Pith. sign in

REVIEW 4 major objections 4 cited by

SpatialFly closes the 2D-to-3D gap in UAV vision-and-language navigation by injecting geometric priors into RGB tokens—no explicit 3D reconstruction required.

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

2026-07-13 21:04 UTC pith:NFFUBLZ2

load-bearing objection The SpatialFly abstract claims a useful RGB-only geometry-guided reparameterization for continuous UAV VLN, but the provided full text is a different paper, so the central empirical claims cannot be audited. the 4 major comments →

arxiv 2603.21046 v2 pith:NFFUBLZ2 submitted 2026-03-22 cs.CV cs.AI

SpatialFly: Implicit 3D Prior-Guided Visual Reparameterization for Continuous UAV Vision-and-Language Navigation

Wen Jiang , Kangyao Huang , Li Wang , Wang Xu , Wei Fan , Jinyuan Liu , Shaoyu Liu , Hanfang Liang
show 4 more authors
Hongwei Duan Bin Xu Xiangyang Ji Huaping Liu
This is my paper
classification cs.CV cs.AI
keywords UAV vision-and-language navigationspatial representationgeometric prior injectionvisual reparameterizationcontinuous trajectory decisionRGB-only navigationcross-modal attention
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.

UAV vision-and-language navigation is hard because the agent sees flat RGB images but must choose continuous paths through 3D space. SpatialFly argues that this structural mismatch, not just weak perception, is what limits spatial reasoning. The framework therefore keeps pure RGB input and never builds an explicit 3D map. Instead it injects scene-level geometric cues into ordinary 2D semantic tokens and then reparameterizes those tokens with geometry-conditioned attention and gated residual fusion. The claim is that this geometry-guided 2D representation is enough to produce better-aligned, smoother trajectories and to beat prior UAV VLN systems on both seen and unseen environments, including a 4.03 m drop in navigation error and a 1.27 % gain in success rate on the hardest unseen split. A sympathetic reader cares because the same lightweight idea could make language-guided drones safer and more reliable without expensive depth sensors or reconstruction pipelines.

Core claim

SpatialFly shows that the structural mismatch between 2D visual perception and continuous 3D trajectory decisions can be mitigated, without explicit 3D reconstruction, by injecting global geometric priors into 2D semantic tokens and then adaptively reparameterizing those tokens with geometry-conditioned cross-modal attention and gated residual fusion; the resulting representation yields lower navigation error, higher success rate, and smoother path alignment than prior UAV VLN baselines in both seen and unseen settings.

What carries the argument

Geometry-guided 2D adaptive representation: a geometric prior injection module that embeds scene-level structural cues into 2D semantic tokens, followed by a geometry-aware reparameterization module that uses geometry-conditioned cross-modal attention and gated residual fusion to reshape the visual tokens for 3D trajectory decisions.

Load-bearing premise

The method assumes that scene-level geometric cues injected into ordinary 2D image tokens are strong and accurate enough, without any true 3D reconstruction, to resolve the mismatch that otherwise blocks good continuous flight decisions.

What would settle it

Run the same models on the unseen Full split with the geometric prior injection and reparameterization modules ablated or replaced by pure semantic tokens; if navigation error no longer drops by roughly 4 m and success rate does not rise, the claim that geometry-guided 2D reparameterization is what drives the gains is falsified.

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

If this is right

  • UAV VLN systems can improve spatial reasoning while remaining RGB-only, avoiding the cost and fragility of online 3D reconstruction.
  • Trajectory smoothness and path alignment become measurable benefits of the same representation change that reduces navigation error.
  • The same prior-injection plus gated reparameterization pattern can be tested as a drop-in visual front-end for other continuous aerial or mobile agents that receive language goals.
  • Gains on the unseen Full split imply better generalization when the agent faces new layouts rather than only new wording of instructions.

Where Pith is reading between the lines

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

  • If the geometric priors are themselves estimated from monocular cues, the method may inherit the same depth-ambiguity failures that explicit reconstruction tries to solve, only more quietly.
  • The approach suggests a broader design pattern: treat the 2D–3D mismatch as a representation reparameterization problem rather than a mapping or mapping-free dichotomy.
  • A natural next stress test is whether the same modules still help when language instructions become long-horizon or when wind and dynamics make the continuous action space far less forgiving.

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

4 major / 0 minor

Summary. The manuscript (as provided under the SpatialFly header) studies the robustness of medical vision–language models (Med-VLMs) on ultrasound multiple-choice QA. It argues that Med-VLMs are sensitive to small, clinically plausible prompt variations and proposes a black-box LLM-driven attack that generates minimal meaning-preserving edits, selected via Monte Carlo Tree Search using the target model’s candidate log-likelihoods and logit margins. On the U2-Bench disease-diagnosis subset, three Med-VLMs (MedGemma, LLaVA-Med, QoQ-Med) are attacked with Qwen-7B, Qwen-30B, and GPT-4.1 mini. Post-attack accuracy falls substantially (e.g., from ~40% pre-attack to the mid-teens with Qwen-7B); smaller attackers can be more effective but less natural; success concentrates on low logit-margin examples; and shallow MCTS depths often suffice. The authors conclude that realistic prompt variability is a deployment risk and sketch mitigation via post-training on successful attacks and margin-aware fine-tuning.

Significance. If the empirical findings hold under broader evaluation, the work is a useful contribution to Med-VLM trustworthiness: it targets realistic clinical phrasing rather than overtly malicious jailbreaks, uses a scalable black-box pipeline (LLM edits + MCTS), and links attack success to logit margin in a way that suggests concrete defenses. Strengths include a clear experimental design on a public ultrasound QA benchmark, multi-attacker and multi-target comparisons, PPL/semantic-similarity filtering for naturalness, and an ablation showing MCTS depth matters. The absolute accuracy drops are large relative to the modest pre-attack baseline and are therefore practically concerning for POCUS-style deployment. The contribution is primarily empirical and evaluation-oriented rather than a new model architecture.

major comments (4)
  1. Identity / content mismatch: the submission is labeled SpatialFly (UAV VLN, arXiv 2603.21046) with an abstract claiming geometry-guided 2D reparameterization and NE/SR gains, but the full manuscript text is an unrelated paper on LLM-driven prompt attacks for ultrasound Med-VLMs (arXiv 2603.21047). No SpatialFly method, equations, or UAV results can be reviewed. This must be corrected before any scientific assessment of SpatialFly is possible.
  2. Table 1 and §3 (Datasets/Experimental Settings): evaluation is restricted to instances the target already answers correctly, and accuracy is reported only on that subset. Absolute post-attack accuracies (e.g., 13.72%) therefore overstate vulnerability relative to the full test distribution and make cross-model comparison sensitive to which items each model got right initially. Report attack success rate on the full set (or at least on all initially correct items with clear denominators) and include initially incorrect items for completeness.
  3. §2 (MCTS Score updates) and RQ3/Fig. 3: the search reward is the confidence gap between top-1 and top-2 under the child question, while success is defined as flipping away from ground truth. When top-1 is already wrong, or when the gap is measured without anchoring to the true label, the UCT objective can favor high-confidence wrong answers rather than minimal meaning-preserving flips of correct answers. Clarify whether the reward is always (true-label logit − best-other) as suggested later in RQ3, and show that the reported shallow-depth successes are not artifacts of an unanchored gap.
  4. RQ1–RQ2 / Table 1: the strongest attacker (Qwen-7B) also produces less natural edits (higher PPL, lower Sim. in Table 2); after PPL<15 filtering the accuracy gap narrows. The central claim that “clinically plausible” minimal edits suffice needs a stricter naturalness protocol (human clinician ratings or a fixed edit-distance/semantic budget applied before measuring ASR), otherwise the headline drops partly reflect non-minimal or off-distribution rewrites.

Circularity Check

0 steps flagged

No circular derivation found for SpatialFly; abstract claims are empirical method + external-baseline comparisons, and the supplied full manuscript is a different paper.

full rationale

The SpatialFly abstract frames a geometry-guided 2D adaptive representation (geometric prior injection into 2D semantic tokens; geometry-aware reparameterization via cross-modal attention and gated residual fusion) and reports NE/SR gains versus SOTA UAV VLN baselines on seen and unseen splits. That is an empirical architecture-plus-benchmark claim, not a first-principles derivation that reduces a 'prediction' to a fitted input by construction. No equations, uniqueness theorems, self-citation load-bearing premises, or fitted-parameter-as-prediction steps appear in the available SpatialFly text. The CACHEABLE full manuscript is not SpatialFly (2603.21046) at all—it is the unrelated medical VLM prompt-attack paper (2603.21047)—so no SpatialFly architecture equations, ablations, or trajectory metrics can be inspected for circular reduction. Under the hard rules, absence of a quotable self-definitional or fitted-input reduction yields score 0 with empty steps. Generic risks (benchmark overfitting, uninspectable priors) are not circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 1 invented entities

Abstract-only review. Load-bearing premises are domain assumptions about UAV VLN and the sufficiency of implicit geometric guidance from RGB. No free parameters or invented physical entities are specified numerically in the abstract; module names are method components, not new ontological entities with independent evidence.

axioms (3)
  • domain assumption A structural representation mismatch between 2D visual perception and the continuous 3D trajectory decision space is a primary bottleneck for UAV VLN performance.
    Stated as the key difficulty motivating SpatialFly in the abstract; if other factors dominate (control, language grounding, sim-to-real), the method's focus is misaligned.
  • ad hoc to paper Global structural/geometric cues can be injected into 2D semantic tokens from RGB alone without explicit 3D reconstruction and still provide useful scene-level geometric guidance.
    Core methodological premise of the geometric prior injection module; not independently justified in the abstract.
  • ad hoc to paper Geometry-conditioned cross-modal attention and gated residual fusion can adaptively reparameterize visual tokens in a way that improves navigation metrics.
    Defines the geometry-aware reparameterization module; success of the central claim depends on this design choice working as described.
invented entities (1)
  • SpatialFly geometry-guided 2D adaptive representation (prior injection + geometry-aware reparameterization) no independent evidence
    purpose: Bridge 2D RGB perception and 3D continuous UAV trajectory decisions without explicit reconstruction.
    Named framework/modules introduced by the paper; independent evidence outside this work is not established in the abstract.

pith-pipeline@v1.1.0-grok45 · 13693 in / 2502 out tokens · 25692 ms · 2026-07-13T21:04:59.516247+00:00 · methodology

0 comments
read the original abstract

UAVs play an important role in applications such as autonomous exploration, disaster response, and infrastructure inspection. However, UAV VLN in complex 3D environments remains challenging. A key difficulty is the structural representation mismatch between 2D visual perception and the 3D trajectory decision space, which limits spatial reasoning. To this end, we propose SpatialFly, a geometry-guided spatial representation framework for UAV VLN. Operating on RGB observations without explicit 3D reconstruction, SpatialFly introduces a geometry-guided 2D adaptive representation mechanism. Specifically, the geometric prior injection module injects global structural cues into 2D semantic tokens to provide scene-level geometric guidance. The geometry-aware reparameterization module then uses geometry-conditioned cross-modal attention and gated residual fusion to adaptively reparameterize the visual tokens. Experimental results show that SpatialFly consistently outperforms state-of-the-art UAV VLN baselines across both seen and unseen environments, reducing NE by 4.03m and improving SR by 1.27% over the strongest baseline on the unseen Full split. Additional trajectory-level analysis shows that SpatialFly produces trajectories with better path alignment and smoother, more stable motion.

discussion (0)

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

Forward citations

Cited by 4 Pith papers

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

  1. DynFly: Dynamic-Aware Continuous Trajectory Generation for UAV Vision-Language Navigation in Urban Environments

    cs.RO 2026-06 unverdicted novelty 6.0

    DynFly bridges high-level UAV navigation reasoning to continuous motion via B-spline trajectory generation with flow matching and UAV-specific dynamic supervision, yielding metric gains on the OpenUAV benchmark.

  2. See-and-Reach: Precise Vision-Language Navigation for UAVs within the Field of View

    cs.CV 2026-06 unverdicted novelty 6.0

    Introduces UAV-VLN-FOV task and 3DG-VLN framework for precise target-visible UAV navigation, reporting 13.82% success rate gain on a new 2,717-trajectory benchmark with code released.

  3. Uni-LaViRA: Language-Vision-Robot Actions Translation for Unified Embodied Navigation

    cs.RO 2026-05 unverdicted novelty 6.0

    A zero-shot unified agent for VLN-CE, ObjectNav, EQA and Aerial-VLN on wheeled, quadruped, humanoid and UAV platforms that translates language and vision inputs into actions via MLLMs plus TDM and SCB mechanisms, matc...

  4. DynFly: Dynamic-Aware Continuous Trajectory Generation for UAV Vision-Language Navigation in Urban Environments

    cs.RO 2026-06 unverdicted novelty 5.0

    DynFly adds a B-spline and flow-matching trajectory layer with UAV-specific dynamic losses to existing UAV-VLN systems, yielding 4.69 NDTW and 4.51 m NE gains on the OpenUAV unseen split.

Reference graph

Works this paper leans on

31 extracted references · 8 linked inside Pith · cited by 3 Pith papers

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

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

    Pith/arXiv arXiv 2023

  2. [2]

    Journal of the American College of Radiology16(9), 1318–1328 (2019)

    Akkus, Z., Cai, J., Boonrod, A., Zeinoddini, A., Weston, A.D., Philbrick, K.A., Erickson, B.J.: A survey of deep-learning applications in ultrasound: Artificial intelligence–powered ultrasound for improving clinical workflow. Journal of the American College of Radiology16(9), 1318–1328 (2019)

    2019

  3. [3]

    Journal of Evaluation in Clinical Practice31(1), e14250 (2025)

    Armitage, R.C.: Implications of large language models for clinical practice: Ethi- cal analysis through the principlism framework. Journal of Evaluation in Clinical Practice31(1), e14250 (2025)

    2025

  4. [4]

    Nature Communications16(1), 1239 (2025)

    Clusmann, J., Ferber, D., Wiest, I.C., Schneider, C.V., Brinker, T.J., Foersch, S., Truhn, D., Kather, J.N.: Prompt injection attacks on vision language models in oncology. Nature Communications16(1), 1239 (2025)

    2025

  5. [5]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025), https: //openreview.net/forum?id=ZwCVFBFUFb

    Dai, W., Chen, P., Ekbote, C., Liang, P.P.: Qoq-med: Building multimodal clin- ical foundation models with domain-aware GRPO training. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025), https: //openreview.net/forum?id=ZwCVFBFUFb

    2025

  6. [6]

    In: International Conference on Learning Representations (2020)

    Ding, G.W., Sharma, Y., Lui, K.Y.C., Huang, R.: Mma training: Direct input space margin maximization through adversarial training. In: International Conference on Learning Representations (2020)

    2020

  7. [7]

    In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

    Gan, E., Zhao, Y., Cheng, L., Yancan, M., Goyal, A., Kawaguchi, K., Kan, M.Y., Shieh, M.: Reasoning robustness of llms to adversarial typographical errors. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. pp. 10449–10459 (2024)

    2024

  8. [8]

    Goktas, P., Gulseren, D., Tobin, A.M.: Large language and vision assistant in dermatology:agamechangerorjusthype?ClinicalandExperimentalDermatology 49(8), 783–792 (2024)

    2024

  9. [9]

    In: Proceedings of the 32nd ACM international conference on multimedia

    Guo, X., Chai, W., Li, S.Y., Wang, G.: Llava-ultra: Large chinese language and vision assistant for ultrasound. In: Proceedings of the 32nd ACM international conference on multimedia. pp. 8845–8854 (2024)

    2024

  10. [10]

    Kocsis, L., Szepesvári, C., Willemson, J.: Improved monte-carlo search. Univ. Tartu, Estonia, Tech. Rep1, 1–22 (2006)

    2006

  11. [11]

    In: The Fourteenth International Conference on Learning Representations (2026), https://openreview.net/forum? id=jU10qDevGg

    Le, A., Liu, H., wangyue, Liu, Z., Zhu, R., Weng, T., Yu, J., Wang, B., Wu, Y., Yan, K., Sun, Q., Jiang, M., Pei, J., Liu, S., Zheng, H., Li, Z., Noble, A., Souquet, J., Guo, X., Lin, M., Guo, H.: U2-BENCH: Benchmarking large vision- language models on ultrasound understanding. In: The Fourteenth International Conference on Learning Representations (2026)...

    2026

  12. [12]

    Advances in Neural Information Processing Systems36, 28541–28564 (2023) 10 Medghalchi et al

    Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems36, 28541–28564 (2023) 10 Medghalchi et al

    2023

  13. [13]

    medRxiv pp

    Li, Y., Liu, Y., Wang, Z., Liang, X., Liu, L., Wang, L., Cui, L., Tu, Z., Wang, L., Zhou, L.: A comprehensive study of gpt-4v’s multimodal capabilities in medical imaging. medRxiv pp. 2023–11 (2023)

    2023

  14. [14]

    arXiv preprint arXiv:2402.11217 (2024)

    Liu, J., Wang, W., Su, Y., Huan, J., Chen, W., Zhang, Y., Li, C.Y., Chang, K.J., Xin, X., Shen, L., et al.: A spectrum evaluation benchmark for medical multi-modal large language models. arXiv preprint arXiv:2402.11217 (2024)

    Pith/arXiv arXiv 2024

  15. [15]

    In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Lu, Y., Bartolo, M., Moore, A., Riedel, S., Stenetorp, P.: Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 8086–8098 (2022)

    2022

  16. [16]

    In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

    Marchisio, K., Ko, W.Y., Bérard, A., Dehaze, T., Ruder, S.: Understanding and mitigating language confusion in llms. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. pp. 6653–6677 (2024)

    2024

  17. [17]

    arXiv preprint arXiv:2404.16873 (2024)

    Paulus, A., Zharmagambetov, A., Guo, C., Amos, B., Tian, Y.: Advprompter: Fast adaptive adversarial prompting for llms. arXiv preprint arXiv:2404.16873 (2024)

    Pith/arXiv arXiv 2024

  18. [18]

    arXiv preprint arXiv:2402.09262 (2024)

    Royer, C., Menze, B., Sekuboyina, A.: Multimedeval: A benchmark and a toolkit for evaluating medical vision-language models. arXiv preprint arXiv:2402.09262 (2024)

    arXiv 2024

  19. [19]

    In: ICASSP 2023 - 2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Sadrizadeh, S., Aghdam, A.D., Dolamic, L., Frossard, P.: Targeted adversarial attacks against neural machine translation. In: ICASSP 2023 - 2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5 (2023). https://doi.org/10.1109/ICASSP49357.2023.10095342

    doi:10.1109/icassp49357.2023.10095342 2023

  20. [20]

    Schechter Vera, H., Dua, S., Zhang, B., et al.: Embeddinggemma: Powerful and lightweight text representations (2025), https://arxiv.org/abs/2509.20354

    Pith/arXiv arXiv 2025

  21. [21]

    arXiv preprint arXiv:2507.05201 (2025)

    Sellergren, A., Kazemzadeh, S., Jaroensri, T., Kiraly, A., Traverse, M., Kohlberger, T., Xu, S., Jamil, F., Hughes, C., Lau, C., et al.: Medgemma technical report. arXiv preprint arXiv:2507.05201 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    arXiv preprint arXiv:2509.14977 (2025)

    She, C., Lu, R., Chen, L., Wang, W., Huang, Q.: Echovlm: Dynamic mixture-of- experts vision-language model for universal ultrasound intelligence. arXiv preprint arXiv:2509.14977 (2025)

    Pith/arXiv arXiv 2025

  23. [23]

    In: The Twelfth International Conference on Learning Representations (2024)

    Sun, J., Shaib, C., Wallace, B.C.: Evaluating the zero-shot robustness of instruction-tuned language models. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  24. [24]

    Team, G.: Gemma 3 (2025), https://goo.gle/Gemma3Report

    2025

  25. [25]

    Team, Q.: Qwen2.5: A party of foundation models (September 2024), https:// qwenlm.github.io/blog/qwen2.5/

    2024

  26. [26]

    In: NeurIPS (2023)

    Wang, B., Chen, W., Pei, H., Xie, C., Kang, M., Zhang, C., Xu, C., Xiong, Z., Dutta, R., Schaeffer, R., et al.: Decodingtrust: A comprehensive assessment of trustworthiness in gpt models. In: NeurIPS (2023)

    2023

  27. [27]

    arXiv preprint arXiv:2310.09909 (2023)

    Wu,C.,Lei,J.,Zheng,Q.,Zhao,W.,Lin,W.,Zhang,X.,Zhou,X.,Zhao,Z.,Zhang, Y., Wang, Y., et al.: Can gpt-4v (ision) serve medical applications? case studies on gpt-4v for multimodal medical diagnosis. arXiv preprint arXiv:2310.09909 (2023)

    Pith/arXiv arXiv 2023

  28. [28]

    Advances in Neural Information Processing Systems37, 140334–140365 (2024)

    Xia, P., Chen, Z., Tian, J., Gong, Y., Hou, R., Xu, Y., Wu, Z., Fan, Z., Zhou, Y., Zhu, K., et al.: Cares: A comprehensive benchmark of trustworthiness in medical vision language models. Advances in Neural Information Processing Systems37, 140334–140365 (2024)

    2024

  29. [29]

    Information Fusion117, 102888 (2025)

    Xiao, H., Zhou, F., Liu, X., Liu, T., Li, Z., Liu, X., Huang, X.: A comprehensive surveyoflargelanguagemodelsandmultimodallargelanguagemodelsinmedicine. Information Fusion117, 102888 (2025)

    2025

  30. [30]

    arXiv preprint arXiv:2505.09388 (2025) Title Suppressed Due to Excessive Length 11

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025) Title Suppressed Due to Excessive Length 11

    Pith/arXiv arXiv 2025

  31. [31]

    Radiology311(2), e233270 (2024)

    Zhou, Y., Ong, H., Kennedy, P., Wu, C.C., Kazam, J., Hentel, K., Flanders, A., Shih, G., Peng, Y.: Evaluating gpt-4v (gpt-4 with vision) on detection of radiologic findings on chest radiographs. Radiology311(2), e233270 (2024)

    2024