acceptodds
Under review as a conference paper at ICLR 2027

ScaleCity: Joint Multi-Scale and Multi-Task Urban Region Representation Learning

12 views

Abstract

Urban region representation learning must accommodate tasks defined over different granularities and regional boundaries. Changes in these spatial configurations alter feature aggregation and target statistics, reflecting the scale and zoning effects of the Modifiable Areal Unit Problem (MAUP). We present ScaleCity, a multimodal urban region representation learning framework that jointly trains one model per city on mixed task–configuration samples. Joint training replaces separately fitted predictors with one model that supports multiple tasks and configurations. Graph-enhanced contrastive pretraining aligns street-view imagery, points-of-interest text, and remote-sensing features on atomic H3 cells. A prompt-conditioned hypernetwork adapts multimodal fusion and prediction to each task–configuration pair, while regional aggregation retains explicit extent information. Across three Chinese megacities, eight tasks, and nine configurations spanning grids, road-network partitions, and administrative regions, ScaleCity outperforms nine baselines in all task- and configuration-averaged R2 comparisons. Further experiments demonstrate zero-shot prediction at unseen configurations within the same city, while few-shot experiments examine adaptation to unseen tasks. These findings support joint multi-scale, multi-task learning of urban region representations. Code is available at https://github.com/far-rele/ScaleCity.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Urban region representation learning, multi-scale learning, prompt-conditioned hypernetwork, multi-modal learning, Modifiable Areal Unit Problem (MAUP), Spatial data mining

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Mp8sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ScaleCity addresses the learning of urban area representations across multiple tasks and spatial configurations, a problem of practical value. The reuse of atomic grids, the generation of conditional parameters, and the explicit modelling of area size are well-targeted; the experiments cover multiple cities, tasks and spatial configurations, and report significant performance improvements. However, the sources of these performance improvements have not yet been fully disentangled, and the evaluation independence and the accuracy–cost trade-off require further verification.

Critical strengths

1: The problem is closely aligned with practical needs, jointly addressing different spatial scales and area boundaries, and thus holds practical value.

2: The atomic grid reuse, conditional parameter generation and area size information are well-designed.

3: The experiments cover multiple cities, tasks and spatial configurations, and report significant performance improvements.

Critical weaknesses

1: Some baseline results differ significantly from those in the corresponding original papers. This paper has modified the implementation of certain baselines, and the rationale for these modifications requires explanation. Furthermore, the inputs for some baselines differ from the original settings; Appendix C.4 reports that the baselines utilised settings from their original papers, or default or fixed parameters, without conducting systematic hyperparameter search. It remains unverified whether these parameters remain applicable following changes to the inputs and implementation.

2: The independence of the evaluation set needs to be strengthened. Appendix C.2 and the open-source code (prompt_learning/train.py) indicate that the paper uses the same holdout fold for both checkpoint selection and final scoring, which may lead to over-optimistic results.

3: This paper and the baselines differ in terms of inputs, supervision methods and prediction architectures. To clarify the contribution of the hypernetwork, the first-stage representation must be fixed; independent predictors, simple multi-task models and the hypernetwork proposed in this paper should be compared under conditions where the same region-size information, aggregation methods and evaluation protocols are used, in order to isolate the contributions of each component.

4: The use of an overlap-weighted average for regional features may result in the loss of intra-region differences and spatial arrangements. The final aggregation may cause regions with different internal structures to receive similar representations.

5: There is a lack of efficiency comparisons under uniform conditions. The time taken per epoch is insufficient to demonstrate the overall cost advantage of the shared model. The total training cost, including pre-training, should be reported and compared with standalone predictors based on the same backbone to clarify the trade-off between accuracy and cost.

Questions for authors

This is a valuable piece of research. From a reviewer’s perspective, whilst I acknowledge its practical value and system design, I currently have reservations regarding the comparability of baselines, the independence of the evaluation, and the sources of performance gains. The aforementioned experiments will directly influence the assessment of the main conclusions.

9 comments
AuthorsOpenReview ↗

W1: Baseline implementations and hyperparameters

W1. Some baseline results differ significantly from those in the corresponding original papers. This paper has modified the implementation of certain baselines, and the rationale for these modifications requires explanation. Furthermore, the inputs for some baselines differ from the original settings; Appendix C.4 reports that the baselines utilised settings from their original papers, or default or fixed parameters, without conducting systematic hyperparameter search. It remains unverified whether these parameters remain applicable following changes to the inputs and implementation.

Response:

We thank the reviewer for asking us to clarify baseline performance and implementation choices.

1. Baseline performance and fairness. Each baseline is trained separately for every task–configuration–fold and evaluated on the same regions, labels, and metrics. Table 1’s Configuration R² (min–max) shows that baselines perform strongly at favorable scales, consistent with their original scale-specific evaluations. Exact scores also depend on cities, data sources, and regional definitions. Averaging across configurations captures weaknesses beyond those favorable scales. ScaleCity sustains performance across all evaluated shapes and scales with a single jointly trained city model.

2. Reasons for implementation adaptations. We adapted the baselines to accept our multi-scale regional inputs (e.g., removing positional encodings tied to fixed grids or replacing fusion modules incompatible with these inputs), while preserving each method’s core architecture. Appendix C.4 documents these implementation choices.

3. Hyperparameter applicability. We used the published, default, or fixed settings in Appendix C.4 without systematic per-method search. Baseline performance at favorable scales remains consistent with the original studies, supporting the applicability of published or default settings under our adapted inputs. The nested cross-validation rerun (W2) also preserves ScaleCity’s overall lead in all three cities.

AuthorsOpenReview ↗

W2: Independent validation

Comment:

W2. The independence of the evaluation set needs to be strengthened. Appendix C.2 and the open-source code (prompt_learning/train.py) indicate that the paper uses the same holdout fold for both checkpoint selection and final scoring, which may lead to over-optimistic results.

Response:

We thank the reviewer for checking the released code and recommending independent validation.

1. Nested protocols. Outer folds are identical to Table 1, allowing direct comparison. Within each outer training portion (80%), 12.5% of L11 anchors form an inner validation set, approximately 10% of all anchors. Scheduling, early stopping, and checkpoint selection use only inner-validation metrics; ScaleCity uses the macro average over 72 task–configuration pairs. After training, the selected checkpoint is loaded for one inference pass on the outer test fold, which is excluded from downstream training and all selection decisions.

2. Baseline handling. All baselines requiring training-time selection—ViT, SpaBERT, GeoVectors, CityFM, HAFusion, and HREP—were retrained under the same nested protocol. ComSRE, Urban2Vec, and UrbanMFM use fixed settings and a single fit per fold. Their Table 1 scores already follow “fit on training folds, evaluate once on the test fold” and therefore remain unchanged.

3. Results (Overall R², submission → nested).

MethodBeijingChengduShenzhen
ScaleCity (ours)0.7833 → 0.72810.6707 → 0.59840.6441 → 0.5715
ViT0.3450 → 0.20250.2707 → 0.06820.2953 → 0.0857
SpaBERT0.2922 → 0.18240.3397 → 0.22420.1814 → 0.0354
GeoVectors0.2712 → 0.12140.2795 → 0.12140.1793 → −0.0770
CityFM0.2200 → 0.14490.1735 → 0.12650.1495 → −0.0288
HAFusion0.2712 → 0.30390.3324 → 0.29250.2396 → 0.2104
HREP0.0978 → 0.03040.1011 → −0.01670.0783 → 0.0038
ComSRE (single-fit, unchanged)0.32550.37750.3348
Urban2Vec (single-fit, unchanged)0.24660.25290.2439
UrbanMFM (single-fit, unchanged)0.23580.13180.0531

ScaleCity remains first in all three cities. Relative gains over each protocol’s strongest baseline change from +127.0% / +77.7% / +92.4% to +123.7% / +58.5% / +70.7% in Beijing / Chengdu / Shenzhen.

4. Supplementary split robustness. Our earlier comparison of three random data splits yielded an R² standard deviation of 0.0332, providing additional evidence of stability across the tested splits.

AuthorsOpenReview ↗

W3: Matched predictor controls

Comment:

W3. This paper and the baselines differ in terms of inputs, supervision methods and prediction architectures. To clarify the contribution of the hypernetwork, the first-stage representation must be fixed; independent predictors, simple multi-task models and the hypernetwork proposed in this paper should be compared under conditions where the same region-size information, aggregation methods and evaluation protocols are used, in order to isolate the contributions of each component.

Response:

We thank the reviewer for asking us to clarify the baseline adaptations and distinguish the contributions of the inputs, joint supervision, and conditional prediction.

1. Matched inputs and training setup. Table 1 baselines use method-specific inputs and separate predictors, adapted to multiple regional configurations as explained in W1 and Appendix C.4. The controls below hold fixed the frozen SVI, POI, and AlphaEarth embeddings, extent covariate, overlap-weighted aggregation, sampling, five-fold splits, and evaluation metric:

  • P: 72 independent MLPs, one per task–configuration pair (0.41M parameters each).
  • J1/J2: a joint MLP (0.42M parameters), using learnable task/configuration ID embeddings or one-hot identifiers.
  • H: a joint shared trunk with one head per task–configuration pair (1.59M parameters).
  • S: the full ScaleCity hypernetwork with descriptive prompts.

Overall R² under the submission protocol (five-fold mean):

MethodBeijingChengduShenzhen
P: independent MLP0.53170.36900.3611
J1: joint MLP (learnable ID embeddings)0.70820.64010.5847
J2: joint MLP (one-hot)0.66540.61840.5693
H: multi-head model0.63210.47110.5862
S: ScaleCity (hypernetwork)0.78330.67070.6441

2. Joint training and conditional prediction. P→J1 gains +0.177 / +0.271 / +0.224 R² in Beijing / Chengdu / Shenzhen. ScaleCity exceeds J1 by +0.075 / +0.031 / +0.059, and also exceeds J2 and H in every city. AlphaEarth and extent are available to every predictor, so their availability does not explain these differences.

3. Sparse regional supervision. The coarsest road-network configuration L1 has 207 / 114 / 102 regions in Beijing / Chengdu / Shenzhen, with only 128 / 58 HFP-labeled regions in Beijing / Chengdu. Each L1 head sees about 41–102 training samples per fold. H has cross-task mean R² −0.62 / 0.03 in Chengdu / Beijing, versus J1's 0.546 / 0.505 and ScaleCity's 0.385 / 0.494, supporting sharing across sparsely labeled task–configuration pairs.

4. Parameter-generation components. The following ablations retain the frozen modalities, extent, descriptive prompts, and joint training. They replace generated projections and gates with shared trainable projections and uniform weights, or replace the generated readout with a shared static head:

SettingBeijingChengduShenzhen
Full ScaleCity0.7830.6710.644
Shared modality projections + uniform gates0.7100.5610.532
Shared static prediction head0.6320.5850.578

Both replacements reduce accuracy, supporting conditional fusion and prediction with the input information and prompt form unchanged.

5. Identity versus descriptive text. Within the same hypernetwork, descriptions and symbolic identifiers achieve three-city mean R² of 0.699 and 0.695 on known conditions (Appendix G). Identity is therefore effective, with a modest descriptive-text benefit. The component controls retain descriptions while changing parameter generation. Descriptions also represent unseen configurations in the transfer experiments.

Reviewer_kuMqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ScaleCity, which jointly predicts multiple urban targets across different spatial configurations using one model per city. It combines multimodal cell representations with a task- and configuration-conditioned hypernetwork for feature fusion and prediction. The reported results are strong, but concerns about the evaluation protocol and baseline comparisons limit my confidence in the conclusions.

Critical strengths

S1. The problem is practically relevant. Supporting multiple tasks and spatial configurations within one model is useful for urban applications.

S2. The design is reasonable. Conditional fusion allows different tasks to use the modalities differently, and explicit region-size information helps predict regional totals.

S3. The experiments cover multiple cities, tasks, and configurations, with useful ablations on the main components.

Critical weaknesses

W1. The evaluation protocol is a major concern. Appendix C.2 states that the same held-out fold is used for model selection and final scoring. This can produce optimistic results. An independent test fold is needed to assess the reported gains reliably.

W2. The need for prompt text is not clearly established. My intuition is that using simple one-hot embeddings, identifiers or learnable embeddings might achieve the same effect, as it might only need some conditions to distinguish tasks. Comparing these alternatives would help clarify whether text semantics provide benefits beyond condition identity.

W3. The choice to freeze the first-stage encoder needs further justification. Sharing cell representations across tasks does not require freezing the encoder, as it could also be jointly fine-tuned with the downstream modules. Since each model is trained within one city, it is unclear whether the computational savings justify keeping the encoder fixed. A comparison with partial or full fine-tuning, reporting both prediction accuracy and total training cost, would clarify this trade-off. Could you also report the time spent on first-stage pretraining, and downstream training time separately?

Questions for authors

W1, W2 and W3

4 comments
AuthorsOpenReview ↗

W1: Independent validation

W1. The evaluation protocol is a major concern. Appendix C.2 states that the same held-out fold is used for model selection and final scoring. This can produce optimistic results. An independent test fold is needed to assess the reported gains reliably.

Response:

We thank the reviewer for recommending independent validation. The submission compared methods using the same folds, regions, labels, and metrics. The nested rerun further separates model selection from final testing, avoiding selection-induced optimism in the absolute scores.

1. Nested protocols. Outer folds are identical to Table 1, allowing direct comparison. Within each outer training portion (80%), 12.5% of L11 anchors form an inner validation set, approximately 10% of all anchors. Scheduling, early stopping, and checkpoint selection use only inner-validation metrics; ScaleCity uses the macro average over 72 task–configuration pairs. After training, the selected checkpoint is loaded for one inference pass on the outer test fold, which is excluded from downstream training and all selection decisions.

2. Baseline handling. All baselines requiring training-time selection—ViT, SpaBERT, GeoVectors, CityFM, HAFusion, and HREP—were retrained under the same nested protocol. ComSRE, Urban2Vec, and UrbanMFM use fixed settings and a single fit per fold. Their Table 1 scores already follow “fit on training folds, evaluate once on the test fold” and therefore remain unchanged.

3. Results (Overall R², submission → nested).

MethodBeijingChengduShenzhen
ScaleCity (ours)0.7833 → 0.72810.6707 → 0.59840.6441 → 0.5715
ViT0.3450 → 0.20250.2707 → 0.06820.2953 → 0.0857
SpaBERT0.2922 → 0.18240.3397 → 0.22420.1814 → 0.0354
GeoVectors0.2712 → 0.12140.2795 → 0.12140.1793 → −0.0770
CityFM0.2200 → 0.14490.1735 → 0.12650.1495 → −0.0288
HAFusion0.2712 → 0.30390.3324 → 0.29250.2396 → 0.2104
HREP0.0978 → 0.03040.1011 → −0.01670.0783 → 0.0038
ComSRE (single-fit, unchanged)0.32550.37750.3348
Urban2Vec (single-fit, unchanged)0.24660.25290.2439
UrbanMFM (single-fit, unchanged)0.23580.13180.0531

ScaleCity remains first in all three cities after an R² decrease of 0.055–0.073. Gains over the strongest baseline under each protocol change from +127.0% / +77.7% / +92.4% to +123.7% / +58.5% / +70.7% in Beijing / Chengdu / Shenzhen.

4. Supplementary split robustness. Our earlier comparison of three random data splits yielded an R² standard deviation of 0.0332, providing additional evidence of stability across the tested splits.

AuthorsOpenReview ↗

W2: Prompt text and identifier alternatives

Comment:

W2. The need for prompt text is not clearly established. My intuition is that using simple one-hot embeddings, identifiers or learnable embeddings might achieve the same effect, as it might only need some conditions to distinguish tasks. Comparing these alternatives would help clarify whether text semantics provide benefits beyond condition identity.

Response:

We thank the reviewer for suggesting these conditioning alternatives.

1. Symbolic identifiers within the same hypernetwork. Appendix G compares six prompt schemes using the same frozen text encoder. Descriptions (A) and symbolic identifiers (E, Table 15) achieve three-city mean Overall R² of 0.699 and 0.695 on known conditions. With architecture fixed, condition identity already provides most of the discrimination, while descriptive wording adds a small improvement.

2. Learnable and one-hot alternatives. The controls share ScaleCity's frozen Stage-1 embeddings, overlap-weighted aggregation, extent covariate, sampling, five-fold splits, and evaluation metric:

  • J1: a shared multi-task MLP (0.42M parameters) with learnable task/configuration ID embeddings.
  • J2: the same MLP with one-hot identifiers.
  • S: ScaleCity's hypernetwork with natural-language prompts.

Overall R² under the submission protocol (five-fold mean):

MethodBeijingChengduShenzhen
J1: joint MLP (learnable ID embeddings)0.70820.64010.5847
J2: joint MLP (one-hot)0.66540.61840.5693
S: ScaleCity (natural-language prompt conditioning)0.78330.67070.6441

ScaleCity outperforms both alternatives in all three cities. These controls compare complete conditioning architectures; A/E separately tests prompt content within the same hypernetwork.

3. Extending to unseen conditions. J1/J2 have trained categorical representations for known configurations; an unseen configuration has no trained identifier representation in these controls. ScaleCity encodes its description without adding parameters, supporting zero-shot prediction at unseen configurations within the same city without inference-time updates (Table 3: R² 0.412–0.937 for HP task). Descriptions thus offer extensibility alongside competitive known-condition accuracy.

AuthorsOpenReview ↗

W3: Frozen encoders — accuracy and total cost

Comment:

W3. The choice to freeze the first-stage encoder needs further justification. Sharing cell representations across tasks does not require freezing the encoder, as it could also be jointly fine-tuned with the downstream modules. Since each model is trained within one city, it is unclear whether the computational savings justify keeping the encoder fixed. A comparison with partial or full fine-tuning, reporting both prediction accuracy and total training cost, would clarify this trade-off. Could you also report the time spent on first-stage pretraining, and downstream training time separately?

Response:

We thank the reviewer for requesting the accuracy–cost comparison and separate pretraining times.

1. Design choice. Sharing representations does not require freezing; joint encoder fine-tuning is possible. We choose freezing based on the measured accuracy–cost trade-off, with task adaptation handled downstream.

2. Partial fine-tuning comparison. Both settings use the protocol with identical folds, validation settings, and downstream hyperparameters. Times cover five folds on 1×A6000.

  • Frozen: Stage 1 is fixed; all downstream modules train jointly.
  • Partial FT: the three GATs, three projection heads, and AE lift train at 1e-4 alongside the 12.85M-parameter downstream model. CLIP/MacBERT remain fixed; all city-cell embeddings are recomputed differentiably at every step.
SettingOverall R² (BJ / CD / SZ)Downstream time (BJ / CD / SZ)
Frozen0.783 / 0.671 / 0.6442.09 / 0.84 / 0.74 h
Partial FT0.792 / 0.679 / 0.6488.19 / 1.56 / 1.49 h

Partial fine-tuning adds +0.009 / +0.008 / +0.004 Overall R² at 3.9× / 1.9× / 2.0× downstream time.

3. Pretraining and total cost. One-time Stage-1 pretraining uses 4×A6000 DDP for 40 epochs, taking 6.97 / 3.49 / 3.73 h in Beijing / Chengdu / Shenzhen. Both settings share this pretraining. Totals below sum the reported pretraining and five-fold downstream stages on their stated hardware:

CityFrozen total elapsed timePartial-FT total elapsed timeFrozen total GPU-hoursPartial-FT total GPU-hours
Beijing9.06 h15.16 h29.9736.07
Chengdu4.33 h5.05 h14.8015.52
Shenzhen4.47 h5.22 h15.6616.41

With stage times in hours, elapsed time is Tpre+TdownT_{\mathrm{pre}}+T_{\mathrm{down}}; GPU-hours are 4Tpre+Tdown4T_{\mathrm{pre}}+T_{\mathrm{down}}.

4. Within-city reuse. One pretraining run serves all 72 task–configuration pairs across five folds. Freezing reuses cached embeddings across these runs and unseen-configuration prediction. Trainable downstream projections, conditioned transformations, fusion gates, and the dynamic head provide task adaptation while avoiding repeated encoder optimization for the small gains observed above.

AuthorsOpenReview ↗

Thank you for your thoughtful review

Dear Reviewer kuMq,

Thank you for your careful review and recognition of ScaleCity's practical relevance, conditional design, and experimental coverage. Below, we report the independent-validation rerun, conditioning controls, and frozen-versus-fine-tuned comparison, including pretraining and total costs. We welcome further questions.

Best regards,

The authors

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

ScaleCity addresses urban prediction across multiple tasks and spatial configurations using one jointly trained model per city. It first learns atomic H3-cell representations by aligning street-view imagery, POI text, and AlphaEarth features with graph-enhanced contrastive learning. A task- and configuration-conditioned hypernetwork generates modality transformations, fusion gates, and prediction parameters. Regional representations combine overlap-weighted mean features with an explicit extent covariate. Experiments cover three Chinese cities, eight tasks, and nine spatial configurations, with additional evaluations of unseen configurations, few-shot task adaptation, and separately trained models in two international cities.

Critical strengths
  1. The problem formulation and architecture are well motivated. Supporting multiple prediction targets and regional boundaries within one city-specific model is practically useful. Reusable atomic representations provide a natural basis for flexible aggregation, while explicitly retaining region extent addresses an important limitation of mean pooling for extensive targets such as population and GDP.
  2. The paper provides extensive component analyses and clearly scoped supplementary experiments. The ablations examine modalities, spatial propagation, conditional computation, and region extent. The distinction between extensive and intensive targets helps explain where extent information matters. The prompt analysis is also informative: symbolic identifiers perform similarly to descriptive prompts on known conditions, while descriptions show a larger benefit in few-shot adaptation.
Critical weaknesses
  1. The evaluation does not provide a clean separation between model selection and final testing, and spatial separation is incomplete. Appendix C.2 explicitly states that the same held-out fold is used for model selection and final scoring; Appendix C.3 specifies validation-driven scheduling, early stopping, and checkpoint selection. Consequently, the reported cross-validation scores are subject to selection bias. Furthermore, assigning regions to folds through their centroid’s L11 anchor does not ensure disjoint geographic support across configurations: coarse training regions can overlap held-out regions and share underlying label sources. The extent of this overlap is not quantified. These issues materially weaken confidence in the reported generalization performance.

  2. The comparisons do not sufficiently isolate the contribution of the proposed conditional architecture. ScaleCity receives joint supervision across tasks and configurations, whereas baselines fit separate predictors and use differing input sources. Several baseline implementations also modify their original architectures, including removing positional encoding or replacing fusion modules. Meanwhile, region extent and AlphaEarth features account for substantial gains, and descriptive prompts offer little advantage over symbolic identifiers on the primary benchmark. A strong control using identical inputs, extent information, joint supervision, and a simpler task/configuration-conditioned predictor is needed to establish the value of ST-HyperNet. The unseen-configuration results also lack comparative baselines and evaluate the same city geography, limiting conclusions about broader spatial generalization.

Questions for authors
  1. Can the main results be repeated with an independent inner validation split for scheduling and checkpoint selection, leaving each outer test fold untouched? How much do the reported gains change?
  2. What fraction of training and test regions share geographic area, atomic H3 cells, or underlying label-source cells across configurations? Can you report results under a geographically disjoint split with boundary-crossing regions removed or appropriately assigned?
  3. How does ScaleCity compare with a jointly trained MLP or multi-head model using the same frozen multimodal embeddings, extent covariate, sampling strategy, and task/configuration identifiers? Does hypernetwork-based parameter generation retain a substantial advantage?
  4. For unseen-configuration prediction, can you compare against simple alternatives such as transferring a source-trained predictor or aggregating atomic predictions, and separately report performance on geography excluded from source-configuration supervision?
6 comments
AuthorsOpenReview ↗

Q1 / W1(a): Independent validation and performance changes

W1(a). The evaluation does not provide a clean separation between model selection and final testing, and spatial separation is incomplete. Appendix C.2 explicitly states that the same held-out fold is used for model selection and final scoring; Appendix C.3 specifies validation-driven scheduling, early stopping, and checkpoint selection. Consequently, the reported cross-validation scores are subject to selection bias.

Q1. Can the main results be repeated with an independent inner validation split for scheduling and checkpoint selection, leaving each outer test fold untouched? How much do the reported gains change?

Response:

We thank the reviewer for requesting an independent evaluation. We reran ScaleCity and all baselines involving training-time model selection with nested cross-validation, reserving each outer test fold for final scoring.

1. Nested protocols. Outer folds are identical to Table 1, allowing direct comparison. Within each outer training portion (80%), 12.5% of L11 anchors form an inner validation set, approximately 10% of all anchors. Scheduling, early stopping, and checkpoint selection use only inner-validation metrics; ScaleCity uses the macro average over 72 task–configuration pairs. After training, the selected checkpoint is loaded for one inference pass on the outer test fold, which is excluded from downstream training and all selection decisions.

2. Baseline handling. All baselines requiring training-time selection—ViT, SpaBERT, GeoVectors, CityFM, HAFusion, and HREP—were retrained under the same nested protocol. ComSRE, Urban2Vec, and UrbanMFM use fixed settings and a single fit per fold. Their Table 1 scores already follow “fit on training folds, evaluate once on the test fold” and therefore remain unchanged.

3. Results (Overall R², submission → nested).

MethodBeijingChengduShenzhen
ScaleCity (ours)0.7833 → 0.72810.6707 → 0.59840.6441 → 0.5715
ViT0.3450 → 0.20250.2707 → 0.06820.2953 → 0.0857
SpaBERT0.2922 → 0.18240.3397 → 0.22420.1814 → 0.0354
GeoVectors0.2712 → 0.12140.2795 → 0.12140.1793 → −0.0770
CityFM0.2200 → 0.14490.1735 → 0.12650.1495 → −0.0288
HAFusion0.2712 → 0.30390.3324 → 0.29250.2396 → 0.2104
HREP0.0978 → 0.03040.1011 → −0.01670.0783 → 0.0038
ComSRE (single-fit, unchanged)0.32550.37750.3348
Urban2Vec (single-fit, unchanged)0.24660.25290.2439
UrbanMFM (single-fit, unchanged)0.23580.13180.0531

ScaleCity remains first in all three cities after an R² decrease of 0.055–0.073. Gains over the strongest baseline under each protocol change from +127.0% / +77.7% / +92.4% to +123.7% / +58.5% / +70.7% in Beijing / Chengdu / Shenzhen.

AuthorsOpenReview ↗

Q2 / W1(b): Geographic overlap and evaluation after removing boundary-crossing regions

W1(b). Furthermore, assigning regions to folds through their centroid's L11 anchor does not ensure disjoint geographic support across configurations: coarse training regions can overlap held-out regions and share underlying label sources. The extent of this overlap is not quantified. These issues materially weaken confidence in the reported generalization performance.

Q2. What fraction of training and test regions share geographic area, atomic H3 cells, or underlying label-source cells across configurations? Can you report results under a geographically disjoint split with boundary-crossing regions removed or appropriately assigned?

Response:

We thank the reviewer for asking us to distinguish cross-scale fold alignment from the geographic support of supervision. The shared L11 partition is designed to keep regions assigned to the same anchor in the same fold across configurations. We clarify the remaining boundary case and report the corresponding filtering experiment below.

1. How crossing can occur despite shared L11 folds. Consider two adjacent L11 anchors: A belongs to training and B to testing. A 3 km region whose centroid lies in A inherits the training fold, but part of its area may extend into B. A finer region located in B belongs to testing and can therefore overlap that coarse training region. For additive targets such as population, the coarse training label may include contributions from the test region’s location.

This overlap stems from the interaction between cross-scale regional boundaries and centroid-based fold assignment. It concerns the common evaluation split, rather than ST-HyperNet’s conditional architecture or the reuse of label-free H3 features. We therefore assess its effect using the same boundary-region filter for all methods.

2. Boundary-crossing statistics and filtering. We reproduce Table 1’s L11-anchored five-fold partition (seed 2026) and calculate each region’s intersection with the opposite fold’s anchor territory. A training or test region is flagged when this intersection exceeds 0.1% of its area. Five-fold averages are:

CityCrossing / training regions per fold (%)Crossing / test regions per fold (%)Mean crossing-area fraction among affected regions: training / test
Beijing1,499 / 15,610 (9.6%)113 / 3,902 (2.9%)5.5% / 7.7%
Chengdu651 / 9,018 (7.2%)49 / 2,254 (2.2%)5.6% / 7.3%
Shenzhen449 / 8,232 (5.5%)53 / 2,058 (2.6%)4.0% / 5.2%

These statistics quantify regional boundary crossings. Changing the threshold from 0.1% to either 0% or 1% changes the crossing counts by approximately 3%–8%.

3. Evaluation after removing crossing regions. Following the reviewer’s suggested removal procedure, we exclude flagged regions from both training and testing. ScaleCity and all baselines are retrained or refitted on the retained regions, using Table 1’s hyperparameters and evaluation procedure. This comparison examines geographic filtering separately from the nested-validation rerun in Q1.

Overall R²: Table 1 → filtered evaluation

MethodBeijingChengduShenzhen
ScaleCity0.7833 → 0.76000.6707 → 0.65090.6441 → 0.6264
Strongest baseline¹0.3450 → 0.32360.3775 → 0.36010.3348 → 0.3193
Relative improvement+127.0% → +134.9%+77.7% → +80.8%+92.4% → +96.2%

¹ ViT in Beijing; ComSRE in Chengdu and Shenzhen. All baselines were rerun; only the strongest baseline per city is shown.

The reported Overall R² retains ScaleCity’s lead in all three cities after filtering.

AuthorsOpenReview ↗

Q3 / W2(a): Isolating the contribution of ST-HyperNet

Comment:

W2(a). The comparisons do not sufficiently isolate the contribution of the proposed conditional architecture. ScaleCity receives joint supervision across tasks and configurations, whereas baselines fit separate predictors and use differing input sources. Several baseline implementations also modify their original architectures, including removing positional encoding or replacing fusion modules. Meanwhile, region extent and AlphaEarth features account for substantial gains, and descriptive prompts offer little advantage over symbolic identifiers on the primary benchmark. A strong control using identical inputs, extent information, joint supervision, and a simpler task/configuration-conditioned predictor is needed to establish the value of ST-HyperNet.

Q3. How does ScaleCity compare with a jointly trained MLP or multi-head model using the same frozen multimodal embeddings, extent covariate, sampling strategy, and task/configuration identifiers? Does hypernetwork-based parameter generation retain a substantial advantage?

Response:

We thank the reviewer for requesting baseline clarification and controls for input information, joint supervision, and conditional prediction.

1. Baseline use and adaptations. Table 1 evaluates baselines with their documented inputs and representation-learning procedures, fitting a separate downstream predictor per task–configuration–fold. Adaptations accommodate our multi-scale regional inputs, including adjustments to spatial encoding and fusion, while preserving each method’s core architecture (Appendix C.4).

2. Common-input predictor comparisons. All variants receive identical frozen SVI, POI, and AlphaEarth embeddings, extent information, overlap-weighted aggregation, sampling, five-fold splits, and evaluation metrics. P fits independent predictors; J1, J2, H, and S train jointly.

  • P: 72 independent MLPs, one per task–configuration pair (0.41M parameters each).
  • J1/J2: a joint MLP (0.42M), using learnable task/configuration ID embeddings or one-hot identifiers, respectively.
  • H: a joint shared trunk with one head per pair (1.59M).
  • S: ScaleCity's full hypernetwork with descriptive prompts.

Overall R² under the submission protocol (five-fold mean):

MethodBeijingChengduShenzhen
P: independent MLP0.53170.36900.3611
J1: joint MLP (learnable ID embeddings)0.70820.64010.5847
J2: joint MLP (one-hot)0.66540.61840.5693
H: multi-head model0.63210.47110.5862
S: ScaleCity (hypernetwork)0.78330.67070.6441

Joint ID-conditioned prediction (P→J1) improves R² by +0.177 / +0.271 / +0.224 in Beijing / Chengdu / Shenzhen. ScaleCity further exceeds J1 by +0.075 / +0.031 / +0.059 and outperforms J2 and H. These comparisons hold AlphaEarth and extent availability fixed.

3. Direct controls for conditional parameter generation. To distinguish parameter generation from prompt content, the following ablations retain the same frozen features, extent, descriptive prompts, and joint training. They replace generated modality projections and gates with shared trainable projections and uniform weights, or replace the generated readout with a shared trainable static head.

SettingBeijingChengduShenzhen
Full ScaleCity0.7830.6710.644
Shared modality projections + uniform gates0.7100.5610.532
Shared static prediction head0.6320.5850.578

Both replacements reduce accuracy in every city, providing direct evidence for conditional fusion and prediction with input information and prompt content fixed.

4. Condition identity versus text semantics. Within the same hypernetwork, descriptive prompts and symbolic identifiers achieve three-city mean R² of 0.699 and 0.695 on known conditions (Appendix G). Condition identity is therefore already effective in this setting. The component controls above support conditional computation beyond descriptive wording; Q4 separately evaluates prediction at unseen configurations.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Follow-up on the author–reviewer discussion

Dear Area Chair,

Thank you for the time and care you have devoted to coordinating the review of our paper for ICLR.

We have posted detailed responses to all three reviews, including additional experiments addressing the reviewers’ main questions. We have not yet received follow-up comments, and we would appreciate your help in encouraging the reviewers to share any remaining concerns before the public discussion closes.

We would be glad to clarify any part of our responses or discuss the additional evidence. Thank you again for your time and support throughout the review process.

Best regards, The Authors

AuthorsOpenReview ↗

Thank you for your feedback and continued discussion

Dear Reviewers and Area Chair,

Thank you for the time and care you have devoted to reviewing our paper and supporting the ICLR review process. We sincerely appreciate your thoughtful feedback, which has helped us clarify our contributions and strengthen the supporting evidence.

We have provided point-by-point responses to the reviews, together with additional experiments and analyses addressing the questions raised. As the discussion period draws to a close, we would welcome your thoughts on these responses and any remaining questions or concerns. We would be happy to provide further clarification and continue the discussion.

Thank you again for your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

Request for assistance with the remaining reviewer discussion

Dear Area Chair,

Thank you for your time and effort in coordinating the review of our submission. We posted our detailed responses on November 16, including additional experiments addressing the reviewers' concerns.

Reviewer Mp8s has since provided helpful follow-up feedback, but we have not yet received responses from Reviewers udfZ and kuMq. Their principal concerns motivated several of the additional experiments reported in our responses: the nested cross-validation reruns requested by both reviewers; the leakage analyses, hypernetwork controls, and unseen-configuration comparisons for Reviewer udfZ; and the prompt-conditioning ablations and the frozen-versus-fine-tuned encoder study for Reviewer kuMq.

As the author–reviewer discussion approaches its close, could you please help invite their feedback on these results? We would welcome the opportunity to clarify any remaining concerns while the discussion is still open.

Thank you again for your help.

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.