Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

AddressVLM: Cross-view Alignment Tuning for Image Address Localization using Large Vision-Language Models

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read A two-stage recipe that matches street views to labeled satellite maps lifts street-and-district address accuracy by 9–12 points.

desk verdict Useful two-stage recipe for street-level address QA in LVLMs; the stage-1 benefit is real but its mechanism is underdetermined by the current experiments. read the letter →

arxiv 2508.10667 v1 pith:IMJ27EZB submitted 2025-08-14 cs.CV cs.AI

classification cs.CVcs.AI
keywords imageaddresslocalizationvision-languagemodelscross-viewalignmentsatelliteimagerystreet-viewvisualquestionansweringgeo-localizationlow-rankadaptation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

AddressVLM is a vision-language model trained to answer address questions about street-view images at street and district level. The paper's central claim is that sparse, viewpoint-limited street views do not by themselves give a model enough of a city's street layout, so direct fine-tuning on street-view VQA underperforms. To fix this, the authors add a first training stage—cross-view alignment tuning—in which a street-view image is grafted onto a labeled satellite map of the same area and the model must explain why the street view matches that map location, with the correct address supplied as a hint in the label-generation prompt. After this stage, the model is fine-tuned on street-view address VQA. On new Pittsburgh and San Francisco benchmarks, this two-stage recipe improves joined street-and-district accuracy by about 9 and 12 points over the direct fine-tuning baseline and beats the adapted prior generative geo-localization method by 11 and 14 points.

What carries the argument

Cross-view alignment tuning: a first-stage training task in which a street-view image is grafted into a labeled satellite map of the surrounding area and the model learns to explain the match, followed by address localization tuning on street-view VQA. The grafting operation combines the two views with a binary mask; the satellite map carries printed street names and the label-generation prompt reveals the true address as a hint. This machinery is what supplies the global street-layout context that street-view-only data lacks.

What would settle it

Remove the address hint from the label-generation prompt during stage 1 and replace the street-name text on the satellite maps with scrambled strings, then measure Asd on street-view-only test sets. If the +9/+12 point gains survive, the claimed transfer of a global street-layout prior via OCR street names is not carrying the result; if they collapse, the mechanism is confirmed.

Watch

Extended reading notes

Core claim

The discovery is that a macro-to-micro alignment stage can put a city-scale street-distribution prior into a generative vision-language model. The mechanism works by grafting a small street-view image into the upper-right corner of a square satellite map whose streets carry printed names, then training the model on automatically generated alignment labels: the label-generation prompt shows the street view, the map, and a text hint with the true address, and an off-the-shelf LVLM produces a reason that ties the visual appearance of the street view to the map location. After this first-stage low-rank adaptation tuning, the model is tuned again on street-view-only address VQA covering generatio

Load-bearing premise

The load-bearing premise is that the first tuning stage teaches a transferable global street-layout prior rather than letting the model latch onto map textures or memorize the address hints that appear in the generated labels.

Editorial extensions

If this is right

  • Direct address localization tuning on street-view VQA alone leaves roughly 9–12 points on the table; a map-alignment first stage recovers them.
  • A single square grafted image works better than stitching two images or feeding two images separately, so single-image vision-language models can absorb cross-view information without extra visual tokens.
  • The trained model answers generation, judgment, and multiple-choice address questions at both street and district levels, not just a fixed candidate list.
  • The pipeline transfers to a second city and to Tokyo, and unified training across cities slightly improves performance, suggesting it can scale to more cities.
  • Joined street-and-district accuracy for an open-set generative model remains below that of a closed-set contrastive baseline, so generative address localization still has headroom.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same grafted-map recipe could be tried for other spatially grounded VQA domains—floor plans, campus maps, mall directories—where a sparse ground-level image needs a global reference frame.
  • Because the paper's ablation shows that removing printed street labels from the satellite maps hurts performance, a clean test is whether the prior is genuinely geometric (road topology) or lexical (OCR of street names); if lexical, the approach may struggle where street signs are absent or use non-Latin scripts.
  • The unified multi-city model slightly beating single-city models hints that adding more cross-view alignment cities could keep improving performance, potentially toward country-scale address localization.
  • A sharper evaluation would measure whether the stage-1 prior transfers across cities without retraining: train stage 2 in a new city and see how much of the +9/+12 point gain survives.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes AddressVLM, a two-stage fine-tuning recipe for city-wide image address localization with an LVLM. Stage 1 ("cross-view alignment tuning") grafts a street-view image onto a satellite map annotated with street-name labels and trains the model to explain the matching, using labels auto-generated by an LVLM prompted with the correct address hint. Stage 2 performs address localization tuning on street-view VQA data. The authors build two VQA datasets from Pitts-IAL and SF-Base-IAL, evaluate generation/judgment/multiple-choice accuracy at district and street levels, and report consistent gains over a direct fine-tuning baseline and over an adapted GeoReasoner baseline, plus qualitative and scalability experiments (multi-city, Tokyo, different backbones).

Significance. The contribution is practically interesting: if the cross-view prior claim holds, it offers a lightweight way to inject global street-distribution knowledge into LVLMs using only LoRA, with released datasets and multiple ablations. The paper is also transparent in including a limitations paragraph and detailed appendices. However, the central attribution of the stage-1 gains to cross-view alignment is not yet established because of confounding factors and unspecified evaluation details. The strengths are the dataset construction, the breadth of ablations (grafting ratio, trainable components, data density, mixed-city, backbone, Tokyo), and the reproducible training setup (xtuner, hyperparameters).

major comments (5)
  1. [§3.2, Tables 2-3] The headline +9.08/+11.83 Asd gains (Table 2) are attributed to the cross-view alignment prior, but the comparison is confounded: the Baseline receives no stage-1 data, so AddressVLM trains on strictly more LoRA data/steps. Table 3 shows that removing the street-name overlay from the satellite map drops Asd by roughly 5 points on both cities, indicating that much of the gain is tied to OCR-style map-label reading. A control with randomly paired (mismatched) street-view and satellite images, or an equivalent stage-1 dataset with correspondence destroyed, is needed to establish that the alignment signal, rather than extra data exposure or map-text reading, causes the transfer. Without such a control the central claim is underdetermined.
  2. [§3.3, Appendix B] The train/test split is described only as "7:2:1" after manual checking. Since the task is address localization, it is essential that test locations (and ideally street names) are disjoint from training locations. If multiple images from the same coordinates or the same address appear in both splits, the model can memorize co-occurrences and the reported generalization is inflated. Please specify whether the split is location-disjoint (e.g., no overlap of GPS coordinates or addresses) and report results under a location-disjoint split if it is not.
  3. [§4.1] The evaluation section defines metrics AG, AJ, AM, and Asd but never specifies how free-form generation outputs are scored. Is a generated street/district considered correct by exact string match, substring containment, or normalization (case, punctuation, "Avenue" vs "Ave")? For Asd, how is "concurrently generate both street and district" parsed and counted? Different matching rules can change all numbers substantially. Please provide the exact matching/normalization procedure, ideally with examples, since the headline accuracy claims depend on it.
  4. [Tables 2-9] All reported accuracies are from single runs with no variance estimates. The central comparison (Table 2) and ablations (Tables 3-5, 8-9) involve differences of a few points; without multiple seeds or at least a stated fixed-seed protocol, it is hard to assess significance. Please provide mean±std over at least three seeds for the main baselines and methods, or justify why fine-tuning is deterministic in this setup.
  5. [§3.2, Fig. 3] In automatic label generation, the LVLM is prompted with the correct address hint, so the generated "reasons" are post-hoc rationalizations of a known answer rather than independent evidence. The paper does not report label quality (e.g., human agreement, fraction of labels that cite actual visible cues) or compare against labels generated without the hint. Given that the stage-1 training image contains street-name text, it is unclear whether the model learns visual correspondence or reads the map label. A label-quality analysis or a control without the hint would strengthen the claim.
minor comments (6)
  1. [§3.2] δ is defined as ∈ [0, 0.5], but Table 9 evaluates δ = 0.7. Reconcile the definition or change the range.
  2. [Abstract/Introduction] "Over 9% and 12%" should be stated as percentage points or relative improvement; the current phrasing is ambiguous.
  3. [§4.3] "Appendix ??" should be replaced with the actual Tokyo appendix/section reference.
  4. [Table 4] The checkmarks are not explained in the caption; please define Variants A-E (which modules are trainable) in the text or caption.
  5. [§3.2] "we have chosen the of 0.5" is a typo; should read "we have chosen δ = 0.5".
  6. [Eq. (3)] Clarify the dimensions and semantics of M and the mask "1" in the grafting equation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: AddressVLM's reported gains are out-of-sample test metrics on street-view-only images without map or address hint; the hint-based stage-1 labels are supervised training data, not a fitted evaluation input.

full rationale

The paper's central claim—that first-stage cross-view alignment tuning improves average street-plus-district accuracy (Asd) by +9.08 and +11.83 points over the stage-2-only Baseline (Table 2)—is an empirical, out-of-sample comparison. The stage-1 training labels are auto-generated with a prompt containing the correct address as a hint (Sec. 3.2, Fig. 3), and these labels are used only for LoRA fine-tuning; evaluation at stage 2 (Sec. 3.3, Fig. 6) uses street-view-only test images with no satellite map and no address hint. No equation or metric in the paper reduces the reported test accuracy to the hint, to a fitted constant, or to the training labels by construction. The ablations (Tables 3–4, Fig. 7) vary grafting ratio, map labels, trainable modules, and data density, and report independent test-set metrics rather than recovering fitted values. The one self-citation that is load-bearing in a data sense is AddressCLIP [4], whose Pitts-IAL and SF-Base-IAL location data underpin the new VQA datasets; that reuse is grounded in external datasets [12, 13], and neither the baseline nor the conclusion is defined in terms of AddressCLIP's own numbers. GeoReasoner [9] is an external baseline reproduced with a parallel hint-based reasoning pipeline (Appendix C). The residual concern—that the +9/+12 gain may partly reflect extra training data or map-text OCR shortcuts rather than a transferable spatial prior—is an experimental-design/soundness question (a random-pairing control would settle it), not a definitional circularity. The derivation chain is therefore self-contained: predictions are genuinely out-of-sample and no result is forced by its own inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities that need independent falsifiable handles; the contributions are a training recipe, a grafting input construction, and two datasets. The key hand-tuned design choices are the overlap ratio and the road-label overlay, both selected by ablation.

free parameters (3)
  • Grafting overlap ratio δ = 0.5
    Chosen by ablation (Tables 3 and 9); smaller ratios obscure street-view detail, larger ratios hide map context.
  • Satellite map road-label overlay = on
    Ablation shows removing street-name labels degrades accuracy, so the map with OCR labels is a hand-selected input component.
  • LoRA rank = 128
    A high-rank LoRA setting used for both stages; it is a chosen hyperparameter rather than a fitted scientific constant.
assumptions (5)
  • domain assumption Satellite images provide a macro-level, perspective-invariant prior that can connect sparse street-view images into a global street distribution.
    Invoked in Sec. 3.2 as the motivation for cross-view alignment tuning; if false, stage 1 cannot supply the claimed prior.
  • domain assumption Grafting a downscaled street view into a 336x336 satellite map preserves enough street-view detail and map layout for the LVLM to learn cross-view matching.
    Central input construction in Eq. (3) and Fig. 3; the ablation on δ tests only specific ratios, not the validity of grafting itself.
  • domain assumption Labels generated by an off-the-shelf LVLM from a prompt containing the correct address hint are accurate enough to train the alignment stage.
    Sec. 3.2 automatic label generation; incorrect generated rationales would train the model with noisy targets.
  • domain assumption The train/test split separates locations so that test images do not come from locations seen in training.
    Sec. 3.3 says the data is split 7:2:1 but does not state whether the split is by location, image, or viewpoint; overlapping locations would inflate the reported gains.
  • domain assumption The adapted GeoReasoner reproduction with emulated reasoning labels is representative of the original GeoReasoner method.
    Appendix C changes the first-stage data source and uses the same VQA data for stage 2; this may not match the published GeoReasoner pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AddressVLM: Cross-view Alignment Tuning for Image Address Localization using Large Vision-Language Models." pith.science (2026). https://pith.science/paper/IMJ27EZB

@misc{pith2026250810667,
  author       = {Pith},
  title        = {Pith review of: AddressVLM: Cross-view Alignment Tuning for Image Address Localization using Large Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IMJ27EZB}},
  note         = {Machine review of arXiv:2508.10667}
}
read the original abstract

Large visual language models (LVLMs) have demonstrated impressive performance in coarse-grained geo-localization at the country or city level, but they struggle with fine-grained street-level localization within urban areas. In this paper, we explore integrating city-wide address localization capabilities into LVLMs, facilitating flexible address-related question answering using street-view images. A key challenge is that the street-view visual question-and-answer (VQA) data provides only microscopic visual cues, leading to subpar performance in fine-tuned models. To tackle this issue, we incorporate perspective-invariant satellite images as macro cues and propose cross-view alignment tuning including a satellite-view and street-view image grafting mechanism, along with an automatic label generation mechanism. Then LVLM's global understanding of street distribution is enhanced through cross-view matching. Our proposed model, named AddressVLM, consists of two-stage training protocols: cross-view alignment tuning and address localization tuning. Furthermore, we have constructed two street-view VQA datasets based on image address localization datasets from Pittsburgh and San Francisco. Qualitative and quantitative evaluations demonstrate that AddressVLM outperforms counterpart LVLMs by over 9% and 12% in average address localization accuracy on these two datasets, respectively.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. CarbonCLIP: Enhance Carbon Prediction from Satellite Imagery via Integrated Street-View Semantics and Temporal Context Training

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A dual-branch contrastive learning framework distills street-view semantics and temporal context into satellite representations, improving monthly carbon emission prediction using only satellite imagery at inference.

Reference graph

Works this paper leans on

37 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: CVPR, pp

    Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: CVPR, pp. 5297–5307 (2016)

  2. [2]

    In: CVPR, pp

    Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N.: Transvpr: Transformer-based place recognition with multi-level attention aggregation. In: CVPR, pp. 13648– 13657 (2022)

  3. [3]

    In: W ACV, pp

    Ali-Bey, A., Chaib-Draa, B., Giguere, P.: Mixvpr: Feature mixing for visual place recognition. In: W ACV, pp. 2998–3007 (2023) 22

  4. [4]

    In: ECCV (2024)

    Xu, S., Zhang, C., Fan, L., Meng, G., Xiang, S., Ye, J.: Addressclip: Empowering vision-language models for city-wide image address localization. In: ECCV (2024)

  5. [5]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

  6. [6]

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. (2023)

  7. [7]

    In: NeurIPS (2024)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2024)

  8. [8]

    arXiv preprint arXiv:2309.17421 9(1), 1 (2023)

    Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.-C., Liu, Z., Wang, L.: The dawn of lmms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)

Show all 37 references
  1. [9]

    In: ICML (2024)

    Li, L., Ye, Y., Jiang, B., Zeng, W.: Georeasoner: Geo-localization with reasoning in street views using a large vision-language model. In: ICML (2024)

  2. [10]

    arXiv preprint arXiv:2106.09685 (2021)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)

  3. [11]

    arXiv preprint arXiv:2406.09722 (2024)

    Durgam, A., Paheding, S., Dhiman, V., Devabhaktuni, V.: Cross-view geo- localization: a survey. arXiv preprint arXiv:2406.09722 (2024)

  4. [12]

    In: CVPR, pp

    Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: CVPR, pp. 883–890 (2013)

  5. [13]

    In: CVPR, pp

    Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large- scale applications. In: CVPR, pp. 4878–4888 (2022)

  6. [14]

    Pattern Recognition 113, 107760 (2021)

    Zhang, X., Wang, L., Su, Y.: Visual place recognition: A survey from deep learning perspective. Pattern Recognition 113, 107760 (2021)

  7. [15]

    In: CVPR, pp

    Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multi- scale fusion of locally-global descriptors for place recognition. In: CVPR, pp. 14141–14152 (2021)

  8. [16]

    IEEE Robotics and Automation Letters (2023)

    Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (2023)

  9. [17]

    In: ECCV, pp

    Seo, P.H., Weyand, T., Sim, J., Han, B.: Cplanet: Enhancing image geolocalization by combinatorial partitioning of maps. In: ECCV, pp. 536–551 (2018)

  10. [18]

    In: ECCV, pp

    Pramanick, S., Nowara, E.M., Gleason, J., Castillo, C.D., Chellappa, R.: Where 23 in the world is this image? transformer-based geo-localization in the wild. In: ECCV, pp. 196–215 (2022)

  11. [19]

    In: CVPR, pp

    Clark, B., Kerrigan, A., Kulkarni, P.P., Cepeda, V.V., Shah, M.: Where we are and what we’re looking at: Query based worldwide image geo-localization using hierarchies and scenes. In: CVPR, pp. 23182–23190 (2023)

  12. [20]

    In: ICCV, pp

    Trivigno, G., Berton, G., Aragon, J., Caputo, B., Masone, C.: Divide&classify: Fine-grained classification for city-wide visual geo-localization. In: ICCV, pp. 11142–11152 (2023)

  13. [21]

    In: ICML, pp

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML, pp. 8748–8763 (2021)

  14. [22]

    arXiv preprint arXiv:2302.00275 (2023)

    Haas, L., Alberti, S., Skreta, M.: Learning generalized zero-shot learners for open- domain image geolocalization. arXiv preprint arXiv:2302.00275 (2023)

  15. [23]

    arXiv preprint arXiv:2309.16020 (2023)

    Cepeda, V.V., Nayak, G.K., Shah, M.: Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization. arXiv preprint arXiv:2309.16020 (2023)

  16. [24]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozi` ere, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

  17. [25]

    arXiv preprint arXiv:2310.06825 (2023)

    Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)

  18. [26]

    arXiv preprint arXiv:2407.10671 (2024)

    Yang, A., Yang, B., Hui, B., Zheng, B., Yu, B., Zhou, C., Li, C., Li, C., Liu, D., Huang, F., et al.: Qwen2 technical report. arXiv preprint arXiv:2407.10671 (2024)

  19. [27]

    arXiv preprint arXiv:2404.14219 (2024)

    Abdin, M., Jacobs, S.A., Awan, A.A., Aneja, J., Awadallah, A., Awadalla, H., Bach, N., Bahree, A., Bakhtiari, A., Behl, H., et al.: Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219 (2024)

  20. [28]

    arXiv preprint arXiv:2312.11805 (2023)

    Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

  21. [29]

    In: CVPR, pp

    Zhu, S., Yang, T., Chen, C.: Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In: CVPR, pp. 3640–3649 (2021)

  22. [30]

    PloS one 18(3), 0283672 (2023)

    Ghanem, A., Abdelhay, A., Salah, N.E., Nour Eldeen, A., Elhenawy, M., Masoud, 24 M., Hassan, A.M., Hassan, A.A.: Leveraging cross-view geo-localization with ensemble learning and temporal awareness. PloS one 18(3), 0283672 (2023)

  23. [31]

    IEEE Transactions on Circuits and Systems for Video Technology 32(2), 867–879 (2021)

    Wang, T., Zheng, Z., Yan, C., Zhang, J., Sun, Y., Zheng, B., Yang, Y.: Each part matters: Local patterns facilitate cross-view geo-localization. IEEE Transactions on Circuits and Systems for Video Technology 32(2), 867–879 (2021)

  24. [32]

    arXiv preprint arXiv:2403.13965 (2024)

    Mi, L., Xu, C., Castillo-Navarro, J., Montariol, S., Yang, W., Bosselut, A., Tuia, D.: Congeo: Robust cross-view geo-localization across ground view variations. arXiv preprint arXiv:2403.13965 (2024)

  25. [33]

    arXiv preprint arXiv:2403.16831 (2024)

    Hao, X., Chen, W., Yan, Y., Zhong, S., Wang, K., Wen, Q., Liang, Y.: Urban- vlp: A multi-granularity vision-language pre-trained foundation model for urban indicator prediction. arXiv preprint arXiv:2403.16831 (2024)

  26. [34]

    https:// github.com/InternLM/xtuner (2023)

    Contributors, X.: XTuner: A Toolkit for Efficiently Fine-tuning LLM. https:// github.com/InternLM/xtuner (2023)

  27. [35]

    In: CVPR, pp

    Lin, J., Yin, H., Ping, W., Molchanov, P., Shoeybi, M., Han, S.: Vila: On pre- training for visual language models. In: CVPR, pp. 26689–26699 (2024)

  28. [36]

    Claude: The claude 3 model family: Opus, sonnet, haiku. (2024). https://api.semanticscholar.org/CorpusID:268232499

  29. [37]

    Qwen: Qwen2-vl (2024) 25

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.