Pith. sign in

REVIEW 4 major objections 4 minor 46 references

Few-Shot Vision-Language Reasoning for Satellite Imagery via Verifiable Rewards

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read One rewarded example gives satellite VLMs double-digit gains

desk verdict First real transfer of 1-shot RLVR to remote sensing VLMs, but the headline gains rest on test-set checkpoint selection and need a clean evaluation before they are believable. read the letter →

arxiv 2507.21745 v2 pith:L55HEBMH submitted 2025-07-29 cs.CV

Aybora Koksal , A. Aydin Alatan This is my paper · ORCID
classification cs.CV
keywords few-shotlearningreinforcementwithverifiablerewardsGRPOvision-languagemodelsremotesensingsatelliteimageryvisualgroundingrule-based
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

The paper sets out to show that a compact vision-language model can be turned into a remote-sensing reasoning specialist without any caption data, using only reinforcement learning with rule-based rewards. Training with as few as one curated example yields double-digit accuracy gains over the unmodified base model across classification, visual question answering, and visual grounding benchmarks. Scaling the training set to 128 examples closes the gap to a model trained on 2,000 annotated images, and even surpasses it on a general remote-sensing knowledge benchmark. If true, this makes heavy annotation pipelines unnecessary for specialized vision-language domains, replacing them with a handful of reward-checkable cases.

What carries the argument

The load-bearing mechanism is GRPO (Group Relative Policy Optimization), a policy-gradient method in which rewards are normalized relative to a group of sampled responses, combined with rule-based verifiable rewards: a format-compliance reward for the <reasoning>/<answer> tags, a binary accuracy reward for closed-ended questions, and a quantized IoU reward for bounding-box grounding (1 for IoU ≥ 0.7, the raw IoU for 0.4–0.7, and 0 below 0.4). Training uses a KL-divergence term with a small coefficient β = 0.001 to preserve language quality, and a deliberately short system prompt that elicits reasoning. The base model is the 2B-parameter Qwen2-VL, and the few-shot examples are sampled randomly from the VHM-Instruct dataset.

What would settle it

Fix the checkpoint rule in advance—for example, stop at a preset step count or select by a held-out validation set—then retrain the 1-, 32-, and 128-shot models and compare accuracy on the same benchmarks; if the gains over the base model largely vanish, the discrepancy shows that test-set checkpoint selection was the load-bearing factor.

Watch

Extended reading notes

Core claim

On the paper's terms, the discovery is that the '1-shot RLVR' phenomenon previously observed in text-only language models transfers to vision-language models operating on satellite imagery. A single curated example, optimized with GRPO under a binary or quantized-IoU reward, unlocks latent reasoning in Qwen2-VL-2B, and performance scales smoothly with the number of examples. The authors find that going from one to 128 examples produces stable, broadly transferable gains, matches a 2000-image fully supervised RL baseline on most metrics, and that overfitting in the extreme one-shot regime is mild and confined to the evaluation split that shares its metric with the training example.

Load-bearing premise

The reported results come from picking each model's best checkpoint on the test set itself, so the measured gains could be inflated by test-set overfitting; without a held-out validation split the few-shot advantage is uncertain.

Editorial extensions

If this is right

  • No caption supervision: domain-specialist VLMs can be aligned using only programmatically checkable rewards, eliminating the need for expensive human or LLM-generated captions.
  • Extreme data efficiency: roughly 32–128 curated examples deliver most of the achievable gain, suggesting a practical sweet spot for data-scarce fields.
  • A 128-example model can match or beat a 2000-example supervised RL system, indicating that the number of reward-checkable examples, not raw data volume, drives the alignment.
  • The one-shot regime carries mild, task-specific overfitting; using 8–32 diverse examples removes this overfitting while retaining strong cross-task generalization.
  • A larger 7B model also benefits from 16-example RLVR, though the relative gain shrinks because its zero-shot performance is already higher.

Reading between the lines

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

  • If the best-checkpoint-per-test-set selection (Section 4.2) is replaced by an a-priori training schedule or a held-out validation split, the headline few-shot gains may shrink; a fixed-protocol replication would settle how much of the effect is test-set optimism.
  • The quantized IoU reward suggests a natural extension to oriented bounding boxes or segmentation masks, where the same coarse-to-fine reward shaping could be applied.
  • Since the paper uses random sampling, deliberately selecting examples with high historical response variance—as suggested by the text-only RLVR literature—could push one-shot gains further, giving a testable variant of the recipe.
  • The prompt-ablation result implies that the technique is sensitive to how the base model was pretrained; re-running RLVR with a prompt matched to a different VLM's pretraining distribution would clarify how portable the recipe is.
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

4 major / 4 minor

Summary. This paper proposes a few-shot reinforcement learning with verifiable rewards (RLVR) framework for satellite imagery. Using GRPO with rule-based binary or IoU rewards on Qwen2-VL-2B, the authors train on 1–128 curated examples and evaluate on remote sensing classification, VQA, visual grounding, and LHRS-Bench. The central claim is that even a single curated example yields double-digit improvements over the untouched base model, and that 128 examples close the gap to a 2000-image RLVR baseline, without any caption supervision.

Significance. If the claims survive rigorous evaluation, the paper offers a highly practical, data-efficient recipe for adapting compact VLMs to specialized domains. The method is well motivated by prior text-only 1-shot RLVR work and the paper includes useful ablations on KL-weight and prompt design, plus comparisons to large-scale RS VLMs. The promise of open-sourcing code and data is a further strength. However, the evaluation protocol currently overstates the quantitative results, so the central empirical claims need to be re-established under a statistically sound protocol before the paper's contribution can be fully credited.

major comments (4)
  1. [Section 4.2, Table 2] The evaluation protocol selects the best-performing checkpoint on each test set: "We evaluate all checkpoints of the trained models and report the ones that achieve the best results." Since the Step column shows different checkpoint numbers across models and likely across metrics, the reported accuracies are maxima over the training trajectory rather than the performance of a single fixed policy. Meanwhile, the Base and VHM-RL rows are single fixed checkpoints, making the comparison asymmetric. Section 4.1's rule of extending training "if the performance continued to improve" appears to use the same test signal. This selection bias inflates the magnitude of the few-shot gains and undermines the "closing the gap" claim. Please report results for a fixed checkpoint (e.g., the last one) or a checkpoint selected on a validation split, and ideally provide mean±std over multiple seeds.
  2. [Abstract and Conclusion] The abstract states that "Scaling to 128 examples matches or exceeds models trained on thousands of annotated samples," but Table 3 does not support this. The 128-shot model (CLS 63.8, VQA 63.9, VG 51.7) is far below GeoChat (67.3, 83.5, 19.7), VHM (85.6, 83.0, 55.9), and TinyRS-R1 (85.6, 76.0, 74.9) on most metrics. It only surpasses VHM-RL-2000 on LHRS-Bench, and VHM-RL-2000 is a 2000-image RLVR baseline, not a model trained on "thousands of annotated samples." Please rephrase the claim to match the evidence, e.g., "approaches or exceeds a 2000-image RLVR baseline on several benchmarks."
  3. [Section 3.5 and Table 2] It is unclear which prompt was used for the Base row in Table 2. Section 3.5 describes modifying all question formats to include "Make your chain of thought reasoning..." and notes that the base model then "sometimes begins with reasoning". If the Base row was evaluated with the original question formats, part of the observed gain could come from the prompt change rather than from RLVR. If the Base row was evaluated with the modified prompt, that should be stated. The paper should report both prompt variants for the base model to isolate the effect of RLVR.
  4. [Section 3.4 and Table 2] The few-shot training sets are claimed to be randomly sampled, but no random seeds or exact training examples are provided, and each configuration is run only once. Given the apparent sensitivity to the chosen example (e.g., π1V versus π1G produce very different classification results), the reported numbers may be highly dependent on the particular random draw. To support the "smooth scaling" and "double-digit gains" claims, please report results averaged over at least three random draws of the training set with different seeds, or make the exact examples available so the experiments are reproducible.
minor comments (4)
  1. [Section 4.1] The phrase "corresponding to 1000 epochs with training sets of 128 examples" is confusing because one GRPO step with a batch of 128 and 128 distinct examples equals one epoch; 1000 steps therefore means 1000 epochs, which seems extreme. Please clarify how examples are duplicated and what is meant by "epoch."
  2. [Table 1] The symbols ✓ and × are not defined in the table caption. Please add an explicit legend.
  3. [Section 3.2] For the quantized IoU reward, the text says "an exact base reward is given for values between 0.4 and 0.7." Please state explicitly that this means the raw IoU value is used as the reward in that range.
  4. [Section 5.3] The sentence "Although the overall performance still improves significantly" is not supported for CLS, which drops by 1.0 point in the 7B model. Please soften the wording to reflect the metric-specific changes.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity; the few-shot RLVR gains are empirical and self-contained, with only minor non-load-bearing self-citations.

full rationale

The paper's central claim—that few-shot RLVR with rule-based rewards unlocks remote-sensing reasoning—is an empirical result, not a derivation. Training uses GRPO with binary/quantized-IoU rewards, and the reported accuracies are measured on benchmark suites; no equation is fitted to the test numbers, and no fitted parameter is renamed as a prediction. The quantized IoU reward is borrowed from the authors' own TinyRS-R1 [20], but [20] is an empirical prior study rather than a theorem, and the present paper's own ablations (beta, prompt, model size) and the 1-to-128-shot scaling curves provide independent content; the central result does not reduce to the citation. The only notable evaluation concern is Section 4.2's best-checkpoint selection ('We evaluate all checkpoints of the trained models and report the ones that achieve the best results'), which is a statistical validity/selection-bias issue, not circularity: the reported accuracies are not defined as the training rewards, and no uniqueness theorem from the authors' prior work is invoked to force the method choice. Verdict: no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities and fits no equations to data; its main load-bearing assumptions are the latent capability of the base VLM and the validity of the evaluation protocol.

free parameters (3)
  • KL penalty coefficient beta = 0.001
    Set to 0.001 based on Wang et al. [44] and confirmed by ablation (Table 4); the default 0.04 leads to instability and lower accuracy.
  • IoU reward quantization thresholds = 0.4, 0.7
    Reward of 1 for IoU >= 0.7, base IoU for 0.4-0.7, 0 below 0.4; adopted from the authors' prior TinyRS-R1 [20] without re-optimization.
  • Group size (responses per image) = 4
    Chosen instead of the default 8 due to H100 memory constraints; may affect stability but not the central claim.
assumptions (4)
  • domain assumption The base Qwen2-VL-2B model possesses latent remote sensing reasoning that can be elicited by RLVR without caption pretraining.
    This premise, taken from Wang et al. [44] for LLMs, underlies the entire method (Section 3.1 and Introduction). If base VLMs lack such latent ability, the few-shot gains cannot occur.
  • domain assumption GRPO with rule-based rewards provides an unbiased policy-gradient signal.
    The paper relies on GRPO's group-relative advantages to shape responses; this is a standard RL method, but the assumption that binary/IoU rewards correlate with desired behavior is a domain assumption.
  • domain assumption The evaluation benchmarks are reliable measures of remote sensing reasoning.
    The paper measures success via standard benchmarks (AID, RSVQA, RSVG-DIOR, LHRS-Bench), assuming these capture the intended capabilities.
  • ad hoc to paper Selecting the best checkpoint on each test set yields a valid estimate of generalization.
    This is not justified and biases results upward; it is a load-bearing assumption for the reported numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Few-Shot Vision-Language Reasoning for Satellite Imagery via Verifiable Rewards." pith.science (2026). https://pith.science/paper/L55HEBMH

@misc{pith2026250721745,
  author       = {Pith},
  title        = {Pith review of: Few-Shot Vision-Language Reasoning for Satellite Imagery via Verifiable Rewards},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L55HEBMH}},
  note         = {Machine review of arXiv:2507.21745}
}
read the original abstract

Recent advances in large language and vision-language models have enabled strong reasoning capabilities, yet they remain impractical for specialized domains like remote sensing, where annotated data is scarce and expensive. We present the first few-shot reinforcement learning with verifiable reward (RLVR) framework for satellite imagery that eliminates the need for caption supervision--relying solely on lightweight, rule-based binary or IoU-based rewards. Adapting the "1-shot RLVR" paradigm from language models to vision-language models, we employ policy-gradient optimization with as few as one curated example to align model outputs for satellite reasoning tasks. Comprehensive experiments across multiple remote sensing benchmarks--including classification, visual question answering, and grounding--show that even a single example yields substantial improvements over the base model. Scaling to 128 examples matches or exceeds models trained on thousands of annotated samples. While the extreme one-shot setting can induce mild, task-specific overfitting, our approach consistently demonstrates robust generalization and efficiency across diverse tasks. Further, we find that prompt design and loss weighting significantly influence training stability and final accuracy. Our method enables cost-effective and data-efficient development of domain-specialist vision-language reasoning models, offering a pragmatic recipe for data-scarce fields: start from a compact VLM, curate a handful of reward-checkable cases, and train via RLVR.

Figures

Figures reproduced from arXiv: 2507.21745 by the authors.

Figure 1
Figure 1. We show that few-shot RLVR achieves strong general [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Qualitative outputs of our proposed method, trained with [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Change of completion length during training for the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 22 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

    arXiv

  2. [2]

    Open-r1v

    Deep Agent. Open-r1v. https://github.com/Deep-Agent/R1- V , 2025. 3

    work page 2025

  3. [3]

    Qwen technical report,

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xi- aodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, Tianhang Zhu, et al. Qwen technical report,

    work page

  4. [4]

    Qwen-vl: A frontier large vision-language model with versatile abilities

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2

    arXiv 2023

  5. [5]

    Rs-llava: A large vision-language model for joint captioning and question an- swering in remote sensing imagery

    Yakoub Bazi, Laila Bashmal, Mohamad Mahmoud Al Rah- hal, Riccardo Ricci, and Farid Melgani. Rs-llava: A large vision-language model for joint captioning and question an- swering in remote sensing imagery. Remote Sensing, 16(9): 1477, 2024. 1, 2, 3

    work page 2024

  6. [6]

    Paligemma: A versatile 3b vlm for trans- fer

    Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexan- der Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for trans- fer. arXiv preprint arXiv:2407.07726, 2024. 2

    arXiv 2024

  7. [7]

    Lan- guage models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Lan- guage models are few-shot learners. Advances in neural in- formation processing systems, 33:1877–1901, 2020. 1, 2

    work page 1901

  8. [8]

    Fine- tuned’small’llms (still) significantly outperform zero-shot generative ai models in text classification

    Martin Juan José Bucher and Marco Martini. Fine- tuned’small’llms (still) significantly outperform zero-shot generative ai models in text classification. arXiv preprint arXiv:2406.08660, 2024. 2

    arXiv 2024

Show all 46 references
  1. [9]

    Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks. In Pro- ceedings of the IEEE/CVF Conference on Computer...

    2024

  2. [10]

    Patch n’pack: Navit, a vision transformer for any aspect ratio and resolution

    Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdul- mohsin, et al. Patch n’pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neu...

    2023

  3. [11]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2

    2018 arXiv

  4. [12]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint a...

    2010 arXiv

  5. [13]

    Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars

    Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh, Nathan Lile, and Noah D Goodman. Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars. arXiv preprint arXiv:2503.01307, 2025. 2

    2025 arXiv

  6. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 2, 3, 4

    2025 arXiv

  7. [15]

    Lora: Low- rank adaptation of large language models

    Edward Hu, Yining Shen, Peter Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Lijuan Wang, and Weizhu Chen. Lora: Low- rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2022. 3

    2022 arXiv

  8. [16]

    A remote sensing vision language model and benchmark

    Y Hu, J Yuan, C Wen, X Lu, and X Li. A remote sensing vision language model and benchmark. arxiv 2023. arXiv preprint arXiv:2307.15266, 2023. 1, 3

    2023 arXiv

  9. [17]

    Huggingface. Open-r1. https://github.com/huggingface/open-r1, 2025. 3

    2025

  10. [18]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel- man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli- hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 2

    2024 arXiv

  11. [19]

    Milchat: Introduc- ing chain of thought reasoning and grpo to a multimodal small language model for remote sensing

    Aybora Koksal and A Aydin Alatan. Milchat: Introduc- ing chain of thought reasoning and grpo to a multimodal small language model for remote sensing. arXiv preprint arXiv:2505.07984, 2025. 2, 3

    2025

  12. [20]

    Tinyrs-r1: Com- pact multimodal language model for remote sensing

    Aybora Koksal and A Aydin Alatan. Tinyrs-r1: Com- pact multimodal language model for remote sensing. arXiv preprint arXiv:2505.12099, 2025. 1, 2, 3, 4, 7

    2025

  13. [21]

    Geochat: Grounded large vision-language model for remote sensing

    Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad Shahbaz Khan. Geochat: Grounded large vision-language model for remote sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27831– 2784...

    2024

  14. [22]

    Lhrs- bot-nova: Improved multimodal large language model for re- mote sensing vision-language interpretation

    Zhenshi Li, Dilxat Muhtar, Feng Gu, Xueliang Zhang, Pengfeng Xiao, Guangjun He, and Xiaoxiang Zhu. Lhrs- bot-nova: Improved multimodal large language model for re- mote sensing vision-language interpretation. arXiv preprint arXiv:2411.09301, 2024. 2, 3

    2024 arXiv

  15. [23]

    Visual instruction tuning, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1

    2023

  16. [24]

    Understanding r1-zero-like training: A critical perspective

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective. arXiv preprint arXiv:2503.20783, 2025. 2

    2025 arXiv

  17. [25]

    Open-r1 multimodal

    LMMs-Lab. Open-r1 multimodal. https://github.com/EvolvingLMMs-Lab/open-r1- multimodal, 2025. 3, 5

    2025

  18. [26]

    Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model

    Dilxat Muhtar, Zhenshi Li, Feng Gu, Xueliang Zhang, and Pengfeng Xiao. Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model. In European Conference on Computer Vision, pages 440–457. Springer, 2024. 1, 2, 3, 5

    2024

  19. [27]

    Quality-driven curation of remote sensing vision-language data via learned scoring models

    Dilxat Muhtar, Enzhuo Zhang, Zhenshi Li, Feng Gu, Yanglangxing He, Pengfeng Xiao, and Xueliang Zhang. Quality-driven curation of remote sensing vision-language data via learned scoring models. arXiv preprint arXiv:2503.00743, 2025. 2, 3, 7

    2025

  20. [28]

    Gpt-3.5 turbo fine-tuning and api updates, 2023

    OpenAI. Gpt-3.5 turbo fine-tuning and api updates, 2023. 2

    2023

  21. [29]

    Gpt-4v(ision) system card, 2023

    OpenAI. Gpt-4v(ision) system card, 2023. 1, 2

    2023

  22. [30]

    Introducing openai o1-preview, 2024

    OpenAI. Introducing openai o1-preview, 2024. 1, 2

    2024

  23. [31]

    Openai o3-mini, 2025

    OpenAI. Openai o3-mini, 2025. 1, 2

    2025

  24. [32]

    Vhm: Versatile and honest vision language model for remote sensing image analysis

    Chao Pang, Xingxing Weng, Jiang Wu, Jiayu Li, Yi Liu, Ji- axing Sun, Weijia Li, Shuai Wang, Litong Feng, Gui-Song Xia, et al. Vhm: Versatile and honest vision language model for remote sensing image analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, pa...

    2025

  25. [33]

    Improving language understanding by gen- erative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by gen- erative pre-training. Technical Report, 2018. 1, 2

    2018

  26. [34]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. arXiv preprint arXiv:2103.00020, 2021. 2

    2021 arXiv

  27. [35]

    Rethinking reflection in pre-training

    Darsh J Shah, Peter Rushton, Somanshu Singla, Mohit Par- mar, Kurt Smith, Yash Vanjani, Ashish Vaswani, Adarsh Chaluvaraju, Andrew Hojel, Andrew Ma, et al. Rethinking reflection in pre-training. arXiv preprint arXiv:2504.04022,

    arXiv

  28. [36]

    Deepseekmath: Pushing the limits of mathe- matical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathe- matical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 2, 3

    2024 arXiv

  29. [37]

    Paligemma 2: A family of versatile vlms for transfer

    Andreas Steiner, André Susano Pinto, Michael Tschannen, Daniel Keysers, Xiao Wang, Yonatan Bitton, Alexey Grit- senko, Matthias Minderer, Anthony Sherbondy, Shangbang Long, et al. Paligemma 2: A family of versatile vlms for transfer. arXiv preprint arXiv:2412.03555, 2024. 2

    2024 arXiv

  30. [38]

    Gemma: Open models based on gemini research and technology,

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Noah Fiedel, Evan Senter, Alek Andreev, Kathleen Kenealy, et al. Gemma: Open models based on gemini research and technology,

  31. [39]

    Internlm: A multilingual language model with progressively enhanced capabilities

    InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities. https : / / github.com/InternLM/InternLM- techreport ,

  32. [40]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599, 2025. 1

    2025 arXiv

  33. [41]

    Llama: Open and efficient foundation lan- guage models, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aure- lien Rodriguez, Armand Joulin, Edouard Grave, and Guil- laume Lample. Llama: Open and efficient foundation lan-...

    2023

  34. [42]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998–6008, 2017. 2

    2017

  35. [43]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 3

    2024 arXiv

  36. [44]

    Reinforcement learning for rea- soning in large language models with one training example

    Yiping Wang, Qing Yang, Zhiyuan Zeng, Liliang Ren, Liyuan Liu, Baolin Peng, Hao Cheng, Xuehai He, Kuan Wang, Jianfeng Gao, et al. Reinforcement learning for rea- soning in large language models with one training example. arXiv preprint arXiv:2504.20571, 2025. 2, 3, 4, 5, 7, 8

    2025 arXiv

  37. [45]

    Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? arXiv preprint arXiv:2504.13837, 2025

    Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Shiji Song, and Gao Huang. Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? arXiv preprint arXiv:2504.13837, 2025. 2

    2025 arXiv

  38. [46]

    Earthgpt: A universal multi-modal large lan- guage model for multi-sensor image comprehension in re- mote sensing domain

    Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multi-modal large lan- guage model for multi-sensor image comprehension in re- mote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2, 3

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.