acceptodds
Under review as a conference paper at ICLR 2027

SpanWorld: A Variable-Span World Model for Vegetation Forecasting

Abstract

World models learn internal representations of environments and their evolution, providing a way to forecast future environmental change. These models need to preserve the predictive value of their internal representations as rollouts progress. Evaluating this capability requires examining continued predictions from generated states alongside direct forecasts from history. To this end, we propose SpanWorld, a remote-sensing world model that learns vegetation dynamics from satellite observations and advances its latent spatial state through weather-conditioned transitions. The model uses given future weather as an external driver, and its subsequent predictions are evaluated against real remote-sensing imagery. SpanWorld combines baseline predictions derived from incomplete historical observations with a dynamic spatial state updated by a shared variable-span transition, supporting direct prediction and partitioned rollout within a finite forecast window. We further introduce a partitioned-path learning mechanism that retains supervision of predictions made directly from the same initial state at every future target date. It also samples temporal partitions and uses observed target images to supervise predictions made by continuing from model-generated intermediate states. On GreenEarthNet's 100-day dense normalized difference vegetation index (NDVI) image forecasting task, we compare SpanWorld with a control using the same architecture. Both models receive supervision of direct predictions at all forecast horizons. Using the same endpoint images, we additionally supervise predictions from partitioned rollouts in SpanWorld and direct predictions in the control. On the IID split, both models roll out to day 50 using one temporal partition unseen during training, then predict the image at each target date from day 55 to day 100 separately from their respective day-50 states. For these subsequent predictions, partitioned-path learning reduces mean squared error (MSE) by approximately 11% relative to the control, while direct-prediction errors over the same period remain similar.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
World Models, Vegetation Forecasting, Remote Sensing, Spatiotemporal Modeling, Variable-Span Transitions, Multi-Step Prediction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tFYaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper proposes SpanWorld, a remote-sensing world model that learns vegetation dynamics from satellite observations and advances its latent spatial state through weather-conditioned transitions. It introduces a partitioned-path learning mechanism that uses observed target images to supervise predictions made by continuing from model-generated intermediate states. On GreenEarthNet, this reduces subsequent-prediction MSE by approximately 11% relative to a same-architecture control, while direct-prediction errors remain similar.

Critical strengths

SpanWorld maintains endpoint accuracy close to direct prediction under unseen temporal partitions, supporting the effectiveness of its shared variable-span transition.

Critical weaknesses
  1. The writing is overly abstract, while key architectural details and comparisons with baselines are deferred to the appendix, making the main text insufficiently informative. These should be summarized in the main text to clarify the method and its performance.

  2. In the metrics reported in Appendix B.2, SpanWorld underperforms Contextformer 6M, whose architecture it adopts for its history encoder, in terms of forecasting metrics across all four splits.

  3. Deep-OSG already includes a training formulation that supervises both direct and composed predictions using observed endpoint states. The paper should clarify the methodological novelty of partitioned-path learning beyond adapting this idea to weather-conditioned vegetation forecasting.

Questions for authors

1.What distinguishes SpanWorld as a world model from a conventional vegetation forecasting model? Beyond this terminology, could you clarify the specific architectural and training contributions introduced over Contextformer?

2.Why does SpanWorld underperform Contextformer 6M, as reported in Appendix B.2? What benefits do the proposed modifications provide that justify this loss in forecasting accuracy?

3.In what application scenarios is a model supporting both direct forecasting and continued prediction from generated intermediate states particularly advantageous? How should one choose between these two prediction modes across different scenarios?

4.How does partitioned-path learning differ from Deep-OSG’s training formulation that supervises both direct and composed predictions using observed endpoint states?

5 comments
AuthorsOpenReview ↗

Response to Reviewer tFYa — Part 1/3: Architecture and contribution

Thank you for your specific comments on the method description, standard forecasting performance, and positioning relative to prior work. We have clarified the connections among state initialization, weather-driven transitions, and prediction outputs, and summarized standard forecasting performance. We also clarify the relationship to Deep-OSG and the respective uses of direct and continued prediction. Below, we address these points with the empirical evidence.

Architecture and empirical contribution

The writing is overly abstract, while key architectural details and comparisons with baselines are deferred to the appendix, making the main text insufficiently informative. These should be summarized in the main text to clarify the method and its performance. What distinguishes SpanWorld as a world model from a conventional vegetation forecasting model? Beyond this terminology, could you clarify the specific architectural and training contributions introduced over Contextformer?

Thank you for pointing out that the main text does not make the architecture and performance sufficiently concrete. A key distinction between SpanWorld and Contextformer is that SpanWorld explicitly constructs and advances a spatial state that can be reused for subsequent prediction, in addition to producing forecasts at target dates from historical inputs. Historical observations establish the initial state; weather over the specified interval drives a shared variable-span transition; and the generated state can then enter the same transition again to predict later observations. The state readout and the baseline forecast jointly form the complete output.

This distinction motivates the central question stated in the Introduction: when training already supervises direct predictions from the same initial state at every future target date, is this sufficient for states generated through partitioned rollout to support accurate subsequent prediction? Equivalently, with direct supervision at all target dates, does the state-transition path through which supervision is applied still affect continued prediction? We therefore evaluate both direct forecasting accuracy and the predictive usefulness of intermediate states formed through different partitions, using subsequent real satellite observations as targets. This explicit modeling and evaluation of state evolution and reuse underlies our world-model framing.

Partitioned-path training retains direct supervision at every forecast date and propagates observed endpoint-image errors through generated states and their subsequent transitions. The matched direct-endpoint control uses the same architecture, initialization, common training settings, and additional supervision endpoints, focusing the comparison on the supervision path. This design distinguishes supervision of a target date from training the generated states along a path through their subsequent predictive use.

Results across three paired seeds further support this design: SpanWorld achieves lower mean continuation MSE on all three evaluated paths, with an 11.94% reduction relative to the direct-endpoint supervision control on the training-unseen 15+35-day partition (Table 1). With directly supervised dates and additional endpoint images held fixed, these results demonstrate improved subsequent prediction from generated states through partitioned-path training, providing concrete empirical support for our state-modeling contribution.

In Section 3.1 of the revised manuscript, we have clarified the connections among historical encoding, state initialization, weather-driven transitions, and the complete prediction output, and added a standard forecasting summary in Section 4.1. Across three seeds, SpanWorld achieves lower mean RMSE on all four test splits than the weather-conditioned ConvGRU and the adapted evaluation of released EO-WM weights. The main text thus presents the concrete method and main findings; full metrics and between-seed variability are provided in Appendix B, and implementation details in Appendix A.

AuthorsOpenReview ↗

Response to Reviewer tFYa — Part 2/3: Forecast accuracy and Deep-OSG

Standard accuracy and the role of the added design

In the metrics reported in Appendix B.2, SpanWorld underperforms Contextformer 6M, whose architecture it adopts for its history encoder, in terms of forecasting metrics across all four splits. Why does SpanWorld underperform Contextformer 6M, as reported in Appendix B.2? What benefits do the proposed modifications provide that justify this loss in forecasting accuracy?

Thank you for asking how standard forecasting accuracy relates to the benefits of the proposed design. To address the accuracy concern, we refined the training configuration and the representation of seasonal and weather conditions within the original state-evolution framework and partitioned-path supervision principle, retrained the models, and evaluated them on the same test splits using the same scoring protocol. The updated three-seed results show that SpanWorld's mean RMSE is close to Contextformer's on all four splits, with relative gaps of approximately 0.4%–2.3%. Full metrics and standard deviations across seeds are reported in Appendix B, Table B2.

The benefit of the proposed design lies in the subsequent use of model-generated states. In the architecture-matched comparison across three paired seeds, SpanWorld has lower mean continuation MSE than the direct-endpoint supervision control on all three evaluated paths, with an 11.94% reduction on the training-unseen 15+35-day path (Table 1). This comparison holds directly supervised dates and additional endpoint images fixed, providing concrete evidence that the supervision path improves continued prediction from generated states.

These two comparisons characterize standard forecasting performance and state-continuation benefits separately; they cannot be directly combined into a single accuracy trade-off. We have not established the specific source of the standard forecasting gap to Contextformer through controlled experiments and therefore do not attribute it to state evolution or partitioned-path training. The revised results show that, at an RMSE level close to Contextformer's, SpanWorld additionally provides a capability for continued prediction from generated states validated by matched experiments. We summarize standard forecasting performance in Section 4.1 and report the core continuation comparison in Section 4.2 and Appendix C.

Relationship to Deep-OSG

How does partitioned-path learning differ from Deep-OSG’s training formulation that supervises both direct and composed predictions using observed endpoint states?

Thank you for raising this comparison. Deep-OSG studies evolution operators for autonomous dynamical systems and includes a formulation that supervises direct and composed predictions using observed endpoint states. SpanWorld differs in its state formulation: it infers latent spatial states from incomplete satellite observation histories, drives a shared variable-span transition with the ordered weather for each interval, and trains the subsequent use of generated states through image prediction. Supervision is applied to the complete image prediction combining the state readout and baseline forecast, while latent intermediate states receive gradients from endpoint-image errors through subsequent transitions.

Partitioned-path learning incorporates this use of states into training. It retains direct supervision at every target date and samples paths with different segment counts and boundaries, training the shared transition to use previously generated states. This design addresses our central question: when all target dates already receive direct supervision, is that sufficient for states generated through partitioned rollout to support accurate subsequent prediction?

We test this question with a matched direct-endpoint control, holding architecture, initialization, training update count, all-date direct supervision, and additional endpoint images fixed, while applying the additional supervision to direct versus partitioned paths. This comparison separates target-date coverage from training generated states through their subsequent use, testing the role of the supervision path within our model.

Results across three paired seeds further validate the effectiveness of this design: SpanWorld achieves lower mean continuation MSE on all three evaluated paths, with an 11.94% reduction relative to the direct-endpoint control on the training-unseen 15+35-day partition (Table 1). The supplementary seed-42 analysis in Appendix C further shows smaller error increases relative to each model’s own direct predictions on all three paths, indicating greater stability in continued prediction from generated states.

In the revised manuscript, we have clarified the relationship to Deep-OSG in Related Work and made the second contribution explicit: training weather-driven state evolution through the subsequent predictive use of generated states.

AuthorsOpenReview ↗

Response to Reviewer tFYa — Part 3/3: Prediction modes

Direct and continued prediction modes

In what application scenarios is a model supporting both direct forecasting and continued prediction from generated intermediate states particularly advantageous? How should one choose between these two prediction modes across different scenarios?

Thank you for asking about application scenarios and mode selection. The two modes allow SpanWorld to support both target-date queries and continuation of a simulated trajectory within the same model. Direct prediction obtains a target-date forecast from the history-initialized state; continued prediction reuses a generated intermediate state and advances it with the weather for the subsequent interval.

One potential application is staged weather-scenario analysis: several trajectories share historical observations and an initial weather sequence, then consider different subsequent conditions, such as persistent drought or rainfall recovery. With direct prediction, each scenario is evaluated from the initial state using its full weather sequence up to the target date. With continued prediction, the latent spatial state formed through the common preceding evolution can be reused and advanced under each subsequent weather sequence. This makes the preceding evolution a reusable simulation result and places subsequent evolution under different weather scenarios on the same generated-state starting point.

Accordingly, direct prediction is appropriate for obtaining target-date outputs under given conditions, whereas continued prediction is appropriate for extending an existing simulation or branching subsequent trajectories from a common state. Our experiments use subsequent real satellite observations to evaluate the predictive ability needed for such state reuse, and matched controls to study how partitioned-path training affects this ability.

Reviewer_T9JTOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes SpanWorld, a weather-conditioned world model for vegetation (NDVI) forecasting on GreenEarthNet. A Contextformer history encoder produces a baseline forecast and an initial state, and a shared transition advances that state over any time span using the interval’s weather. The main idea, partitioned-path learning, trains the model not only to predict each future date directly but also to reach a few endpoints through chained multi-step rollouts. Compared with a control trained only on direct jumps, SpanWorld’s predictions from its own day-50 state are about 11% better, while direct prediction stays the same.

Critical strengths
  1. A clean controlled comparison: SpanWorld and the control share the same architecture, initialization, training budget, and endpoint images, and the statistics (paired geographic bootstrap) are careful.
  2. A reasonable question and transparently studied: Whether a model’s own intermediate states stay useful for further prediction is a legitimate question. The state and weather replacement tests probe it sensibly, and the authors are transparent about case selection (even showing a failure example).
Critical weaknesses

Limited contribution

  • In simple terms, the task is finite-horizon forecasting: a fixed 100-day window where every date is directly supervised.
  • The world-model framing seems too broad. It predicts only NDVI, not the full image, and it’s not clear why only NDVI.
  • The technique is not new, as Deep-OSG already supervises both direct and composed predictions.
  • The empirical result is somewhat expected, since a model never trained on its own outputs will degrade when chained.
  • Multi-step prediction only matches direct prediction and never improves on it.
  • The empirical validation rests on one in-house control, with no other multi-step methods or VegeDiff/VegSim for comparison.
  • It scores below Contextformer and PredRNN on the official metrics, even though Contextformer is its own backbone.

The setting doesn’t support the world-model framing or real-world use

  • No real use case is tested: no new observations mid-forecast, and no forecasting beyond day 100.
  • The weather data seems to be observed rather than forecast, and the paper does not say where it comes from.
  • Inputs and targets both come from the preprocessed benchmark. This raises concerns about real-world impact, since real-time forecasting would face noisy satellite imagery.

Additional issues

  • The internal representations highlighted in the motivation are never analyzed directly, either quantitatively or qualitatively.
  • There is no real ablation study, only test-time changes to one trained model.
  • The weather analysis uses only 84 samples, and the paper’s own test shows no extra benefit under hot-dry conditions.
  • The paper offers no ecological insight as a physical-sciences paper.
  • The writing is repetitive and hard to follow.
Questions for authors
  1. How does partitioned-path learning differ from Deep-OSG? A comparison with Deep-OSG, or with other multi-step training methods like scheduled sampling or latent overshooting, would help to see what is new here.
  2. In exp. setup, multi-step prediction only matches direct prediction. Is there a case where continuing from a generated state actually helps, for example when a new observation arrives at day 50, or when forecasting beyond day 100?
  3. Where does the future weather come from? It looks like observed data rather than forecasts. If so, how would performance change with real weather forecasts, or with noise added to the weather?
  4. The paper motivates the work through internal representations, but I couldn’t find any direct analysis of them. Could you show how close the states reached in steps are to those reached directly, for both SpanWorld and the control?
  5. Why does SpanWorld score below Contextformer on the official metrics when Contextformer is its own backbone? Retraining without individual components would also help clarify which parts of the design matter.
  6. Why focus only on NDVI? Since the paper frames SpanWorld as a remote-sensing world model, have you tried predicting the full multispectral image?
8 comments
AuthorsOpenReview ↗

Response to Reviewer T9JT — Part 1/6: Contribution and training formulation

Thank you for your specific comments on the methodological contribution, empirical evidence, and practical use. We have added comparisons with recent external methods, training-objective and component-retraining ablations, and latent-state analysis, and expanded the main weather evaluation from 84 samples to 1,054 OOD-T samples. Below, we use this evidence to address the design contribution, predictive value of generated states, and scope of the conclusions.

Contribution and related training formulations

How does partitioned-path learning differ from Deep-OSG? A comparison with Deep-OSG, or with other multi-step training methods like scheduled sampling or latent overshooting, would help to see what is new here.

Thank you for highlighting the need for methodological comparison and a clearer interpretation of the evidence. Deep-OSG includes observed-endpoint supervision of direct and composed predictions. SpanWorld studies how latent spatial states inferred from incomplete satellite observations can support subsequent image prediction under interval-specific weather. Partitioned-path learning trains this use of states by propagating endpoint-image errors through generated states and subsequent transitions, while retaining direct supervision at every target date.

To clarify the role of this design, we examine the supervision path and the objective applied to partitioned predictions separately. First, the core control tests whether applying additional supervision to direct or partitioned paths affects continuation from generated states when every target date already receives direct supervision and the additional endpoint observations are identical. This comparison holds architecture, initialization, training update count, directly supervised dates, and additional endpoint images fixed, separating the effect of the supervision path from the effects of adding supervised dates or additional images. Degradation relative to each model’s own direct-prediction error further separates direct forecasting ability from the subsequent predictive usefulness of generated states.

Results across three paired seeds further support the role of partitioned-path training in state continuation: SpanWorld has lower mean continuation MSE than the direct-endpoint supervision control on all three evaluated paths. On the training-unseen 15+35-day path, continuation MSE is reduced by 11.94%, with SpanWorld and the control achieving 0.043272±0.000514 and 0.049138±0.001003, respectively (Table 1; mean ± standard deviation across seeds). These results show that, even when every target date already receives direct supervision, applying supervision through partitioned paths can improve prediction of subsequent vegetation changes from generated states.

Second, we added a consistency-training control with the same architecture and sampled paths to test whether a simpler objective can achieve comparable performance while retaining the partitioned paths. This control retains direct observed-image supervision at every target date and replaces additional observed-endpoint supervision along partitioned paths with agreement to stop-gradient direct predictions at the same dates. Under matched common training settings, the seed-42 comparison after 40 epochs shows that the full model achieves the lowest continuation MSE on all three evaluated paths. On the training-unseen 15+35-day path, the full model and consistency alternative achieve MSEs of 0.04174650 and 0.04311991, respectively, a relative reduction of 3.19% (Appendix I, Table I1A). This ablation further supports that, with the same partitioned paths retained, supervising partitioned predictions using observed endpoints is more beneficial for continued prediction from generated states than merely matching direct predictions.

These two comparisons examine the supervision path and objective separately, providing evidence beyond the general expectation that multi-step training benefits multi-step prediction. The consistency control is a matched test of the supervision objective, not a full reproduction of Deep-OSG, scheduled sampling, or latent overshooting. We clarify the relationship to prior methods in Related Work and the contribution statement, and use these comparisons to examine how our modeling design affects the subsequent predictive usefulness of generated states.

AuthorsOpenReview ↗

Response to Reviewer T9JT — Part 2/6: Applications and ecological interpretation

Usefulness of continued prediction

In exp. setup, multi-step prediction only matches direct prediction. Is there a case where continuing from a generated state actually helps, for example when a new observation arrives at day 50, or when forecasting beyond day 100?

Thank you for highlighting the distinction between being able to continue prediction and demonstrating its practical value. We recognize that showing partitioned predictions close to direct predictions does not by itself establish benefits for applications such as incorporating new observations or forecasting over longer horizons.

The demonstrated benefit is that partitioned-path training improves continued prediction when the model reuses its own generated states (Table 1) and reduces error growth relative to each model's own direct predictions (Appendix C). Thus, partitioned prediction remaining close to direct prediction reflects the model's ability to retain subsequent predictive capability after its own rollout, beyond the observation that the two prediction modes yield similar errors.

Incorporating a new observation at day 50 and forecasting beyond day 100 are concrete directions for testing further application value. We have not evaluated these two settings. Our current experiments examine state continuation within the benchmark's 100-day forecast window without additional remote-sensing observations. This setting isolates states formed through the model's own evolution, without introducing information gains from new observations.

This ability provides a basis for extending an existing simulation or developing subsequent weather trajectories from a common intermediate state. Our contribution is to establish this state-evolution and training formulation and evaluate the usefulness of generated states against subsequent real satellite observations; the proposed application extensions require further validation.

Preprocessing and application scope

Inputs and targets both come from the preprocessed benchmark. This raises concerns about real-world impact, since real-time forecasting would face noisy satellite imagery.

Thank you for asking how results obtained with preprocessed data relate to practical remote-sensing applications. We follow GreenEarthNet's standard quality-control procedures to distinguish usable vegetation observations from observations affected by clouds and other factors. Historical inputs still contain missing observations, and the model uses observation masks to form predictive representations from incomplete satellite histories. On the target side, validity masks allow supervision and scoring to reflect vegetation change rather than cloud contamination and other observation artifacts. The input-processing, supervision, and evaluation rules are described in Appendices A.1, A.2, and B.1, respectively.

In this setting, the model must still advance its own state without new satellite observations and predict subsequent real vegetation changes. Standard quality control enables this capability to be evaluated against valid observations; subsequent satellite images are used only for supervision and evaluation, not as image inputs during prediction. The study thus provides empirical evidence for the model’s ability to support continued vegetation prediction from incomplete observations.

The current results support state evolution and continuation under these standard quality-control conditions. The effects of different observation quality, mask errors, or additional image noise require dedicated evaluation.

Ecological interpretation

The paper offers no ecological insight as a physical-sciences paper.

Thank you for raising the question of ecological insight. Our study provides a concrete finding for vegetation dynamics modeling: direct supervision at target dates is not sufficient to ensure that states generated through the model's own rollout are equally effective for subsequent prediction. How these states are trained through their subsequent predictive use affects their value for predicting real vegetation changes.

The matched experiments show that partitioned-path training improves continued prediction from generated states and reduces prediction degradation along the evaluated paths (Table 1 and Appendix C). State-replacement and weather-replacement analyses further examine the predictive roles of sample-specific states and corresponding weather (Appendices E and F).

For vegetation models intended for continued evolution, these findings suggest evaluating not only direct forecasting accuracy but also whether generated states retain information needed for continued prediction, using subsequent real satellite observations. We frame this as a methodological finding for vegetation dynamics modeling, rather than identification of a specific ecological mechanism. Its value lies in providing empirical evidence to inform state training and evaluation in weather-driven models.

AuthorsOpenReview ↗

Response to Reviewer T9JT — Part 3/6: External methods and latent states

External methods

The empirical validation rests on one in-house control, with no other multi-step methods or VegeDiff/VegSim for comparison.

Thank you for pointing out the lack of external comparisons. We added a comparison with EO-WM, a recent weather-conditioned remote-sensing forecasting method, undertook a reproduction of VegeDiff, and examined VegSim's task setting. The added external comparison and three-seed paired experiments assess standard forecasting performance and continued prediction from generated states, respectively.

To provide an external forecasting reference, we evaluated publicly released EO-WM weights on our GreenEarthNet test splits using the same pixel-level evaluation protocol. EO-WM also studies weather-conditioned multispectral forecasting, and its EarthNet data setting is closely related to GreenEarthNet in satellite observations, meteorological drivers, and dataset design. Under the shared NDVI evaluation protocol, SpanWorld achieves lower mean RMSE than EO-WM on all four test splits, demonstrating its competitiveness on the vegetation forecasting task studied here. Full results and standard deviations across three seeds are included in Appendix B, Table B2 of the revised manuscript, together with weight provenance, required input adaptations, and differences in training conditions.

For VegeDiff, we implemented and trained an adaptation based on the official code. We did not find directly usable autoencoder weights in the official resources we inspected, so we used the publicly released EarthNet-finetuned EO-VAE and further adapted it on the GreenEarthNet training data. We trained the diffusion predictor for 100 epochs and selected the checkpoint using the validation set. The current adapted implementation obtains RMSE values of 0.4043, 0.3882, 0.4031, and 0.4030 on IID, OOD-T, OOD-S, and OOD-ST, respectively. The large gap from other reference methods makes us cautious about whether these reproduction results represent the original method. Since the effects of adaptation factors, including pretrained weights and training and inference configurations, have not been fully resolved, we report these results transparently but exclude them from the formal method ranking and do not use them to claim superiority over VegeDiff.

We also examined VegSim. It is closely related to our work in weather-driven latent vegetation dynamics, but its released task predicts minicube-level NDVI sequences and quantiles, whereas our evaluation measures pixel-level spatial forecasts. Directly juxtaposing its published metrics would not constitute a comparison on the same task. We explain this methodological relationship and difference in output granularity in Appendix B of the revised manuscript rather than rank metrics from different tasks together.

The external comparison demonstrates SpanWorld's competitiveness on the vegetation forecasting task studied here, while the paired experiments further test the role of its state-evolution design. Across three paired seeds, SpanWorld reduces subsequent prediction MSE by 11.94% relative to the direct-endpoint supervision control after the training-unseen 15+35-day rollout (Table 1). Together, these results support both the model's ability to predict future vegetation observations and the benefit of partitioned-path training when generated states are reused for prediction.

Analysis of internal states

The paper motivates the work through internal representations, but I couldn’t find any direct analysis of them. Could you show how close the states reached in steps are to those reached directly, for both SpanWorld and the control?

Thank you for highlighting the need to analyze the internal representations directly. We have added a comparison of the states of SpanWorld and the direct-endpoint supervision control in Appendix E.3, Table E3. On the same 2,837 IID samples, we extract the latent states reached at day 50 directly and through the 25+25-day and 15+35-day partitioned paths. Within each model, we compute the cosine similarity between each partitioned state and its own direct-path state. Both models use validation-selected checkpoints from seed 42.

For the 25+25-day and 15+35-day paths, SpanWorld achieves mean cosine similarities of 0.9270 and 0.9094, respectively, compared with 0.6889 and 0.6990 for the direct-endpoint supervision control. On both evaluated paths, including the training-unseen 15+35-day partition, SpanWorld’s partitioned states are closer to its own direct-path states.

This analysis directly provides the internal-representation evidence requested in your question. Together with the existing continuation evaluation, it examines the role of partitioned-path learning through both the relationships between states and their use in subsequent prediction.

Reviewer_Pr4XOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SpanWorld is a weather-conditioned world model for vegetation forecasting on GreenEarthNet. Given incomplete historical Sentinel-2 imagery and future weather, it predicts 100 days of NDVI. A shared variable-span transition advances a latent spatial state, supporting direct prediction and partitioned rollouts. The main contribution is partitioned-path learning, which adds endpoint image supervision on predictions reached through model-generated intermediate states, while keeping direct supervision at all horizons. A matched control receives the same endpoint supervision applied to direct predictions.

Critical strengths
  1. The paper distinguishes accuracy at a target date reached directly from accuracy when a generated intermediate state is reused. Evaluating these separately makes the continued-prediction problem concrete.

  2. SpanWorld and the direct-endpoint control share architecture, initialization, update count, direct supervision at all horizons, and endpoint images. The only difference is the supervision path, so the comparison isolates that factor.

  3. State replacement, removing the state correction, and replacing the transition with an identity map each degrade performance, which supports the claim that the dynamic state carries useful information.

  4. The paper uses paired bootstrap intervals for the main differences, shows a degradation example (Figure G2), and acknowledges that PredRNN and Contextformer have higher R².

Critical weaknesses
  1. The central claim rests on a single seed and one split. The 11% continued-prediction improvement comes from one training run per model on the IID split. Stability across seeds and generalization to the OOD splits are not shown for this comparison.

  2. The control may not isolate the mechanism. The paper does not test simpler alternatives, such as a consistency loss on the same rollout paths or noise injection into intermediate states, that might produce the same effect.

Questions for authors
  1. Can you report the core continued-prediction comparison (Table 1) over at least three seeds, with variability across seeds?
  2. How does the partitioned-path control compare with simpler alternatives, such as training on the same rollout paths with only a consistency loss, or adding noise to intermediate states?
  3. Can you report inference cost for partitioned versus direct prediction?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Pr4X — Part 1/2: Paired seeds and consistency control

Thank you for your suggestions on the stability of the core result, training controls, and computational cost. We have expanded the core continuation comparison to three paired random seeds and added a consistency-training ablation on the same paths, together with inference-cost measurements for direct and partitioned prediction. Below, we report the findings, their statistical scope, and the corresponding revisions.

Paired-seed continuation evaluation

The central claim rests on a single seed and one split. The 11% continued-prediction improvement comes from one training run per model on the IID split. Stability across seeds and generalization to the OOD splits are not shown for this comparison. Can you report the core continued-prediction comparison (Table 1) over at least three seeds, with variability across seeds?

Thank you for suggesting an assessment of the core result across random seeds. We have extended the core comparison in Table 1 to three paired random seeds and report means and between-seed standard deviations for each model's prediction errors and the paired differences.

The three-seed paired results further support the central finding: SpanWorld has lower mean continuation MSE than the direct-endpoint supervision control on all three evaluated paths. On the training-unseen 15+35-day path, SpanWorld and the control achieve MSEs of 0.043272±0.000514 and 0.049138±0.001003, respectively, a relative reduction of 11.94%. The paired MSE difference is −0.005866±0.001470 (Table 1; mean ± standard deviation across seeds). Thus, the continuation benefit observed in the original submission receives further support from the expanded three-seed comparison.

In addition to between-seed variability, we report paired geographic-cluster bootstrap intervals. The between-model difference intervals lie below zero for all three continuation paths, providing complementary evidence for the continuation advantage within the evaluated sample set. These intervals hold the models from all three seeds fixed; they characterize sample-level variability, whereas the between-seed standard deviations characterize variability across training runs. Calculation details are provided in Appendix C.

Consistency-training alternative

How does the partitioned-path control compare with simpler alternatives, such as training on the same rollout paths with only a consistency loss, or adding noise to intermediate states?

Thank you for raising this alternative explanation. To test whether consistency training on the same partitioned paths can achieve comparable performance, we added a training-objective comparison in Appendix I, Table I1A. All models retain direct observed-image supervision at every target date. The consistency alternative uses the same partitioned paths and replaces additional observed-endpoint supervision with agreement to the direct prediction at the same date, with gradients stopped through that direct prediction. This directly compares two objectives: matching partitioned predictions to direct predictions and supervising partitioned predictions with observed endpoints.

Under matched common training settings, we compare all models after 40 epochs of training, using seed 42 throughout. The full model achieves the lowest continuation MSE on all three evaluated paths. On the training-unseen 15+35-day path, its MSE is 0.04174650, compared with 0.04311991 for the consistency alternative, a relative reduction of 3.19%. Thus, in this comparison, matching partitioned predictions to direct predictions alone does not attain the continuation accuracy achieved with observed-endpoint supervision.

We selected consistency training as a targeted control because it directly tests the supervision objective while retaining the same partitioned paths. We did not include an intermediate-state noise control in this comparison, and our conclusion is limited to the evaluated alternative. Section 4.4 summarizes the comparison, and Appendix I provides the full settings and results.

AuthorsOpenReview ↗

Response to Reviewer Pr4X — Part 2/2: Inference cost

Inference cost

Can you report inference cost for partitioned versus direct prediction?

Thank you for suggesting that we report inference costs. We have added a comparison of latency and peak GPU memory for direct and partitioned prediction with the same inputs and target dates in Appendix H, Table H2. Both end-to-end measurements include computation of the historical state, baseline predictions, and weather conditions.

On the same shared RTX 4090 with batch size 1, direct and partitioned prediction have median latencies of 122.44 ms and 122.15 ms, respectively, and both have a peak allocated memory of 595.85 MiB. Their measured end-to-end costs are similar. The table also reports latency percentiles and the measurement settings.

We additionally distinguish continuation from an existing day-50 state and cached baseline predictions from the corresponding preparation cost: their median latencies are 5.21 ms and 117.38 ms, respectively. This reports both the full computational cost starting from historical inputs and the subsequent cost when existing states and cached predictions are reused.

AuthorsOpenReview ↗

Follow-up on the additional experiments

Dear Reviewer ,

Thank you for your careful review and constructive suggestions. Following your feedback, we have added three-seed paired continuation experiments, a consistency-training comparison on the same rollout paths, and inference-cost measurements to our responses and revised manuscript.

If these additions have adequately addressed your main concerns, we would appreciate it if you could reassess the paper in light of the additional evidence and consider updating your score accordingly.

Thank you for your time and for helping us strengthen this work.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Contributions, Revisions, and Additional Experiments

Dear Reviewer,

We thank the Area Chair and reviewers for their constructive feedback. The reviews recognized the research question, controlled comparisons, and performance under unseen temporal partitions. Two reviewers subsequently acknowledged that the revisions addressed many or most of their concerns.

SpanWorld models vegetation change through weather-driven evolution of reusable spatial states. As a remote-sensing world model, it infers latent states from incomplete satellite histories and advances them with shared variable-span transitions. Partitioned-path learning trains state formation and subsequent use through observed endpoint imagery while retaining direct supervision at every target date. Our controlled experiments show that supervision paths matter for subsequent prediction from generated states, even when all target dates are already directly supervised.

1. Robust continuation gains across seeds (Pr4X, T9JT). We expanded the architecture-matched comparison to three paired seeds. On 2,837 IID samples, each model reaches day 50 through the training-unseen 15+35-day path, then predicts each target date from days 55 to 100 independently from its fixed day-50 state. Mean continuation MSE is 0.043272 for SpanWorld versus 0.049138 for the direct-endpoint control, an 11.94% reduction. SpanWorld has lower mean continuation MSE on all three tested paths, with similar direct-prediction errors. Initialization, training updates, supervised dates, and additional endpoint images are held fixed (Section 4.2, Table 1; Appendix C).

2. Benefits beyond consistency training (Pr4X, T9JT). We added a consistency alternative using the same architecture and partitioned paths, replacing observed-endpoint supervision with agreement to stop-gradient direct predictions. In the matched seed-42, 40-epoch comparison, the full model achieves lower continuation MSE on all three paths, including a 3.19% reduction on the 15+35-day path. This supports the value of training state reuse against observed endpoints (Section 4.4; Appendix I, Table I1A).

3. Evidence on state representations and component contributions (T9JT). Using seed-42 checkpoints, cosine similarity between partitioned and each model's own direct-path states is 0.9094 versus 0.6990 on the 15+35-day path, favoring greater agreement in SpanWorld. In the seed-42 retraining ablation, the full model has higher R² on all four splits than the model without the dynamic state branch, with numerically close RMSE (Appendix E.3, Table E3; Appendix I, Table I1B).

4. Forecasting performance and computational cost (Pr4X, T9JT, tFYa). Updated three-seed results show standard forecasting RMSE close to Contextformer's. SpanWorld has lower mean RMSE on all four splits than the weather-conditioned ConvGRU and the adapted evaluation of released EO-WM weights. Direct and partitioned prediction also have similar measured end-to-end latency and peak memory (Section 4.1; Appendix B, Table B2; Appendix H, Table H2).

5. Stronger weather evidence and clearer methodology (T9JT, tFYa). We expanded the main weather analysis from 84 to 1,054 matched OOD-T samples. Both weather replacements increase prediction error, supporting the predictive value of sample-specific weather. We clarified the observed-weather inputs, architecture, and relationship to Deep-OSG, making the contribution to weather-driven latent-state modeling more explicit (Sections 2-3 and 4.4; Appendix F, Table F1).

Together, the model, training mechanism, and controlled experiments make the subsequent predictive usefulness of generated states an explicit target of world-model learning and evaluation. With supervision coverage held fixed, the findings establish the value of training states through their subsequent use in the evaluated setting. We believe this contribution and the strengthened evidence provide a strong case for acceptance.

Thank you for your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.