Pith. sign in

REVIEW 3 major objections 5 minor 59 references

Geo-Embed beats all 15 compared baseline models across 45 urban tasks, the paper reports.

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 ↗

GeoMEB unifies 45 urban embedding tasks into a ranking protocol, and Geo-Embed, an instruction-conditioned vision-language embedder fine-tuned on it, tops the leaderboard.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A genuinely useful 45-task urban embedding benchmark, but the headline model comparison is unfair: Geo-Embed is fine-tuned on the benchmark's own training data while all baselines are zero-shot. the 3 major comments →

arxiv 2608.03826 v1 pith:CPSVVGWY submitted 2026-08-04 cs.CV cs.LG

Geo-Embed: Towards Unified Multimodal Embeddings for Urban Understanding

Jiapeng Li , Yong Li , Junjie Zhou , Fan Zhang , Yu Liu This is my paper
classification cs.CV cs.LG
keywords multimodal embeddingsurban understandinggeospatial benchmarkinstruction-conditioned rankingcontrastive learningcross-view geo-localizationchange detectionvisual grounding
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.

The reading

The paper claims that the many different comparison tasks urban systems need—matching street photos to satellite imagery, answering questions about a scene, locating a described object, spotting what changed over time, classifying land cover—can be handled by one embedding model, as long as each task is recast as ranking a target candidate against an instructed query. To support this, it introduces GeoMEB, a benchmark that standardizes 45 urban tasks into a common instruction-conditioned ranking protocol with 1.32 million training examples and 286,000 evaluation queries, and Geo-Embed, a 2-billion-parameter model that encodes each query and candidate separately with side-specific instructions and trains with contrastive learning. On GeoMEB, Geo-Embed scores 25.48 on the task-count-weighted average, a 15.3% relative improvement over the strongest baseline, Qwen3-VL-Embedding-8B at 22.09. The paper's deeper claim is that geospatial embedding research should organize training and evaluation around explicit query-target relations—semantic, cross-view, region-level, and temporal—rather than around generic image-text matching.

Core claim

The central discovery claimed is that a single instruction-conditioned vision-language embedding model, fine-tuned on urban query-target pairs, outperforms every compared CLIP-style, instruction-tuned, and embedding-specialized model on a unified 45-task urban benchmark. Geo-Embed treats retrieval, visual question answering, classification, grounding, and change detection all as matching an instructed query to candidates in a task-specific pool, using the same backbone for both sides. It reaches first place on classification, retrieval, and grounding, and second on VQA and change detection. The paper also reports that the gains are uneven across task types: image-caption matching and land-co

What carries the argument

The load-bearing mechanism is instruction-conditioned query-target matching. Every query and every candidate is serialized with its own side-specific instruction, then encoded by a shared vision-language backbone; the last-layer hidden state of the final token, after L2 normalization, becomes the embedding, and ranking is done by cosine similarity. Training uses the InfoNCE contrastive objective with in-batch negatives. Side-specific instructions let the same raw content receive different embeddings when it plays different roles—query versus candidate, or in one task relation versus another—and this is what lets one model handle single images, multiple images, text, regions, and masks within

Load-bearing premise

The headline ranking assumes that a task-count-weighted average—giving each task equal weight regardless of how many queries or candidates it has—is a meaningful measure of overall model quality.

What would settle it

Recompute the overall score with query-count weighting, macro-median, or bootstrap confidence intervals from the per-task scores in Appendix Tables 8–12; if Geo-Embed's lead over Qwen3-VL-Embedding-8B disappears or reverses, the strongest-performance claim is unsupported. A second check: evaluate all models on a held-out set of urban query-target tasks not in GeoMEB and see whether the same ranking holds.

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

If this is right

  • A single embedding interface could serve city-scale applications that currently need separate models: urban search, geo-localization, disaster monitoring, and socioeconomic sensing.
  • Generic multimodal ability does not guarantee strong urban retrieval; instruction-tuned VLMs score low unless they are trained for retrieval.
  • Multi-task supervision is not automatically additive: grounding performance can drop when retrieval or VQA data is mixed in, suggesting interference between structurally different query-target relations.
  • The hard task types—cross-view geolocalization, score-binning urban perception, region grounding, and change-mask matching—are concrete stress tests for future geospatial embedding models.
  • Training and evaluation should be organized around explicit query-target relations rather than generic image-text alignment.

Where Pith is reading between the lines

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

  • Not tested in the paper: because the headline average weighs each task equally regardless of query count (for example OSF Recovery has 137 queries and 4 candidates, while CityLens tasks have 1,000), the 15.3% margin may shrink or flip under query-count weighting or per-task confidence intervals; a bootstrap re-analysis of Appendix Tables 8–12 would settle this.
  • The instruction-conditioned ranking protocol is not inherently urban; the same standardization could be applied to medical imaging, embodied navigation, or document grounding, with GeoMEB as a template.
  • The observed interference between grounding and retrieval suggests a possible curriculum or mixture-weighting strategy that the paper does not explore.
  • Future work could test whether Geo-Embed's gains persist on urban tasks constructed after the benchmark's release, or whether it partly overfits GeoMEB's template distribution.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper introduces GeoMEB, a geospatial multimodal embedding benchmark standardizing 45 urban tasks—retrieval, visual question answering, change detection, classification, and visual grounding—into an instruction-conditioned ranking protocol, with 1.32M training examples and 286K evaluation queries. It also presents Geo-Embed, initialized from Qwen3-VL-Embedding-2B and LoRA-fine-tuned with an InfoNCE objective, encoding queries and candidates with side-specific instructions into a shared space. The main empirical claim is that Geo-Embed achieves the strongest overall performance on GeoMEB (25.48 vs. 22.09 for Qwen3-VL-Embedding-8B, a 15.3% relative improvement), with ablations on sub-batch sampling, task-modality interactions, instruction-induced embedding shifts, LoRA rank, and batch size.

Significance. The benchmark and the instruction-conditioned query-target formulation are potentially useful contributions: GeoMEB covers under-studied geospatial relations (cross-view, temporal, region-level) and is assembled from diverse public datasets; the task taxonomy and standardized ranking protocol could support future comparisons. The paper also provides unusually broad baseline coverage and a genuinely informative instruction-shift analysis (Fig. 6) that shows instructions do not uniformly improve alignment. However, the central performance claim is currently under-supported: Geo-Embed is fine-tuned on training splits of the same datasets used for evaluation, while all baselines are evaluated zero-shot, and the aggregate metric used for the headline number is sensitive to benchmark design choices. The contribution would be significant if the comparison were made symmetric and the aggregate ranking shown to be robust.

major comments (3)
  1. [§5, Table 2, Appendix E] The headline comparison is asymmetric. Geo-Embed is initialized from Qwen3-VL-Embedding-2B and LoRA-fine-tuned on GeoMEB's 1.32M training collection (§4, Appendix E), which is drawn from the same datasets as the 45 evaluation tasks (Table 1: UAV-GeoLoc, VRSBench, SV Localization, CityLens, etc.). No baseline is fine-tuned on that training data, and the paper does not state this asymmetry in the main-results section. The claim that Geo-Embed 'achieves the strongest overall performance among representative multimodal embedders' therefore conflates training-data access with model/method quality. The load-bearing comparison must include an equally trained baseline (e.g., Qwen3-VL-Embedding-2B or -8B fine-tuned on the same GeoMEB training collection under the same protocol), or at minimum a zero-shot evaluation of Geo-Embed before fine-tuning. Without this, the 15.3% margin does not support t
  2. [Table 2, Table 1] The aggregate metric is not robust enough to carry the headline claim. The 'Overall Avg' column is an unweighted mean of the five meta-task means, not a 45-task mean; this gives the 7 grounding+change-detection tasks the same aggregate weight as the 26 retrieval+VQA tasks. Recomputing from Appendix Tables 8–12, a simple 45-task mean shrinks the margin substantially. In addition, OSF Recovery has only 137 evaluation queries and a 4-candidate pool (Table 1) yet contributes a large share of the change-detection meta-average (Table 11: e.g., 42.34 for Seed1.6-Embedding, 37.96 for Geo-Embed). With such small query counts and tiny candidate pools, the ranking differences may reflect noise or benchmark design rather than embedding quality. The paper should report query-weighted or per-task-level aggregates, confidence intervals or significance tests, and show that the overall ranking is stable
  3. [§6, Fig. 6] The instruction-conditioned query-target design is not isolated as the cause of the gains. The paper's motivation emphasizes side-specific instructions, but the main comparison varies both the method and the training data; no ablation trains a strong baseline without instructions, or trains Geo-Embed without side-specific instructions. Fig. 6 shows that instruction-induced shifts are not uniformly target-oriented (e.g., Qwen3-VL-Embedding-2B Text→SV is -0.0448, RS→SV is -0.0103), which weakens the claim that instructions per se drive the improvement. A controlled ablation that fixes the training data and varies only the instruction design is needed to support the central architectural claim.
minor comments (5)
  1. [Table 11] The handling of '–' (counted as 0 in change-detection averages) penalizes CLIP-family models that cannot accept multi-image inputs. This is a modeling limitation, but reporting it as 0 in the same aggregate as models that support multi-image inputs may distort the ranking; consider reporting separate averages or a footnote in Table 2.
  2. [Figure 4] The caption says the fine-tuned panel is 'after VRSBench fine-tuning', while the main model is trained on the full GeoMEB collection. Clarify which checkpoint is visualized.
  3. [Eq. (5), Appendix Table 7] The temperature τ is a key hyperparameter (set to 0.02) but is not included in the hyperparameter sensitivity study. At minimum, state whether results are stable across reasonable τ values.
  4. [Appendix B] The provenance of the Im2GPS3k caption annotations and SV Localization captions should be described more precisely: were they human-annotated, generated, or adapted from prior work? This matters for the benchmark's reliability.
  5. [Table 2 caption] The caption says 'task-count-weighted averages over fine-grained task types and meta-task groups', but the Overall Avg column is an unweighted mean of the five meta-task averages. The wording should be changed to avoid ambiguity.

Circularity Check

0 steps flagged

No significant circularity: Geo-Embed's results are empirical evaluations, not derivations reducing to their inputs.

full rationale

Geo-Embed is a contrastively fine-tuned embedding model; its performance on GeoMEB is measured on held-out evaluation queries after training on a disjoint 1.32M-example collection. The paper contains no derivation chain in which a prediction is defined in terms of the fitted parameter or in which the claimed result is equivalent by construction to the benchmark definition. The closest self-referential elements are (i) GeoMEB is introduced by the same authors and Geo-Embed is evaluated on it, and (ii) the SV Localization task derives from the authors' prior Unified Urban Tuning work (Li et al. 2026b) and is used as both training and evaluation data. These are data/benchmark construction choices, not load-bearing circular arguments: the benchmark's protocol (instruction-conditioned ranking) does not by itself guarantee Geo-Embed's ranking, and the model's advantage over zero-shot baselines remains an empirical, potentially confounded observation. The asymmetry that Geo-Embed is fine-tuned on GeoMEB training data while baselines are evaluated zero-shot is a fairness/correctness concern for the 15.3% headline, but it is not circularity because the outcome is not forced by construction. No self-citation supplies an unverified premise that the paper's conclusions depend on.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on benchmark design choices and standard hyperparameters rather than on derived constants. All numeric training parameters are disclosed in Appendix E. No new physical or formal entities are postulated.

free parameters (4)
  • InfoNCE temperature tau = 0.02
    Temperature scaling in Eq. 5; chosen by hand, affects the sharpness of the contrastive distribution.
  • LoRA rank = 16 (alpha 32)
    Selected from {4, 8, 16, 32} on a random 10% GeoMEB subset; rank 16 gives the best overall average in Table 4.
  • Batch size = 1024
    Selected from {64, 256, 512, 1024}; larger batches improve the overall score, especially on grounding and change detection (Table 4).
  • Learning rate / weight decay / warmup ratio = 2e-5 / 0.1 / 0.05
    Standard fine-tuning hyperparameters listed in Appendix E; not swept, but they determine the training trajectory.
axioms (4)
  • domain assumption Ranking over a task-specific candidate pool with cosine similarity preserves the semantics of VQA, grounding, change detection, and classification.
    Section 3 meta-task design: every task is cast as matching an instructed query to candidates. If this reformulation distorts task semantics, GeoMEB scores do not measure the underlying abilities.
  • domain assumption The 45 tasks and the task-count-weighted average are a valid aggregate measure of urban embedding quality.
    Table 2 uses task-count-weighted averages; tiny tasks (OSF Recovery: 137 queries, 4 candidates) receive the same weight as large ones (SV Localization: 39,887 queries).
  • domain assumption Train/evaluation splits prevent target-instance leakage.
    Section 3 states official splits are preserved or random splits ensure no target instances are shared across splits; no external verification is provided.
  • domain assumption Instruction templates correctly specify the query-target relation for each task.
    Equation 2 uses side-specific instructions p_q^r and p_c^r; the wording of these templates is a hand-designed choice, and Section 6 shows instructions move embeddings unevenly across directions.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Geo-Embed: Towards Unified Multimodal Embeddings for Urban Understanding." pith.science (2026). https://pith.science/paper/CPSVVGWY

@misc{pith2026260803826,
  author       = {Pith},
  title        = {Pith review of: Geo-Embed: Towards Unified Multimodal Embeddings for Urban Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CPSVVGWY}},
  note         = {Machine review of arXiv:2608.03826}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Geospatial and urban applications increasingly require models to compare heterogeneous evidence across street-view imagery, remote-sensing observations, text descriptions, region proposals, and temporal change cues. However, existing multimodal embedding models and benchmarks are still largely designed and evaluated around general-purpose image-text matching, leaving unclear whether unified embedding space can support heterogeneous geospatial tasks involving spatial relationships, fine-grained semantics, and temporal changes. To address this gap, we make three key contributions. First, we introduce GeoMEB, a large-scale multimodal embedding benchmark that standardizes 45 urban evaluation tasks across retrieval, visual question answering, change detection, classification, and visual grounding, together with training collections comprising 1.32M examples and 286K evaluation queries. Second, we present Geo-Embed, a unified embedding model that adapts a shared vision-language backbone to instruction-conditioned query-target matching over heterogeneous geospatial inputs, including single images, multiple images, text, regions, and masks. On GeoMEB, Geo-Embed achieves the strongest overall performance among representative multimodal embedders, with a 15.3% relative improvement over the strongest baseline. These results motivate future geospatial embedders that organize training and evaluation around explicit query-target relations, including semantic, cross-view, region-level, and temporal correspondence.

Figures

Figures reproduced from arXiv: 2608.03826 by Fan Zhang, Jiapeng Li, Junjie Zhou, Yong Li, Yu Liu.

Figure 1
Figure 1. Figure 1: Overview of GeoMEB and Geo-Embed. GeoMEB unifies classification, visual question answering, change detection, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: GeoMEB task taxonomy and input composition. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of Geo-Embed. The query and its [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Feature space visualization: (a) Qwen3-VL [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison of models trained with [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Instruction-induced alignment shifts, reporting the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 1
Figure 1. Figure 1: Annotated aligned VIGOR triplets used for instruction-shift analysis. [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: A qualitative example of the SV Localization image-to-text and text-to-image tasks. The annotated caption provides a [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A qualitative example of the Im2GPS3k caption image-to-text and text-to-image tasks. The annotated description [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

59 extracted references · 45 canonical work pages · 4 internal anchors

  1. [1]

    Illuminating Visual Identity in Universal Multimodal Embeddings , booktitle =

    Cao, Jiawei and Feng, Junyi and Hua, Jiashen and Huang, Ziheng and Deng, Bing and Wu, Kaijie and Gu, Chaochen and Ye, Jieping , year = 2026, pages =. Illuminating Visual Identity in Universal Multimodal Embeddings , booktitle =

    work page 2026

  2. [2]

    2604.23321 , primaryclass =

    Huang, Haohang and Lu, Xuan and Su, Mingyi and Zhang, Xuan and Jiang, Ziyan and Nie, Ping and Zou, Kai and Pfister, Tomas and Chen, Wenhu and Zhang, Wei and Shen, Xiaoyu and Meng, Rui , year = 2026, number =. 2604.23321 , primaryclass =

    Pith/arXiv arXiv 2026

  3. [3]

    Jiang, Ziyan and Meng, Rui and Yang, Xinyi and Yavuz, Semih and Zhou, Yingbo and Chen, Wenhu , year = 2024, language =. The

    work page 2024

  4. [4]

    Liu, Tianhui and Pang, Hetian and Zhang, Xin and Ouyang, Tianjian and Zhang, Zhiyuan and Feng, Jie and Li, Yong and Hui, Pan , year = 2025, language =. The

    work page 2025

  5. [5]

    Meng, Rui and Jiang, Ziyan and Liu, Ye and Su, Mingyi and Yang, Xinyi and Fu, Yuepeng and Qin, Can and Thirukovalluru, Raghuveer and Zhang, Xuan and Chen, Zeyuan and Xu, Ran and Xiong, Caiming and Zhou, Yingbo and Chen, Wenhu and Yavuz, Semih , year = 2025, journal =

    work page 2025

  6. [6]

    Proceedings of the

    Xiao, Chenghao and Chung, Isaac and Kerboua, Imene and Stirling, Jamie and Zhang, Xin and Kardos, M. Proceedings of the

    work page

  7. [7]

    2509.26378 , primaryclass =

    Zhou, Junjie and Liu, Ze and Xiong, Lei and Yao, Jin-Ge and Wang, Yueze and Xiao, Shitao and Lin, Fenfen and Chen, Miguel Hu and Dou, Zhicheng and Bao, Siqi and Lian, Defu and Xiong, Yongping and Liu, Zheng , year = 2025, number =. 2509.26378 , primaryclass =

    arXiv 2025

  8. [8]

    and Kazmierski, Michal R

    Brown, Christopher F. and Kazmierski, Michal R. and Pasquarella, Valerie J. and Rucklidge, William J. and Samsikova, Masha and Zhang, Chenhui and Shelhamer, Evan and Lahera, Estefania and Wiles, Olivia and Ilyushchenko, Simon and Gorelick, Noel and Zhang, Lihui Lydia and Alj, Sophia and Schechter, Emily and Askay, Sean and Guinan, Oliver and Moore, Rebecc...

    Pith/arXiv arXiv 2025

  9. [9]

    and Burns, Benjamin and

    Chen, Ziru and Chen, Shijie and Ning, Yuting and Zhang, Qianheng and Wang, Boshi and Yu, Botao and Li, Yifei and Liao, Zeyi and Wei, Chen and Lu, Zitong and Dey, Vishal and Xue, Mingyi and Baker, Frazier N. and Burns, Benjamin and. The

  10. [10]

    Jina-Embeddings-v4:

    G. Jina-Embeddings-v4:. Proceedings of the 5th

    work page

  11. [11]

    Li, Mingxin and Zhang, Yanzhao and Long, Dingkun and Chen, Keqin and Song, Sibo and Bai, Shuai and Yang, Zhibo and Xie, Pengjun and Yang, An and Liu, Dayiheng and Zhou, Jingren and Lin, Junyang , year = 2026, number =. Qwen3-. 2601.04720 , primaryclass =

    Pith/arXiv arXiv 2026

  12. [12]

    Shanbhogue, Madhuri and Li, Zhe and Zhang, Shanfeng and. Gemini. arXiv , language =:2605.27295 , primaryclass =

    Pith/arXiv arXiv

  13. [13]

    Bridging

    Zhang, Xin and Zhang, Yanzhao and Xie, Wen and Li, Mingxin and Dai, Ziqi and Long, Dingkun and Xie, Pengjun and Zhang, Meishan and Li, Wenjie and Zhang, Min , year = 2025, pages =. Bridging. 2025

    work page 2025

  14. [14]

    Proceedings of the 63rd

    Zhou, Junjie and Xiong, Yongping and Liu, Zheng and Liu, Ze and Xiao, Shitao and Wang, Yueze and Zhao, Bo and Zhang, Chen Jason and Lian, Defu , editor =. Proceedings of the 63rd

    work page

  15. [15]

    Zhou, Dewei and Huang, Xinyu and Wang, Xun and Xie, Ji and Zhang, Yabo and Li, Liang and Li, Kunchang and Yang, Zongxin and Yang, Yi , year = 2026, publisher =

    work page 2026

  16. [16]

    Cui, Xuanming and Cheng, Jianpeng and Chen, Hong-You and Shukla, Satya Narayan and Awasthi, Abhijeet and Pan, Xichen and Ahuja, Chaitanya and Mishra, Shlok Kumar and Tian, Taipeng and Guo, Qi and Lim, Ser-Nam and Singh, Aashu and Fan, Xiangjun , year = 2025, language =. Think. The

    work page 2025

  17. [17]

    2604.02073 , primaryclass =

    He, Chenwei and Hao, Xiangzhao and Yang, Tianyu and Ma, Yuxiang and Jia, Yuheng and Wu, Lingxiang and Zhao, Chaoyang and Guo, Haiyun and Wang, Jinqiao , year = 2026, number =. 2604.02073 , primaryclass =

    Pith/arXiv arXiv 2026

  18. [18]

    Jiang, Haonan and Wang, Yuji and Zhu, Yongjie and Lu, Xin and Qin, Wenyu and Wang, Meng and Wan, Pengfei and Tang, Yansong , year = 2026, number =. Embed-. 2602.13823 , primaryclass =

    arXiv 2026

  19. [19]

    Lan, Zhibin and Niu, Liqiang and Meng, Fandong and Zhou, Jie and Su, Jinsong , year = 2025, language =. The

    work page 2025

  20. [20]

    2604.06156 , primaryclass =

    Wang, Yuchi and Yu, Haiyang and Bian, Weikang and Long, Jiefeng and Liang, Xiao and Feng, Chao and Li, Hongsheng , year = 2026, number =. 2604.06156 , primaryclass =

    Pith/arXiv arXiv 2026

  21. [21]

    Wu, Peixi and Mei, Ke and Ma, Feipeng and Chai, Bosong and Lan, Zhibin and Zhao, Chenxi and Yan, Shannan and Chen, Jie and Hu, Zhangchi and Peng, Yansong and Lin, Bo and Zhou, Junjie and Yin, Dacheng and Wang, Tianyi and Rao, Fengyun and Lyu, Jing and Li, Hebei and Sun, Xiaoyan , year = 2026, number =. Beyond. 2604.22280 , primaryclass =

    Pith/arXiv arXiv 2026

  22. [22]

    Zhang, Longxiang and Dai, Weilong and Zhang, Guanghao and Jiang, Hao and Huang, Pipei , year = 2026, number =. Think. 2605.14448 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  23. [23]

    Learning

    Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya , year = 2021, pages =. Learning. Proceedings of the 38th

    work page 2021

  24. [24]

    Autonomous Robots , volume =

    Street-View Change Detection with Deconvolutional Networks , author =. Autonomous Robots , volume =

    work page

  25. [25]

    Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and Chen, Keqin and Chen, Xionghui and Cheng, Zesen and Deng, Lianghao and Ding, Wei and Gao, Chang and Ge, Chunjiang and Ge, Wenbin and Guo, Zhifang and Huang, Qidong and Huang, Jie and Huang, Fei and Hui, Binyuan and Jiang, Shutong and Li, Zhaohai and Li, Mingsheng and Li, Mei and Li, Kaixin and Lin, Zicheng a...

    Pith/arXiv arXiv 2025

  26. [26]

    Chen, Hao and Shi, Zhenwei , year = 2020, journal =. A

    work page 2020

  27. [27]

    Cordts, Marius and Omran, Mohamed and Ramos, Sebastian and Rehfeld, Timo and Enzweiler, Markus and Benenson, Rodrigo and Franke, Uwe and Roth, Stefan and Schiele, Bernt , year = 2016, pages =. The. Proceedings of the

    work page 2016

  28. [28]

    Hu, Edward J and Shen, Yelong and Wallis, Phillip and. Lora:. Iclr , volume =

    work page

  29. [29]

    Nature , volume =

    Built Environment Disparities Are Amplified during Extreme Weather Recovery , author =. Nature , volume =

    work page

  30. [30]

    Jia, Pengyue and Liu, Yiding and Li, Xiaopeng and Zhao, Xiangyu and Wang, Yuhao and Du, Yantong and Han, Xiao and Wei, Xuetao and Wang, Shuaiqiang and Yin, Dawei , year = 2024, language =. G3:. The

    work page 2024

  31. [31]

    Proceedings of the

    Kuckreja, Kartik and Danish, Muhammad Sohail and Naseer, Muzammal and Das, Abhijit and Khan, Salman and Khan, Fahad Shahbaz , year = 2024, pages =. Proceedings of the

    work page 2024

  32. [32]

    Li, Xiang and Ding, Jian and Elhoseiny, Mohamed , year = 2024, journal =

    work page 2024

  33. [33]

    Liu, Fan and Chen, Delong and Guan, Zhangqingyun and Zhou, Xiaocong and Zhu, Jiale and Ye, Qiaolin and Fu, Liyong and Zhou, Jun , year = 2024, journal =

    work page 2024

  34. [34]

    Exploring

    Lu, Xiaoqiang and Wang, Binqiang and Zheng, Xiangtao and Li, Xuelong , year = 2018, journal =. Exploring

    work page 2018

  35. [35]

    Representation

    van den Oord, Aaron and Li, Yazhe and Vinyals, Oriol , year = 2019, number =. Representation. 1807.03748 , primaryclass =

    Pith/arXiv arXiv 2019

  36. [36]

    Salesses, Mark Philip , year = 2012, eprint =. Place

    work page 2012

  37. [37]

    Shi, Qian and Liu, Mengxi and Li, Shengchen and Liu, Xiaoping and Wang, Fei and Zhang, Liangpei , year = 2022, journal =. A

    work page 2022

  38. [38]

    2502.14786 , primaryclass =

    Tschannen, Michael and Gritsenko, Alexey and Wang, Xiao and Naeem, Muhammad Ferjad and Alabdulmohsin, Ibrahim and Parthasarathy, Nikhil and Evans, Talfan and Beyer, Lucas and Xia, Ye and Mustafa, Basil and H. 2502.14786 , primaryclass =

    Pith/arXiv arXiv

  39. [39]

    Revisiting

    Vo, Nam and Jacobs, Nathan and Hays, James , year = 2017, pages =. Revisiting. Proceedings of the

    work page 2017

  40. [40]

    Wang, Zhecheng and Prabha, Rajanie and Huang, Tianyuan and Wu, Jiajun and Rajagopal, Ram , year = 2024, journal =

    work page 2024

  41. [41]

    Wu, Rouwan and Deng, Jiacheng and Mou, Mingyu and He, Xingyi and Zhang, Maojun and Liu, Yu and Yan, Shen , year = 2025, journal =

    work page 2025

  42. [42]

    Xia, Gui-Song and Hu, Jingwen and Hu, Fan and Shi, Baoguang and Bai, Xiang and Zhong, Yanfei and Zhang, Liangpei and Lu, Xiaoqiang , year = 2017, journal =

    work page 2017

  43. [43]

    Ertler, Christian and Mislej, Jerneja and Ollmann, Tobias and Porzi, Lorenzo and Neuhold, Gerhard and Kuang, Yubin , year = 2020, pages =. The

    work page 2020

  44. [44]

    2023 , eprint=

    Data Filtering Networks , author=. 2023 , eprint=

    work page 2023

  45. [45]

    Xu, Mengyao and Zhou, Wenfei and Babakhin, Yauhen and Moreira, Gabriel and Ak, Ronay and Osmulski, Radek and Liu, Bo and Oldridge, Even and Schifferer, Benedikt , year = 2025, number =. Omni-. 2510.03458 , primaryclass =

    arXiv 2025

  46. [46]

    Zhang, Zilun and Zhao, Tiancheng and Guo, Yulong and Yin, Jianwei , year = 2024, journal =

    work page 2024

  47. [47]

    Zhang, YiFan and Zhang, Huanyu and Tian, Haochen and Fu, Chaoyou and Zhang, Shuangqing and Wu, Junfei and Li, Feng and Wang, Kun and Wen, Qingsong and Zhang, Zhang and Wang, Liang and Jin, Rong , year = 2025, journal =

    work page 2025

  48. [48]

    Zhou, Baichuan and Yang, Haote and Chen, Dairong and Ye, Junyan and Bai, Tianyi and Yu, Jinhua and Zhang, Songyang and Lin, Dahua and He, Conghui and Li, Weijia , year = 2025, journal =

    work page 2025

  49. [49]

    Zhu, Sijie and Yang, Taojiannan and Chen, Chen , year = 2021, pages =. 2021

    work page 2021

  50. [50]

    2025 , month = jun, day =

    Built on Seed1.6-Flash, Seed-1.6-Embedding Launched , author =. 2025 , month = jun, day =

    work page 2025

  51. [51]

    ObjEmbed: Towards Universal Multimodal Object Embeddings

    Fu, Shenghao and Su, Yukun and Rao, Fengyun and Lyu, Jing and Xie, Xiaohua and Zheng, Wei-Shi , year = 2026, number =. 2602.01753 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  52. [52]

    Relational Visual Similarity , booktitle =

    Nguyen, Thao and Mo, Sicheng and Singh, Krishna Kumar and Wang, Yilin and Shi, Jing and Kolkin, Nicholas and Shechtman, Eli and Lee, Yong Jae and Li, Yuheng , year = 2026, pages =. Relational Visual Similarity , booktitle =

    work page 2026

  53. [53]

    Tang, Changli and Xiao, Qinfan and Mei, Ke and Wang, Tianyi and Rao, Fengyun and Zhang, Chao , year = 2025, language =. The

    work page 2025

  54. [54]

    The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric

    Wang, Sheng-Yu and Nitzan, Yotam and Hertzmann, Aaron and Zhu, Jun-Yan and Shechtman, Eli and Efros, Alexei A. and Zhang, Richard , year = 2026, number =. The. 2607.18237 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  55. [55]

    Wei, Yiyang and Song, Tingyu and Zhang, Siyue and Zhao, Yilun , editor =. A. Proceedings of the 64th

    work page

  56. [56]

    Weller, Orion and Boratko, Michael and Naim, Iftekhar and Lee, Jinhyuk , year = 2025, language =. On the. The

    work page 2025

  57. [57]

    Bringing Agentic Search to Earth Observation Data Discovery

    Yu, Minghan and Sun, Youran and Yi, Chugang and Wen, Yixin and Yang, Haizhao , year = 2026, number =. Bringing. 2607.02387 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  58. [58]

    Yunzeliu/Awesome-Unified-Embedding , author =

    work page

  59. [59]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Unified Urban Tuning: Co-Enhancing Satellite and Street View Reasoning with a Progressive Tuning Framework , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.