acceptodds
Under review as a conference paper at ICLR 2027

Understanding Before Forecasting: Multi-City Urban Dynamics Prediction with Vision-Language Model

Abstract

Multi-city urban dynamics prediction requires capturing fine-grained spatial-temporal patterns across heterogeneous cities while effectively transferring knowledge across diverse urban environments. Vision-language models (VLMs), with their large-scale pretraining, multimodal representation, and generalization capabilities, offer a potential opportunity for this task. However, VLMs lack domain-specific knowledge of spatial-temporal urban dynamics and an understanding of flexible prediction tasks that vary in target dynamics, spatial regions and scales, and prediction times. Moreover, heterogeneous city distributions can induce conflicting optimization objectives during joint learning, hindering effective cross-city knowledge sharing. In this paper, we propose UrbanBind, a unified visual-language framework for multi-city urban dynamics prediction with three tightly coupled components: (i) QA-Guided Spatial-Temporal Knowledge Grounding, which adapts a pretrained VLM to urban dynamics and flexible prediction-task semantics through structured question-answering supervision; (ii) Multi-City Urban Dynamics Encoding, which explicitly learns spatial-temporal representations across cities, complementing QA-guided grounding with spatial-temporal knowledge and task understanding; and (iii) Nash-Balanced Multi-City Optimization, which balances conflicting city-wise updates to shared parameters while preserving city-specific optimization. Extensive experiments on multiple real-world urban dynamics datasets demonstrate the effectiveness of UrbanBind for multi-city forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Urban Dynamics Prediction, Spatial Temporal Forecasting, Vision-Language Models, Multi-City Learning

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

This paper proposes UrbanBind, a unified vision-language framework for multi-city urban dynamics forecasting. The framework aims to address two challenges: the limited spatial-temporal understanding of pretrained vision-language models (VLMs) and optimization conflicts arising from heterogeneous urban data distributions. UrbanBind consists of three main components: (1) QA-Guided Spatial-Temporal Knowledge Grounding, which adapts a pretrained VLM to urban dynamics through structured understanding-oriented and prediction-oriented QA tasks; (2) Multi-City Urban Dynamics Encoding, which explicitly models spatial-temporal dependencies and cross-city interactions, and integrates numerical representations with VLM features through a conditional multi-projector and gated fusion mechanism; and (3) Nash-Balanced Multi-City Optimization, which balances city-specific gradients when optimizing shared parameters. Experiments on three urban dynamics datasets demonstrate improvements over several representative baselines. Additional analyses, including QA grounding evaluations, ablation studies, and few-shot adaptation to a held-out city, are provided to validate the effectiveness of the proposed framework.

Critical strengths
  1. The paper addresses a meaningful challenge in applying pretrained VLMs to heterogeneous urban dynamics. The proposed QA-guided grounding strategy is interesting, as it attempts to equip the VLM with explicit spatial-temporal knowledge before downstream forecasting. The design of 18 understanding-oriented and 8 prediction-oriented QA task families, together with textual, numerical, and dense-map supervision, provides a systematic approach to adapting general-purpose VLMs to urban prediction tasks. The grounding analysis in Table 2 shows improvements in QA performance and representation quality, providing empirical support for the effectiveness of this training strategy.

  2. UrbanBind achieves the best reported MAE and RMSE across all three main forecasting datasets in Table 1. The authors also provide ablation studies, representation analyses, input perturbation experiments, and few-shot transfer results on the held-out BIKEDC dataset. These evaluations help demonstrate the contributions of different components and provide evidence that the proposed framework can benefit multi-city urban dynamics prediction.

Critical weaknesses
  1. Although UrbanBind integrates a VLM with several spatial-temporal modeling components, the experiments do not clearly isolate the VLM's contribution. Table 3 shows that removing QA grounding causes relatively modest performance degradation, whereas removing spatial-temporal encoding leads to substantially larger errors. This raises questions about whether the improvements primarily come from the numerical encoder rather than the VLM. A controlled comparison with a numerical-only variant retaining the spatial-temporal encoder, cross-city interaction, and optimization strategy is needed to justify the additional complexity of the VLM.

  2. The main experiments involve only three datasets from two cities, with few-shot generalization evaluated on a single held-out city. Moreover, the adaptation procedure for unseen cities remains unclear, particularly given the city-specific encoders, decoders, and concatenated multi-city representations. The fairness of baseline comparisons also requires clarification, as UrbanGPT's original Vicuna-7B backbone is replaced with Qwen2.5-1.5B. Finally, the absence of multi-run variance, statistical significance, and computational cost analysis makes it difficult to assess the reliability of the relatively modest improvements and whether they justify the increased model complexity.

Questions for authors
  1. Could the authors provide a numerical-only baseline that retains the proposed spatial-temporal encoder, cross-city interaction mechanism, and Nash-balanced optimization, but replaces the VLM with a lightweight prediction network? A comparison under similar computational budgets would help establish whether the VLM contributes complementary information beyond what can be learned through explicit numerical modeling.

  2. Since the architecture contains city-specific encoders and decoders, and the multi-city projection operates on concatenated city representations, how is the architecture instantiated for BIKEDC? Are new city-specific modules initialized from scratch, transferred from existing cities, or jointly fine-tuned? Does introducing a new city require modifying or retraining the shared projection layers? Clarifying the adaptation protocol and the trainable parameters would strengthen the cross-city generalization claims.

  3. The derivation in Appendix A.5 assumes that a feasible common direction exists with a positive inner product with every city gradient. However, this condition may not always hold when city-wise objectives are strongly conflicting. How does the implementation handle infeasible or numerically unstable cases? Could the authors provide an analysis of gradient conflicts during training and comparisons with alternative multi-task gradient balancing approaches, such as PCGrad?

  4. Could the authors report mean and standard deviation over multiple random seeds, together with training/inference costs and parameter counts? Additionally, how sensitive are the baseline results to the architectural modifications described in Appendix A.6, particularly the replacement of UrbanGPT's original language-model backbone? These details would help assess whether the performance improvements consistently justify the additional complexity of UrbanBind.

4 comments

We thank the reviewer for the careful assessment and constructive feedback. We address the concerns and questions below.

W1 & Q1: VLM Contribution Beyond Numerical Modeling

To isolate the contribution of VLM representations, we evaluate two additional variants: (1) Numerical-only, which removes VLM conditioning while retaining the spatial-temporal encoder, cross-city interaction (CCI), dense decoder, and Nash optimization; and (2) 2-Layer Transformer, which replaces the Stage-2 language backbone with a lightweight Transformer while retaining Stage-1 QA-grounded representations.

ModelNYCTAXI MAE/RMSEBIKECHI MAE/RMSENYC-BIKE MAE/RMSE
UrbanBind0.169 / 1.8920.649 / 2.7252.540 / 5.538
Numerical-only0.186 / 2.0700.671 / 2.8622.757 / 6.096
2-Layer Transformer0.192 / 2.5510.730 / 3.5282.628 / 5.909

UrbanBind achieves the lowest MAE and RMSE across all three datasets. The Numerical-only comparison demonstrates the complementary contribution of VLM representations beyond numerical forecasting, while the Transformer replacement supports the effectiveness of the Stage-2 language backbone.

W2 & Q2: Held-Out-City Adaptation

Our primary task is unified multi-city forecasting, with BIKEDC serving as an additional held-out-city few-shot evaluation.

For BIKEDC, we initialize its city-specific encoder and decoder from NYC-BIKE, as both datasets use 16×816 \times 8 spatial grids. Only these target-specific modules are fine-tuned on BIKEDC support samples; the QA-grounded VLM, source-city branches, CCI, and shared projection/fusion modules remain frozen.

BIKEDC replaces the NYC-BIKE slot in the three-city interface, with NYCTAXI and BIKECHI providing source-city historical context. The concatenated representation retains its original dimensionality, requiring no modification or retraining of the shared projection layers. This enables few-shot adaptation while preserving shared multi-city knowledge.

W2 & Q3: Nash Feasibility and Gradient Conflicts

The ideal Nash solution assumes a feasible common-descent direction, which may not always exist. Our implementation handles numerical instability through warm-started optimization, coefficient normalization, and positive lower bounds. If the solver fails, previously valid weights (initially uniform) are reused, ensuring a well-defined update. Such fallback updates do not guarantee strict common descent.

To examine inter-city conflicts, we analyze 24 gradient probes over shared parameters:

City PairMean Gradient CosineNegative Cosine Ratio (%)
NYCTAXI–BIKECHI0.01650.000
NYCTAXI–NYC-BIKE−0.03537.500
BIKECHI–NYC-BIKE−0.02062.500

The negative cosine ratios (37.5%–62.5%) indicate frequent conflicting optimization directions across cities.

We further compare Equal Weighting, Nash, and PCGrad (Yu et al., 2020) using eight additional matched gradient probes from a fixed checkpoint, without parameter updates. Nash uses the numerical reference solution to Eq. (11), while PCGrad results are averaged over 16 random projection orders per probe.

MethodAll-City Descent (%) ↑Mean Worst-City Cosine ↑
Equal Weighting62.5000.212
Nash (Eq. 11)100.0000.546
PCGrad100.0000.289

The all-city descent ratio measures first-order descent across all cities, while the mean worst-city cosine measures alignment with the least-aligned city gradient.

Both Nash and PCGrad achieve 100% all-city descent on these probes, while Nash achieves substantially stronger worst-city alignment (0.546 vs. 0.289). These gradient-direction diagnostics, together with the w/o NBO forecasting ablation (Table 3), support Nash's effectiveness in balancing competing city-wise objectives. The direction comparison does not involve separately trained forecasting models.

W2 & Q4: Baseline Fairness, Robustness, and Computational Cost

Backbone fairness. As disclosed in Appendix A.6, UrbanGPT's original Vicuna-7B backbone was replaced with frozen Qwen2.5-1.5B due to GPU memory constraints. We additionally evaluate UrbanGPT with Qwen2.5-3B, closer in scale to UrbanBind's Qwen2.5-VL-3B, while preserving UrbanGPT's spatial-temporal encoder, alignment mechanism, and forecasting head.

ModelNYCTAXI MAE/RMSEBIKECHI MAE/RMSENYC-BIKE MAE/RMSE
UrbanGPT (1.5B)0.176 / 1.9300.725 / 3.3523.386 / 8.064
UrbanGPT (3B)0.173 / 1.9390.692 / 2.9852.802 / 6.969
UrbanBind0.169 / 1.8920.649 / 2.7252.540 / 5.538

UrbanBind maintains lower errors across all three datasets under this closer backbone-size comparison, despite differences in model modality and computational requirements.

Few-shot robustness. We repeat BIKEDC support-set sampling with four random seeds, keeping the source checkpoint and adaptation training seed fixed.

ShotsMAE (Mean ± SD)RMSE (Mean ± SD)
1002.025 ± 0.03211.042 ± 0.510
2001.911 ± 0.07710.471 ± 0.661

These results quantify support-set variability and show improved average performance with additional target samples. All methods use the same few-shot training budgets and 205 separate labeled BIKEDC validation windows for model selection, without using these windows for parameter updates.

Training efficiency. We report per-epoch wall-clock time, including training and validation. For UrbanBind, the reported 8.85 h/epoch combines the average per-epoch costs of both training stages, compared with 9.23 h/epoch for UrbanGPT (3B), 2.79 h/epoch for UrbanGPT (1.5B), and 33 min/epoch for PLM-CUP (0.6B). These measurements reflect the respective architectures and training procedures on the same source datasets.

Thank you again for your thoughtful review and constructive feedback. We would be grateful if you could let us know whether our response has addressed your concerns. We would be happy to provide any further clarification.

As the discussion period draws to a close, we would like to thank you again for your constructive feedback. We hope our responses and additional experiments have addressed your concerns. If any concerns remain, we would be happy to provide further clarification.

Reviewer_Ez8xOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents UrbanBind, a two-stage framework for multi-city urban dynamics forecasting. It first adapts a pretrained VLM using structured understanding and prediction QA tasks, then combines its frozen representations with a numerical spatial-temporal encoder through cross-city attention, learned projection, and gated fusion. Nash-balanced optimization coordinates updates to shared parameters, and experiments report improvements on three source datasets and competitive few-shot adaptation to the held-out BIKEDC dataset.

Critical strengths
  1. The paper addresses a meaningful challenge: general-purpose VLMs may not adequately capture quantitative urban dynamics or forecasting-task semantics. The structured QA design provides a concrete way to adapt them across temporal, spatial, and cross-city tasks.

  2. The framework preserves native city grids while combining explicit numerical modeling with VLM representations. Separating shared-parameter optimization from city-specific updates is also a sensible design for heterogeneous datasets.

  3. The grounding analysis goes beyond forecasting accuracy by including temporal shuffling, city permutation, blank-image tests, and frozen-representation probing. These experiments provide useful evidence that the adapted VLM uses visual context and learns more prediction-relevant features.

Critical weaknesses
  1. The claimed task flexibility is not sufficiently supported by the forecasting architecture and evaluation. The model uses city-specific encoders and decoders, a concatenation-based projection tied to the city set, and a fixed forecasting horizon in the main experiments. These choices support multi-dataset forecasting, but do not establish that language instructions enable flexible prediction over new regions, spatial scales, or horizons.

  2. The rationale for sample-level cross-city interaction needs stronger justification. Cross-city attention conditions each prediction on historical windows from other cities, yet the datasets cover different periods and the forecasting-window pairing strategy is unclear.

  3. The claimed mitigation of inter-city optimization conflicts is not directly demonstrated. The w/o NBO ablation shows that the chosen update rule improves forecasting, but does not establish whether it reduces gradient conflicts or offers advantages over simpler loss-balancing strategies.

Questions for authors

Please refer to the weaknesses.

3 comments

We thank the reviewer for the thoughtful review and constructive feedback.

W1: Task Flexibility and Forecasting Architecture

UrbanBind addresses task flexibility through QA-guided representation learning, while Stage 2 evaluates these representations under a unified forecasting setting. As detailed in Section 3.2 and Appendix A.4, Stage 1 incorporates 26 QA families spanning different prediction objectives, spatial granularities, and spatial-temporal reasoning tasks. This enables the VLM to learn task-aware representations across diverse urban dynamics tasks.

Stage 2 uses a fixed four-step forecasting horizon for consistent multi-city evaluation, with city-specific encoders and decoders preserving heterogeneous native grids. Our task-flexibility claim concerns the diversity of tasks supported by QA grounding, while downstream experiments focus on unified multi-city forecasting. The grounding analyses (Table 2) and forecasting ablations (Table 3) support the effectiveness of these learned representations.

W2: Cross-City Interaction and Temporal Alignment

Cross-City Interaction (CCI) captures transferable spatial-temporal patterns across cities without requiring temporally synchronized observations. Each city independently encodes its historical spatial and temporal dynamics, which are aggregated into a city-level representation through readout attention. CCI then applies self-attention across these representations to facilitate cross-city information exchange.

During training, historical windows are independently shuffled and sampled within each city, then combined into joint multi-city batches without matching absolute timestamps. Thus, CCI operates on learned city-level representations rather than synchronized events or spatially aligned grid cells.

The w/o CCI ablation (Table 3) further supports this design, showing consistently higher forecasting errors across all three datasets when explicit cross-city attention is removed.

W3: Inter-City Gradient Conflicts and Nash-Balanced Optimization

To examine inter-city optimization conflicts, we analyze city-wise gradients over shared parameters using 24 gradient probes.

City PairMean Gradient CosineNegative Cosine Ratio (%)
NYCTAXI–BIKECHI0.01650.000
NYCTAXI–NYC-BIKE−0.03537.500
BIKECHI–NYC-BIKE−0.02062.500

Negative gradient cosine similarities indicate conflicting city-wise optimization directions. The negative cosine ratios (37.500%–62.500%) confirm that such conflicts occur frequently among the evaluated city gradients.

We further compare Equal Weighting, Nash, and PCGrad (Yu et al., 2020) using eight additional matched gradient probes from a fixed UrbanBind checkpoint, without parameter updates. Nash uses the numerical reference solution to Eq. (11), while PCGrad results are averaged over 16 random projection orders per probe.

MethodAll-City Descent (%) ↑Mean Worst-City Cosine ↑
Equal Weighting62.5000.212
Nash (Eq. 11)100.0000.546
PCGrad100.0000.289

The all-city descent ratio measures how often an update direction provides first-order descent for every city, while the mean worst-city cosine measures its average alignment with the least-aligned city gradient.

Both Nash and PCGrad achieve 100% all-city descent on the evaluated probes, outperforming Equal Weighting (62.5%). Nash further achieves the strongest worst-city alignment (0.546), compared with PCGrad (0.289) and Equal Weighting (0.212).

These findings provide direct evidence that Nash bargaining produces more balanced update directions under conflicting city-wise objectives, complementing the forecasting improvements demonstrated by the w/o NBO ablation (Table 3). The comparison evaluates gradient directions rather than separately trained forecasting models.

Thank you again for your thoughtful review and constructive feedback. We would be grateful if you could let us know whether our response has addressed your concerns. We would be happy to provide any further clarification.

As the discussion period draws to a close, we would like to thank you again for your constructive feedback. We hope our responses and additional experiments have addressed your concerns. If any concerns remain, we would be happy to provide further clarification.

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

UrbanBind studies multi-city urban flow forecasting with a vision-language model. It first adapts a VLM using structured questions about visualized flow histories, including spatial and temporal understanding and future prediction. It then combines the frozen, adapted VLM with numerical spatial-temporal encoders, cross-city attention, and a gated multi-projector module. Nash-based weighting balances city-wise gradients on shared parameters. Experiments use three source datasets and a held-out target dataset, with additional visual perturbation tests, representation probes, and component ablations.

Critical strengths
  1. The combination of visual task representations and numerical dynamics modeling is well motivated. Preserving each dataset's native spatial grid is useful when learning across heterogeneous urban datasets, and the method improves both MAE and RMSE over the reported baselines on all three source datasets.

  2. The evaluation goes beyond the main forecasting table. The image perturbation tests, representation probes, and held-out-city experiment provide useful evidence about what the model learns. I also appreciate the detailed QA taxonomy and the chronological splitting and training-only normalization in the released implementation.

Critical weaknesses
  1. The experiments do not yet separate the proposed VLM understanding contribution from the numerical forecasting machinery and forecasting pretraining. There are two related gaps here.

First, Section 3.3.2 describes the final prediction as a decoder applied to the VLM output r. In the released code, however, the decoder also receives dense numerical history features and raw flow histories. It includes temporal attention, spatial refinement, and a learned history-based prediction added to the output. The training loss also includes a spatial-gradient term that is not described in the paper's forecasting objective. These are reasonable design choices, but they seem important enough to document and control for. A numerical-only comparison retaining the same encoder, decoder, and training setup would help establish what the VLM adds; w/o QG still retains a base VLM and does not answer this question.

Second, Stage 1 includes direct future-value and future-map prediction tasks alongside understanding QA. Better forecasting and probe results therefore do not isolate the benefit of understanding supervision from ordinary forecasting pretraining. A forecasting-only QA control, ideally alongside an understanding-only variant under comparable training budgets, would make the central claim more convincing. The existing QG ablation is useful, but some differences are small: NYC-BIKE RMSE changes from 5.581 without QG to 5.538 with it.

  1. The comparison and transfer results need a little more support before attributing the gains to the proposed design. Appendix A.6 replaces UrbanGPT's original backbone with a frozen Qwen2.5-1.5B model, while UrbanBind uses Qwen2.5-VL-3B. I understand the computational constraint and appreciate that the change is disclosed, but a closer capacity or compute comparison would help. The held-out-city results are also quite close: in Table 5, UrbanBind improves MAE over the best baseline by 0.01-0.02, while RMSE is tied at the reported precision in all three settings. Repeating the support-set sampling and reporting variability would be more informative here than adding many more benchmarks. The released few-shot code also uses 205 target-city validation windows for checkpoint selection; this additional label budget and its consistency across methods should be stated explicitly.
Questions for authors
  1. Could you confirm whether the released dense decoder and spatial-gradient loss were used for the reported results, and clarify their role in Section 3.3.2? A comparison that retains this numerical forecasting path but removes the VLM conditioning would directly address the first concern.

  2. How much of the final forecasting gain remains when Stage 1 uses only forecasting QA? If available, a comparison with understanding-only and full QA under similar training budgets would help distinguish the two sources of improvement.

  3. Were the baseline and few-shot experiments matched in target-city validation data and model-selection procedure? For the close BIKEDC results, could you report variation across support-set samples or seeds? A capacity-matched UrbanGPT comparison would also help, if feasible.

3 comments

W1 & Q1: Forecasting Architecture and VLM Contribution

We thank the reviewer for the constructive suggestions. We confirm that all submitted results were obtained using the released dense decoder and spatial-gradient loss.

UrbanBind integrates VLM-conditioned representations with dense numerical forecasting through the multi-projector and gated fusion mechanisms. The decoder combines a learned historical-flow baseline with residual corrections predicted using horizon-specific temporal attention, VLM-based semantic conditioning, and spatial refinement. The spatial-gradient L1 loss, weighted by 0.05, further encourages consistency between predicted and ground-truth spatial variations.

To isolate the contribution of VLM representations, we additionally evaluate a numerical-only ablation that retains the numerical encoder, cross-city interaction, dense decoder, and Nash optimization while removing VLM conditioning.

ModelNYCTAXI MAE/RMSEBIKECHI MAE/RMSENYC-BIKE MAE/RMSE
UrbanBind0.169 / 1.8920.649 / 2.7252.540 / 5.538
Numerical-only0.186 / 2.0700.671 / 2.8622.757 / 6.096

UrbanBind consistently achieves lower MAE and RMSE across all three datasets, demonstrating that VLM-conditioned representations provide complementary benefits beyond the numerical forecasting pathway.

W1 & Q2: Understanding vs. Prediction QA Supervision

To distinguish the contributions of understanding and forecasting supervision, we evaluate two additional Stage-1 variants: prediction-only QA (P1–P8) and understanding-only QA (U1–U18), alongside Full QA.

QA SupervisionNYCTAXI MAE/RMSEBIKECHI MAE/RMSENYC-BIKE MAE/RMSE
Full QA (U+P)0.169 / 1.8920.649 / 2.7252.540 / 5.538
Prediction-only0.178 / 1.9840.651 / 2.7962.549 / 5.621
Understanding-only0.179 / 1.8990.659 / 2.8142.554 / 5.560

Full QA achieves the lowest MAE and RMSE across all three datasets, consistently outperforming prediction-only supervision. These results indicate that understanding-oriented QA provides additional benefits beyond forecasting QA alone, supporting the complementary roles of spatial-temporal understanding and prediction-oriented grounding.

W2 & Q3: Baseline Fairness and Held-Out-City Evaluation

(1) UrbanGPT backbone fairness. To assess the impact of backbone size, we additionally evaluate UrbanGPT with a Qwen2.5-3B backbone, closer in parameter scale to UrbanBind's Qwen2.5-VL-3B, while preserving UrbanGPT's spatial-temporal encoder, alignment mechanism, and forecasting head.

ModelNYCTAXI MAE/RMSEBIKECHI MAE/RMSENYC-BIKE MAE/RMSE
UrbanGPT (1.5B)0.176 / 1.9300.725 / 3.3523.386 / 8.064
UrbanGPT (3B)0.173 / 1.9390.692 / 2.9852.802 / 6.969
UrbanBind (3B)0.169 / 1.8920.649 / 2.7252.540 / 5.538

Increasing UrbanGPT's backbone size generally improves its forecasting performance, particularly on BIKECHI and NYC-BIKE. Nevertheless, UrbanBind maintains lower MAE and RMSE across all three datasets under this closer backbone-size comparison, providing further evidence of the effectiveness of its proposed framework.

(2) Held-out-city transfer and robustness. Our primary objective is unified multi-city forecasting across NYCTAXI, BIKECHI, and NYC-BIKE, with BIKEDC serving as an additional held-out-city transfer evaluation.

All methods use the same target-city data splits, few-shot training budgets, and 205 labeled BIKEDC validation windows for checkpoint selection. These validation windows do not contribute to parameter updates.

To assess sensitivity to target-city sample selection, we repeat BIKEDC support-set sampling with four random seeds, keeping the source checkpoint and adaptation training seed fixed. Results for the 100- and 200-shot settings are summarized below.

BIKEDC ShotsMAE (Mean ± SD)RMSE (Mean ± SD)
1002.025 ± 0.03211.042 ± 0.510
2001.911 ± 0.07710.471 ± 0.661

These results characterize the variability across few-shot support sets and show improved average forecasting performance with additional target-city supervision. They further support UrbanBind's ability to transfer learned urban dynamics knowledge to an unseen city under limited labeled data. The complete robustness analysis is provided in Appendix B.7.

The numerical-only run answers my main question. With the same numerical encoder and decoder, removing VLM conditioning made forecasts worse on all three source datasets. That’s enough for me to move to 3.

Thank you for your thoughtful feedback and for reconsidering your rating. We appreciate your time and are glad that the additional experiment addressed your main concern.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.