Pith. sign in

REVIEW 3 major objections 4 minor 73 references

A general vision-language model, retrained on diverse remote sensing data, can match or beat specialized models without any new architecture.

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

2026-08-01 21:51 UTC pith:KBYQH2AP

load-bearing objection Real benchmark gains and a useful scaling study, but the headline claim that data scale beats architecture is not supported by the paper's own analyses. the 3 major comments →

arxiv 2607.15942 v1 pith:KBYQH2AP submitted 2026-07-17 cs.CV cs.LG

More with Less: a Large Scale Remote Sensing VLM with a Simple Recipe

Stefan Maria Ailuro , Mario Markov , Mohammad Mahdi , Luc Van Gool , Danda Pani Paudel (INSAIT , Sofia University "St. Kliment Ohridski") This is my paper
classification cs.CV cs.LG
keywords remote sensingvision-language modelmulti-task reinforcement learningzero-shot generalizationdata scalingsegmentationdetectionvisual question answering
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 paper argues that a strong general-purpose vision-language model, trained on a large and diverse mix of remote sensing tasks through multi-task reinforcement learning, can match or beat specialized remote sensing models without introducing any new architecture. It introduces MLRS, a model that answers questions in text or calls a segmentation tool, and reports top-two zero-shot results on 10 of 12 benchmarks, including a 7.6x detection gain on one benchmark. The authors claim the driver is data scale and per-task data diversity rather than architectural novelty, supported by learning curves where source-rich domains keep improving while source-poor domains plateau or decline. The paper also reports honest failures on temporal reasoning, where gains peak early and then fall below the base model. A sympathetic reader would care because the result redirects effort from designing new encoders and fusion modules toward scaling diverse supervision.

Core claim

MLRS is a remote sensing vision-language model built without modifying a general vision-language backbone: a single language policy either answers directly in text or emits a structured segmentation call that is parsed into prompts for a segmentation model, with both components trained jointly using reinforcement learning with adaptive per-task rewards. The paper's central claim is that this simple recipe is sufficient: given enough diverse remote sensing supervision, a general VLM reaches competitive or state-of-the-art zero-shot performance across optical, SAR, multi-temporal, multi-view, and ultra-high-resolution benchmarks. The authors further claim that performance improves consistently

What carries the argument

A single language policy paired with a segmentation tool and trained by multi-task reinforcement learning with adaptive rewards. The policy is prompted to either answer in text or emit a structured localization call (a noun phrase plus bounding boxes) that is converted into prompts for a segmentation model; both the policy and the tool are updated jointly. The reward has a shared form combining a format-validity term with a task-specific score: set IoU for multiple-choice VQA, embedding cosine similarity for free-form answers and captions, box IoU for detection, and mask IoU for segmentation. This one interface covers all tasks and input types without task-specific encoders or fusion modules

Load-bearing premise

The load-bearing premise is that out-of-distribution gains are caused by per-task training-source diversity rather than by the base model simply having more room to improve; the paper's own analysis says the two are correlated and the seven-domain test is underpowered.

What would settle it

Train the same base model on the same task composition but replace a source-rich domain's data with an equal number of samples from a single source, holding base score roughly constant: if out-of-distribution gain does not drop, the diversity claim is wrong. Alternatively, if a single-source domain with a low base score (for example, detection from one dataset) shows the same large gains, headroom alone explains the results.

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

If this is right

  • Specialized remote sensing encoders, alignment modules, and task-specific fusion mechanisms may be unnecessary; a strong general VLM plus diverse multi-task reinforcement learning can reach state-of-the-art zero-shot results.
  • Scaling diverse training data yields consistent out-of-distribution gains on most task families: detection precision on GEOBench-VLM more than doubles, XLRS-Bench detection improves 7.6x, and segmentation IoU roughly doubles on held-out segmentation benchmarks.
  • Domains backed by few independent data sources—temporal VQA and multi-view VQA—plateau or decline with more training, indicating that source diversity, not raw sample count, is the limiting factor.
  • Temporal reasoning can regress during multi-task training: on VLRS-Bench and RSHR-Bench the final model falls below the base model, a failure the paper attributes to overfitting on a narrow temporal distribution.
  • Fine-tuning the segmentation tool during reinforcement learning helps most segmentation benchmarks but hurts zero-shot generalization on GEOBench-VLM, so adapting tool-specific components requires caution.

Where Pith is reading between the lines

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

  • If the diversity conclusion holds, the practical recipe for the field is to invest in gathering many independent annotation sources per task type rather than in bespoke architectures; benchmark suites should begin reporting per-domain source diversity.
  • The early-peak-then-decline trajectory on temporal reasoning suggests a forgetting or overfitting dynamic that could be tested directly by adding diverse temporal data or replay during training—this is an inference, not a claim the paper makes.
  • The paper's own statistics show that source diversity and base-model headroom are correlated and the domain-level test is underpowered, so the central scaling conclusion could partly be an artifact of regression to the mean; a decisive test would vary source count while holding task type and base model fixed.
  • The large detection gains may largely reflect the base model's initially poor detection ability, implying that the 'data scale' story is about unlocking existing capabilities of a strong base model rather than about diversity per se.

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 introduces MLRS, a remote sensing vision-language model built from InternVL3.5-8B as a language policy, SAM3 as an external localization tool, and GRTO reinforcement learning with adaptive task rewards. The model is trained on 80k curated samples covering MC-VQA, free-form VQA, captioning, detection, and segmentation across RGB, SAR, multi-temporal, multi-view, and ultra-high-resolution inputs. The paper reports zero-shot results on 12 OOD benchmarks, claiming first or second place on 10 and first on 7, and presents scaling experiments showing OOD gains growing with training data, with per-domain gains correlated with the number of training sources. The appendix contains a detailed statistical analysis of these trends, including permutation tests, leave-one-domain-out checks, cluster-robust regressions, and seed/scale robustness studies.

Significance. If the central claims hold, the paper would be significant: it would show that a general-purpose VLM plus a simple tool interface and multi-task RL can match or beat specialized RS-VLM architectures, and that scaling diverse multi-task data can substitute for architectural novelty. The paper is also commendable for honestly reporting failures (temporal VQA degradation below base, GRTO vs GRPO trade-offs) and for including sensitivity analyses (LODO, permutation tests, robust regressions) that are often absent in this literature. However, the headline scientific claim about the causal role of data diversity is not yet secured; the benchmark results are plausible and useful, but the interpretation is overstated relative to the evidence in the paper's own appendix.

major comments (3)
  1. [Abstract, §5, Appendix A.4/A.6/A.7] The abstract and conclusion claim that OOD gains 'correlate with per-task data diversity' and that 'data scale is more important than architectural novelty.' This causal reading is not supported by the paper's own statistical analysis. A.4 reports a domain-level Spearman ρ=0.54 with permutation p=0.24 at n=7, and the partial rank correlation controlling for base performance drops to ρ=0.26. A.6 states that 'because n_src and s_0 are themselves anti-correlated in our mix... the two coefficients share variance,' and A.7 says collinearity 'prevents a fully clean attribution between H3 and H4.' The observed pattern—largest gains in detection (near-zero base), no gains in RGB caption (saturated base)—is equally consistent with regression to the headroom. The Limitations section explicitly says the diversity trends 'should be interpreted as suggestive rather than conclusive.' The abstract and
  2. [Abstract, §5, Limitations] The claim that 'data scale is more important than architectural novelty' is not tested by the experiments. No architecture is varied in a controlled way: the scaling runs use only InternVL-family backbones (2B and 8B), and the comparison against specialized architectures in Tables 2–3 relies on published models trained with different data, backbones, and pipelines. The Limitations section acknowledges that 'architecture- and data-centric approaches are yet to be directly compared under single controlled environment.' Thus the relative-importance claim is an interpretation, not an experimental result. The paper should either add an architectural control or rephrase the claim to what the data actually show: a strong general-purpose backbone, when trained with this recipe, reaches competitive results.
  3. [§4.1, Tables 2–3] The state-of-the-art rankings may mix incomparable evaluation protocols. The metrics paragraph states that baselines were re-evaluated with G-Eval only for SARLANG-1M and XLRS-Bench captioning; for other free-form VQA tasks it refers to 'dataset-specific GPT-style scoring rubrics and prompts,' and for detection/segmentation it follows benchmark protocols. It is unclear whether the baseline numbers in Tables 2 and 3 were produced by the same evaluation code and prompts as MLRS, or taken from the original papers. Since the 'competitive or state-of-the-art' claim is central, the paper must specify per column which baselines were re-evaluated under identical conditions, or restrict the ranking claims to columns where this is the case.
minor comments (4)
  1. [Appendix A.1] Typo: 'we further anayse effect' should be 'we further analyse the effect.' Also the sentence beginning 'We further anayse effect of random seed...' is missing a verb complement.
  2. [Figure 3 caption] Grammar: 'Following metrics are used' should be 'The following metrics are used.' Also clarify in the caption that one training step corresponds to 16 unique samples, making the x-axis directly interpretable as data scale.
  3. [References] Reference [15] contains 'oxi Wu' as an author name, which appears to be a typo. Several other references have formatting issues (e.g., [16], [49]) that should be cleaned.
  4. [Table 2] The asterisk marker for fine-tuned baselines is explained, but in the Qwen row values like '3.90*' and '73.33*' could be misread as confidence values. Consider marking the model name rather than the score.

Circularity Check

0 steps flagged

No significant circularity: the headline numbers come from held-out benchmarks with an evaluation judge distinct from the training reward; self-citations (GRTO/B-GRTO) are method references with the full objective re-derived in Appendix C. The data-diversity claim is statistically confounded by base-model headroom, but this is an identification problem, not an equation-level circularity.

full rationale

The central derivation chain is empirical rather than circular. MLRS is trained with the multi-task reward of Eq. (1)/(4), whose task scores are defined directly (set-IoU for MC-VQA, Qwen3-Embedding cosine for free-form VQA and captioning, box/mask IoU for detection/segmentation); these reward definitions do not contain the OOD benchmark scores that are later reported. Free-form and captioning evaluations use G-Eval with Qwen2.5-32B-Instruct (Sec. 4.1 Metrics), a different judge from the training reward model, so there is no fitted-input-called-prediction loop. GRTO and the SAM3 auxiliary loss are cited from the authors' B-GRTO [26], but Appendix C fully specifies the objective (Eqs. 14-22, 18-22), so the self-citation is a method reference rather than an unverified load-bearing premise. The paper's own Limitations section and Appendix A.7 concede the key weakness: 'the collinearity between source diversity and base headroom in this particular mix ... prevents a fully clean attribution between H3 and H4' and 'our diversity-related trends should be interpreted as suggestive rather than conclusive.' That is a statistical confounding problem (nsrc anti-correlated with base score, small n=7 domains), not circularity: the diversity effect and the headroom effect are competing explanations of the same empirical correlation, not definitions of each other. The normalized-gain metric Delta_norm = (s_t - s0)/(1 - s0) partly encodes headroom, which explains the negative H4 correlation, but the diversity conclusion is an empirical association rather than a tautology. No step reduces to its own input by construction, and no fitted parameter is renamed as a prediction.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The paper introduces no physical or architectural entities: no new encoders, no new tokens, no new forces. Its load-bearing inputs are (i) the RL objectives GRPO/GRTO, one of which is self-cited prior work, (ii) embedding and judge models treated as unbiased reward/evaluation oracles, (iii) hand-assigned covariates (n_src, mixture proportions, reward weights) that drive the scaling narrative, and (iv) the assumption that OOD benchmarks are clean. The counts here are moderate for an empirical ML paper, but the n_src covariate and the fixed-80k 'data scale' design deserve emphasis because the headline conclusion rests on them.

free parameters (5)
  • Reward weights λ_fmt / λ_task = 0.1 / 0.9
    Hand-chosen in Eq. (1)/Appendix B; shapes the balance between format validity and task score across all behaviors the model learns.
  • GRPO KL weight β = 0.01
    Training hyperparameter (Section 4.1); standard GRPO coefficient, chosen by hand.
  • LoRA rank, group size, learning rates = rank 64, group 8, LR 1e-6 (VLM) / 1e-4 (SAM3)
    Section 4.1 training setup; not tuned against benchmarks, but they set the optimization regime.
  • Per-domain source counts n_src = Detection 19, Segmentation 13, RGB Caption 11, Multi-modal VQA 5, UHR VQA 3, Temporal 2, Multi-view 1
    Hand-assigned covariate (Appendix A.1) that carries the paper's central diversity claim. Includes judgment calls: detection inherits 13 segmentation sources; 16 nominal multi-modal sources collapsed to 5. The headline gain-diversity correlation depends on these counts.
  • 80k training-mixture proportions = e.g., 3333 samples per MC-VQA source, 2500 per free-form VQA source, 4000 per segmentation source
    Appendix D: mixture balanced by hand from the 2.3M pool; the curation procedure is not algorithmically specified, so the exact data exposure is not reproducible.
axioms (6)
  • standard math GRPO policy objective with clipped surrogate (Shao et al., DeepSeekMath)
    Appendix C, Eq. (16); adopted as background for the policy update.
  • domain assumption GRTO joint policy-tool optimization objective (B-GRTO)
    Appendix C, Eq. (18); the authors' own prior work [26] defines the joint VLM-SAM3 update; the paper relies on its correctness without re-deriving it.
  • domain assumption SAM3 semantic head can be fine-tuned with BCE + soft-IoU surrogate loss
    Appendix C, Eqs. (20)-(22); follows B-GRTO [26] and assumes the surrogate gradient actually improves tool behavior.
  • domain assumption Qwen3-Embedding-0.6B cosine similarity is a valid semantic reward and Qwen2.5-32B/G-Eval is an unbiased judge
    Section 3 reward (Eq. 8) and Section 4.1 metrics; no human-agreement analysis is reported for the remote sensing setting, and the claim that G-Eval aligns better with humans than reference-overlap metrics is imported from [22].
  • domain assumption OOD benchmarks share no images or annotations with training data
    Section 4.1 asserts zero-shot cleanliness of XLRS-Bench, RSHR, VLRS, UrBench, GEOBench-VLM, GeoSeg-Bench2; not independently verifiable from the paper.
  • domain assumption InternVL3.5-8B/2B are strong general VLMs and LoRA adaptation suffices
    Section 4.1; all scaling runs use one model family, which the paper lists as a limitation for generalizing the trend.

pith-pipeline@v1.3.0-alltime-deepseek · 31370 in / 19901 out tokens · 206384 ms · 2026-08-01T21:51:16.517131+00:00 · methodology

0 comments
read the original abstract

Remote sensing vision-language models are increasingly expected to support open-ended reasoning over Earth Observation data and a variety of tasks. Most recent progress in this area has been driven by remote-sensing-specific architectural designs, often introducing new encoders, alignment modules, or task-specific fusion mechanisms. In this work, we challenge the necessity of such architectural specialization. We show that a generally capable vision-language model can achieve competitive or state-of-the-art performance at challenging remote sensing benchmarks, provided that it is trained at sufficient scale across diverse data and tasks. Our model uses a single language policy that can either answer directly in text or invoke a localization tool for segmentation and grounding. To train this heterogeneous behaviour, we employ a multi-task reinforcement learning framework with adaptive task rewards covering multiple-choice VQA, free-form VQA, captioning, detection, and segmentation across a large variety of input types. Our approach achieves competitive results across a broad set of benchmarks, including high-resolution, multi-temporal, multi-modal and multi-view tasks. Further, as training data scales, our experiments show consistent improvements across most tasks both in and out of distribution, which correlate with per-task data diversity. These findings suggest that, for remote sensing VLMs, data scale is more important than architectural novelty.

Figures

Figures reproduced from arXiv: 2607.15942 by Danda Pani Paudel (INSAIT, Luc Van Gool, Mario Markov, Mohammad Mahdi, Sofia University "St. Kliment Ohridski"), Stefan Maria Ailuro.

Figure 1
Figure 1. Figure 1: Diverse multi-task training drives out-of-distribution generalization in remote sensing. (a) OOD gain over the base model during multi-task RL training, aggregated per task-domain. For robustness test the training is performed a second time with different seed and paraphrased prompts. (b) Final OOD gain vs. number of training sources per domain; for visualization, detection excludes a x7.6 gain outlier, XL… view at source ↗
Figure 2
Figure 2. Figure 2: A single language policy (InternVL [43]) handles all input types — ultra-high￾resolution, multi-modal, multi-view, multi-temporal, visual prompting. The policy rea￾sons in text, answers directly, or invokes a SAM3 segmentation tool via a textual prompt; components communicate only through text, with no task-specific encoders or fusion modules. Both components are optimized end-to-end with GRTO [26], using … view at source ↗
Figure 3
Figure 3. Figure 3: Performance vs training steps. Following metrics are used: aggregated accu￾racy for multi-choice VQA, G-Eval for free-form VQA and caption, aggregated IoU for segmentation, Precision@0.5 for GEOBench-VLM detection, and Accuracy@0.5 for XLRS-Bench detection. Effects of scaling training data [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance comparisons. (a) Zero-shot performance: percentage difference between MLRS and the best-performing existing baseline for each dataset and task. For every comparison, we report the competing model and the evaluation metric used. (b) GRTO vs. GRPO ablation: IoU percentage difference between MLRS, where SAM3 is fine-tuned (GRTO), and keeping SAM3 frozen (GRPO) for segmentation datasets [PITH_FULL… view at source ↗
Figure 5
Figure 5. Figure 5: Model size ablation. OOD performance vs training steps of the main data scale training, reproduced for InternVL3.5-2B. The following metrics are used: aggregated accuracy for MC VQA, G-Eval for free-form VQA and captioning, aggregated IoU for segmentation, Precision@0.5 for GEOBench-VLM detection, and Accuracy@0.5 for XLRS-Bench detection. form VQA and detection, and Tables 3a and 3b for MC VQA and segment… view at source ↗
Figure 6
Figure 6. Figure 6: Example outputs of MLRS for five general tasks: captioning, detection, free￾form VQA, segmentation, multiple-choice VQA. For detection, the bounding box is visualized on the input image. For segmentation, SAM3 is prompted with the VLM’s output, and the resulting mask is visualized on the input image [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: OOD normalized-gain trajectories per task domain (panels ordered by de￾creasing nsrc). Lines are means over the benchmarks in each domain, bands are min– max ranges across benchmarks; colors denote the three runs. The three qualitative behaviours are visible: sustained improvement in source-rich domains (Detection, Seg￾mentation, Multi-modal VQA), early peak followed by stagnation or decline in source￾poor… view at source ↗
Figure 8
Figure 8. Figure 8: Domain-level normalized gain at step 3000 (mean ± bootstrap 95% CI over benchmarks × 3 runs) against the two training-mix covariates. Left (H1): no depen￾dence on sampling volume. Right (H3): positive dependence on the number of unique sources; the visible off-trend point is RGB Caption, whose negative gain is explained by base-model saturation ( [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Normalized gain at step 3000 against base-model score. Small markers: individ￾ual benchmarks per run (circle/square/triangle = 8B-s1/8B-s2/2B); large diamonds: domain means with bootstrap 95% CIs; dashed line: OLS fit on benchmark means. Detection benchmarks (near-zero base) and RGB Caption (base ≈ 0.88 BERTScore) anchor the two extremes of the headroom effect. significant (peak retention vs. nsrc: ρ = 0.5… view at source ↗
Figure 10
Figure 10. Figure 10: (a) Seed comparison: 8B seed-1 vs. seed-2 scores over 25 benchmarks at steps 1k–3k (color = step); points lie on the identity line. (b) Scale comparison: 2B vs. 8B (seed-mean) normalized gains at 3k per OOD benchmark (colors = domains as in [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Example outputs of MLRS for five general tasks: captioning, detection, free￾form VQA, segmentation, multiple-choice VQA. For detection, the bounding box is visualized on the high-resolution input image and zoomed in. For segmentation, SAM3 is prompted with the VLM’s output, and the resulting mask is visualized on the input image [PITH_FULL_IMAGE:figures/full_fig_p032_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Example outputs of MLRS for five general tasks: captioning, detection, free￾form VQA, segmentation, multiple-choice VQA. The captioning prompt and model output are shortened. For detection, the bounding box is visualized on the input image. For segmentation, SAM3 is prompted with the VLM’s output, and the resulting mask is visualized on the input image. The MC VQA example prompts visually: the target is z… 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

73 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Ailuro, S.M., Markov, M., Mahdi, M., Boychev, D., Gool, L.V., Paudel, D.P.: Osmda: Openstreetmap-based domain adaptation for remote sensing vlms (2026), https://arxiv.org/abs/2603.118044

    arXiv 2026

  2. [2]

    Blushtein-Livnon, R., Rafaeli, O., Ioffe, D., Boger, A., Esquenazi, K.S., Svoray, T.: On the effectiveness of textual prompting with lightweight fine-tuning for sam3 remote sensing segmentation (2026),https://arxiv.org/abs/2512.155645

    Pith/arXiv arXiv 2026

  3. [3]

    Cai, M., Wang, G., Zhang, W., Zhou, G., Zhuang, Y., Zhang, T., Wang, H., Chen, H., Li, J.: Earth-onevision: Extending remote sensing multimodal large language models to more sensor modalities and tasks (2026),https://arxiv.org/abs/ 2606.108192, 4

    Pith/arXiv arXiv 2026

  4. [4]

    Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Suris, D., Ryali, C., Alwala, K.V., Khedr, H., Huang, A., Lei, J., Ma, T., Guo, B., Kalla, A., Marks, M., Greer, J., Wang, M., Sun, P., Rädle, R., Afouras, T., Mavroudi, E., Xu, K., Wu, T.H., Zhou, Y., Momeni, L., Hazra, R., Ding, S., Vaze, S., Porcher, F., Li, F., Li, S., Kamath, A., Cheng, H.K., ...

    Pith/arXiv arXiv 2025

  5. [5]

    Chen, T., Cao, R., Yu, X., Zhu, L., Ding, C., Ji, D., Chen, C., Zhu, Q., Xu, C., Mao, P., Zang, Y.: SAM3-Adapter: Efficient adaptation of segment anything 3 for cam- ouflage object segmentation, shadow detection, and medical image segmentation (2025),https://arxiv.org/abs/2511.194255

    arXiv 2025

  6. [6]

    In: NeurIPS (2021) 4

    Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: NeurIPS (2021) 4

    2021

  7. [7]

    In: Forty-second International Conference on Machine Learn- ing (2025),https://openreview.net/forum?id=dYur3yabMj4

    Chu, T., Zhai, Y., Yang, J., Tong, S., Xie, S., Schuurmans, D., Le, Q.V., Levine, S., Ma, Y.: SFT memorizes, RL generalizes: A comparative study of foundation model post-training. In: Forty-second International Conference on Machine Learn- ing (2025),https://openreview.net/forum?id=dYur3yabMj4

    2025

  8. [8]

    arXiv preprint arXiv:2512.17319 (2025) 7, 12, 35

    Dang, Y., Zhu, M., Wang, D., Zhang, Y., Yang, J., Fan, Q., Yang, Y., Li, W., Miao, F., Gao, Y.: A benchmark for ultra-high-resolution remote sensing mllms. arXiv preprint arXiv:2512.17319 (2025) 7, 12, 35

    arXiv 2025

  9. [9]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Danish, M., Munir, M.A., Shah, S.R.A., Kuckreja, K., Khan, F.S., Fraccaro, P., Lacoste, A., Khan, S.: Geobench-vlm: Benchmarking vision-language models for geospatial tasks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7132–7142 (October 2025) 7, 8, 9

    2025

  10. [10]

    Doerksen, K., Kerner, H.: Earthshift: a benchmark for measuring robustness to real-world distribution shifts in earth observation (2026),https://arxiv.org/ abs/2605.293304

    Pith/arXiv arXiv 2026

  11. [11]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Feng, J., Wang, S., Liu, T., Xi, Y., Li, Y.: Urbanllava: A multi-modal large lan- guage model for urban intelligence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 6209–6219 (October 2025) 3, 4

    2025

  12. [12]

    Ailuro et al

    Fiaz, M., Debary, H., Fraccaro, P., Paudel, D., Gool, L.V., Khan, F., Khan, S.: Geovlm-r1: Reinforcement fine-tuning for improved remote sensing reasoning (2025),https://arxiv.org/abs/2509.250263, 4 16 S.M. Ailuro et al

    arXiv 2025

  13. [13]

    In: International Con- ference on Learning Representations (2022),https://openreview.net/forum?id= nZeVKeeFYf98

    Hu, E.J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Con- ference on Learning Representations (2022),https://openreview.net/forum?id= nZeVKeeFYf98

    2022

  14. [14]

    In: International Conference on Learning Representations (2025) 3, 4

    Irvin, J.A., Liu, E.R., Chen, J.C., Dormoy, I., Kim, J., Khanna, S., Zheng, Z., Er- mon, S.: Teochat: A large vision-language assistant for temporal earth observation data. In: International Conference on Learning Representations (2025) 3, 4

    2025

  15. [15]

    Jiang, L., Pei, Y., oxi Wu, Zhao, Y., Wu, T., Yu, S., Zhang, L., Cai, D.: Geoseg: Training-free reasoning-driven segmentation in remote sensing imagery (2026), https://arxiv.org/abs/2603.039837

    arXiv 2026

  16. [16]

    kelu, Y., Nuo, X., Rong, Y., Yingying, X., Zhuoyan, G., Titinunt, K., yi, R., Pu, Z., Jin, W., Ning, W., Chao, L.: Falcon: A remote sensing vision-language foundation model (2025) 4

    2025

  17. [17]

    arXiv:2304.02643 (2023) 4

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) 4

    Pith/arXiv arXiv 2023

  18. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Kuckreja, K., Danish, M.S., Naseer, M., Das, A., Khan, S., Khan, F.S.: Geochat: Grounded large vision-language model for remote sensing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 27831–27840 (June 2024) 2, 3, 4

    2024

  19. [19]

    https://doi.org/10.48550/arXiv.2504.09644,https://arxiv.org/abs/2504

    Li, K., Xin, Z., Pang, L., Pang, C., Deng, Y., Yao, J., Xia, G., Meng, D., Wang, Z., Cao, X.: SegEarth-R1: Geospatial pixel reasoning via large language model (2025). https://doi.org/10.48550/arXiv.2504.09644,https://arxiv.org/abs/2504. 096444, 5, 7, 8, 28

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.09644 2025

  20. [20]

    ISPRS Journal of Photogrammetry and Remote Sensing 227, 539–550 (sep 2025).https://doi.org/10.1016/j.isprsjprs.2025.06.003 2, 3, 4

    Li, Z., Muhtar, D., Gu, F., He, Y., Zhang, X., Xiao, P., He, G., Zhu, X.: Lhrs- bot-nova: Improved multimodal large language model for remote sensing vision- language interpretation. ISPRS Journal of Photogrammetry and Remote Sensing 227, 539–550 (sep 2025).https://doi.org/10.1016/j.isprsjprs.2025.06.003 2, 3, 4

    work page doi:10.1016/j.isprsjprs.2025.06.003 2025

  21. [21]

    Liu, J., Sun, L., Fu, R., Yang, B.: Towards faithful reasoning in remote sensing: A perceptually-grounded geospatial chain-of-thought for vision-language models (2026),https://arxiv.org/abs/2509.222214

    arXiv 2026

  22. [22]

    In: Bouamor, H., Pino, J., Bali, K

    Liu, Y., Iter, D., Xu, Y., Wang, S., Xu, R., Zhu, C.: G-Eval: NLG evaluation using GPT-4 with better human alignment. In: Bouamor, H., Pino, J., Bali, K. (eds.) Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 2511–2522. Association for Computational Linguistics, Singapore (Dec 2023).https://doi.org/10.18653/v1/2...

    doi:10.18653/v1/2023.emnlp- 2023

  23. [23]

    arXiv preprint arXiv:2406.10100 (2024).https://doi.org/10.48550/arXiv.2406.101002, 3, 4

    Luo, J., Pang, Z., Zhang, Y., Wang, T., Wang, L., Dang, B., Lao, J., Wang, J., Chen, J., Tan, Y., Li, Y.: Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding. arXiv preprint arXiv:2406.10100 (2024).https://doi.org/10.48550/arXiv.2406.101002, 3, 4

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.10100 2024

  24. [24]

    In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV)

    Luo, J., Zhang, Y., Yang, X., Wu, K., Zhu, Q., Liang, L., Chen, J., Li, Y.: When large vision-language model meets large remote sensing imagery: Coarse-to-fine text-guided token pruning. In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV). pp. 9206–9217 (October 2025) 2, 3, 4

    2025

  25. [25]

    Luo, Z., Wang, D., Wang, H., Guo, H., Zhang, J., Du, B.: Vlrs-bench: A vision- languagereasoningbenchmarkforremotesensing.arXivpreprintarXiv:2602.07045 (2026) 7, 36 More with Less: a Large Scale RS-VLM with a Simple Recipe 17

    Pith/arXiv arXiv 2026

  26. [26]

    Markov, M., Ailuro, S.M., Mahdi, M., Van Gool, L., Paudel, D.P.: B-grto: Boot- strappedgrouprelativetooloptimizationforreferringsegmentation.arXivpreprint arXiv:2605.23500 (2026).https://doi.org/10.48550/arXiv.2605.235005, 6, 14, 26, 27

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.23500 2026

  27. [27]

    In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G

    Muhtar, D., Li, Z., Gu, F., Zhang, X., Xiao, P.: Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model. In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G. (eds.) Computer Vision – ECCV 2024. pp. 440–457. Springer Nature Switzerland, Cham (2025) 3

    2024

  28. [28]

    Ni, S., Wang, D., Chen, H., Guo, H., Zhang, N., Zhang, J.: Unigeoseg: Towards unified open-world segmentation for geospatial scenes (2025),https://arxiv.org/ abs/2511.233323, 4, 5, 7, 8, 28, 38

    Pith/arXiv arXiv 2025

  29. [29]

    IEEE Geoscience and Remote Sensing Magazine13(3), 324–337 (2025).https://doi.org/10.1109/ MGRS.2025.35602933, 4, 5

    Ou, R., Hu, Y., Zhang, F., Chen, J., Liu, Y.: Geopix: A multimodal large language model for pixel-level image understanding in remote sensing. IEEE Geoscience and Remote Sensing Magazine13(3), 324–337 (2025).https://doi.org/10.1109/ MGRS.2025.35602933, 4, 5

    arXiv 2025

  30. [30]

    Pang, C., Weng, X., Wu, J., Li, J., Liu, Y., Sun, J., Li, W., Wang, S., Feng, L., Xia, G.S., He, C.: Vhm: Versatile and honest vision language model for remote sens- ing image analysis. Proceedings of the AAAI Conference on Artificial Intelligence 39(6), 6381–6388 (Apr 2025).https://doi.org/10.1609/aaai.v39i6.32683, https://ojs.aaai.org/index.php/AAAI/art...

    work page doi:10.1609/aaai.v39i6.32683 2025

  31. [31]

    In: The Thirty- ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2026),https://openreview.net/forum?id=X7CxMmmgkb3, 4, 5

    Quenum, J., Hsieh, W.H., Wu, T.H., Gupta, R., Darrell, T., Chan, D.M.: LISAt: Language-instructed segmentation assistant for satellite imagery. In: The Thirty- ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2026),https://openreview.net/forum?id=X7CxMmmgkb3, 4, 5

    2026

  32. [32]

    arXiv preprint arXiv:2408.00714 (2024),https://arxiv.org/ abs/2408.007144

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollár, P., Feichtenhofer, C.: SAM 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024),https://arxiv.org/ abs/2408.007144

    Pith/arXiv arXiv 2024

  33. [33]

    arXiv preprint arXiv:2501.13925 (2025) 3, 4, 5

    Shabbir, A., Zumri, M., Bennamoun, M., Khan, F.S., Khan, S.: Geopixel: Pixel grounding large multimodal model in remote sensing. arXiv preprint arXiv:2501.13925 (2025) 3, 4, 5

    Pith/arXiv arXiv 2025

  34. [34]

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Zhang, M., Li, Y., Wu, Y., Guo, D.: DeepSeekMath: Pushing the limits of mathematical reasoning in open language models (2024),https://arxiv.org/abs/2402.033006

    Pith/arXiv arXiv 2024

  35. [35]

    arXiv preprint arXiv:2506.01667 (2025).https: //doi.org/10.48550/arXiv.2506.016672, 3, 4, 5

    Shu, Y., Ren, B., Xiong, Z., Paudel, D.P., Van Gool, L., Demir, B., Sebe, N., Rota, P.: Earthmind: Towards multi-granular and multi-sensor earth observation with large multimodal models. arXiv preprint arXiv:2506.01667 (2025).https: //doi.org/10.48550/arXiv.2506.016672, 3, 4, 5

    doi:10.48550/arxiv.2506.01667 2025

  36. [36]

    In: IEEE conference on Computer Vision and Pattern Recognition (2026) 3, 4, 5

    Shu, Y., Ren, B., Xiong, Z., Zhu, X.X., Demir, B., Sebe, N., Rota, P.: Terras- cope: Pixel-grounded visual reasoning for earth observation. In: IEEE conference on Computer Vision and Pattern Recognition (2026) 3, 4, 5

    2026

  37. [37]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Soni, S., Dudhane, A., Debary, H., Fiaz, M., Munir, M.A., Danish, M.S., Frac- caro, P., Watson, C.D., Klein, L.J., Khan, F.S., Khan, S.: Earthdial: Turning multi-sensory earth observations to interactive dialogues. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14303–14313 (June 2025).https://doi.org/10.1...

    arXiv 2025

  38. [38]

    Wang, D., Liu, S., Jiang, W., Wang, F., Liu, Y., Qin, X., Luo, Z., Zhou, C., Guo, H., Zhang, J., Du, B., Tao, D., Zhang, L.: Geozero: Incentivizing reasoning from scratch on geospatial scenes (2026),https://arxiv.org/abs/2511.226454, 7, 28

    arXiv 2026

  39. [39]

    In: The Thirty- ninthAnnualConferenceonNeuralInformationProcessingSystems(2026),https: //openreview.net/forum?id=LTgUInLTbP2, 3, 4, 8, 28

    Wang, F., Chen, M., Li, Y., Wang, D., Wang, H., Guo, Z., Wang, Z., Boqi, S., Lan, L., Wang, Y., Wang, H., Yang, W., Du, B., Zhang, J.: GeoLLaVA-8k: Scaling remote-sensing multimodal large language models to 8k resolution. In: The Thirty- ninthAnnualConferenceonNeuralInformationProcessingSystems(2026),https: //openreview.net/forum?id=LTgUInLTbP2, 3, 4, 8, 28

    2026

  40. [40]

    arXiv preprint arXiv:2602.14201 (2026) 4

    Wang, F., Chen, M., Li, Y., Yang, Y., Zhang, Y., Lan, L., Yang, X., Sun, H., Wang, Y., Wang, D., et al.: Geoeyes: On-demand visual focusing for evidence- grounded understanding of ultra-high-resolution remote sensing imagery. arXiv preprint arXiv:2602.14201 (2026) 4

    arXiv 2026

  41. [41]

    Wang, F., Wang, H., Guo, Z., Wang, D., Wang, Y., Chen, M., Ma, Q., Lan, L., Yang, W., Zhang, J., Liu, Z., Sun, M.: Xlrs-bench: Could your multimodal llms understand extremely large ultra-high-resolution remote sensing imagery? In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR). pp. 14325–14336 (June 2025) 7, 9, 12, 34

    2025

  42. [42]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2026), https://openreview.net/forum?id=sQO1ZEQGqX7, 8, 28

    Wang, J., Xuan, W., Qi, H., Liu, Z., Liu, K., Wu, Y., Chen, H., Song, J., Xia, J., Zheng, Z., Yokoya, N.: Disasterm3: A remote sensing vision-language dataset for disaster damage assessment and response. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2026), https://openreview.net/forum?id=sQO...

    2026

  43. [43]

    arXiv preprint arXiv:2508.18265 (2025).https://doi

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025).https://doi. org/10.48550/arXiv.2508.18265,https://arxiv.org/abs/2508.182655, 8

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.18265 2025

  44. [44]

    Wang, Y., Liu, Z., Wang, Z., Hu, H., Liu, P., Rao, Y.: Geovista: Web-augmented agentic visual reasoning for geolocalization (2025),https://arxiv.org/abs/2511. 157053

    2025

  45. [45]

    IEEE Transactions on Geoscience and Remote Sensing64, 1–20 (2026).https: //doi.org/10.1109/TGRS.2026.36520997, 8, 9, 28

    Wei, Y., Xiao, A., Ren, Y., Zhu, Y., Chen, H., Xia, J., Yokoya, N.: Sarlang- 1m: A benchmark for vision–language modeling in sar image understanding. IEEE Transactions on Geoscience and Remote Sensing64, 1–20 (2026).https: //doi.org/10.1109/TGRS.2026.36520997, 8, 9, 28

    arXiv 2026

  46. [46]

    Xin, Z., Li, K., Chen, L., Li, W., Xiao, Y., Qiao, H., Zhang, W., Meng, D., Cao, X.: Segearth-r2: Towards comprehensive language-guided segmentation for remote sensing images (2025),https://arxiv.org/abs/2512.200134, 5, 7, 8, 28, 37

    arXiv 2025

  47. [47]

    In: The Thirty-ninth Annual Conference on Neural Information ProcessingSystemsDatasetsandBenchmarksTrack(2026),https://openreview

    Xuan, W., Wang, J., Qi, H., Chen, Z., Zheng, Z., Zhong, Y., Xia, J., Yokoya, N.: DynamicVL: Benchmarking multimodal large language models for dynamic city understanding. In: The Thirty-ninth Annual Conference on Neural Information ProcessingSystemsDatasetsandBenchmarksTrack(2026),https://openreview. net/forum?id=zubCrOvUZ43, 4, 5, 7, 8, 12, 28

    2026

  48. [48]

    arXiv preprint arXiv:2412.15115 (2024) 9

    Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., Lin, H., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Lin, J., Dang, K., Lu, K., Bao, K., Yang, K., Yu, L., Li, M., Xue, M., Zhang, P., Zhu, Q., Men, R., Lin, R., Li, T., Xia, T., Ren, X., Ren, X., Fan, Y., Su, Y., Zhang, Y., Wan, Y., Liu, Y., Cui...

    Pith/arXiv arXiv 2024

  49. [49]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Yao, L., Liu, F., Lu, H., Zhang, C., Min, R., Xu, S., Di, S., Peng, P.: Remoterea- soner: Towards unifying geospatial reasoning workflow. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 11883–11891 (2026) 4, 5, 6 More with Less: a Large Scale RS-VLM with a Simple Recipe 19

    2026

  50. [50]

    Ye, J., Lin, H., Ou, L., Chen, D., Wang, Z., Zhu, Q., He, C., Li, W.: Where am i? cross-view geo-localization with natural language descriptions (2025),https: //arxiv.org/abs/2412.170073, 8, 28

    Pith/arXiv arXiv 2025

  51. [51]

    Yuan, Z., Mou, L., Hua, Y., Zhu, X.X.: Rrsis: Referring remote sensing image segmentation (2024),https://arxiv.org/abs/2306.086255

    Pith/arXiv arXiv 2024

  52. [52]

    ISPRS Journal of Pho- togrammetry and Remote Sensing221, 64–77 (2025).https://doi.org/https: //doi.org/10.1016/j.isprsjprs.2025.01.020,https://www.sciencedirect

    Zhan, Y., Xiong, Z., Yuan, Y.: Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. ISPRS Journal of Pho- togrammetry and Remote Sensing221, 64–77 (2025).https://doi.org/https: //doi.org/10.1016/j.isprsjprs.2025.01.020,https://www.sciencedirect. com/science/article/pii/S09242716250002062, 4

    doi:10.1016/j.isprsjprs.2025.01.020 2025

  53. [53]

    In: International Conference on Learning Repre- sentations (2020),https://openreview.net/forum?id=SkeHuCVFDr8

    Zhang*, T., Kishore*, V., Wu*, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Eval- uating text generation with bert. In: International Conference on Learning Repre- sentations (2020),https://openreview.net/forum?id=SkeHuCVFDr8

    2020

  54. [54]

    IEEE Transactions on Geoscience and Remote Sensing63, 1–21 (2025).https://doi.org/10.1109/TGRS.2025

    Zhang, W., Cai, M., Ning, Y., Zhang, T., Zhuang, Y., Lu, S., Chen, H., Li, J., Mao, X.: Earthgpt-x: A spatial mllm for multilevel multisource remote sensing imagery understanding with visual prompting. IEEE Transactions on Geoscience and Remote Sensing63, 1–21 (2025).https://doi.org/10.1109/TGRS.2025. 36269413, 4

    doi:10.1109/tgrs.2025 2025

  55. [55]

    IEEE Transac- tions on Geoscience and Remote Sensing (2024) 2, 3, 4

    Zhang, W., Cai, M., Zhang, T., Zhuang, Y., Li, J., Mao, X.: Earthmarker: A visual prompting multi-modal large language model for remote sensing. IEEE Transac- tions on Geoscience and Remote Sensing (2024) 2, 3, 4

    2024

  56. [56]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–20 (2024)

    Zhang, W., Cai, M., Zhang, T., Zhuang, Y., Mao, X.: Earthgpt: A universal multi- modallargelanguagemodelformultisensorimagecomprehensioninremotesensing domain. IEEE Transactions on Geoscience and Remote Sensing62, 1–20 (2024). https://doi.org/10.1109/TGRS.2024.34096243, 4

    arXiv 2024

  57. [57]

    ISPRS Journal of Photogrammetry and Remote Sens- ing237, 217–235 (2026).https://doi .org/https://doi.org/10.1016/ j

    Zhang, X., Ge, J., Zheng, Y., Guo, K., Liang, J.: Bridging semantics and ge- ometry: A decoupled lvlm–sam framework for reasoning segmentation in op- tical remote sensing. ISPRS Journal of Photogrammetry and Remote Sens- ing237, 217–235 (2026).https://doi .org/https://doi.org/10.1016/ j. isprsjprs.2026.04.036,https://www.sciencedirect.com/science/article/...

    2026

  58. [58]

    arXiv preprint arXiv:2506.05176 (2025)

    Zhang, Y., Li, M., Long, D., Zhang, X., Lin, H., Yang, B., Xie, P., Yang, A., Liu, D., Lin, J., Huang, F., Zhou, J.: Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176 (2025). https://doi.org/10.48550/arXiv.2506.051767, 25

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.05176 2025

  59. [59]

    In: Advances in Neural Information Processing Systems

    Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging LLM-as-a-judge with MT-bench and chatbot arena. In: Advances in Neural Information Processing Systems. vol. 36, pp. 46595–46623 (2023) 8

    2023

  60. [60]

    AAAI (2025) 3, 7, 37

    Zhou, B., Yang, H., Chen, D., Ye, J., Bai, T., Yu, J., Zhang, S., Lin, D., He, C., Li, W.: Urbench: A comprehensive benchmark for evaluating large multimodal models in multi-view urban scenarios. AAAI (2025) 3, 7, 37

    2025

  61. [61]

    org/abs/2605.144753, 4 20 S.M

    Zhu, J., Fu, R., Hu, J., et al., X.N.: Geovista: Visually grounded active perception for ultra-high-resolution remote sensing understanding (2026),https://arxiv. org/abs/2605.144753, 4 20 S.M. Ailuro et al. A Trends analysis A.1 Setup, metrics, and covariates We analyze three training runs: an 8B model trained to 5k steps, a second 8B run with a different...

    Pith/arXiv arXiv 2026

  62. [62]

    Segment the

    and GeoSeg-Bench2 (2B base0.004) would dominate any ratio-based analy- sis; however, the normalized-gain metric∆norm bounds their influence, and all results were re-checked excluding detection benchmarks and under Huber robust regression, without sign changes. The main limitations are: statistical power – seven task domains cap the resolutionofalldomain-l...

  63. [63]

    noun phrase

    If the task requires segmenting/localizing objects in the image, call the segmentation tool by outputting: <segmentation>{"noun phrase": "...", "objects": [{"bbox": [x1, y1, x2, y2]}]}</segmentation> The segmentation JSON must contain: - "noun phrase": exactly one referential noun phrase identifying the target. - "objects": a list of objects, each with on...

  64. [64]

    Ailuro et al

    If the task asks for a textual answer and does not require segmentation, output: <answer>your textual answer</answer> 30 S.M. Ailuro et al. Use <think>...</think> before the final tag when reasoning is useful. Do not use <answer> for segmentation. Do not use <segmentation> for non-segmentation answers. G-Eval prompt for captioning. You are an expert judge...

  65. [65]

    Analyze the Ground Truth for core objects and counts

  66. [66]

    Imaginary Objects

    Check the Prediction for "Imaginary Objects" (Hallucinations) not present in the GT

  67. [67]

    Verify if object counts and spatial relationships match the GT

  68. [68]

    Scoring Rubric: 1 (Critical Failure): Major hallucination (imaginary objects) or completely wrong scene classification

    Assign a strict score from 1-5 using the rubric below. Scoring Rubric: 1 (Critical Failure): Major hallucination (imaginary objects) or completely wrong scene classification. 2 (Poor): Correct scene type, but severe errors in object counting or wrong object attributes. 3 (Fair): Captures the main gist, but has minor hallucinations or noticeable counting e...

  69. [69]

    Analyze the Question to understand what information is required

  70. [70]

    More with Less: a Large Scale RS-VLM with a Simple Recipe 31

    Examine the Ground Truth Answer for key facts, values, and constraints. More with Less: a Large Scale RS-VLM with a Simple Recipe 31

  71. [71]

    Check the Predicted Answer for hallucinations (information not supported by the GT)

  72. [72]

    Verify correctness, precision, and completeness of the Predicted Answer

  73. [73]

    Scoring Rubric: 1 (Critical Failure): Incorrect answer or major hallucination; does not address the question

    Assign a strict score from 1-5 using the rubric below. Scoring Rubric: 1 (Critical Failure): Incorrect answer or major hallucination; does not address the question. 2 (Poor): Partially related but mostly incorrect; major factual errors or missing key elements. 3 (Fair): Captures the general idea but contains minor errors, ambiguity, or incomplete details....

    arXiv 2024