Pith. sign in

REVIEW 3 major objections 8 minor 47 references

A training-free parse-search-confirm pipeline with structured spatial memory lets multimodal models navigate high-altitude UAV dialog tasks as well as several finetuned systems.

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 · grok-4.5

2026-07-14 04:57 UTC pith:AMP5DOPO

load-bearing objection Solid first training-free AVDN system: PSC + SSM scaffolding gets off-the-shelf MLLMs to match several finetuned baselines on ANDH/ANDH-Full, with clean ablations. the 3 major comments →

arxiv 2607.11529 v1 pith:AMP5DOPO submitted 2026-07-13 cs.CV

Parse, Search, and Confirmation: Training-Free Aerial Vision-and-Dialog Navigation with Chain-of-Thought Reasoning and Structured Spatial Memory

Yu Qi , Hongyu Li , Shaofei Huang , Tianrui Hui , Yaxiong Wang , Lechao Cheng , Zhun Zhong , Si Liu
show 1 more author
Meng Wang
This is my paper
classification cs.CV
keywords Aerial Vision-and-Dialog Navigationtraining-free navigationchain-of-thought reasoningstructured spatial memorymultimodal large language modelsUAV navigationhigh-altitude remote sensingANDH benchmark
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.

High-altitude UAV navigation from dialog is hard for off-the-shelf multimodal language models: they mishandle abstract directions such as “10 o’clock,” cannot reliably spot tiny texture-sparse landmarks from above, and forget where they have already looked. This paper shows that the failure is not inevitable. By first converting dialogue into clean geometric heading and destination cues, then running a stepwise Search chain-of-thought and a fine-grained Confirmation chain-of-thought, and by feeding the model an explicit Structured Spatial Memory of multi-scale views, past trajectory crops, and a live semantic grid map, the same models can complete Aerial Vision-and-Dialog Navigation without any task-specific training. On the standard ANDH and ANDH-Full benchmarks the resulting system sets the training-free state of the art and matches or beats several supervised finetuning baselines. Resource-constrained aerial agents can therefore inherit modern multimodal reasoning without the cost of re-annotation and re-training for every new environment.

Core claim

The authors establish that a tightly coupled three-stage Parsing-Search-Confirmation reasoning pipeline plus Structured Spatial Memory is sufficient to turn general-purpose multimodal language models into competitive training-free Aerial Vision-and-Dialog Navigation agents, reaching new state-of-the-art numbers among training-free methods on ANDH and ANDH-Full while matching or surpassing several finetuned systems.

What carries the argument

PSC-AVDN: a Parsing stage that normalizes ambiguous dialog into absolute heading and destination text, a Search Chain-of-Thought that progressively localizes candidates under multi-scale views, a Confirmation Chain-of-Thought that verifies spatial relations around those candidates, and a Structured Spatial Memory that continuously supplies multi-scale crops, trajectory-fused visual memory, and an N×N semantic reference grid.

Load-bearing premise

The claim rests on the premise that today’s off-the-shelf multimodal models can already generate accurate semantic grids and perform fine-grained spatial verification on small, weak high-altitude landmarks without any aerial-domain training or external detectors.

What would settle it

Replace the backbone multimodal model with a weaker or differently trained one while keeping the same prompts and memory modules; if success rate and SPL on the ANDH unseen validation set collapse below the simple iterative-search baseline, the claim that the pipeline alone unlocks reliable aerial navigation is falsified.

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

If this is right

  • High-altitude UAV dialog navigation can be deployed without collecting new labeled trajectories or running expensive finetuning for each environment.
  • Abstract clock-face and compass directions become usable once an explicit heading-resolution step converts them into absolute angles.
  • Long-horizon drift is reduced by continuously updating a fused visual trajectory canvas and a live semantic grid map inside the reasoning loop.
  • The same three-stage plus memory pattern can be swapped onto stronger future multimodal backbones and immediately raise training-free aerial performance.
  • Ablations show each stage and each memory component contributes measurable gains, so practitioners can prune the pipeline for compute budgets while retaining most of the benefit.

Where Pith is reading between the lines

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

  • Because the method never relies on external segmentation or GIS priors, it may transfer more cleanly to real UAV hardware where those services are unavailable or expensive.
  • The same structured-memory idea could be tested on continuous low-altitude urban aerial navigation, where landmarks are larger but occlusion and dynamic objects dominate.
  • If future multimodal models improve native spatial grounding, the Confirmation stage might be simplified or even dropped, turning the system into a lighter two-stage parser-search agent.
  • Failure cases on texture-sparse or densely clustered landmarks suggest that a lightweight open-vocabulary detector could be inserted only at the Confirmation step without sacrificing the training-free claim.

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 / 8 minor

Summary. The paper proposes PSC-AVDN, a training-free framework for Aerial Vision-and-Dialog Navigation (AVDN) that couples a three-stage Parsing–Search–Confirmation pipeline with a Structured Spatial Memory (SSM). Parsing uses an LLM plus a Heading Resolution module to convert ambiguous dialogue into absolute heading and destination cues; Search Chain-of-Thought (S-CoT) performs stepwise high-altitude localization; Confirmation Chain-of-Thought (C-CoT) does fine-grained disambiguation. SSM supplies multi-scale visual observation (MVO), spatial visual memory (SVM), and structured geometric memory (SGM via N×N semantic reference grids). On ANDH and ANDH-Full, the method reports new training-free state-of-the-art SPL/SR/GP and matches or exceeds several supervised baselines (Table 1), with ablations of each stage and SSM component (Tables 2–5).

Significance. If the reported numbers hold under independent re-implementation, this is a solid systems contribution: the first training-free AVDN pipeline that is competitive with supervised methods on high-altitude dialog navigation. The three-stage decoupling of directional grounding from localization, together with an explicit spatial memory that does not rely on external detectors (Grounded-SAM, GIS, etc.), is a clear and useful design pattern for MLLM-based aerial agents. Strengths include complete seen/unseen comparisons on both ANDH and ANDH-Full, stage-wise and component-wise ablations, and grid-size/scale-factor sweeps. The promised public code release would further raise impact for resource-efficient UAV navigation research.

major comments (3)
  1. Section 4.2 and Table 1: Primary results depend on proprietary APIs (DeepSeek-V3 for parsing, Qwen-VL-Max for S-CoT/C-CoT). For a training-free SOTA claim that is meant to be resource-efficient and reusable, the main paper should report at least one fully open-weight MLLM configuration with fixed decoding settings (temperature, seed, model version). The brief note that results improve “across different MLLMs (see Appendix 5)” is insufficient without those numbers and prompts in the main text or a released artifact, because closed-model drift makes the Table 1 numbers hard to audit.
  2. Section 3.3 (Eqs. 7–8) and Table 3: SGM is a load-bearing part of SSM—the model is prompted to label an N×N grid with a fixed set of 12 dialogue-derived categories, and the ablation credits SGM with a non-trivial gain (SPL 16.6→17.8). There is no quantitative check of grid-label accuracy, consistency across steps, or failure rate on small/texture-sparse landmarks. Without even a small human-judged sample or agreement metric on generated grids, it remains unclear whether the reported gains come from genuine structured geometry or from incidental prompt effects. A short diagnostic (e.g., label accuracy on 50 frames) would substantially strengthen the central SSM claim.
  3. Tables 1–5: All metrics are single point estimates with no variance, multiple runs, or significance tests, despite stochastic MLLM decoding and a max-3-step search policy. The training-free vs. supervised gaps are large enough that the ranking is unlikely to reverse, but for journal-level SOTA claims the authors should report mean±std over a small number of seeds (or temperature settings) on at least the Unseen Val. splits, or justify why single-run API results are stable.
minor comments (8)
  1. Throughout: “A VDN” / “A VLN” appear with an erroneous space (e.g., Abstract, §1, §2.2); standardize to AVDN/AVLN.
  2. Figure 2: “Confirmtaion” is misspelled; also “l Generate…” bullets look like OCR/prompt artifacts and should be cleaned.
  3. §4.4 / Table 4 caption area: “Gird size” → “Grid size”.
  4. §3.2.1: notation “s dir” / “dir” is inconsistent; define dir and des once and reuse.
  5. Table 1: several supervised methods show “—” on ANDH-Full; briefly state whether those methods were not evaluated on full trajectories or could not be re-run, so the comparison scope is clear.
  6. §4.1: GP is described only qualitatively; a one-line formal definition (or citation to the exact formula in Fan et al.) would help interpret negative GP for GPT-4o.
  7. §3.3: the set C of 12 semantic categories is never listed; include the list (or a pointer to the appendix) so SGM is reproducible.
  8. Figure 4: trajectory visualizations are useful but hard to read in grayscale; add step indices and a short success/failure caption for each panel.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper whose SOTA claims rest on held-out trajectory metrics, not on any equation or self-citation that reduces a prediction to its own inputs.

full rationale

PSC-AVDN is a training-free engineering pipeline (Parsing via DeepSeek-V3 + Heading Resolution, S-CoT/C-CoT via Qwen-VL-Max, SSM with multi-scale crops / trajectory canvas / N×N semantic grid). The central claim is purely empirical: higher SPL/SR/GP than prior training-free baselines (and some supervised ones) on the ANDH / ANDH-Full Unseen Val and Test splits (Table 1). Ablations (Tables 2–5) select grid size, scale factors and stage contributions on Unseen Val; those choices are disclosed and do not redefine the success metrics on held-out trajectories. Formalisms (Eqs. 1–8) are system definitions (instruction-to-bbox mapping, angle wrap, memory update), not first-principles derivations that recover fitted quantities. Related-work citations are external or non-load-bearing; no uniqueness theorem, ansatz, or self-citation chain forces the reported numbers. Standard hyper-parameter selection on validation does not constitute circularity under the stated criteria. Score 0 is therefore the correct, proportionate finding.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 3 invented entities

The central empirical claim rests on a small set of hand-chosen hyper-parameters, standard assumptions about commercial MLLM capabilities, and three engineered memory constructs that are defined inside the paper. No free parameters are fitted to the final test metrics; the invented memory modules are pure algorithmic scaffolding rather than new physical entities.

free parameters (3)
  • reference grid size N = 5×5
    Chosen by ablation (Table 4); 5×5 selected as best. Directly affects SGM and CoT reasoning load.
  • multi-scale factors = (3,5,7)
    Chosen by ablation (Table 5); (3,5,7) selected. Controls the hierarchical visual crops fed to the MLLM.
  • maximum execution steps = 3
    Set to 3 in Implementation Details; limits search horizon per dialogue round.
axioms (3)
  • domain assumption Off-the-shelf MLLMs possess sufficient high-altitude visual-spatial reasoning once given multi-scale crops and an explicit CoT scaffold.
    Core premise of the entire training-free approach (Introduction and §3.2); never proven, only measured.
  • ad hoc to paper A fixed set of 12 high-frequency semantic categories derived from dataset dialogues is adequate to label every grid cell for SGM.
    Stated in §3.3 Structured geometric memory; categories are dataset-specific and not shown to transfer.
  • domain assumption Homography warping of the real-world FOV quadrilateral into a canonical 768×768 frame preserves all geometric relations needed for localization.
    Implementation Details; standard remote-sensing assumption but critical for pixel-to-geo mapping.
invented entities (3)
  • Structured Spatial Memory (SSM) with MVO + SVM + SGM no independent evidence
    purpose: Supply the global spatial context and history that native MLLMs lack during multi-step aerial search.
    Defined in §3.3; purely algorithmic construct with no external physical referent.
  • Search Chain-of-Thought (S-CoT) and Confirmation Chain-of-Thought (C-CoT) no independent evidence
    purpose: Decompose high-altitude target localization into interpretable, progressive reasoning steps that reduce visual ambiguity.
    Defined in §3.2.2–3.2.3; prompt templates rather than new theoretical objects.
  • Heading Resolution (HR) module no independent evidence
    purpose: Convert heterogeneous directional phrases (clock, degrees, compass) into a single absolute angle.
    Rule-based parser in §3.2.1; simple engineering utility.

pith-pipeline@v1.1.0-grok45 · 18793 in / 2947 out tokens · 33867 ms · 2026-07-14T04:57:05.633911+00:00 · methodology

0 comments
read the original abstract

In this paper, we tackle the Aerial Vision-and-Dialog Navigation (AVDN) task in the training-free setting for resource-efficient high-altitude UAV navigation.Naively applying MLLMs leads to unreliable navigation due to weak directional grounding and the lack of explicit spatial memory.To address these issues, we propose PSC-AVDN, a training-free framework that tightly couples a three-stage Parsing-Search-Confirmation reasoning pipeline with a Structured Spatial Memory (SSM).The parsing stage uses an LLM to convert ambiguous dialogue instructions into stable geometric directional and destination cues.A Search Chain-of-Thought (S-CoT) then performs stepwise target exploration under high-altitude observations, and a Confirmation Chain-of-Thought (C-CoT) conducts fine-grained verification around candidate regions to resolve visual ambiguity.Meanwhile, SSM integrates three complementary sources of spatial cues, including multi-scale visual observation, spatial visual memory, and structured geometric memory to provide global spatial context and long-horizon consistency.Extensive experiments on ANDH and ANDH-Full show that PSC-AVDN establishes new state-of-the-art performance in the training-free setting, matching or surpassing several finetuned methods.Code will be publicly available at: https://github.com/QY6616/PSC-AVDN

Figures

Figures reproduced from arXiv: 2607.11529 by Hongyu Li, Lechao Cheng, Meng Wang, Shaofei Huang, Si Liu, Tianrui Hui, Yaxiong Wang, Yu Qi, Zhun Zhong.

Figure 1
Figure 1. Figure 1: Motivation of our method. (a) The MLLM baseline [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall architecture of our proposed Parsing-Search-Confirmation framework for Aerial Vision-and-Dialog Navigation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: SSM Module diagram. A concrete case is presented to [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of navigation trajectories from our PSC-AVDN. (a) represents a two-round dialogue case, and (b) represents a [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 12 linked inside Pith

  1. [1]

    Anderson, Q

    P. Anderson, Q. Wu, D. Teney, J. Bruce, M. Johnson, N. S ¨underhauf, I. Reid, S. Gould, and A. van den Hen- gel. Vision-and-language navigation: Interpreting visually- grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2

    2018

  2. [2]

    S. Bai, K. Chen, X. Liu, et al. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6

    Pith/arXiv arXiv 2025

  3. [3]

    L. M. Schulze Buschoff et al. Visual cognition in multimodal large language models.Nature Machine Intelligence, 7(1),

  4. [4]

    B. Chen, Z. Xu, S. Kirmani, et al. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3

    2024

  5. [5]

    A. C. Cheng, H. Yin, Y . Fu, et al. Spatialrgpt: Grounded spatial reasoning in vision-language models.Advances in Neural Information Processing Systems (NeurIPS), 37, 2024. 3

    2024

  6. [6]

    Dong et al

    Y . Dong et al. Insight-v: Exploring long-chain visual reason- ing with multimodal large language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), 2025. 3

    2025

  7. [7]

    Y . Fan, W. Chen, T. Jiang, and et al. Aerial vision-and-dialog navigation. InFindings of the Association for Computational Linguistics (ACL Findings), 2023. 1, 3, 6

    2023

  8. [8]

    Y . Gao, Z. Wang, P. Han, et al. Exploring spatial representa- tion to enhance llm reasoning in aerial vision-language nav- igation.arXiv preprint arXiv:2410.08500, 2024. 3

    Pith/arXiv arXiv 2024

  9. [9]

    Y . Gao, C. Li, Z. You, et al. Openfly: A comprehensive plat- form for aerial vision-language navigation.arXiv preprint arXiv:2502.18041, 2025. 3, 6

    arXiv 2025

  10. [10]

    Greff, R

    K. Greff, R. K. Srivastava, J. Koutn ´ık, B. R. Steunebrink, and J. Schmidhuber. Lstm: A search space odyssey.IEEE Transactions on Neural Networks and Learning Systems, 28 (10), 2016. 6

    2016

  11. [11]

    J. Han, K. Gong, Y . Zhang, et al. Onellm: One framework to align all modalities with language. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2

    2024

  12. [12]

    Y . Hu, W. Shi, X. Fu, et al. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models.Ad- vances in Neural Information Processing Systems(NeurIPS), 37, 2024. 3

    2024

  13. [13]

    Huang, R

    S. Huang, R. Ling, H. Li, et al. Unleashing the temporal- spatial reasoning capacity of gpt for training-free audio and language referenced video object segmentation. InPro- ceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 3

    2025

  14. [14]

    H. Ju, S. Huang, S. Liu, et al. Video2bev: Transforming drone videos to bevs for video-based geo-localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision(ICCV), 2025. 1

    2025

  15. [15]

    H. Ju, S. Huang, H. Li, et al. From instruction to event: Sound-triggered mobile manipulation.arXiv preprint arXiv:2601.21667, 2026. 2

    Pith/arXiv arXiv 2026

  16. [16]

    Krantz, E

    J. Krantz, E. Wang, Q. Wu, A. Clegg, N. Snavely, D. Batra, P. Anderson, and S. Lee. Beyond the nav-graph: Vision-and- language navigation in continuous environments. InPro- ceedings of the European Conference on Computer Vision (ECCV), 2020. 3

    2020

  17. [17]

    A. Ku, P. Anderson, R. Patel, E. Ie, and J. Baldridge. Room- across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. InProceedings of the 2020 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP), 2020. 2

    2020

  18. [18]

    Kucharczyk and C

    M. Kucharczyk and C. H. Hugenholtz. Remote sensing of natural hazard-related disasters with small drones: Global trends, biases, and research opportunities.Remote Sensing of Environment, 264, 2021. 1

    2021

  19. [19]

    J. Lee, T. Miyanishi, S. Kurita, K. Sakamoto, D. Azuma, Y . Matsuo, and N. Inoue. Citynav: A large-scale dataset for real-world aerial navigation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025. 1, 3

    2025

  20. [20]

    C. Li, W. Wu, H. Zhang, et al. Imagine while reasoning in space: Multimodal visualization-of-thought.arXiv preprint arXiv:2501.07542, 2025. 3

    Pith/arXiv arXiv 2025

  21. [21]

    H. Li, J. Chen, Z. Wei, et al. Llava-st: A multimodal large language model for fine-grained spatial-temporal un- derstanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2

    2025

  22. [22]

    K. Lin, X. E. Wang, L. Zhou, Z. Yu, Z. Wang, Y . Chen, and J. Gao. Learning vision-and-language navigation from youtube videos. InProceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), 2023. 3

    2023

  23. [23]

    A. Liu, B. Feng, B. Xue, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 2, 6

    Pith/arXiv arXiv 2024

  24. [24]

    Y . Liu, F. Yao, Y . Yue, G. Xu, X. Sun, and K. Fu. Navagent: Multi-scale urban street view fusion for uav embodied vision-and-language navigation.arXiv preprint arXiv:2411.08579, 2024. 3

    Pith/arXiv arXiv 2024

  25. [25]

    Y . Long, X. Li, W. Cai, et al. Discuss before moving: Visual language navigation via multi-expert discussions. InPro- ceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2024. 3

    2024

  26. [26]

    G. Luo, X. Yang, W. Dou, et al. Mono-internvl: Pushing the boundaries of monolithic multimodal large language mod- els with endogenous visual pre-training. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2

    2025

  27. [27]

    Moudgil, A

    A. Moudgil, A. Majumdar, K. Krishna, S. Lee, P. Anderson, and D. Batra. Soat: A scene- and object-aware transformer for vision-and-language navigation. InAdvances in Neural Information Processing Systems (NeurIPS), 2021. 2

    2021

  28. [28]

    Pashevich, C

    A. Pashevich, C. Schmid, and C. Sun. Episodic transformer for vision-and-language navigation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6

    2021

  29. [29]

    Ping et al

    Y . Ping et al. Multimodal large language models-enabled uav swarm: Towards efficient and intelligent autonomous aerial systems.arXiv preprint arXiv:2506.12710, 2025. 1

    Pith/arXiv arXiv 2025

  30. [30]

    Y . Qiao, W. Lyu, H. Wang, et al. Open-nav: Exploring zero- shot vision-and-language navigation in continuous environ- ment with open-source llms. InProceedings of the IEEE In- ternational Conference on Robotics and Automation (ICRA),

  31. [31]

    G. Raja, A. Manoharan, and H. Siljak. Ugen: Uav and gan- aided ensemble network for post-disaster survivor detection through oran.IEEE Transactions on Vehicular Technology, 73(7), 2024. 1

    2024

  32. [32]

    T. Ren, S. Liu, A. Zeng, and et al. Grounded sam: As- sembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159, 2024. 3

    Pith/arXiv arXiv 2024

  33. [33]

    Shahriar, B

    S. Shahriar, B. Lund, N. R. Mannuru, M. Arshad, K. Hayawi, R. Varma Kumar Bevara, A. Mannuru, and L. Batool. Putting gpt-4o to the sword: A comprehensive evaluation of lan- guage, vision, speech, and multimodal proficiency.arXiv preprint arXiv:2407.09519, 2024. 6

    Pith/arXiv arXiv 2024

  34. [34]

    H. Shao, S. Qian, H. Xiao, et al. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning.Advances in Neural Information Processing Systems(NeurIPS), 37, 2024. 3

    2024

  35. [35]

    Y . Su, D. An, Y . Xu, et al. Target-grounded graph-aware transformer for aerial vision-and-dialog navigation.arXiv preprint arXiv:2308.11561, 2023. 1, 3, 6

    Pith/arXiv arXiv 2023

  36. [36]

    Y . Su, D. An, K. Chen, W. Yu, B. Ning, Y . Ling, Y . Huang, and L. Wang. Learning fine-grained alignment for aerial vision-dialog navigation. InProceedings of the AAAI Con- ference on Artificial Intelligence (AAAI 2025), 2025. 1, 3, 6

    2025

  37. [37]

    Y . Tian, F. Lin, Y . Li, et al. Uavs meet llms: Overviews and perspectives towards agentic low-altitude mobility.Informa- tion Fusion, 122, 2025. 1

    2025

  38. [38]

    T. Wang, Y . Shi, F. Deuser, et al. The 3rd workshop on uavs in multimedia: Capturing the world from a new perspective. InProceedings of the 3rd International Workshop on UAVs in Multimedia: Capturing the World from a New Perspective,

  39. [39]

    X. Wang, D. Yang, Z. Wang, H. Kwan, J. Chen, W. Wu, H. Li, Y . Liao, and S. Liu. Towards realistic uav vision-language navigation: Platform, benchmark, and methodology. InPro- ceedings of the International Conference on Learning Rep- resentations (ICLR), 2025. 3

    2025

  40. [40]

    J. Xiao, Y . Sun, Y . Shao, et al. Uav-on: A benchmark for open-world object goal navigation with aerial agents. InPro- ceedings of the 33rd ACM International Conference on Mul- timedia (ACM MM), 2025. 1

    2025

  41. [41]

    H. Xu, Y . Hu, C. Gao, et al. Geonav: Empowering mllms with explicit geospatial reasoning abilities for language-goal aerial navigation.arXiv preprint arXiv:2504.09587, 2025. 3

    arXiv 2025

  42. [42]

    J. Yang, S. Yang, A. W. Gupta, et al. Thinking in space: How multimodal large language models see, remember, and recall spaces. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3

    2025

  43. [43]

    F. Yu, Q. Deng, S. Tang, Y . Li, and L. Cheng. Open-world 3d scene graph generation for retrieval-augmented reason- ing. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2026. 3

    2026

  44. [44]

    Zhang, Y

    H. Zhang, Y . Yu, L. Wang, et al. Advancing audio-visual navigation through multi-agent collaboration in 3d environ- ments.arXiv preprint arXiv:2509.22698, 2025. 3

    arXiv 2025

  45. [45]

    B. Zhao, J. Fang, Z. Dai, et al. Urbanvideo-bench: Benchmarking vision-language models on embodied intel- ligence with video data in urban spaces.arXiv preprint arXiv:2503.06157, 2025. 1

    arXiv 2025

  46. [46]

    Zhong, C

    L. Zhong, C. Gao, Z. Ding, et al. Topv-nav: Unlocking the top-view spatial reasoning potential of mllm for zero-shot object navigation.arXiv preprint arXiv:2411.16425, 2024. 3

    Pith/arXiv arXiv 2024

  47. [47]

    D. Zhu, J. Chen, X. Shen, et al. Minigpt-4: Enhancing vision-language understanding with advanced large language models.arXiv preprint arXiv:2304.10592, 2023. 2

    Pith/arXiv arXiv 2023