acceptodds
Under review as a conference paper at ICLR 2027

GEWorld: A Genotype-Environment World Model for Crop Phenotype Prediction and Rollout

14 views

Abstract

Phenotype prediction aims to predict traits from genotype and environmental information. Most existing methods learn static mappings from aggregated data, making it difficult to dynamically adapt to new observations or changing environments. Here, we propose , a genotype–environment world model that casts phenotype prediction as the evolution of a latent physiological state: it fuses genotypes with an environmental embedding through a gated, multiplicative genotype-by-environment interaction to initialize this state, assimilates observed phenotype prefixes through a GRU-based observation module to correct the state as new measurements arrive, and evolves the state open-loop through a Mamba-2-inspired state-space transition under future environments and elapsed time to project how the trajectory unfolds. GEWorld decodes static terminal traits from the initial state and residual dynamic trajectories from the rolled states; environmental counterfactuals are obtained by rolling the same assimilated state under alternative future weather. Across static and dynamic benchmarks spanning five crops, GEWorld achieves state-of-the-art predictive accuracy, while its forecasted phenotype trajectories closely track observed development. The code is publicly available at https://anonymous.4open.science/r/GEWorld-E070/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Genotype × environment interaction, phenotype prediction, dynamic trajectory, counterfactual intervention, world model

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

This paper proposes GEWorld, a stateful genotype-by-environment (G ×\times E) modeling framework in which genotype state and time-varying environmental factors drive state transitions, while sparse phenotypic observations are assimilated to update latent states conditioned on genotype and environment. The framework supports static phenotype prediction and dynamic trajectory forecasting. Experiments across multiple crops, including factual and counterfactual forecasting settings, demonstrate the effectiveness of GEWorld. Further analyses provide insight into how genotype and environment influence phenotypic trajectories.

Critical strengths
  1. The paper addresses an important challenge in genotype-to-phenotype prediction by modeling dynamic G×E interactions and enabling state updates under changing environmental conditions.
  2. The paper evaluates both static and dynamic phenotype prediction across multiple crops and further investigates counterfactual forecasting using crop simulations, demonstrating the effectiveness of GEWorld and providing understanding about how genotype and environmental factors interact to influence crop phenotypes.
Critical weaknesses
  1. Insufficient justification and ablation analysis of the stateful modeling framework. The architectural novelty of GEWorld lies mainly in the “Gated G ×\times E Initialization” and “Observation Assimilation and State Transition” modules. However, it remains unclear why a Mamba-based state-space model is preferred over conventional temporal models, such as RNNs or Transformers in general world models.

  2. The paper does not clearly specify which mentioned datasets are utilized for train/val/test split. In addition, the counterfactual evaluation relies on APSIM simulations, and Table S8 reports substantially higher sign agreement for counterfactual forecasting than for factual forecasting. Without a clear description of the data splits and evaluation protocol, it is difficult to assess whether these results reflect generalizable environmental responses or overfitting to the training data.

Questions for authors
  1. Why was Mamba selected for the state transition model? How does it compare with conventional temporal models under comparable settings?
  2. Why is sign agreement substantially higher in APSIM-based counterfactual forecasting than in factual forecasting (Table S8)? Could the authors clarify the evaluation protocols and further assess whether the model accurately captures intervention-induced changes in phenotype?
  3. How should the embedding visualization in Figure S4 be interpreted? Maize shows less distinct clustering by phenotype than rice, despite achieving better phenotype prediction performance. Could the authors explain this discrepancy?
4 comments
AuthorsOpenReview ↗

Rebuttal Part 1/2

We thank the reviewer for the careful reading and constructive questions. We will address each of them in turn.


W1. Insufficient justification of the stateful framework

The Model esign is biologically motivated. Within-season phenotype is better viewed as a partially observed physiological process: genotype sets response capacity, weather drives development over time, and field traits are sparse, noisy readouts of that latent state. A stateful assimilate–rollout loop matches this structure: initialize from G×E, correct when measurements arrive, then open-loop under future environments, unlike a static G×E map.

We therefore use a Mamba-inspired selective SSM as the transition: its input-dependent gating lets current weather and time control what developmental memory is retained, aligning with stage- and stress-dependent crop dynamics. Whether this Mamba-style transition outperforms GRU/RNN/Transformer is answered empirically in Q1.


W2. Unclear splits and concern that Table S8 overstates environmental responses

We fully agree with the reviewer’s concerns. In our setting, train/val/test are genotype-grouped: held-out genotypes in known environments, not unseen sites. More importantly, Table S8’s “factual” and “counterfactual” rows are not matched scores on the same samples,they mix panels and contrasts, so direction agreement alone can look strong. We therefore clarify the split protocol, reinterpret Table S8 as cross-environment consistency, with an environment-mean baseline and magnitude errors, and add a fixed-history APSIM weather benchmark that changes only post-cutoff weather; protocols and numbers are in Q2.


Q1. Why was Mamba selected, and how does it compare under comparable settings?

A selective SSM lets environmental/time inputs modulate state evolution via input-dependent decay, update, and readout, while keeping a recurrent interface for assimilation. The module is a Mamba-inspired, one-step selective SSM, with internal memory rebuilt from the shared latent state each step.

We compare the selective SSM, GRU, RNN, and causal Transformer on sorghum TERRA-REF using a 70% observed prefix. All four models use a latent width of D=128D{=}128 and share the encoders, genotype FiLM, assimilation GRU, decoder, optimizer, and early-stopping settings. Each model is evaluated with the seasonal curve enabled and disabled. We use five independent genotype-disjoint splits, each with a 70/15/1570/15/15 train/validation/test ratio. The main-text experiments vary random seeds under a fixed genotype holdout, whereas these controls assess whether the models’ relative performance remains robust across different held-out genotype sets.

TransitionCover, onHeight, onCover, offHeight, off
Selective SSM0.5317 ± 0.10460.8315 ± 0.01530.4569 ± 0.15260.8155 ± 0.0136
GRU0.4782 ± 0.17510.8216 ± 0.01500.3432 ± 0.27570.8046 ± 0.0149
RNN0.4864 ± 0.15100.8218 ± 0.01740.3780 ± 0.18800.8036 ± 0.0116
Transformer0.4318 ± 0.18220.8175 ± 0.02180.2904 ± 0.24190.7938 ± 0.0264

The selective SSM achieves the highest mean test R2R^2 in all four settings. Its cover gains over GRU, RNN, and Transformer are 0.05360.0536, 0.04530.0453, and 0.09990.0999 with the seasonal curve enabled, and 0.11370.1137, 0.07900.0790, and 0.16650.1665 with it disabled. Its advantage with the curve enabled also suggests that the learned dynamics complement the seasonal prior.

We have summarized the transition comparison in Section 4.2 (RQ4), and provided the details in Appendix C.3 of our revision.

AuthorsOpenReview ↗

Rebuttal Part 2/2

Q2. Why is sign agreement higher for counterfactual forecasting in Table S8, and does the model capture intervention-induced changes?

Table S8's factual and counterfactual rows use different panels and contrasts, so their scores are not directly comparable. Direction agreement can also reflect the average environmental response. We will clarify this interpretation. We therefore added a controlled APSIM weather experiment. For wheat and sorghum, paired runs share identical simulator settings and the first 84 days, verified by deterministic replay. From day 85, only future weather changes: rainfall is halved or maximum/minimum temperatures increase by 10°C. The benchmark contains 3,000 trajectories and 2,000 intervention pairs, split by weather year into 70/15/15 training/validation/test years, keeping all branches and initial-water conditions of each year together. Each model assimilates the common history once, copies its state, and predicts both branches using future weather only. We compare the predicted intervention-minus-factual change with the APSIM change.

We evaluate intervention effects using MAE across both interventions and all 12 forecast weeks. For GEWorld, we average predictions from three initializations before computing MAE. Lower values indicate more accurate effect estimates.

Crop / traitGEWorldZero effectTraining mean effectRidge
Wheat biomass20.90786.50151.18149.099
Wheat grain7.66324.10613.85913.881
Wheat LAI0.1620.3770.3250.307
Sorghum biomass71.302131.723127.131120.060
Sorghum grain35.82071.49656.52053.680
Sorghum LAI0.2530.3540.3320.325

GEWorld has lower effect MAE than all three baselines in all six pooled comparisons. This supports prediction of intervention magnitudes in this controlled simulation benchmark. Improvement over the zero-effect baseline indicates that the predicted responses capture meaningful intervention signals. The gains over the training-mean and ridge baselines further suggest that GEWorld captures variation in response magnitudes beyond an average effect or a linear approximation.

We have summarized these results in Section 4.2 (RQ3), and provided the full evaluation in Appendix C.6 of our revision.


Q3. How should Figure S4 be interpreted when maize clusters less but predicts better?

Maize can be predicted accurately without forming distinct clusters because the task is continuous trait regression. Predictions can follow a smooth trait gradient through the representation: nearby genotypes can have gradually changing phenotypes, with no visible boundary between groups. Conversely, visibly separated rice clusters need not separate samples along the directions most relevant to the target trait. Figure S4 further compresses the initialized state s0s_0 into two UMAP dimensions; UMAP emphasizes neighborhood structure, without optimizing preservation of phenotype-predictive information.

The maize panel also has a strong predictable environmental component. A training-only environment-mean baseline already reaches R2=0.5746R^2=0.5746 in maize, compared with 0.01690.0169 in rice. Thus, much of maize's overall predictability is accessible from known environmental groups without requiring visually distinct genotype clusters. This provides a concrete explanation for why maize can predict better despite weaker clustering; it does not imply that maize has stronger genetic separation.

We report test R2R^2 as mean ± sample SD across five genotype splits. Scaling and PCA are fitted on training data, and ridge regularization is selected on validation data.

PredictorMaizeRice
Stored single-model decoder0.6404 ± 0.00470.4699 ± 0.0657
Full-state ridge0.6344 ± 0.01060.4854 ± 0.1711
2D PCA ridge0.5617 ± 0.00540.4140 ± 0.1030
Training-only environment mean0.5746 ± 0.01820.0169 ± 0.0121

The full-state ridge also outperforms the 2D PCA probe in maize by 0.073 R², supporting predictive information beyond a two-dimensional linear projection. We will clarify in Figure S4's caption that cluster visibility illustrates representation geometry and should not be used to rank predictive accuracy across crops.

We have clarified this interpretation in Section 4.2 (RQ4), and added the predictive probes in Appendix C.4 of our revision.

Thanks for the authors' response. Most of my concerns have been addressed. I'd like to maintain my original socre for a weak acceptance.

We sincerely thank the reviewer for their valuable feedback and for the time and effort devoted to carefully evaluating our responses. The insightful comments have helped us further improve the clarity and presentation of our manuscript. We also greatly appreciate the reviewer’s positive assessment and continued support of our work.

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

This paper proposes a genotype-environment world model, GEWorld, that fuses genotype and environment for crop trait prediction and dynamic trajectory forecasting. An end-to-end training framework and rollout learning strategy are used. From evaluation on five crop benchmarks, GEWorld achieved effective results on trajectory and counterfactual forecasing.

Critical strengths
  • The problem studied in this paper is clearly-motivated. Dynamical phenotype prediction adaptable to new observations and changing environments is fairly practical but not sufficiently investigated.

  • The results on both real-world and simulated datasets verify the effectiveness of the proposed method.

  • This paper is clearly presented and easy to follow.

Critical weaknesses

The proposed method has some crucial weaknesses. These following main issues are central to the credibility of the paper’s main claims and methodological contributions.

  • The counterfactual intervention is not aligned with the evaluation reference, and the information boundary of the latent state needs clarification. The current evaluation can support the consistency of predicted cross-environment response trends, but it does not directly establish the accuracy of intervention effects conditional on a fixed current state.

  • The conditional incremental value of genetic information in dynamic prediction and its pathways of influence have not been sufficiently isolated. GEWorld uses SNP information both to initialize the latent state and to condition subsequent state transitions. However, early phenotypic observations already encode some information about genetic effects. It is therefore important to determine whether the dynamic prediction gains arise from explicit SNP inputs or primarily from the indirect representation of genetic effects in the observed growth history. The existing static-input ablation does not directly address this question in the dynamic setting. Similarly, the dynamic-module ablation does not separate the contributions of genotype information entering through initialization from those entering through state transitions.

  • The ability of latent states to represent history-dependent processes lacks direct functional validation. The paper interprets the hidden state as a latent state that integrates growth history. However, the primary evidence for this interpretation comes from trajectory prediction performance and module ablations. Meanwhile, the decoder in Equations (9)–(11) also uses the previous time step’s trait values and stage-wise means computed from the training set. These inputs can provide strong anchors for the current state and encode shared seasonal growth patterns. Consequently, close agreement between predicted and observed trajectories does not, by itself, establish what historical information the latent state retains beyond current observations and common growth trends. The static-initialization representation analysis in the appendix also does not directly validate the historical memory function of the dynamic state.

  • Performance gains attributable to the core architecture and the controls used in baseline comparisons remain insufficiently established. The current “w/o Transition” ablation freezes the hidden state and demonstrates that state evolution contributes to prediction. However, it does not establish the additional advantage of the Mamba-style state-space model (SSM) over conventional recurrent transitions. Furthermore, the seasonal prior incorporated into the adaptive decoder may contribute to the shape and stability of the predicted trajectories. The advantage of the full model over other temporal methods therefore cannot yet be clearly attributed to the specific state-transition architecture.

  • Consistency between the main experiments and ablation results, as well as the statistical reliability of performance gains, requires further clarification. Table 2 and Table S6 report different results for the full GEWorld model on apparently identical tasks. But the manuscript does not explain whether these discrepancies arise from different data splits, model configurations, ensembling strategies, or statistical aggregation procedures. This makes it difficult to connect the performance gains reported in the main experiments to the contributions attributed to individual modules in the ablations.

Questions for authors

The main questions lie in the weaknesses pointed above. Accordingly, I suggest the authors to provide additional answers and evidences from following aspects:

  • A counterfactual reference based on an identical complete simulator state, with only future weather replaced, to directly evaluate the intervention effects defined by the model.

  • Dynamic ablations separating SNP-based initialization from persistent genotype-conditioned transitions, accompanied by genotype-level evaluations that directly assess genetic-specific responses.

  • A direct test of history-dependent prediction, using samples with similar current observable states but different histories under identical future weather conditions.

  • Controlled transition-model comparisons under matched assimilation and decoder settings, together with a clear reconciliation of the main experimental results and ablation results.

3 comments
AuthorsOpenReview ↗

Rebuttal Part 1/3

We thank the reviewer for raising these questions. We will address each of them in turn.


W1. Counterfactual reference and latent-state information boundary

A common difficulty in evaluating counterfactual growth predictions is that multi-environment observations do not provide paired outcomes with identical growth histories and different future weather. Our original evaluation therefore compares response trends across environments to assess whether GEWorld captures genotype-dependent environmental responses. Following the reviewer’s suggestion, we add in Q1 a fixed-history APSIM weather benchmark that changes only future weather (rainfall halved or warming after a shared cutoff), and directly evaluate intervention magnitudes via paired effect differences against zero-effect, training-mean, and ridge baselines. The experimental results are reported in Q1 below.


W2. Conditional value of explicit SNPs and their pathways

We thank the reviewer for this suggestion. We further examine how genotype information is transmitted in GEWorld. Genetic information can influence predictions through three pathways: (i) indirectly through assimilated phenotypes; (ii) explicitly through genotype-conditioned initialization (zg→s0z_g \rightarrow s_0); and (iii) explicitly through genotype FiLM that continually modulates state updates driven by weather and elapsed time.

After the cutoff, the model no longer assimilates phenotypes, but genetic information contained in the initialization and phenotype history can still propagate through the latent state. Genotype-related dynamics after the cutoff therefore cannot be attributed entirely to continued FiLM conditioning. In Q2 below, we control phenotype history, separate initialization from transition FiLM, and add genotype-level MSE contrasts plus SNP identity probes to assess genetic-specific responses beyond panel-average R2R^2.


W3. History-dependent memory of the latent state

We thank the reviewer for highlighting the need to distinguish the contribution of historical information in the latent state from that of the decoder’s current phenotype and seasonal mean. We therefore (i) retrain with truncated phenotype–weather windows under fixed current anchors and future weather, and (ii) run the reviewer’s matched-current / different-history test, swapping earlier history while matching current observables and future weather. Results are in Q3 below.


W4. Gains from the SSM vs state evolution and the decoder

We select a Mamba-inspired selective SSM on biological grounds: within-season phenotype behaves as a partially observed physiological process driven by genotype and time-varying weather, so the transition should let current environment and elapsed time gate what developmental memory is retained or overwritten. Freezing the transition only shows that state evolution matters, and the seasonal decoder prior can also shape trajectories; neither isolates an SSM-specific gain. We therefore compare the selective SSM with GRU, RNN, and Transformer under matched assimilation/decoder settings and the same latent width; experimental evidence is reported in Q4 below.


W5. Consistency between the main experiments and ablation results

We apologize for the inconsistency between the Full results in the main experiments and the ablations. Upon checking, we found that the sorghum numbers reported in Table 2 of the main text were computed with an older model checkpoint. We therefore re-ran the main-table results under the same configuration as the ablation Full, and updated the main text accordingly. The updated numbers are reported in Q4 below.


AuthorsOpenReview ↗

Rebuttal Part 2/3

Q1. Fixed-history counterfactual reference (identical state; only future weather changes)

For the wheat and sorghum datasets, we use APSIM simulations that share a complete initial configuration and an identical first 84 days. From day 85, we introduce only future weather changes (rainfall ×0.5 or warming), yielding a fixed-history benchmark with 3,000 trajectories / 2,000 intervention pairs, split by year into 70/15/15. GEWorld assimilates the shared history once, copies that state, and forecasts both branches under weather-only future inputs. We evaluate the model against APSIM using the paired effect Δy^=y^int−y^fact\widehat{\Delta y}=\hat y_{\mathrm{int}}-\hat y_{\mathrm{fact}}. Effect MAE over both interventions and all 12 forecast weeks:

Crop / traitGEWorldZero effectTraining mean effectRidge
Wheat biomass20.90786.50151.18149.099
Wheat grain7.66324.10613.85913.881
Wheat LAI0.1620.3770.3250.307
Sorghum biomass71.302131.723127.131120.060
Sorghum grain35.82071.49656.52053.680
Sorghum LAI0.2530.3540.3320.325

GEWorld’s improvements over the training-mean and ridge baselines further suggest that it captures variation in intervention magnitudes across conditions beyond an average effect or a linear approximation.

We have summarized these results in Section 4.2 (RQ3), and provided the full evaluation in Appendix C.6 of our revision.


Q2. How genotype information contributes to dynamic prediction

On sorghum TERRA-REF at a 50% observed prefix, G00 disables both SNP routes, G10 enables initialization only, G01 enables transition FiLM only, and G11 enables both. We test each with the current phenotype alone or the full prefix, keeping weather, anchors, targets, and training budgets fixed. Disabled routes use a shared zero-SNP encoder reference to match capacity. The eight settings use five genotype splits and three initializations per split, totaling 120 runs. We report mean test R2R^2 ± sample SD across splits. Since G00 can still receive genetic information through phenotypes, these controls measure the added value of explicit SNP inputs.

Phenotype historySNP routeCover R2R^2Height R2R^2
Current onlyG000.7880 ± 0.01520.9136 ± 0.0117
Current onlyG100.7929 ± 0.01470.9231 ± 0.0054
Current onlyG010.7811 ± 0.02020.9260 ± 0.0047
Current onlyG110.7824 ± 0.02430.9256 ± 0.0053
Full prefixG000.7833 ± 0.01510.9180 ± 0.0079
Full prefixG100.7770 ± 0.02390.9189 ± 0.0065
Full prefixG010.7477 ± 0.02420.9261 ± 0.0049
Full prefixG110.7384 ± 0.02420.9242 ± 0.0049

For height, initialization and transition FiLM improve mean R2R^2 by 0.0095 and 0.0124 with only the current phenotype, with both gains positive in all five splits. With the full prefix, the initialization gain shrinks to 0.0010, while transition FiLM retains a mean gain of 0.0082, positive in all five splits. Cover shows no consistent benefit and declines under the full prefix. These results suggest that the additional value of explicit SNP information depends on the trait, phenotype history, and route used.

We further conduct a genotype-level response analysis. Aggregating MSE over the five splits for each held-out genotype (225 genotypes), explicit SNP routes improve height for a majority of genotypes:

Phenotype historyComparisonGenotypes with lower MSE (%)Mean MSE reduction
Current onlyG10 vs G0058.7%45.5
Current onlyG01 vs G0062.7%58.4
Current onlyG11 vs G0066.2%56.3
Full prefixG10 vs G0052.0%—
Full prefixG01 vs G0056.4%—
Full prefixG11 vs G0057.3%—

An identity-sensitivity probe that zeros or permutes SNP inputs on the trained G11 model also worsens height for most trajectories, indicating that genetic identity is not fully absorbed by the phenotype channel alone:

Probe on trained G11Trajectories with increased height error (%)
Zero SNP inputs60.8%
Permute SNP inputs61.7%

Together, these findings support an incremental contribution of explicit SNPs through both initialization and transition, beyond genetic information implicit in observed growth history.

We have summarized these findings in Section 4.2 (RQ4), and provided the detailed analyses in Appendix C.3 of our revision.


AuthorsOpenReview ↗

Rebuttal Part 3/3

Q3. Testing the contribution of earlier history

(i) History-window retraining

At a 50% cutoff on sorghum TERRA-REF, we retrain with the full prefix or only its last 1/2/4 steps. Phenotype and weather are truncated together; the current phenotype, current environment, and future weather stay fixed. Seasonal curve on/off; mean ± SD over five genotype splits:

SeasonalHistoryCover R2R^2Height R2R^2
OnLast-10.7057 ± 0.04520.9188 ± 0.0021
OnLast-20.6893 ± 0.09040.9171 ± 0.0118
OnLast-40.7227 ± 0.05310.9234 ± 0.0047
OnFull0.7439 ± 0.04590.9239 ± 0.0086
OffLast-10.6258 ± 0.06650.9177 ± 0.0044
OffLast-20.6308 ± 0.07340.9172 ± 0.0086
OffLast-40.6443 ± 0.06310.9211 ± 0.0079
OffFull0.6978 ± 0.04710.9237 ± 0.0086

Full history beats last-1 on cover by +0.038 with curve on and +0.072 with curve off, positive in all five splits; wheat gains are +0.038/+0.040 under the same protocol. Height changes little, and longer windows are not strictly better. So joint history helps beyond the current anchors.

(ii) Matched current, different history

We compare forecasts using each test trajectory’s own history (Own history) or replacing its early history with that of a trajectory with a similar current phenotype but a different earlier history (Substituted history). The trained model, genotype, current phenotype, current environment, and future weather are held fixed. ΔR2\Delta R^2 is Own history − Substituted history. “Own better” reports the percentage of trajectories with lower forecast MSE using their own history.

DatasetSeasonalTraitOwn historySubstituted historyΔR2\Delta R^2Own better
SorghumOnCover0.7439±0.04590.7439\pm0.04590.7008±0.04720.7008\pm0.0472+0.04359%
SorghumOnHeight0.9239±0.00860.9239\pm0.00860.9118±0.01260.9118\pm0.0126+0.01263%
SorghumOffCover0.6978±0.04710.6978\pm0.04710.5639±0.08820.5639\pm0.0882+0.13469%
SorghumOffHeight0.9237±0.00860.9237\pm0.00860.9120±0.01280.9120\pm0.0128+0.01260%
WheatOnCover0.8611±0.00280.8611\pm0.00280.7407±0.01780.7407\pm0.0178+0.12072%
WheatOnHeight0.9765±0.00150.9765\pm0.00150.8585±0.03250.8585\pm0.0325+0.11868%
WheatOffCover0.8618±0.00550.8618\pm0.00550.7326±0.01140.7326\pm0.0114+0.12971%
WheatOffHeight0.9756±0.00270.9756\pm0.00270.8389±0.04150.8389\pm0.0415+0.13768%

Swapping early history while matching the current state and future weather lowers R2R^2 in every cell, and own history wins on most trajectories. That is hard to explain by the decoder anchors alone: the latent state is using earlier history.

We have added the history controls in Appendix C.2.


Q4. Transition comparison, and reconciliation of main vs ablation Full

(i) Transition comparison (W4)

Same-width replacement of only the transition on sorghum TERRA-REF at a 70% prefix, with shared D=128D=128, encoders, genotype conditioning, assimilation, and decoder; five genotype splits; seasonal curve on/off. Test R2R^2 is reported as mean ± sample SD over five genotype splits, with one initialization per split:

TransitionCover, onHeight, onCover, offHeight, off
Selective SSM0.5317 ± 0.10460.8315 ± 0.01530.4569 ± 0.15260.8155 ± 0.0136
GRU0.4782 ± 0.17510.8216 ± 0.01500.3432 ± 0.27570.8046 ± 0.0149
RNN0.4864 ± 0.15100.8218 ± 0.01740.3780 ± 0.18800.8036 ± 0.0116
Transformer0.4318 ± 0.18220.8175 ± 0.02180.2904 ± 0.24190.7938 ± 0.0264

At this cutoff, SSM has the highest mean R2R^2 for both traits with the seasonal curve either on or off. Cover margins over GRU/RNN/Transformer are +0.0536/+0.0453/+0.0999 with curve on and +0.1137/+0.0790/+0.1665 with curve off.

(ii) Reconciliation of main vs ablation Full (W5)

To remove any remaining ambiguity between the two Full rows, we re-ran the main-table experiments under the same protocol as the ablation Full and updated the main text. Updated sorghum test R2R^2:

TraitPrefixUpdated GEWorld R2R^2
Height25%0.959±0.0050.959\pm0.005
Height50%0.921±0.0110.921\pm0.011
Height75%0.788±0.0170.788\pm0.017
Cover25%0.953±0.0040.953\pm0.004
Cover50%0.772±0.0080.772\pm0.008
Cover75%0.487±0.1400.487\pm0.140

We have summarized the transition comparison in Section 4.2 (RQ4), detailed it in Appendix C.3, and updated Table 2 of our revision.

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

The paper proposed a Genotype-Environment World Model for Crop Phenotype Prediction, which casts phenotype prediction as the evolution of a latent physiological state. The framework is used for prediction of endpoint and growth trajectories. Comprehensive experiments are conducted to show the superior performance.

Critical strengths
  1. The motivation to represent the Crop Phenotype Prediction through Genotype-Environment World Model is very reasonable. It provides a unified prediction framework which supports endpoint prediction, forecasting of growth trajectories, and simulation under alternative future weather conditions within a shared framework.
  2. The experiments for the four research questions are comprehensive, the strong empirical performance is convincing.
Critical weaknesses
  1. The methodology is hard to follow. The definition of Δtt\Delta t_t, its relationship to Δt\Delta_t, definition of ee and relationship to ete_t, and the definitions and dimensions of BtB_t, CtC_t, and Alog⁡A_{\log} are unclear. How SSM works with states, environment, genotype is also not clearly stated with formula. Some sentences in experiment section seem to be generated by AI and are hard to understand as well.
  2. The Figure 1b does not match with the formulas. e.g FiLM is not applied to generate s_0 according to formula.
Questions for authors
  1. In Figure 2, why biomass, LAI and even height could go up and down?
  2. According to Table S1, Sorghum dataset contains only 569 samples but there are 2,048 markers, how is overfitting controlled? Please report the model’s parameter count, regularization settings, and sensitivity to different train/test splits.
  3. Could you further illustrate the modules for assimilation, transition and Decode more explicitly?
2 comments
AuthorsOpenReview ↗

Rebuttal Part 1/2

We thank the reviewer for the valuable comments. We address each point below.

W1. Notation and SSM operation

We thank the reviewer for pointing this out. We will improve the readability of the method description and clarify the notation. The core idea of GEWorld is to initialize a latent state from genotype and an environmental summary, then use an SSM to advance this state under step-specific environmental conditions and elapsed time, with genotype conditioning the transition through FiLM.

Here, ee is a trial-level summary of the environmental sequence used for initialization, whereas ete_t contains the environmental covariates at step tt. Δtt\Delta t_t denotes the observed time interval between consecutive steps, whereas Δt\boldsymbol\Delta_t is a learned positive SSM update scale. The former influences the latter through the transition input.

We will define these symbols and dimensions at first use and expand the transition equations. We will polish the unclear experimental descriptions and will carefully proofread the entire manuscript to ensure clear, precise, and consistent wording.

We have clarified the notation and state-transition formulation in Sections 3.1–3.3 of our revision.

W2. Figure 1b and FiLM

Thank you for pointing out the inconsistency between the figure and the text. We will revise Figure 1b to follow the equations step by step, from genotype/environment encoding and gated G×E fusion to initialization, transition, assimilation, and decoding. Each block will use the same symbols as the text and reference.

For initialization, we will explicitly draw the conditioning arrows from zez_e to environment-conditioned FiLM and from zgz_g to genotype-conditioned FiLM, placing LayerNorm between the two operations. We will refine the equation details in the main text and clarify the methodology.

For the dynamic pathway, we will label the environment/time inputs to transition, phenotype/mask inputs to assimilation, and state-to-phenotype output of decoding, distinguishing the observed prefix from future rollout. The caption will explain these connections so that the figure, equations, and execution order are consistent.

We have made the FiLM initialization order and the transition–assimilation execution order explicit in Sections 3.2–3.3.

Q1. Why can biomass, LAI, and height decrease?

We thank the reviewer for the careful observation. The fluctuations in Figure 2 have several sources: (1) the curves average trajectories of different lengths and observation availability, so changes in the contributing samples across forecast horizons can produce dips even when individual trajectories continue increasing; (2) for simulated sorghum, preprocessing combined records from four planting densities within each genotype–environment sequence, causing oscillations as consecutive records switch between density conditions, whereas biomass within each density is nondecreasing; (3) LAI can decline naturally as leaves senesce; and (4) field height records also fluctuate, potentially reflecting measurement variability or changes in canopy structure.

We will clarify this point in the main text.

AuthorsOpenReview ↗

Rebuttal Part 2/2

Q2. Overfitting with 569 samples and 2,048 markers

We thank the reviewer for this insightful comment. High-dimensional genomic inputs and limited training samples increase the risk of overfitting in phenotype prediction., as genome-wide markers are abundant while phenotyping across genotypes and environments is costly. Our panel contains 569 genotype–season records from 296 genotypes. We use five genotype-disjoint splits, each with 207/44/45 train/validation/test genotypes; marker selection and normalization use training data only. With seven initializations averaged per split, the manuscript reports test R2=0.560±0.080R^2=\mathbf{0.560\pm0.080} across splits. GEWorld has 5,559,173 registered parameters. We use dropout 0.05, weight decay 10−410^{-4}, and validation-MSE early stopping (patience 20; maximum 100 epochs).

Additional matched controls use the same five splits and three initializations per setting:

SettingTrain R2R^2Test R2R^2 (mean ± SD)
Full model (d=256d=256), regularization + early stopping0.8840.556 ± 0.066
Same model without early stopping (epoch 100)0.9980.478 ± 0.122
No dropout or weight decay, with early stopping0.8940.525 ± 0.078
Smaller model (d=64d=64), regularization + early stopping0.9100.541 ± 0.070

Early stopping and dropout/weight decay improve mean test R2R^2 by 0.078 and 0.031, respectively; reducing width provides no improvement in mean test performance. Without early stopping, training R2R^2 reaches 0.998 while test performance declines and its variability increases, consistent with overfitting. The smaller model’s lower mean test performance further suggests that reducing capacity alone does not improve generalization under these settings. These controls support the effectiveness of our overfitting safeguards.

We have incorporated these results in Appendix C.5.2 of our revision.

Q3. Roles of assimilation, transition, and decoding

We thank the reviewer for this suggestion. The three modules have complementary roles in tracking observed development and forecasting future growth:

  1. Assimilation corrects the model’s estimate of the plant’s current condition. As phenotype measurements become available, this module incorporates them into the latent state while retaining information from earlier observations. This keeps the state aligned with observed development and allows new measurements to correct errors accumulated during state evolution. Assimilation is skipped when no measurements are available.

  2. Transition models how the plant’s condition evolves over time. It carries developmental information forward and updates it as environmental conditions and elapsed time change. Genotype conditioning allows different genotypes to respond differently to the same weather. This module enables the model to continue forecasting growth after phenotype observations stop.

  3. Decoding translates the latent condition into predicted traits. It turns the learned developmental representation into quantities such as height, canopy cover, biomass, and LAI. For dynamic forecasting, it combines a predicted change from the preceding phenotype with a reference based on the training-set seasonal pattern and a bounded correction. The first component captures changes specific to the current state, while the reference helps limit accumulated drift over longer forecasts.

During the observed prefix, the first measurement directly corrects the initialized state; subsequent steps alternate between transition and assimilation when measurements are available. After the cutoff, transition advances the state and decoding generates the phenotype forecast at each step. Assimilation stops because future phenotype measurements are unavailable.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.