Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

NeSy-Route claims that current multimodal models fail at constrained route planning in satellite imagery, and locates the failure in perception and planning rather than textual constraint understanding.

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

2026-08-02 18:02 UTC pith:76CFI2EM

load-bearing objection A large, thoughtfully built benchmark whose central planning metric is vacuous: the Task 3 evaluator fills in the dense path with A* on the ground-truth cost map, so a model that outputs only start and end points receives perfect planning scores. the 3 major comments →

arxiv 2603.16307 v3 pith:76CFI2EM submitted 2026-03-17 cs.AI

NeSy-Route: A Neuro-Symbolic Benchmark for Constrained Route Planning in Remote Sensing

Ming Yang , Zhi Zhou , Shi-Yu Tian , Kun-Yang Yu , Lan-Zhe Guo , Yu-Feng Li This is my paper
classification cs.AI
keywords neuro-symbolic benchmarkconstrained route planningremote sensingmultimodal large language modelsA* searchland cover classificationhierarchical evaluationoptimal trajectories
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 introduces NeSy-Route, a benchmark for constrained route planning in aerial and satellite imagery. Its central claim is that planning can be decomposed into three hierarchical stages — reading written constraints, anchoring them to land-cover regions in an image, and generating an optimal waypoint path — and that current multimodal large language models handle the first stage but largely fail the second and third. The benchmark generates 10,821 planning samples with provably optimal ground-truth routes by running A* search over a cost map built from symbolic traversability and preference rules. A sympathetic reading is that if this claim holds, the field gains a scalable, automated way to measure whether models can make grounded spatial decisions, not just recognize objects or answer questions about images.

Core claim

On the paper's own terms, the discovery is that constrained route planning is a distinct capability that today's multimodal models lack, and that it can be measured objectively. NeSy-Route formalizes a route as a sequence of pixels minimizing a cumulative cost W(p), where each of eight land-cover classes gets a traversability bit and a preference rank; A* with an admissible Euclidean heuristic finds the true optimal path, giving every sample a gold-standard trajectory. Across evaluated models, textual constraint understanding often succeeds, but text–image alignment and route planning collapse, with adherence rates below a third and average cost ratios far above the optimal 1.0. The authors

What carries the argument

The load-bearing object is the pixel cost map W(p), defined by two symbolic vectors: V_trav marks each land-cover class as traversable or blocked, and V_pref ranks traversable classes by the routing objective (shortest, fastest, safest, comfortable). W(p) is infinite on blocked classes and equals max(pref) − pref(class) + 1 elsewhere, so A* search with the Euclidean heuristic is provably optimal. The same cost map drives the evaluator: a model's sparse waypoints are connected by A* to produce a dense reconstructed path, whose adherence, cost ratio, and Chamfer distance to the gold route are then scored. This closed neuro-symbolic loop — symbolic knowledge base, neural model, classical search

Load-bearing premise

The Task 3 evaluator assumes that A*-reconstructing a dense path from sparse waypoints on the ground-truth cost map measures the model's planning, but this assumption fails for a model that outputs only the start and end points, which scores perfectly.

What would settle it

Run the benchmark's Task 3 with a baseline that always returns just the two given endpoint coordinates; if the formulas yield AR=100%, CR=1, and CD=0 on feasible pairs, the planning scores are inflated by oracle assistance and the reported model rankings would need to be recomputed under a stricter protocol.

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

If this is right

  • If the benchmark's finding is right, improving a model's language understanding of constraints will not by itself produce good routes; models must also ground those constraints in visual regions and perform global search-like reasoning.
  • The three-level protocol lets developers pinpoint whether a failure comes from perception, constraint alignment, or planning, which should make targeted training or prompting easier.
  • The automated generation pipeline can be reused to build similar planning benchmarks for other geographies or agent types without manual route annotation.
  • Every sample has a provably optimal trajectory, so the benchmark supplies a clean reference for comparing future planning-capable systems.
  • The reported baselines quantify how far current systems are from usable autonomous route planning in remote sensing.

Where Pith is reading between the lines

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

  • One consequence not drawn in the paper: the Task 3 evaluator may over-credit planning, because a model that emits only the start and end waypoints receives perfect adherence, cost ratio 1, and zero Chamfer distance whenever the pair is feasible — the A* reconstruction does the planning for it.
  • A stronger evaluation would compare the model's own full path against the optimum, or require a minimum number of waypoints before reconstruction is allowed.
  • A testable extension would change the traversability of a single land-cover class after the first query and ask the model to re-plan; a system with genuine planning should re-route, whereas one that pattern-matches text may not.
  • The same neuro-symbolic recipe could be applied to dynamic settings, such as flood or fire spread, where the cost map changes over time and the model must re-plan under evolving constraints.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. NeSy-Route is a remote-sensing benchmark for constrained route planning, built on OpenEarthMap with 10,821 planning samples organized as three hierarchical tasks: textual constraint understanding (Task 1), text-image constraint alignment (Task 2), and constrained route planning (Task 3). The authors propose an automated generation pipeline that combines KB-derived traversability/preference vectors, LLM-written queries with self-inference verification, morphological processing of semantic masks, and A* search over a pixel cost map. Task 3 is evaluated by reconstructing dense paths from predicted sparse waypoints using A* on the ground-truth cost map, then computing AR, CR, VR, and CD. Experiments on 11 MLLMs report strong Task 1 performance but significant deficiencies in perception and planning. The central contribution is a large, automated, symbolically verifiable benchmark with hierarchical failure attribution.

Significance. If the evaluation protocol were valid, the scale and hierarchical decomposition would be a useful contribution: the KB-based generation is transparent, the dataset is large, Tasks 1-2 metrics are standard consistency checks, and the dataset/code are released. However, the headline planning claim (Abstract; Tables 4-6) rests on the Task 3 evaluator, and that evaluator is not a valid measure of model planning. Because the evaluator fills in dense paths with the same A* procedure that generated the ground truth, the reported planning metrics reflect oracle-assisted completion rather than model-generated routing. The current benchmark can at most support claims about waypoint selection, not about constrained route planning. This is an internal correctness issue, not a disagreement with consensus.

major comments (3)
  1. [§4.2, Eqs. (13)–(16)] The Task 3 planning metrics are vacuous as a measure of model planning. For any sample in S_A (all predicted waypoints traversable), the evaluator reconstructs the dense path by A* on the ground-truth cost map W(p) between adjacent waypoints. Since the query supplies S and E and a feasible path exists by construction (§3.3), a model that outputs only S and E is placed in S_A, and the reconstructed \hatτ is exactly the same A*-generated trajectory used to define τ*. Therefore AR=100%, CR=1, and CD=0 for every such sample, independent of the model. This is an oracle-assisted completion metric, not a route-planning metric. The headline claim in the Abstract and Table 6 that MLLMs show significant deficiencies in planning is not supported. The protocol must score the model's own dense output directly, or evaluate its waypoint sequence without A* completion (e.g., against landmarks of τ*).
  2. [§3.3, Eq. (7)] The proof that A* is guaranteed to find the optimal trajectory is incorrect as written. The admissibility argument uses only W(p)≥1, but with 8-neighbor moves on the image grid, Euclidean distance can exceed the minimum path cost: a diagonal run from (0,0) to (100,100) has a path of at least 101 pixels with per-pixel cost ≥1, while the Euclidean heuristic is about 141.4. Unless the A* implementation uses 4-connectivity or a different admissible heuristic (e.g., Chebyshev distance with appropriate per-pixel costs), the 'provably optimal solutions' claim (Abstract; §4.3; Table 3) is not established. Please specify the connectivity and re-derive the heuristic bound.
  3. [§3.2 vs §4.2] The task specification is internally inconsistent. Eq. (2) asks the model to produce a trajectory τ that minimizes cumulative cost, and Fig. 1 shows a dense coordinate path as the expected answer; §4.2 instead evaluates 'predicted sparse waypoints' and reconstructs the dense path. If sparse waypoints are the intended output, the benchmark evaluates waypoint selection, not route planning; if dense output is intended, the evaluator discards the model's trajectory and substitutes its own. This ambiguity must be resolved before the benchmark can support planning claims.
minor comments (5)
  1. [Tables 4–5, §5] 'Close-source Models' should be 'Closed-source Models'.
  2. [Fig. 1 caption] The caption says 'sparse waypoint trajectory' but the shown expected answer is a dense path; align terminology with the Task 3 definition.
  3. [Eq. (10)] PR is undefined when M<2, e.g., Boat/Fastest has a single traversable class; specify a convention for degenerate cases.
  4. [References] Some references are incomplete (e.g., [10] lacks venue information); check model citation consistency and dataset version details.
  5. [Conclusion / affiliations] Typo 'preception' in the conclusion; 'Artifical' in the author affiliation.

Circularity Check

1 steps flagged

Task 3 planning metric is self-definitional: the evaluator's A* reconstruction on the ground-truth cost map makes endpoint-only outputs score as perfectly optimal.

specific steps
  1. self definitional [§4.2 Symbolized Evaluator, Task 3 (Eqs. 13–16); with §3.2 Eq. (1)–(2) and §3.3 constrained trajectory generation]
    "For samples in S_A, the dense trajectory τ̂ is reconstructed by connecting adjacent waypoints using the A-Star search algorithm based on the cost map W(p). ... CR = (1/|S_A|) Σ_{k∈S_A} ( Σ_{p∈τ̂_k} W(p) / Σ_{q∈τ*_k} W(q) )"

    The benchmark gives the model the start and end points and asks it to minimize Σ W(p) (Eq. 2); the ground truth τ* is itself computed by A* on W(p) (Eq. 1, §3.3). Since a feasible S→E path exists by construction, a model emitting only {S,E} is in S_A, and the evaluator re-runs A* on the same cost map to produce τ̂ = τ*. Eq. 14 then gives CR = 1 and Eq. 16 gives CD = 0 for every such sample. Thus perfect planning scores are achieved without planning: the oracle evaluator does the planning. The headline conclusion that MLLMs have planning deficiencies and the reported CR/CD numbers are therefore not evidence about model-generated routes.

full rationale

The only substantive circularity is in the Task 3 evaluator. The ground-truth trajectory is obtained by A* on the author-defined cost map, and the evaluator reconstructs the model's dense path with the same A* on the same cost map, then compares the two. Hence the optimality metrics (CR, CD) measure the evaluator's own optimization, not the model's route, whenever the model's sparse waypoints are traversable; the model need not emit any interior waypoint. This is a reduction by construction of the central planning measurement. There is no fitted-parameter-called-prediction issue, and the symbolic vectors and A* ground truths are deterministic consequences of the stated KB rather than fitted to model outputs. The self-referential use of DeepSeek/Gemini to author and verify queries is a data-quality concern, not a circular derivation. The central benchmark claim — that NeSy-Route evaluates constrained route planning and reveals MLLM planning deficiencies — is therefore partially circular/unsupported in its planning component, though Tasks 1–2 remain independent.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The benchmark is a synthetic test harness; it introduces no new physical entities. Its validity rests on the accuracy of OpenEarthMap labels, the author-defined KB rules, the unspecified pixel-adjacency details of the A* proof, and especially the questionable assumption that oracle-assisted route completion measures model planning.

free parameters (3)
  • Complexity score weights λ1..λ4 (Eq. 8) = not reported
    Hand-chosen weights for difficulty stratification of Task 3; values are not given and affect the Easy/Medium/Hard split.
  • Morphological filtering and erosion parameters = not reported
    Design choices for filling small regions and eroding large land-cover patches are described qualitatively but exact iterations/thresholds are unspecified.
  • Cost scale offset in W(p)=max(V_pref)-V_pref+1 = +1
    The +1 offset is an arbitrary design choice that sets the minimum pixel cost; it does not change optimality ordering but affects CR magnitudes.
axioms (5)
  • domain assumption OpenEarthMap semantic masks are accurate ground-truth land cover.
    All cost maps and GT trajectories inherit errors from these masks; no independent verification of mask accuracy is provided.
  • ad hoc to paper The KB traversability and priority rankings (Tables 1-2) are a valid formalization of real-world route constraints.
    The agent-specific rules are invented by the authors; all GT labels and evaluations are relative to these arbitrary rules.
  • standard math The A* Euclidean heuristic is admissible and consistent on the pixel grid used.
    The proof requires W(p)>=1 and an implicit movement model; the paper never specifies 4- vs 8-connectivity or per-step costs, so the 'provably optimal' claim is under-specified.
  • domain assumption LLM-generated queries faithfully express the symbolic constraints.
    DeepSeek self-inference and Gemini verification reduce but do not eliminate ambiguity or hallucination in the generated text.
  • ad hoc to paper Oracle-assisted dense-path completion from sparse waypoints measures the model's planning ability.
    This is the weakest assumption: the evaluator's A* reconstruction with the ground-truth cost map means a model can output only endpoints and still get perfect planning scores.

pith-pipeline@v1.3.0-alltime-deepseek · 13428 in / 13949 out tokens · 150739 ms · 2026-08-02T18:02:40.415243+00:00 · methodology

0 comments
read the original abstract

Remote sensing underpins crucial applications such as disaster relief and ecological field surveys, where systems must understand complex scenes and constraints and make reliable decisions. Current remote-sensing benchmarks mainly focus on evaluating perception and reasoning capabilities of multimodal large language models (MLLMs). They fail to assess planning capability, stemming either from the difficulty of curating and validating planning tasks at scale or from evaluation protocols that are inaccurate and inadequate. To address these limitations, we introduce NeSy-Route, a large-scale neuro-symbolic benchmark for constrained route planning in remote sensing. Within this benchmark, we introduce an automated data-generation framework that integrates high-fidelity semantic masks with heuristic search to produce diverse route-planning tasks with provably optimal solutions. This allows NeSy-Route to comprehensively evaluate planning across 10,821 route-planning samples, nearly 10 times larger than the largest prior benchmark. Furthermore, a three-level hierarchical neuro-symbolic evaluation protocol is developed to enable accurate assessment and support fine-grained analysis on perception, reasoning, and planning simultaneously. Our comprehensive evaluation of various state-of-the-art MLLMs demonstrates that existing MLLMs show significant deficiencies in perception and planning capabilities. We hope NeSy-Route can support further research and development of more powerful MLLMs for remote sensing. The dataset and code are available at https://mingyang1010.github.io/NeSy-Route/.

Figures

Figures reproduced from arXiv: 2603.16307 by Kun-Yang Yu, Lan-Zhe Guo, Ming Yang, Shi-Yu Tian, Yu-Feng Li, Zhi Zhou.

Figure 1
Figure 1. Figure 1: A typical example from our NeSy-Route Benchmark. NeSy-Route evaluates MLLMs through a hierarchical reasoning pipeline consisting of three integrated tasks. Task 1 involves extracting symbolic traversability and cost vectors from the provided hiker mission scenario. Task 2 focuses on anchoring these symbolic constraints to spe￾cific identified regions within the remote sensing image. Task 3 assesses the cap… view at source ↗
Figure 2
Figure 2. Figure 2: An overview of automated data generation framework. The pipeline integrates dual-LLM logical verification for query synthesis, morphological erosion for semantic visual grounding, and constrained A-Star search algorithm for deriving mathematically optimal trajectories under symbolic rules. clude Gemini-3-Pro [7], GPT-5.1 [31], and Qwen3-VL-Plus [2], which demon￾strate strong capabilities in perception and … view at source ↗

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. LAST: Leveraging Tools as Hints to Enhance Spatial Reasoning for Multimodal Large Language Models

    cs.CV 2026-04 unverdicted novelty 6.0

    LAST augments MLLMs with a tool-abstraction sandbox and three-stage training to deliver around 20% gains on spatial reasoning tasks, outperforming closed-source models.

Reference graph

Works this paper leans on

43 extracted references · 9 linked inside Pith · cited by 1 Pith paper

  1. [1]

    In: arXiv (2025)

    An, X., Xie, Y., Yang, K., Zhang, W., Zhao, X., Cheng, Z., Wang, Y., Xu, S., Chen, C., Wu, C., Tan, H., Li, C., Yang, J., Yu, J., Wang, X., Qin, B., Wang, Y., Yan, Z., Feng, Z., Liu, Z., Li, B., Deng, J.: Llava-onevision-1.5: Fully open framework for democratized multimodal training. In: arXiv (2025)

    2025

  2. [2]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    Barrow, H.G., Tenenbaum, J.M., Bolles, R.C., Wolf, H.C.: Parametric correspon- dence and chamfer matching: Two new techniques for image matching. Tech. rep. (1977)

    1977

  4. [4]

    ai: Dataset for automatic mapping of buildings, woodlands, water and roads from aerial imagery

    Boguszewski, A., Batorski, D., Ziemba-Jankowska, N., Dziedzic, T., Zambrzycka, A.: Landcover. ai: Dataset for automatic mapping of buildings, woodlands, water and roads from aerial imagery. In: CVPR. pp. 1102–1110 (2021)

    2021

  5. [5]

    In: Seminal graphics: pioneering efforts that shaped the field, pp

    Bresenham, J.E.: Algorithm for computer control of a digital plotter. In: Seminal graphics: pioneering efforts that shaped the field, pp. 1–6 (1998)

    1998

  6. [6]

    Proceedings of the IEEE100(10), 2876–2890 (2012)

    Dell’Acqua, F., Gamba, P.: Remote sensing and earthquake damage assessment: Experiences, limits, and perspectives. Proceedings of the IEEE100(10), 2876–2890 (2012)

    2012

  7. [7]

    Model Card (2025),https:// storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model- Card.pdf, accessed: 2026-01-28

    Google DeepMind: Gemini 3 pro: Model card. Model Card (2025),https:// storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model- Card.pdf, accessed: 2026-01-28

    2025

  8. [8]

    In: CVPR

    Guo, X., Lao, J., Dang, B., Zhang, Y., Yu, L., Ru, L., Zhong, L., Huang, Z., Wu, K., Hu, D., He, H., Wang, J., Chen, J., Yang, M., Zhang, Y., Li, Y.: Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In: CVPR. pp. 27672–27683 (June 2024)

    2024

  9. [9]

    The European Physical Journal Special Topics174(1), 65–79 (2009)

    Hamed, K.: Effect of persistence on the significance of kendall’s tau as a measure of correlation between natural time series. The European Physical Journal Special Topics174(1), 65–79 (2009)

    2009

  10. [10]

    Hart, P.E., Nilsson, N.J., Raphael, B.: A formal basis for the heuristic determina- tion of minimum cost paths4(2), 100–107 (1968)

    1968

  11. [11]

    IEEE TCSVT (2026)

    He, X., Wang, J., Su, Y., Liu, D., Zhao, J., Lu, G.: Monotonic rank knowledge distillation via kendall correlation. IEEE TCSVT (2026)

    2026

  12. [12]

    IET Intelligent Transport Systems18(8), 1477–1493 (2024)

    Hu, R., Bai, S., Wen, W., Xia, X., Hsu, L.T.: Towards high-definition vector map construction based on multi-sensor integration for intelligent vehicles: Systems and error quantification. IET Intelligent Transport Systems18(8), 1477–1493 (2024)

    2024

  13. [13]

    Frontiers in Genetics15, 1511456 (2024)

    Huang, T., Niu, S., Zhang, F., Wang, B., Wang, J., Liu, G., Yao, M.: Correlating gene expression levels with transcription factor binding sites facilitates identifica- tion of key transcription factors from transcriptome data. Frontiers in Genetics15, 1511456 (2024)

    2024

  14. [14]

    Biometrika30(1-2), 81–93 (1938)

    Kendall, M.G.: A new measure of rank correlation. Biometrika30(1-2), 81–93 (1938)

    1938

  15. [15]

    In: 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: CVPR. pp. 27831– 27840 (2024)

    2024

  16. [16]

    Remote Sensing17(22), 3774 (2025)

    Lee, D.H., Hong, J.H., Seo, H.W., Oh, H.: Kfgod: A fine-grained object detection dataset in kompsat satellite imagery. Remote Sensing17(22), 3774 (2025)

    2025

  17. [17]

    arXiv preprint arXiv:2410.23828 (2024) 16 M

    Li, K., Dong, F., Wang, D., Li, S., Wang, Q., Gao, X., Chua, T.S.: Show me what and where has changed? question answering and grounding for remote sensing change detection. arXiv preprint arXiv:2410.23828 (2024) 16 M. Yang et al

    Pith/arXiv arXiv 2024

  18. [18]

    ISPRS journal of photogrammetry and remote sensing159, 296–307 (2020)

    Li, K., Wan, G., Cheng, G., Meng, L., Han, J.: Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing159, 296–307 (2020)

    2020

  19. [19]

    Li, T., Wang, C., de Silva, C.W., et al.: Coverage sampling planner for uav-enabled environmentalexplorationandfieldmapping.In:IEEEIROS.pp.2509–2516.IEEE (2019)

    2019

  20. [20]

    NeurIPS37, 3229–3242 (2024)

    Li, X., Ding, J., Elhoseiny, M.: Vrsbench: A versatile vision-language benchmark dataset for remote sensing image understanding. NeurIPS37, 3229–3242 (2024)

    2024

  21. [21]

    2: Pushing the frontier of open large language models

    Liu, A., Mei, A., Lin, B., Xue, B., Wang, B., Xu, B., Wu, B., Zhang, B., Lin, C., Dong, C., et al.: Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    In: CVPR

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: CVPR. pp. 26296–26306 (2024)

    2024

  23. [23]

    IEEE TGRS58(12), 8555–8566 (2020)

    Lobry, S., Marcos, D., Murray, J., Tuia, D.: Rsvqa: Visual question answering for remote sensing data. IEEE TGRS58(12), 8555–8566 (2020)

    2020

  24. [24]

    arxiv 2024

    Luo, J., Pang, Z., Zhang, Y., Wang, T., Wang, L., Dang, B., Lao, J., Wang, J., Chen, J., Tan, Y., et al.: Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding. arxiv 2024. arXiv preprint arXiv:2406.10100

    Pith/arXiv arXiv 2024

  25. [25]

    In: 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: ICCV. pp. 9206–9217 (2025)

    2025

  26. [26]

    arXiv preprint arXiv:2412.04204 (2024)

    Marsocci, V., Jia, Y., Bellier, G.L., Kerekes, D., Zeng, L., Hafner, S., Gerard, S., Brune, E., Yadav, R., Shibli, A., et al.: Pangaea: A global and inclusive benchmark for geospatial foundation models. arXiv preprint arXiv:2412.04204 (2024)

    Pith/arXiv arXiv 2024

  27. [27]

    In: ECCV

    Muhtar, D., Li, Z., Gu, F., Zhang, X., Xiao, P.: Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model. In: ECCV. pp. 440–

  28. [28]

    In: ICCV

    Palladin, E., Brucker, S., Ghilotti, F., Narayanan, P., Bijelic, M., Heide, F.: Self- supervised sparse sensor fusion for long range perception. In: ICCV. pp. 27498– 27509 (2025)

    2025

  29. [29]

    Qwen Team: Qwen3.5: Towards native multimodal agents (February 2026),https: //qwen.ai/blog?id=qwen3.5

    2026

  30. [30]

    In: GLOBECOM 2023-2023 IEEE Global Communications Conference

    Shi, G., Wu, C.C., Hsu, C.H.: Error concealment of dynamic lidar point clouds for connected and autonomous vehicles. In: GLOBECOM 2023-2023 IEEE Global Communications Conference. pp. 5409–5415. IEEE (2023)

    2023

  31. [31]

    arXiv preprint arXiv:2601.03267 (2025)

    Singh, A., Fry, A., Perelman, A., Tart, A., Ganesh, A., El-Kishky, A., McLaughlin, A., Low, A., Ostrow, A., Ananthram, A., et al.: Openai gpt-5 system card. arXiv preprint arXiv:2601.03267 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    Wang, F., Wang, H., Guo, Z., Wang, D., Wang, Y., Chen, M., Ma, Q., Lan, L., Yang, W., Zhang, J., et al.: Xlrs-bench: Could your multimodal llms understand ex- tremely large ultra-high-resolution remote sensing imagery? In: CVPR. pp. 14325– 14336 (2025)

    2025

  33. [33]

    arXiv preprint arXiv:2505.21089 (2025)

    Wang, J., Xuan, W., Qi, H., Liu, Z., Liu, K., Wu, Y., Chen, H., Song, J., Xia, J., Zheng, Z., et al.: Disasterm3: A remote sensing vision-language dataset for disaster damage assessment and response. arXiv preprint arXiv:2505.21089 (2025)

    arXiv 2025

  34. [34]

    In: AAAI

    Wang, J., Zheng, Z., Chen, Z., Ma, A., Zhong, Y.: Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering. In: AAAI. vol. 38, pp. 5481–5489 (2024) NeSy-Route: Neuro-Symbolic Route Planning Benchmark 17

    2024

  35. [35]

    arXiv preprint arXiv:2110.08733 (2021)

    Wang, J., Zheng, Z., Ma, A., Lu, X., Zhong, Y.: Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733 (2021)

    Pith/arXiv arXiv 2021

  36. [36]

    5: Advancing open-source multimodal models in versatility, reasoning, and efficiency

    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)

    Pith/arXiv arXiv 2025

  37. [37]

    IEEE TGRS55(7), 3965–3981 (2017)

    Xia, G.S., Hu, J., Hu, F., Shi, B., Bai, X., Zhong, Y., Zhang, L., Lu, X.: Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE TGRS55(7), 3965–3981 (2017)

    2017

  38. [38]

    In: WACV

    Xia, J., Yokoya, N., Adriano, B., Broni-Bediako, C.: Openearthmap: A benchmark dataset for global high-resolution land cover mapping. In: WACV. pp. 6254–6264 (2023)

    2023

  39. [39]

    arXiv preprint arXiv:2311.01017 (2023)

    Zhang, L., Xiong, Y., Yang, Z., Casas, S., Hu, R., Urtasun, R.: Copilot4d: Learning unsupervised world models for autonomous driving via discrete diffusion. arXiv preprint arXiv:2311.01017 (2023)

    Pith/arXiv arXiv 2023

  40. [40]

    IEEE TGRS62, 1–20 (2024)

    Zhang, W., Cai, M., Zhang, T., Zhuang, Y., Mao, X.: Earthgpt: A universal multi- modallargelanguagemodelformultisensorimagecomprehensioninremotesensing domain. IEEE TGRS62, 1–20 (2024)

    2024

  41. [41]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing9(9), 4117–4128 (2016)

    Zhang, X., Sun, Y., Shang, K., Zhang, L., Wang, S.: Crop classification based on feature band set construction and object-oriented approach using hyperspec- tral images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing9(9), 4117–4128 (2016)

    2016

  42. [42]

    IEEE TGRS60, 1–14 (2021)

    Zheng, X., Wang, B., Du, X., Lu, X.: Mutual attention inception network for remote sensing visual question answering. IEEE TGRS60, 1–14 (2021)

    2021

  43. [43]

    Remote sensing of environment228, 164–182 (2019)

    Zhu, Z., Zhou, Y., Seto, K.C., Stokes, E.C., Deng, C., Pickett, S.T., Taubenböck, H.: Understanding an urbanizing planet: Strategic directions for remote sensing. Remote sensing of environment228, 164–182 (2019)

    2019