- 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
-
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.
-
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
-
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.
-
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
-
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.
-
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.
-
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?
-
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
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.
| Model | NYCTAXI MAE/RMSE | BIKECHI MAE/RMSE | NYC-BIKE MAE/RMSE |
|---|---|---|---|
| UrbanGPT (1.5B) | 0.176 / 1.930 | 0.725 / 3.352 | 3.386 / 8.064 |
| UrbanGPT (3B) | 0.173 / 1.939 | 0.692 / 2.985 | 2.802 / 6.969 |
| UrbanBind | 0.169 / 1.892 | 0.649 / 2.725 | 2.540 / 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.
| Shots | MAE (Mean ± SD) | RMSE (Mean ± SD) |
|---|---|---|
| 100 | 2.025 ± 0.032 | 11.042 ± 0.510 |
| 200 | 1.911 ± 0.077 | 10.471 ± 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.
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.
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 16×8 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:
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.
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.