/test splits, and the number of samples in each split?

  2. Could the authors compare their method with a non-LLM model using the same environmental information and graph features?

  3. Have the authors considered evaluating the static and dynamic GNNs separately?

  4. Could the authors provide additional results on other ocean regions to demonstrate that the method generalizes beyond the South China Sea?

1 comment
AuthorsOpenReview ↗

Response to Reviewer MiLi

Thank you for finding the motivation clear, the combination of environmental text and spatial information reasonable, and the forecasting results competitive.

W1 / Q3. Separate static and dynamic graphs

The table reports daily results for dual-graph Ours and the static-only model.

ModelMetricDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10Mean
Ours (dual graph)RMSE0.2120.3220.3810.4240.4570.4850.5120.5350.5530.5660.445
MAE0.1310.2160.2690.3060.3350.3580.3780.3960.4130.4270.323
R²0.9910.9790.9700.9630.9560.9500.9430.9370.9310.9270.955
Static-onlyRMSE0.1710.2740.3470.4040.4530.4960.5330.5670.5970.6240.447
MAE0.1170.1990.2580.3050.3450.3790.4090.4360.4610.4840.339
R²0.9930.9810.9700.9600.9500.9400.9300.9210.9130.9040.946

On the ten-day means, Ours has RMSE 0.445 versus 0.447 for static-only (0.002, or 0.447%, lower), MAE 0.323 versus 0.339 (0.016, or 4.720%, lower), and R² 0.955 versus 0.946 (0.009 higher).

The dynamic-only and no-graph experiments are in progress; their results will be included in the revised version.

W2 / Q2. The LLM and other regions

We added the same-input non-LLM control E1. The daily results are:

ModelMetricDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10Mean
OursRMSE0.2120.3220.3810.4240.4570.4850.5120.5350.5530.5660.445
MAE0.1310.2160.2690.3060.3350.3580.3780.3960.4130.4270.323
R²0.9910.9790.9700.9630.9560.9500.9430.9370.9310.9270.955
E1 (non-LLM regressor)RMSE0.5790.5680.5540.5520.6000.7250.8020.8580.8340.7820.685
MAE0.4940.4710.4470.4370.4770.5940.6660.7150.6830.6250.561
R²0.9180.9210.9240.9250.9110.8710.8420.8200.8290.8500.881

On the reported ten-day means, Ours outperforms E1 on all three metrics: RMSE is lower by 0.240 (35.036%), MAE by 0.238 (42.424%), and R² is higher by 0.074. Ours is also better on each of the ten days for all three metrics.

The East China Sea evaluation is in progress; its results will be included in the revised version.

Q1. Dataset splits and sampling

Each sample combines a forecast origin and an ocean cell. It contains ten historical days and ten target days. The full 20-day window stays within its split.

SplitObservation dates
Training1981-09-01 to 2016-12-31
Validation2017-01-01 to 2020-12-31
Test2021-01-01 to 2025-02-16

The training, validation, and test sets contain 212,350, 23,280, and 5,308,285 samples, respectively.

Training and validation use feature-coverage sampling. The features include spatial blocks, coastal/open-ocean location, cyclone influence, temperature, and trends. Each day, a greedy procedure selects cells to cover these feature groups.

The purpose is to reduce training and validation costs. Adjacent cells and overlapping time windows contain similar information, so exhaustive enumeration produces many similar examples. Feature-coverage sampling aims to retain different regions, cyclone conditions, and temperature-change patterns while reducing repeated computation.

Q4. East China Sea evaluation

The East China Sea domain is defined as latitude [28°,32°] and longitude [123.25°,129.25°] at 0.25° resolution, containing 384 ocean cells in total. The time splits are the same as for the South China Sea.

The East China Sea experiment is in progress; results will be included in the revised version.

Reviewer_nrs8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an LLM-based framework for regional multi-step sea surface temperature (SST) forecasting, formulated as conditional numerical generation for one target cell and one forecast origin at a time. The prompt contains the SST and anomaly histories of the target cell, a static ocean-knowledge text and date-aligned environmental records, while the regional spatial state is supplied through two graphs over the grid cells. The static graph has distance-based edge weights and node features given by the coordinates and the SST climatology, and the date-specific dynamic graph mixes a top-kk SST correlation graph with a localized tropical-cyclone influence graph. Two independent two-layer GCNs encode these graphs, and the target-node representation is mapped by a factorized MLP projector to a short sequence of continuous prefix tokens that are appended to the prompt embeddings. The GCNs, the projector, and LoRA adapters of Qwen2.5-7B-Instruct are trained jointly with the standard autoregressive objective, and the future SST values are generated as text. A separate rule-based module then matches the directions of the forecast and of four factors with a fixed set of knowledge entries to return source-linked post-hoc explanations. Experiments on OISST data over the South China Sea, report the best mean MAE and R2R^2 and a slightly lower mean RMSE than eight baselines, together with a comparison across four LLM backbones, ablations of the text sources and of the spatial branch, and two qualitative explanation examples.

Critical strengths

S1. The proposed method fuses different data modalities into one framework, and the authors are able to engineer it in a neat way. The roles of the two conditioning channels are clearly separated, i.e., date-aligned environmental records and static ocean knowledge enter as text, while the regional spatial state enters through the dual-graph modeling, which avoids serializing the SST field as text and keeps the prompt length independent of the number of grid cells. The localized graph is also a good way to encode the spatial footprint of storms, which the text alone does not provide.

S2. The paper is well-written and easy to follow. The method section gives explicit formulas for every graph and layer together with the main fine-tuning hyperparameters.

Critical weaknesses

W1. Fusing spatiotemporal information with the use of GNNs and different neural networks (e.g., CNNs, RNNs, Transformers, etc.) is an interesting avenue for tackling problems where there are signals in multiple axes (space, time, image, text, etc.). Nonetheless, this is not a new idea, nor are the authors able to design a new method/framework where the gains are notable or there is a novel analysis. The pipeline is mostly assembled from existing building blocks, e.g., applying GCNs over distance, correlation, or environmental phenomena graphs, soft prompts computed from a GNN representation as in GraphToken (Perozzi et al., 2024), a prompt organized following Wang et al. (2024), and LoRA fine-tuning (Hu et al., 2022). Works such as Cao et al. (2020), Netto et al. (2022), and Barros et al. (2024) have already been working with such ideas and demonstrated the effectiveness and robustness of such modeling, providing important theoretical and empirical analysis. Spatiotemporal GNN forecasting also has a long history, e.g., Li et al. (2018) and Wu et al. (2019), including for geophysical fields (Lam et al., 2023). But apart from Kipf and Welling (2017), every work discussed in Section 2 appeared in 2024 or later, most of them in 2025 or 2026. It would be worth having discussions on such papers, as well as others that appeared before 2025, in the related work section.

W2. Though the authors claim to achieve better performance compared to the tested baselines, this is hard to verify given that there are no standard deviations, and the results apparentely come from a single seed. That is important, especially given that the results are fairly close when compared with the baselines, and we cannot guarantee any statistical significance. The ablations of Table 3 raise a similar concern, since removing the static knowledge text, which, according to the text in the main body (l.209), is the same text KK for every sample and thus carries no sample-specific information, even improves RMSE. This suggests that the differences in Table 3 may be within run-to-run variability. Finally, the controls that would calibrate the gains are missing. The No GNN variant of Table 3, which receives no spatial or environmental information beyond the histories and coordinates of the target cell, already outperforms every baseline except PCFNet in mean RMSE and MAE, so persistence and climatology baselines are needed to show how much of the skill comes from local persistence, as is an LLM-free variant of the same pipeline to show what the LLM adds. LLM-based forecasters discussed in Section 2 are not compared either.

W3. Many details are left out of the paper. For instance, what is the observation/acquisition frequency of the data? What is the time interval between prediction steps? It is also not explained how the sources are aligned in time. Neither a complete prompt nor the text KK is shown, and the decoding strategy, the handling of malformed generations, and the training and inference costs are never described. Several constants of the graph construction, as well as the kernel parameters (Eq. (5)), are fixed without justification or sensitivity analysis, and the correlations (Eq. (4)) are estimated from only ten values per node, with the neighbors selected among all ocean nodes, which makes the dynamic graph very sensitive to noise. The notation could also be improved, since Fs\mathbf{F}^s and Ftd\mathbf{F}^d_t are used in l.154-155 but only defined in Section 3.3, at,i,ja_{t,i,j} in Eq. (10) is never defined, α\alpha is a scalar in l.282 but belongs to R48\mathbb{R}^{48} in Figure 1, and Figure 1 uses TT and MM where the text uses LL and NN, while also using LL for the number of Transformer blocks and MM for the prompt length. There are also typos, e.g., "esponse" (l.289), "Figure. 1" (l.190), and missing spaces in l.266-267.

Questions for authors

Q1. Could the authors add, with several seeds and on the same test samples for all rows, persistence and climatology baselines, an LLM-free variant in which the dual-GNN representation and the numerical histories feed a small regression head, and at least one of the LLM-based forecasters of Section 2, such as Time-LLM or OKG-LLM? The persistence baseline is particularly relevant given that the RMSE of the full model is well below that of every baseline at the first lead but grows much faster up to the second one.

Q2. What do the authors regard as the main methodological novelty, and what does the language model contribute? Dynamic graphs for SST forecasting are already used by DDSTGNN (Jin et al., 2025), soft prompts computed from GNN representations follow GraphToken (Perozzi et al., 2024), and OKG-LLM (Yang et al., 2026) already aligns ocean knowledge with SST observations in an LLM-based predictor. Could the authors explain which parts of the framework are new with respect to these works and to those in W1, and why this information requires a large language model rather than a spatiotemporal network that receives the same covariates?

References

  • Barros et al. “Early Detection of Extreme Storm Tide Events Using Multimodal Data Processing.” Proceedings of the AAAI Conference on Artificial Intelligence, 38(20), 21923–21931. (2024). https://doi.org/10.1609/aaai.v38i20.30194

  • Cao et al. “Spectral Temporal Graph Neural Network for Multivariate Time-series Forecasting.” In Proceedings of the 34th International Conference on Neural Information Processing Systems (NIPS '20). Curran Associates Inc., Red Hook, NY, USA, Article 1491, 17766–17778. (2020).

  • Lam et al. “Learning Skillful Medium-Range Global Weather Forecasting.” Science, 382(6677), 1416–1421. (2023).

  • Li et al. “Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting.” International Conference on Learning Representations (ICLR). (2018).

  • Netto et al. “Modeling Oceanic Variables with Dynamic Graph Neural Networks.” AI: Modeling Ocean and Climate Change @ International Joint Conference on Artificial Intelligence (IJCAI-ECAI). (2022)

  • Wu et al. “Graph WaveNet for Deep Spatial-Temporal Graph Modeling.” Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI), 1907–1913. (2019).

2 comments
AuthorsOpenReview ↗

Response to Reviewer nrs8

Thank you for highlighting the value of spatiotemporal and multimodal forecasting, and for identifying relevant earlier work. Your comments help us clarify the method's contribution and experimental evidence.

W1. Methodological contribution and relation to earlier work

We do not claim a new GNN operator. Related designs include graph diffusion and learned adjacency (DCRNN, Graph WaveNet), spectral graph-temporal learning (StemGNN), dynamic graphs for irregular ocean observations (Netto et al.), graph fusion of data streams (Barros et al.), mesh-based weather forecasting (GraphCast), and GNN soft prompts for graph QA (GraphToken).

Our contribution is an SST-specific framework combining a static geographic/climatological graph with a date-dependent graph of recent SST correlations and cyclone influence. It maps the target cell's dual-graph representation to 32 continuous prefix tokens and jointly conditions SST generation on this spatial state, historical SST, and environmental text; the GNNs, projector, and LoRA adapters are trained together. A detailed comparison with these works will be given in the revision.

W2. Controls and statistical evidence

We added a same-input non-LLM control, adapted OKG-LLM and Time-LLM, and report the submitted static-text ablation. Their ten-day means:

Model / settingRMSEMAER²
Ours0.4450.3230.955
E1 (non-LLM regressor)0.6850.5610.881
OKG-LLM0.4720.3560.940
Time-LLM0.4710.3550.940
No static text0.4370.3300.949

Ours beats E1 by 0.240 RMSE (35.036%) and 0.238 MAE (42.424%), with R² 0.074 higher. The two adapted LLMs are close and both outperform E1, while Ours has the best displayed means.

W3. Data and implementation details

SST is daily, with one day between forecast steps. Cyclone best tracks are aligned by UTC date, and ENSO and regional 850-hPa winds use complete-month values assigned to each day in that month.

Current strict evaluation retains the raw text and scores ten finite values as generated; incomplete or invalid outputs prevent a complete-test score. Historical processing could select a numerical window, replace or pad values, and round predictions.

Graph weights are 0.6/0.4, with k = 10. For C = πR_E/180 ≈ 111.195 km, the static scale is C and cutoff 1.5C. For wind speed v in knots, R = C·clip(2√(v/64), 1.5, 5) and σ² = −R²/(2 ln 0.1). Wind uses the 99th percentile, 140 kt; each UTC day retains the maximum-wind record. These settings control neighbors, spatial weights, and wind normalization. We will define symbols at first use, distinguish prefix scaling, prompt length, and Transformer layer count, and correct the listed typos.

Complete inference-prompt example

Due to the character limit here, the complete dual-graph text input will be provided in the revised manuscript.

Q1. Matched controls and the day-2 error increase

E1 and the two adapted LLM baselines are in W2. The adapted LLMs have lower RMSE on days 1–3 whereas Ours leads on days 4–10; Ours leads on MAE and R² from day 3 onward.

For Ours, RMSE / MAE / R² move from 0.212 / 0.131 / 0.991 on day 1 to 0.322 / 0.216 / 0.979 on day 2: RMSE rises 0.110 (51.887%), MAE 0.085 (64.885%), and R² falls 0.012. OKG-LLM and Time-LLM show similar RMSE increases (0.104, 0.111), while E1's RMSE drops 0.011. The three LLM forecasts generate the full sequence in one call, so each later day is conditioned on earlier values; the longer lead and possible propagation of earlier errors are plausible contributors.

Q2. Differences from DDSTGNN, GraphToken, and OKG-LLM

  • DDSTGNN uses a dynamic graph for SST forecasting. We combine a geographic/climatological static graph with a date-specific SST-correlation and cyclone graph, then condition an LLM with environmental text.
  • GraphToken projects graph representations into a frozen LLM for GraphQA. We use the target-cell readout from two graphs and jointly train the GNNs, projector, and LoRA for daily SST generation.
  • OKG-LLM aligns SST patches with structural and textual embeddings from an ocean knowledge graph for global weekly prediction. Our inputs include cyclone, ENSO, wind context, and date-specific SST-grid relations, and we generate daily temperatures as text tokens.
Reviewer_nrs8OpenReview ↗

Reply to Authors

Thank you for the response. I appreciate the authors’ efforts to clarify their work, but my main concerns remain. In particular, it still lacks a discussion of prior work, and the stated contribution does not fully address my concerns about novelty. The experimental results also do not yet establish statistical significance. Finally, the proposed revisions are described as future changes, but are not reflected in the current version. For these reasons, I will maintain my original assessment.

Reviewer_KP3SOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work presents a regional sea surface temperature (SST) forecasting model using a finetuned large language model (LLM) with a dual graph neural network (GNN) multimodal input adaptor. Additionally, a rule-based module is used to provide text explanations of forecasts by matching temperature slopes and other factors against 52 pre-written knowledge entries.

Critical strengths

This work engineered a way to use LLM with a GNN adaptor for regional SST forecasting.

Critical weaknesses
  • Unfair comparison with baseline models: this work uses extra information (tropical cyclone record and ENSO context) compared with baseline models. And it is very likely, although not possible to verify due to lack of source code, the 7B-LLM model used in this work is much larger than other baseline models. The advantage of this work shown in Table 1 may come from a) extra information , or b) more parameters, instead of from architectural improvements.
  • The rule-based explanation module is a confusing addition: it applies to any SST forecast and it is not evaluated at all. Section 4.5 and Figure 2 only present some examples of the explanations and contribute little to the model and benchmark.
  • Practicality issue: this work tries to do 10 day regional (South China Sea) SST forecasting without specifying boundary conditions. As Tropical Cyclone (TC) can greatly affect SST, we can imagine a case where a TC is 1000km away (well outside the 3deg buffer) from the forecast region at t0 and move into the region during the 10day period. Therefore, this work will have a hard time to correctly predict SST without the knowledge of TC far away at t0. This issue is not just limited TC: other upwind weather phenomena can also greatly affect SST in the forecast region.
  • Unclear definition of symbols: A, F, c_t, L, H, K, a_tij, w_tij.
Questions for authors

I have a hard time in understanding the formulation of autoregressive forecast in both token level and timestep level. How does the LLM output decode to the SST at each cell i?

2 comments
AuthorsOpenReview ↗

Response to Reviewer KP3S

Thank you for recognizing our GNN–LLM SST forecasting design and identifying where the comparisons and methods need clarification.

W1. Input information and model capacity

We use a pretrained LLM to incorporate prior oceanographic knowledge through environmental text and jointly process it with SST history and graph-derived spatial information. The choice of an LLM is motivated by this role, rather than by its parameter count. Our contribution is an integrated framework for regional SST forecasting, not a new neural-network architecture.

W2. Interpretability module

Users of ocean forecasts need to compare predicted changes against physical understanding to judge whether the results are reasonable. Clare et al. compare SHAP and LRP explanations of ocean-dynamics predictions with physical theory, and discuss the role of explanations in scientific judgment and credibility. Clare et al., 2022.

Similar interpretability analyses already exist in SST forecasting. Liu et al. use LRP to locate the sources of North Atlantic SST predictability. Hao et al. use LRP to analyze the contributions of SST, heat flux, and wind stress at different forecast leads. Liu et al., 2023, Hao et al., 2026.

After forecasting, our interpretability module matches temperature trends and environmental factors to 52 knowledge entries with literature sources. It provides traceable oceanographic context, helping researchers compare predicted changes with processes such as mixing, advection, and air–sea exchange.

The interpretability module is a supplementary component; our main contribution is a forecasting framework that improves SST prediction accuracy.

W3. Earlier cyclone information

To examine the effect of using earlier cyclone information, we keep the SST history and forecast period fixed, and use only the cyclone observation from one day before the history window begins. We have run this experiment; its result is not yet available and will be reported when finalized.

W4. Symbols

  • A and F denote adjacency matrices and node features. s and d identify static and dynamic graphs.
  • t is the forecast origin. c_i is cell i's coordinate.
  • L and H are history and forecast lengths, both ten days. K is the shared static text.
  • P is the prefix length, set to 32. α scales it and starts at 1.
  • w_{t,i,j} is token j of sample (t,i). a_{t,i,j} is its binary loss mask.

Q1. From generated tokens to daily SST

One greedy call generates ten comma-separated SST values for the target cell. Each value may span several tokens. The h-th value represents day t+h.

For a July 10 origin, the first two values represent July 11 and 12. Tokens are generated autoregressively in order, so later values can attend to earlier ones. The entire ten-day sequence is generated in a single call.

Strict scoring saves raw text and requires ten finite values. The general response provides the complete prompt and shared static text. The 32 continuous graph tokens enter separately. Response is empty at inference.

Thank you for your response. My concern on boundary conditions remains unsolved. And the paper lacks an ablation study to decouple the effect of using current architecture and using more input data. Therefore, I will maintain my rating.

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

The paper studies regional multi-step sea surface temperature (SST) forecasting in the South China Sea as a conditional numerical generation task using a large language model (LLM). Historical SST and anomaly sequences for a target cell, date-aligned environmental records, and static ocean knowledge form a textual prompt. Regional spatial state is not serialized as text; instead, a static geographic–climatological graph and a date-specific dynamic graph are encoded by two independent 2-layer GCNs. The target-node vector is mapped by a factorized projector to continuous prefix tokens, which are concatenated with prompt embeddings for autoregressive generation using Qwen2.5-7B-Instruct with rank-8 LoRA. A separate rule-based module maps forecast slope and factor directions to 52 knowledge entries for post-hoc, source-linked explanations. Evaluation on the 0.25° OISST dataset shows that the proposed method achieves good performance.

Critical strengths

S1. A clean, time-aligned text–graph interface that avoids serializing the full grid. By presenting per-date TC/ENSO/circulation information together with static mixing/advection/upwelling knowledge as text, and by injecting only the target node’s dual-GNN state as 32 prefix tokens, the prefix length remains independent of N. This is a consequential and well-scoped solution to the question of how to jointly supply forecast-date environmental context and date-varying spatial relations to an LLM.

S2. A date-specific dynamic graph that combines correlation and TC influence, with honest lead-wise reporting. Encoding synchronous history-window correlation together with a finite-support TC kernel is domain-appropriate for the South China Sea, and the paper transparently reports both the lead-dependent pattern and the backbone trade-off.

Critical weaknesses

W1. The reported gains are marginal, internally inconsistent, and lack statistical support. The central SOTA claim therefore does not hold as stated. Relative to PCFNet, the proposed method reduces mean RMSE only from 0.448 to 0.445 (−0.003, −0.7%) and increases R² only from 0.953 to 0.955 (+0.002), while PCFNet attains lower RMSE at 5 of the 10 lead times (steps 2–6). Moreover, Table 3 shows that No static text achieves RMSE 0.437, better than Full at 0.445; i.e., adding static knowledge improves MAE/R² but degrades RMSE. This directly contradicts the abstract/conclusion claims of the “lowest mean RMSE, MAE and highest R² among compared methods” and that the “complete configuration achieves best MAE and R².”

W2. Several core design choices are heuristic and lack empirical validation, and the claimed interpretability of the “explanation” module is not verified. The graph weights 0.6/0.4, k = 10, l_s, the R_o formula, the v_99/130 kt fallback, and the seven dynamic features are all fixed without sensitivity analysis. No ablation is provided for static-GCN-only vs. dynamic-GCN-only vs. both, or for the projector, and no sensitivity analysis is reported for P = 32, d_m = 64, or α initialized at 1. In addition, the most relevant baseline, OKG-LLM, is discussed but not numerically compared, while the remaining baselines are not current LLM-based forecasters (Time-LLM, TimeCMA, LangTime).

Questions for authors
  1. Please explain why No static text achieves a lower RMSE (0.437) than Full (0.445). For Tables 1–3, report mean ± std across at least three random seeds and include paired significance tests. In addition, disaggregate the results by TC-active vs. TC-inactive dates and by coastal vs. open-ocean cells. Does static knowledge improve only MAE/R² while worsening RMSE? If so, where does this occur?
  2. Please report: (i) static-GCN-only vs. dynamic-GCN-only vs. both; (ii) the same text and graph features with a lightweight non-LLM decoder (e.g., an MLP/RNN head) vs. Qwen/Llama; and (iii) sensitivity to prefix length P and α. Also report inference time, memory usage, and trainable-parameter count compared with PCFNet.
  3. Please justify the 0.6/0.4 weighting, k = 10, and the R_o/σ_o formulas. Why retain only the maximum-wind observation per date? Provide performance conditioned on storm distance and intensity to demonstrate that the branch is necessary.
  4. Please add direct comparisons with OKG-LLM and at least one LLM-based time-series method (e.g., Time-LLM, TimeCMA, or LangTime) under the identical 10-in/10-out South China Sea protocol. Also clarify the source of the “corresponding SST benchmark table” used for Table 1.
1 comment
AuthorsOpenReview ↗

Response to Reviewer mqou

Thank you for recognizing the text–graph interface, target-cell prefix design, cyclone-aware dynamic graph, and lead-wise reporting.

W1. Size and consistency of the gains

The improvement is more pronounced on day 1. Using the values displayed in submitted Table 1, RMSE falls from PCFNet's 0.273°C to 0.212°C, a 22.344% reduction. MAE falls from 0.207°C to 0.131°C, a 36.715% reduction.

Mean RMSE / MAE / R² over ten days are 0.445 / 0.323 / 0.955 for Ours and 0.448 / 0.344 / 0.953 for PCFNet. Ours has lower RMSE on days 1 and 7–10, while PCFNet has lower RMSE on days 2–6. Ours has lower MAE on nine days, with a tie on day 3. Its R² is higher on days 1 and 6–10, tied on day 5, and lower on days 2–4.

In the submitted backbone comparison, Llama 3.1 achieves 0.434 / 0.326 / 0.950. Its mean RMSE is lower, while Qwen has better mean MAE and R². We will state conclusions by comparison, metric.

W2. Design choices and supplementary control

The 0.6/0.4 weights on the dynamic graph correspond to recent-SST-correlation edges and cyclone-event edges, respectively. This is a reasonable setting that balances temperature correlations with cyclone disturbances. The planned 0.5/0.5 control keeps other conditions unchanged to test performance after adjusting the mixing weights. Its result will be added when completed.

We added a same-input non-LLM control, adapted OKG-LLM and Time-LLM, and report the submitted static-text ablation. Their ten-day means:

Model / settingRMSEMAER²
Ours0.4450.3230.955
No static text0.4370.3300.949
OKG-LLM (adapted)0.4720.3560.940

Overall, Ours achieves the best performance and outperforms OKG-LLM significantly across all three metrics (RMSE and MAE are lower by 0.027 and 0.033 respectively, while R² is higher by 0.015). Meanwhile, No static text also surpasses OKG-LLM in all metrics.

Seven features describe SST level, trends, anomalies, and variability. They are mean, standard deviation, trend, seven-day mean, full-window anomaly mean, seven-day anomaly mean, and range.

P = 32 sets graph-prefix length. d_m = 64 limits factorized projection width. α starts at 1 and learns prefix amplitude. These are the submitted settings.

Q1. Static text and robustness

Static text supplies general process context without sample-specific observations. It also changes prompt length and context. Existing results do not isolate these effects.

Q2. Graph contributions, LLM comparison, and cost

All configurations retain full text. Scores are ten-day means.

ModelGCNRMSEMAER²
QwenBoth0.4450.3230.955
QwenStatic0.4470.3390.946
QwenDynamic0.4500.3410.945
Llama 3.1Both0.4340.3260.950
E1 (direct MLP)Both0.6850.5610.881

Static-only RMSE is 0.003 lower than dynamic-only. With the same cases and input information, Llama's RMSE is 0.251 lower than E1's. Ours leads on MAE and R².

P=32. We learn α, initialized at 1. Qwen Full has 20,690,977 trainable parameters (503,425 in the projector). E1 has 535,274.

Q3. Graph constants and cyclone records

The mix is 0.6/0.4. k = 10 sets correlation-neighbor count.

Let C = πR_E/180 ≈ 111.195 km. The static distance scale is C, with cutoff 1.5C. For wind speed v in knots, R = C·clip(2√(v/64), 1.5, 5) and σ² = −R²/(2 ln 0.1). The Gaussian kernel equals 0.1 at R.

Wind speed is normalized by the event table's 99th percentile, 140 kt. Selecting the maximum-wind observation per UTC day represents the daily peak. This design is intended to reduce computational cost and simplify the cyclone input: keeping one record per UTC day limits the amount of cyclone data used in graph construction while retaining the peak-intensity information. It discards other storms and intraday track changes.

The planned 0.5/0.5 control tests graph weights. Other constants are explained through their computational roles above.

Q4. Additional baselines and the source of Table 1

The mean RMSE / MAE / R² are 0.472 / 0.356 / 0.940 for adapted OKG-LLM and 0.471 / 0.355 / 0.940 for adapted Time-LLM.

Both use our full text, SST history, and dual graphs for the South China Sea 10-in/10-out task. They retain their respective temporal reprogramming, normalization, numerical heads, and denormalization paths. The OKG adaptation replaces the external knowledge-graph resources with the project inputs.

The eight non-LLM baselines in submitted Table 1 come from PCFNet's Table II. We will explicitly mark the source.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Additional Results

We thank all reviewers for their careful reading and constructive feedback. We appreciate the positive assessments of our motivation, text–graph design, and SST forecasting results.

Completed additions are static-only, the non-LLM regressor E1, and adaptations of OKG-LLM and Time-LLM using our inputs. Pending experiments cover dynamic-only, no-graph, 0.5/0.5 weights, earlier cyclone information, and the East China Sea. Results will be added as completed.

Table R1. Supplementary SST forecasting results

The task uses 10 historical days to predict the next 10 days. RMSE and MAE are in °C. Mean is the arithmetic average of the ten daily metrics.

ModelMetricDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10Mean
OursRMSE0.2120.3220.3810.4240.4570.4850.5120.5350.5530.5660.445
MAE0.1310.2160.2690.3060.3350.3580.3780.3960.4130.4270.323
R²0.9910.9790.9700.9630.9560.9500.9430.9370.9310.9270.955
Static graph onlyRMSE0.1710.2740.3470.4040.4530.4960.5330.5670.5970.6240.447
MAE0.1170.1990.2580.3050.3450.3790.4090.4360.4610.4840.339
R²0.9930.9810.9700.9600.9500.9400.9300.9210.9130.9040.946
E1 (non-LLM regressor)RMSE0.5790.5680.5540.5520.6000.7250.8020.8580.8340.7820.685
MAE0.4940.4710.4470.4370.4770.5940.6660.7150.6830.6250.561
R²0.9180.9210.9240.9250.9110.8710.8420.8200.8290.8500.881
OKG-LLMRMSE0.1800.2840.3690.4290.4790.5220.5620.6020.6330.6630.472
MAE0.1250.2070.2740.3210.3590.3940.4260.4590.4840.5080.356
R²0.9920.9800.9670.9550.9440.9330.9230.9110.9020.8920.940
Time-LLMRMSE0.1750.2860.3690.4270.4830.5220.5620.5980.6290.6580.471
MAE0.1200.2090.2750.3190.3630.3940.4260.4550.4800.5040.355
R²0.9920.9800.9660.9550.9430.9330.9230.9120.9030.8940.940

E1 uses the same SST history, environmental text, and dual graphs. OKG-LLM and Time-LLM are adapted to these inputs while retaining their numerical prediction paths. The OKG-LLM adaptation uses the project's graphs in place of the original external knowledge graph.

Static-only retains the full text and only the static graph. Dynamic-only and no-graph results will follow.

Among the values reported in Table R1, the submitted Ours reference has the best mean metrics: RMSE / MAE / R² of 0.445 / 0.323 / 0.955. Its reported mean RMSE is lower than the tabulated E1, adapted OKG-LLM, and adapted Time-LLM values by 35.036%, 5.720%, and 5.520%, and its mean MAE by 42.424%, 9.270%, and 9.014%, respectively.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.