Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

UrbanFusion claims that a single fused encoder, trained with stochastic masking and reconstruction, beats state-of-the-art GeoAI models across 41 urban prediction tasks while accepting any subset of modalities.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 09:42 UTC pith:QW7M2TKG

load-bearing objection A broad, honest benchmark for multimodal geo-embeddings, but the headline fusion gains are confounded by the extra modalities UrbanFusion alone sees. the 3 major comments →

arxiv 2510.13774 v2 pith:QW7M2TKG submitted 2025-10-15 cs.LG cs.CV

UrbanFusion: Stochastic Multimodal Fusion for Contrastive Learning of Robust Spatial Representations

Dominik J. M\"uhlematter , Lin Che , Ye Hong , Martin Raubal , Nina Wiedemann This is my paper
classification cs.LG cs.CV
keywords GeoAIfoundation modelmultimodal learningcontrastive learninglocation encodingurban predictionpartial information decompositionstochastic multimodal fusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

UrbanFusion sets out to show that one spatial representation model can absorb four heterogeneous urban data streams—street view, satellite imagery, cartographic maps, and points of interest—and do so more faithfully than the pairwise-contrastive methods used by earlier GeoFMs. Its Stochastic Multimodal Fusion (SMF) training pairs contrastive alignment of two randomly masked complementary modality subsets with a latent reconstruction objective, which the paper argues provably retains redundant, unique, and synergistic information needed by downstream tasks. On 41 tasks spanning housing, health, energy, land use, and perception, it reports first-place results on the majority of settings against baselines trained on the same data, and maintains this when applied to cities never seen in training. The practical payoff claimed is flexibility: the encoder accepts whatever subset of modalities is available, so heterogeneous and incomplete real-world geodata can be reused without perfect alignment.

Core claim

The central claim is that the SMF training objective, not any single modality, is what carries the performance. The paper argues that pure contrastive alignment (InfoNCE) captures only redundant information shared across views, while the reconstruction head plus random masking forces the fused embedding to keep modality-unique and synergistic content. It supports this with Lemma 1, which states that under Assumption 1 the SMF loss maximizes a lower bound on the mutual information I(m_1,...,m_K;Y), decomposed via partial information decomposition as redundant plus synergistic plus the per-modality unique terms, and with a synthetic-data experiment isolating each information type. Empirically,

What carries the argument

Stochastic Multimodal Fusion (SMF). At each training step, the model randomly splits the available modalities into two complementary subsets, encodes each through frozen per-modality encoders, and fuses them with a single-block Transformer. A symmetric InfoNCE loss aligns the two fused views (contrastive location alignment), while a latent reconstruction head predicts the latent features of all modalities from each fused view (reconstruction loss). The paper's theoretical vehicle is Lemma 1, which uses the Barber-Agakov lower bound on mutual information and the InfoNCE bound to argue that this joint loss maximizes a lower bound on I(m_1,...,m_K;Y) = R+S+sum U_i, the redundant plus synergisti

Load-bearing premise

The proof that SMF provably preserves task-relevant information rests on Assumption 1: for each downstream task there exists a subset of modalities whose prediction is at least as demanding as predicting the task itself; the paper states this assumption but does not verify it, and its own ablation shows the reconstruction term changes headline scores by at most ~0.5 points.

What would settle it

Run the full 41-task evaluation with the reconstruction loss removed (CL-only) and with all modalities fixed to the same set; if performance on tasks typically dependent on unique or synergistic signals (e.g., fine-grained land use, urban perception) does not drop beyond what the paper's Table 6 shows (<=0.5 points), the claimed mechanism is not driving the results. Alternatively, estimate I(A\S_Y;Y) vs I(A\S_Y;S_Y) on at least one real task (e.g., housing prices) to check whether Assumption 1 holds.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If SMF works as claimed, switching to this training objective should improve any multimodal GeoFM, not just UrbanFusion's exact architecture.
  • The model can be pretrained and used with any subset of modalities, and the ablation shows even bimodal (coordinates + one modality) retains ~99% of full-model performance, so practitioners could reuse heterogeneous geospatial datasets without fully aligning all modalities.
  • Cross-regional results imply that multimodal fusion with non-coordinate modalities (satellite, street view, maps, POIs) is what generalizes to unseen cities; coordinate-only encoders degrade sharply there.
  • The synthetic PID experiment suggests a concrete diagnostic: contrastive-only methods lose modality-unique and synergistic signals, whereas SMF retains them, predicting which types of downstream tasks benefit most.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported gains over PP2-M-trained contrastive baselines are modest (0.1-3 R^2/F1 points), and the paper's own loss ablation shows the reconstruction term moves headline scores by at most ~0.5 points; the practical win may come mostly from simply having more modalities and a fused encoder, not from the SMF mechanism per se.
  • A testable extension is to apply SMF to a non-spatial multimodal benchmark (e.g., vision-language) to see whether the same retention of unique and synergistic signals appears—the paper claims SMF is model-agnostic but only demonstrates it on geographic data.
  • Assumption 1 (the existence of a proxy modality subset for every downstream task) is stated but never validated; estimating the mutual-information terms on one or two real tasks would show whether the final proof step actually binds.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes UrbanFusion, a multimodal geospatial foundation model that fuses coordinates, street-view imagery, remote sensing, OpenStreetMap basemaps, and POIs through a single-block Transformer. The training objective, Stochastic Multimodal Fusion (SMF), combines a symmetric InfoNCE contrastive loss between randomly masked and complementary modality subsets with a latent reconstruction loss. The authors claim that this objective provably retains redundant, unique, and synergistic information, and they support the claim with a formal lemma (Appendix D.2). Experiments cover 41 downstream tasks across 56 cities in three settings: coordinate-only encoding, multimodal inference, and cross-regional generalization to held-out cities. The paper reports competitive or best performance on many tasks, releases the enriched PP2-M dataset, code, and weights, and includes extensive ablations (loss terms, modality subsets, SatCLIP order, incomplete-modality pretraining). The empirical scaffolding is substantial and mostly transparent, with disclosed data exclusions and honest notes on PDFM's in-sample nature.

Significance. If the headline claims held, UrbanFusion would be a valuable, flexible GeoFM: it accepts arbitrary modality subsets at both pretraining and inference, and the released PP2-M dataset plus code would support reproducibility. The evaluation breadth is a real strength, as are the ablations and the explicit discussion of data exclusions. However, the central multimodal and cross-regional comparisons are confounded: UrbanFusion is evaluated with validation-selected OSM/POI inputs that no baseline receives, the GAIR baseline is reimplemented without its INR module, and the theoretical justification rests on an unvalidated assumption while the loss ablation shows the reconstruction term is not clearly responsible for the reported gains. These issues prevent accepting the 'outperforms prior models' claim as evidence for the fusion mechanism itself, although the paper's resources and evaluation methodology remain useful.

major comments (3)
  1. [§4.3/§4.4, Tables 3/4 vs 18/23] Multimodal and cross-regional comparisons are confounded by input modality availability. In Tables 3/4, UrbanFusion's validation-selected subsets include OSM and/or POI for most tasks (Tables 18/23, e.g., 'SV+RS+OSM+POI+Coords' for many ZIP tasks), whereas GAIR receives only SV+RS, GeoCLIP only SV, and SatCLIP only RS. Reported wins such as Crime 88.5 vs 85.4 and ZIP average 75.1 vs 70.5 may therefore reflect the predictive content of OSM/POI rather than SMF or the fusion architecture. A controlled comparison—UrbanFusion restricted to exactly the baseline modality set, or baselines augmented with OSM/POI features—is needed before the headline 'outperforms prior models' in the multimodal setting can be attributed to the method.
  2. [Appendix D.2, Assumption 1 (Eq. A); Table 6] Lemma 1 rests on unvalidated Assumption 1: for every downstream task Y there exists a proxy subset S_Y with I(A\S_Y;Y) ≤ I(A\S_Y;S_Y). This is stated but never checked for any of the 41 tasks; if it fails, the proof's final step collapses. Moreover, the mechanism the lemma is meant to secure—the reconstruction loss—is not visibly driving the empirical results: Table 6 shows CL-only matches or exceeds CL+rec. on several multimodal tasks (Crime 89.6 vs 88.5; ZIP 75.5 vs 75.1), and Rec-only wins on some cross-regional tasks. The theory should be reframed as conditional on Assumption 1, or its role as an explanation of the experimental gains should be withdrawn.
  3. [Appendix F.3.3; Table 4] GAIR is not evaluated as published: the reimplementation omits GAIR's INR module, a core component, because official code/weights were unavailable. The comparison 'UrbanFusion outperforms GAIR' is therefore against an ablated GAIR. This is consequential because GAIR is the closest multimodal competitor; in Table 4, GAIR already beats UrbanFusion on the ZIP-code weighted average (62.5 vs 56.7), a fact the main text's 'ranking first on 5 out of 6 tasks' obscures. The authors should either compare with the official GAIR or explicitly label the baseline as 'GAIR without INR' and temper the state-of-the-art claim.
minor comments (4)
  1. [Appendix C.1] The text refers to 'Table 4.6' but the table is numbered Table 5 in the appendix; the table formatting is also garbled.
  2. [Main text, Contributions] The phrase 'the first foundation model to flexibly integrate...' is a strong novelty claim; given the number of recent GeoFMs, it should be softened or backed by a concrete survey comparison of input flexibility.
  3. [Tables 2/3/4] Linear probing results are reported as point estimates with no error bars or significance testing. Many margins are at or below 0.5 R²/F1 points, so it is hard to assess whether differences are meaningful; reporting repeated-run variability or at least seed-averaged MLP numbers for the linear setting would help.
  4. [Abstract and §4.1] The claim '41 tasks' is not transparently enumerated from the tables (8+6+29 plus perception/land-use breakdowns). A clear task list or count reconciliation would aid reproducibility and prevent miscounting.

Circularity Check

1 steps flagged

Synthetic PID 'validation' is self-definitional: the unique/synergistic probes are exactly UrbanFusion's own reconstruction targets; the external benchmarks are independent, so the circularity is partial (6/10).

specific steps
  1. self definitional [Sec. 4.5 + App. C.2.3 (empirical PID validation); Sec. 3.1 (L_recon)]
    "To measure unique (modality-specific) information, we focus on the third feature dimension of each modality ... The task is to reconstruct this feature dimension from the shared location representation. (App. C.2.3) ... we construct synthetic data with unique random values per modality, location-specific values as redundant information, and predicting the sum of the unique values as synergistic information (Sec. 4.5) ... The second head performs Latent Modality Reconstruction: ... g_m is trained to reconstruct the latent vector h_m for all modalities m ∈ A (Sec. 3.1)."

    The 'unique-information' probe in Figure 3 is ridge regression of the shared representation onto exactly the latent features h_m that UrbanFusion's L_recon minimizes (MSE between g_m(z) and h_m). The 'synergistic' probe is the sum of those same third dimensions, so it is also reconstructible once L_recon succeeds. The paper presents this as empirical validation of the claimed PID property, but the result is enforced by the loss construction: a model trained with L_recon must score well on a task that is L_recon, while contrastive-only baselines are not given that objective. Thus, the claim that UrbanFusion 'reliably captures all mutual information components' is a restatement of the training target rather than an independent test, making the information-preservation 'validation' circular b

full rationale

The large downstream evaluation is not circular: the 41 tasks use independent external labels (housing, crime, CDC health measures, land cover, energy) at held-out locations and entirely held-out cities, and neither the fitted parameters nor the loss objective define those target variables. I explicitly flag that Lemma 1 rests on the unvalidated Assumption 1 (App. D.2) and that the proof jumps from per-modality reconstruction bounds to a bound on I(m_1,...,m_K;Y); however, that is a soundness/proof-gap concern rather than a circular reduction, so it does not by itself raise the score. The circular content is localized to the synthetic PID validation (Fig. 3, App. C.2): the unique and synergistic probes are defined as reconstructing the same latent features that the SMF reconstruction loss directly optimizes, so the observed advantage over contrastive-only baselines is settled by the loss design. The comparison confound in Tables 3/4/19—UrbanFusion may consume OSM/POI at inference while GAIR/GeoCLIP/SatCLIP receive fewer modalities—is a fairness/attribution issue rather than a definitional circularity, and therefore does not contribute additional points here. Overall: the practical benchmark contribution stands independently, but the paper's headline theoretical claim about capturing unique/synergistic information is validated by a self-defined experiment, giving partial circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central engineering claim rests on standard SSL bounds (InfoNCE, Barber–Agakov) plus four fixed encoders; the new theoretical claim rests on Assumption 1, which is stated but not validated. Free parameters are few and mostly standard practice (λ, τ₀, RFF scales), plus a basemap encoder trained on the same dataset. No invented entities: SMF is a training procedure, and the PID components R/S/U are analytic categories from the cited literature, not new ontological commitments.

free parameters (5)
  • reconstruction-loss weight λ = 0.0625
    Set by hand so contrastive and reconstruction loss magnitudes match early in training (§F.2); the SMF objective the paper claims retains unique/synergistic information is a weighted sum with this coefficient.
  • InfoNCE temperature initial value τ₀ = 0.07 (learnable)
    Initial temperature of the contrastive loss; jointly learned during pretraining (§F.2). Standard setting, but a fitted parameter of the alignment objective underlying Lemma 1.
  • RFF scales σ = 1, 16, 256
    Random Fourier Feature bandwidths for coordinate encoding at three spatial scales, taken from GeoCLIP/GAIR (§F.1); hand-chosen hyperparameters on which the coordinate-only results depend.
  • basemap MAE encoder = ViT-B/16 fine-tuned on PP2-M training split (mask ratio 0.6, 100 epochs, AdamW lr 3e-4)
    The OSM modality encoder is not an external frozen feature extractor; it is trained on the same PP2-M data used for pretraining and evaluation (§F.1), so its features are fitted, not independent.
  • POI context budget = 15 nearest POIs within ≤200 m
    Hand-chosen limits on the text-prompt context per location (§E.1); fixed constants, not fitted to downstream data.
axioms (5)
  • ad hoc to paper Assumption 1 (Eq. A): for each downstream task Y there is a proxy subset S_Y ⊆ A with I(A∖S_Y; Y) ≤ I(A∖S_Y; S_Y)
    Load-bearing premise of Lemma 1's conclusion that L_total retains R+S+ΣU_i; introduced and used only in this paper's argument, never validated (App. D.2).
  • standard math InfoNCE with batch negatives lower-bounds the mutual information between the two masked-view representations (van den Oord et al. 2018)
    First step of the Lemma 1 proof (App. D.2).
  • standard math Minimizing the Gaussian-decoder MSE maximizes the Barber–Agakov lower bound on I(z; h_m)
    Second step of the Lemma 1 proof, Eqs. 6–10 (App. D.2).
  • domain assumption Frozen pretrained encoders (CLIP ViT-L/14, MoCo ViT-S/16, MAE ViT-B/16, bge-small) preserve all task-relevant information from each raw modality
    UrbanFusion never sees raw inputs; any information lost by these encoders is unavailable to the model. No analysis quantifies this loss (§3.2).
  • domain assumption The 56 PP2-M cities plus the 7 held-out cities are representative of 'urban environments' for the generalization claims
    All training and evaluation locations are large cities; rural and small-town generalization is untested and noted as a limitation (App. E.1, §5).

pith-pipeline@v1.3.0-alltime-deepseek · 56333 in / 28159 out tokens · 224884 ms · 2026-08-04T09:42:37.127176+00:00 · methodology

0 comments
read the original abstract

Forecasting urban phenomena such as housing prices and public health indicators requires the effective integration of various geospatial data. Current methods primarily utilize task-specific models, while recent generic models for spatial representations often support only limited modalities and lack multimodal fusion capabilities. To overcome these challenges, we present UrbanFusion, a spatial representation model that features Stochastic Multimodal Fusion (SMF). The framework employs modality-specific encoders to process different types of inputs, including street view imagery, remote sensing data, cartographic maps, and points of interest (POIs) data. These multimodal inputs are integrated via a Transformer-based fusion module that learns unified representations. An extensive evaluation across 41 tasks in 56 cities worldwide demonstrates UrbanFusion's strong generalization and predictive performance compared to state-of-the-art GeoAI models. Specifically, it 1) outperforms prior models on location-encoding, 2) allows multimodal input during inference, and 3) generalizes well to regions unseen during training. UrbanFusion can flexibly utilize any subset of available modalities for a given location during both pretraining and inference, enabling broad applicability across diverse data availability scenarios.

Figures

Figures reproduced from arXiv: 2510.13774 by Dominik J. M\"uhlematter, Lin Che, Martin Raubal, Nina Wiedemann, Ye Hong.

Figure 1
Figure 1. Figure 1: UrbanFusion. Pretrained modality-specific encoders extract features projected into to￾kens. After random token masking, a Transformer fuses the tokens. The output feeds into two heads: one for Contrastive Location Alignment (CL), the other for Latent Modality Reconstruction (Rec.). For downstream tasks, coordinates or available modalities are input into the frozen encoder (green arrows) to obtain feature v… view at source ↗
Figure 2
Figure 2. Figure 2: Visual comparison of Multimodal location embeddings, all trained on the PP2-M dataset. Embeddings are grouped into 10 clusters using k-means. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Empirical results on synthetic data analyzing the preservation of [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: RGB composite image of the top three principal components of location representa￾tions computed globally for the Portland area using only coordinates. (part 1 of 2) 18 [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: RGB composite image of the top three principal components of location representa￾tions computed globally for the Portland area using only coordinates. (part 2 of 2) 19 [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: KMeans clustering (k=10) results for New York City, based on Cross-Regional Gen￾eralization. The full names of all modality abbreviations are provided in the Appendix A. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Designing feature dimensions that capture solely [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: Synthetic data generation. Each location includes two modalities, each represented by a three-dimensional vector. The first two dimensions contain localization-relevant information. The third dimension consists of random values during inference, and during training, is batch-augmented to remain constant across samples. This enforces zero mutual information with location, effectively decoupling unique infor… view at source ↗
Figure 7
Figure 7. Figure 7: Urban areas covered by the PP2-M dataset and the corresponding data splits. For pretraining, we leverage the Place Pulse 2.0 (PP 2.0) dataset (Dubey et al., 2016), which com￾prises 110’988 locations, each with associated geographic coordinates and street view images. We enrich this dataset with additional modalities, referring to the resulting extended version as PP2-M. The dataset spans 56 cities across 2… view at source ↗
Figure 8
Figure 8. Figure 8: Example points of interest (POIs) text prompt provided as input to a language model for a single location. The example corresponds to coordinates 40.025, −75.223 in Philadelphia. E.2 DOWNSTREAM TASKS This study addresses prediction tasks in urban environments. We evaluate methods for Coordinate￾Only Spatial Encoding, where models take only raw geographic coordinates as input, and explore Multimodal Spatial… view at source ↗
Figure 9
Figure 9. Figure 9: Cartographic basemap reconstruction using MAE on the validation set. The masked view is used as input to the encoder, while the reconstructed view is the output of the decoder. Since the decoder is trained to reconstruct only the masked tokens (not the visible ones), we additionally present a combined view that merges the input tokens with the reconstructed tokens for better visu￾alization. 38 [PITH_FULL_… view at source ↗
Figure 10
Figure 10. Figure 10: Cartographic basemap reconstruction using MAE for generalization to regions un￾seen during training. The masked view is input to the encoder, and the decoder outputs the re￾construction of the masked tokens. For improved visualization, we also show a combined view that merges the original input tokens with the reconstructed ones. 39 [PITH_FULL_IMAGE:figures/full_fig_p039_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Training curves for UrbanFusion: (a) Training and validation loss, and (b) validation batch Top-5 retrieval accuracy. Validation Top-5 batch retrieval accuracy is computed on the similar￾ity matrix between two masked views. Accuracy is averaged over both query→key and key→query retrieval directions within each batch, and then across all validation batches To reduce memory usage and improve training speed,… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Multi-Agent Collaborative Reasoning with Tool-Augmented Evidence for Urban Region Profiling

    cs.AI 2026-07 conditional novelty 6.0

    A multi-agent LLM framework with reinforcement-learned tool use outperforms prior urban profiling models on GDP, population, and carbon estimation, including on unseen cities.

  2. Enhancing the Socioeconomic Understanding of Foundation Models with Urban Mobility

    cs.SI 2026-06 unverdicted novelty 5.0

    MobFusion fuses mobility networks into foundation models via three designs and reports improved performance on income, density, and crime prediction tasks using data from three U.S. metropolitan areas.

Reference graph

Works this paper leans on

81 extracted references · 1 canonical work pages · cited by 2 Pith papers

  1. [1]

    General geospatial inference with a population dynamics foundation model

    Mohit Agarwal, Mimi Sun, Chaitanya Kamath, Arbaaz Muslim, Prithul Sarker, Joydeep Paul, Hector Yee, Marcin Sieniek, Kim Jablonski, Yael Mayer, David Fork, Sheila de Guia, Jamie McPike, Adam Boulanger, Tomer Shekel, David Schottlander, Yao Xiao, Manjit Chakravarthy Manukonda, Yun Liu, Neslihan Bulut, Sami Abu-El-Haija , Arno Eigenwillig, Parth Kothari, Bry...

    Pith/arXiv arXiv 2024

  2. [2]

    Optuna: A next-generation hyperparameter optimization framework

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2019

    2019

  3. [3]

    Crime open database (code)

    Matthew P J Ashby. Crime open database (code). https://osf.io/zyaqn, 2017. Accessed 8 July 2025

    2017

  4. [4]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  5. [5]

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016

    Pith/arXiv arXiv 2016

  6. [6]

    The im algorithm: a variational approach to information maximization

    David Barber and Felix Agakov. The im algorithm: a variational approach to information maximization. Advances in neural information processing systems, 16 0 (320): 0 201, 2004

    2004

  7. [7]

    Pattern recognition and machine learning, volume 4

    Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer, 2006

    2006

  8. [8]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    2020

  9. [9]

    Unsupervised urban land use mapping with street view contrastive clustering and a geographical prior

    Lin Che, Yizi Chen, Tanhua Jin, Martin Raubal, Konrad Schindler, and Peter Kiefer. Unsupervised urban land use mapping with street view contrastive clustering and a geographical prior. arXiv preprint arXiv:2504.17551, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning (ICML), 2020

    2020

  11. [11]

    Functional map of the world

    Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  12. [12]

    Urban atlas land cover land use 2018 (vector), europe

    Copernicus Land Monitoring Service & European Environment Agency . Urban atlas land cover land use 2018 (vector), europe. https://land.copernicus.eu/en/products/urban-atlas/urban-atlas-2018, 2021. Accessed 8 July 2025

    2018

  13. [13]

    The use of urban analytics in strategic planning – a case study of the greater sydney region plan

    Claire Daniel and Chris Pettit. The use of urban analytics in strategic planning – a case study of the greater sydney region plan. Computers, Environment and Urban Systems, 117: 0 102249, 2025

    2025

  14. [14]

    Data commons 2024, cdc places, electronic dataset

    Data Commons . Data commons 2024, cdc places, electronic dataset. https://datacommons.org, 2024. Accessed 29 May 2024

    2024

  15. [15]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2009

    2009

  16. [16]

    Postcode level all domestic meters electricity 2023

    Department for Energy Security and Net Zero . Postcode level all domestic meters electricity 2023. https://www.gov.uk/government/statistics/postcode-level-electricity-statistics-2023, 2024. Accessed 8 July 2025

    2023

  17. [17]

    Range: Retrieval augmented neural fields for multi-resolution geo-embeddings

    Aayush Dhakal, Srikumar Sastry, Subash Khanal, Adeel Ahmad, Eric Xing, and Nathan Jacobs. Range: Retrieval augmented neural fields for multi-resolution geo-embeddings. arXiv preprint arXiv:2502.19781, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Climplicit: Climatic implicit embeddings for global ecological tasks

    Johannes Dollinger, Damien Robert, Elena Plekhanova, Lukas Drees, and Jan Dirk Wegner. Climplicit: Climatic implicit embeddings for global ecological tasks. In Proceedings of the International Conference on Learning Representations (ICLR), 2025

    2025

  19. [19]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representatio...

    2021

  20. [20]

    Drusch, U

    M. Drusch, U. Del Bello , S. Carlier, O. Colin, V. Fernandez, F. Gascon, B. Hoersch, C. Isola, P. Laberinti, P. Martimort, A. Meygret, F. Spoto, O. Sy, F. Marchese, and P. Bargellini. Sentinel-2: Esa's optical high-resolution mission for gmes operational services. Remote Sensing of Environment, 120: 0 25--36, 2012

    2012

  21. [21]

    Abhimanyu Dubey, Nikhil Naik, Devi Parikh, Ramesh Raskar, and C \'e sar A. Hidalgo. Deep learning the city: Quantifying urban perception at a global scale. In European Conference on Computer Vision (ECCV), pp.\ 196--212, 2016

    2016

  22. [22]

    What to align in multimodal contrastive learning? In Proceedings of the International Conference on Learning Representations (ICLR), 2025

    Benoit Dufumier, Javiera Castillo Navarro, Devis Tuia, and Jean-Philippe Thiran. What to align in multimodal contrastive learning? In Proceedings of the International Conference on Learning Representations (ICLR), 2025

    2025

  23. [23]

    Uk postcodes with latitude and longitude

    Free Map Tools . Uk postcodes with latitude and longitude. https://www.freemaptools.com/download-uk-postcode-lat-lng.htm, 2024

    2024

  24. [24]

    SimCSE : Simple contrastive learning of sentence embeddings

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. SimCSE : Simple contrastive learning of sentence embeddings. In Empirical Methods in Natural Language Processing (EMNLP), 2021

    2021

  25. [25]

    Imagebind: One embedding space to bind them all

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  26. [26]

    Street view static api

    Google LLC . Street view static api. https://developers.google.com/maps/documentation/streetview, 2007 . Accessed 14 August 2025

    2007

  27. [27]

    Gorelick, M

    N. Gorelick, M. Hancher, M. Dixon, S. Ilyushchenko, D. Thau, and R. Moore. Google earth engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 202: 0 18--27, 2017

    2017

  28. [28]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  29. [29]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  30. [30]

    Gaussian error linear units ( GELUs )

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units ( GELUs ). arXiv preprint arXiv:1606.08415, 2016

    Pith/arXiv arXiv 2016

  31. [31]

    Trueskill: a bayesian skill rating system

    Ralf Herbrich, Tom Minka, and Thore Graepel. Trueskill: a bayesian skill rating system. In Advances in Neural Information Processing Systems (NeurIPS), 2006

    2006

  32. [32]

    Evaluating geospatial context information for travel mode detection

    Ye Hong, Emanuel Stüdeli, and Martin Raubal. Evaluating geospatial context information for travel mode detection. Journal of Transport Geography, 113: 0 103736, 2023

    2023

  33. [33]

    Lo RA : Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lo RA : Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2022

    2022

  34. [34]

    Johannes Jakubik, Sujit Roy, Christopher Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniel Szwarcman, Carlos Gomes, Gabby Nyirjesy, Blair Edwards, Daiki Kimura, Naomi Simumba, Linsong Chu, S. Karthik Mukkavilli, Devyani Lambhate, Kamal Das, Ranjini Bangalore, Dario Oliveira, Michal Muszynski, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha ...

    Pith/arXiv arXiv 2023

  35. [35]

    Mixlight: Borrowing the best of both spherical harmonics and gaussian models

    Xinlong Ji, Fangneng Zhan, Shijian Lu, Shi-Sheng Huang, and Hua Huang. Mixlight: Borrowing the best of both spherical harmonics and gaussian models. arXiv preprint arXiv:2404.12768, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015

    2015

  37. [37]

    Satclip: Global, general-purpose location embeddings with satellite imagery

    Konstantin Klemmer, Esther Rolf, Caleb Robinson, Lester Mackey, and Marc Rußwurm. Satclip: Global, general-purpose location embeddings with satellite imagery. Proceedings of the AAAI Conference on Artificial Intelligence, 39 0 (4): 0 4347--4355, 2025

    2025

  38. [38]

    Challenges in data-driven geospatial modeling for environmental research and practice

    Diana Koldasbayeva, Polina Tregubova, Mikhail Gasanov, Alexey Zaytsev, Anna Petrovskaia, and Evgeny Burnaev. Challenges in data-driven geospatial modeling for environmental research and practice. Nature Communications, 15 0 (1): 0 10700, 2024

    2024

  39. [39]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pp.\ 9459--9474, 2020

    2020

  40. [40]

    Factorized contrastive learning: Going beyond multi-view redundancy

    Paul Pu Liang, Zihao Deng, Martin Ma, James Zou, Louis-Philippe Morency, and Ruslan Salakhutdinov. Factorized contrastive learning: Going beyond multi-view redundancy. In Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  41. [41]

    Gair: Improving multimodal geo-foundation model with geo-aligned implicit representations

    Zeping Liu, Fan Zhang, Junfeng Jiao, Ni Lao, and Gengchen Mai. Gair: Improving multimodal geo-foundation model with geo-aligned implicit representations. arXiv preprint arXiv:2503.16683, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations (ICLR), 2019

    2019

  43. [43]

    Csp: Self-supervised contrastive spatial pre-training for geospatial-visual representations

    Gengchen Mai, Ni Lao, Yutong He, Jiaming Song, and Stefano Ermon. Csp: Self-supervised contrastive spatial pre-training for geospatial-visual representations. In International Conference on Machine Learning (ICML), 2023

    2023

  44. [44]

    Towards the next generation of geospatial artificial intelligence

    Gengchen Mai, Yiqun Xie, Xiaowei Jia, Ni Lao, Jinmeng Rao, Qing Zhu, Zeping Liu, Yao-Yi Chiang, and Junfeng Jiao. Towards the next generation of geospatial artificial intelligence. International Journal of Applied Earth Observation and Geoinformation, 136: 0 104368, 2025

    2025

  45. [45]

    M\"uhlematter, Michelle Halbheer, Alexander Becker, Dominik Narnhofer, Helge Aasen, Konrad Schindler, and Mehmet Ozgur Turkoglu

    Dominik J. M\"uhlematter, Michelle Halbheer, Alexander Becker, Dominik Narnhofer, Helge Aasen, Konrad Schindler, and Mehmet Ozgur Turkoglu. LoRA -ensemble: Efficient uncertainty modelling for self-attention networks. arXiv preprint arXiv:2405.14438, 2024

    arXiv 2024

  46. [46]

    Mühlematter, Sebastian Schweizer, Chenjing Jiao, Xue Xia, Magnus Heitzler, and Lorenz Hurni

    Dominik J. Mühlematter, Sebastian Schweizer, Chenjing Jiao, Xue Xia, Magnus Heitzler, and Lorenz Hurni. Probabilistic road classification in historical maps using synthetic data and deep learning. arXiv preprint arXiv:2410.02250, 2024 a

    Pith/arXiv arXiv 2024

  47. [47]

    Mühlematter, Nina Wiedemann, Yanan Xin, and Martin Raubal

    Dominik J. Mühlematter, Nina Wiedemann, Yanan Xin, and Martin Raubal. Spatially-aware station based car-sharing demand prediction. Journal of Transport Geography, 114: 0 103765, 2024 b

    2024

  48. [48]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  49. [49]

    Planet dump retrieved from https://planet.osm.org

    OpenStreetMap contributors . Planet dump retrieved from https://planet.osm.org. https://www.openstreetmap.org , 2017

    2017

  50. [50]

    Oshan, Ziqi Li, Wei Kang, Levi J

    Taylor M. Oshan, Ziqi Li, Wei Kang, Levi J. Wolf, and A. Stewart Fotheringham. mgwr: A python implementation of multiscale geographically weighted regression for investigating process spatial heterogeneity and scale. ISPRS International Journal of Geo-Information, 8 0 (6), 2019

    2019

  51. [51]

    PyTorch : an imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K\" o pf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch : an imperative style, high-per...

    2019

  52. [52]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in P ython. Journal of Machine Learning Research, 12: 0 2825--2830, 2011

    2011

  53. [53]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021

    2021

  54. [54]

    A stochastic approximation method

    Herbert Robbins and Sutton Monro. A stochastic approximation method. The Annals of Mathematical Statistics, 22 0 (3): 0 400--407, 1951

    1951

  55. [55]

    E. Rolf, J. Proctor, T. Carleton, I. Bolliger, V. Shankar, M. Ishihara, B. Recht, and S. Hsiang. A generalizable and accessible approach to machine learning with global satellite imagery. Nature Communications, 12 0 (1): 0 4392, 2021

    2021

  56. [56]

    Geographic location encoding with spherical harmonics and sinusoidal representation networks

    Marc Rußwurm, Konstantin Klemmer, Esther Rolf, Robin Zbinden, and Devis Tuia. Geographic location encoding with spherical harmonics and sinusoidal representation networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  57. [57]

    M. Sun, C. Kamath, M. Agarwal, A. Muslim, H. Yee, D. Schottlander, S. Bavadekar, N. Efron, S. Shetty, and G. Prasad. Community search signatures as foundation features for human-centered geospatial modeling. arXiv preprint arXiv:2410.22721, 2024

    Pith/arXiv arXiv 2024

  58. [58]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  59. [59]

    World Urbanization Prospects: The 2018 Revision

    United Nations . World Urbanization Prospects: The 2018 Revision. United Nations, New York, 2019

    2018

  60. [60]

    World Cities Report 2024: Cities and Climate Action

    United Nations . World Cities Report 2024: Cities and Climate Action. United Nations, Nairobi, 2024

    2024

  61. [61]

    Geological Survey, Earth Resources Observation and Science Center

    U.S. Geological Survey, Earth Resources Observation and Science Center . Annual national land cover database (nlcd) collection 1.1, 1985–2023. https://www.usgs.gov/centers/eros/science/annual-national-land-cover-database, 2024. Accessed 8 July 2025

    1985

  62. [62]

    The iNaturalist species classification and detection dataset

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The iNaturalist species classification and detection dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  63. [63]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  64. [64]

    Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization

    Vicente Vivanco, Gaurav Kumar Nayak, and Mubarak Shah. Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization. In Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  65. [65]

    Combining deep learning and street view imagery to map smallholder crop types

    Sherrie Wang , Jordi Laguarta , and Thomas Friedel . Combining deep learning and street view imagery to map smallholder crop types . In AGU Fall Meeting Abstracts, 2023

    2023

  66. [66]

    Multi-modal contrastive learning of urban space representations from poi data

    Xinglei Wang, Tao Cheng, Stephen Law, Zichao Zeng, Lu Yin, and Junyuan Liu. Multi-modal contrastive learning of urban space representations from poi data. Computers, Environment and Urban Systems, 120: 0 102299, 2025

    2025

  67. [67]

    Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation

    Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. arXiv preprint arXiv:2211.07044, 2022

    Pith/arXiv arXiv 2022

  68. [68]

    Urban2vec: Incorporating street view imagery and pois for multi-modal urban neighborhood embedding

    Zhecheng Wang, Haoyuan Li, and Ram Rajagopal. Urban2vec: Incorporating street view imagery and pois for multi-modal urban neighborhood embedding. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020

    2020

  69. [69]

    Spatio–temporal effects in GeoAI : From predictability to evaluation

    Nina Wiedemann. Spatio–temporal effects in GeoAI : From predictability to evaluation . PhD dissertation, ETH Zurich, Zürich, Switzerland, 2025. URL https://doi.org/10.3929/ethz-b-000738075

    work page doi:10.3929/ethz-b-000738075 2025

  70. [70]

    Williams and Randall D

    Paul L. Williams and Randall D. Beer. Nonnegative decomposition of multivariate information. arXiv preprint arXiv:1004.2515, 2010

    Pith/arXiv arXiv 2010

  71. [71]

    London house price data

    Jake Wright. London house price data. https://www.kaggle.com/datasets/jakewright/house-price-data, 2025. Accessed 8 July 2025

    2025

  72. [72]

    C-pack: Packaged resources to advance general chinese embedding

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. C-pack: Packaged resources to advance general chinese embedding. arXiv preprint arXiv:2309.07597, 2023

    Pith/arXiv arXiv 2023

  73. [73]

    Mapping fine scale urban housing prices by fusing remotely sensed imagery and social media data

    Yao Yao, Jinbao Zhang, Ye Hong, Haolin Liang, and Jialv He. Mapping fine scale urban housing prices by fusing remotely sensed imagery and social media data. Transactions in GIS, 22 0 (2): 0 561--581, 2018

    2018

  74. [74]

    Gps2vec: Towards generating worldwide gps embeddings

    Yifang Yin, Zhenguang Liu, Ying Zhang, Sheng Wang, Rajiv Ratn Shah, and Roger Zimmermann. Gps2vec: Towards generating worldwide gps embeddings. In Proceedings of the 27th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, pp.\ 416--419, 2019

    2019

  75. [75]

    Learning multi-context aware location representations from large-scale geotagged images

    Yifang Yin, Ying Zhang, Zhenguang Liu, Yuxuan Liang, Sheng Wang, Rajiv Ratn Shah, and Roger Zimmermann. Learning multi-context aware location representations from large-scale geotagged images. In Proceedings of the 29th ACM International Conference on Multimedia, pp.\ 899--907, 2021

    2021

  76. [76]

    Multimodal contrastive training for visual representation learning

    Xin Yuan, Zhe Lin, Jason Kuen, Jianming Zhang, Yilin Wang, Michael Maire, Ajinkya Kale, and Baldo Faieta. Multimodal contrastive training for visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  77. [77]

    The equal earth map projection

    Bojan Šavrič, Tom Patterson, and Bernhard Jenny. The equal earth map projection. International Journal of Geographical Information Science, 33 0 (3): 0 454--465, 2019

    2019

  78. [78]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  79. [79]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  80. [80]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

Showing first 80 references.