Pith. sign in

REVIEW 3 major objections 4 minor 38 references

A frozen vision-language model can navigate a UAV more reliably by generating many candidate waypoints, self-correcting each one, and choosing by a safety-weighted score — no retraining needed.

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 12:50 UTC pith:ACH4F2EC

load-bearing objection The test-time scaling recipe is sensible, but the SOTA claim hinges on an undisclosed p_tar that may be the ground-truth goal; that must be fixed before the results mean anything. the 3 major comments →

arxiv 2607.19288 v1 pith:ACH4F2EC submitted 2026-07-21 cs.CV cs.RO

No Training, Better Flights: Test-Time Scaled VLMs for UAV Navigation

Feinan Cheng , Dongliang Xu , Wenli Nong , Zhiheng Zhang , Ang Liu , Tianyu Wang , Yue Yao This is my paper
classification cs.CV cs.RO
keywords test-time scalingvision-language navigationUAV navigationself-correctionparallel explorationmulti-criteria scoringTravelUAV benchmarkfrozen VLM
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 aims to show that test-time scaling — spending more compute during inference — can substitute for extra training in aerial vision-language navigation. Its method wraps a frozen navigation VLM in three stages: sample N candidate waypoints in parallel, refine each through a 'Wait, let's reconsider' self-correction prompt, then pick the winner with a weighted score combining safety, goal alignment, and forward progress. On the three TravelUAV test sets, this configuration beats the previous best method (TravelUAV) and improves with both the number of candidates and the number of refinement rounds. The significance is practical: if correct, operators can get better and safer drone flight plans from the same pretrained model simply by letting it think longer and pick deliberately.

Core claim

On the paper's own terms, the core claim is that a single-pass navigation VLM leaves substantial performance on the table, and that a deliberate Explore–Refine–Select loop recovers it. The authors report that with three parallel candidates and two refinement rounds, their frozen-model pipeline reaches 24.96% success rate (SR), 47.39% oracle success rate (OSR), and 20.93% success-weighted path length (SPL) on the TravelUAV TS test set, against 22.94%, 44.92%, and 19.50% for the TravelUAV baseline; similar smaller gains appear on unseen-object and unseen-map sets. They further report a monotone relationship between inference token budget and success rate, which they take as evidence that compu

What carries the argument

The load-bearing mechanism is the three-stage Explore–Refine–Select pipeline, with the scoring function doing the final arbitration: S_total = 0.5·S_obs + 0.3·S_tar + 0.2·S_prog, where S_obs is a logarithmically scaled minimum depth over the central regions of five depth views (safety), S_tar is the cosine alignment between the candidate direction and the direction to the global target (goal alignment), and S_prog is a tanh-scaled distance to the current position (forward progress). The self-reflective prompt ('Wait, let's reconsider...') is what turns a single forward pass into a self-correcting loop; the parallel sampling broadens the search, the scoring collapses it.

Load-bearing premise

The Stage-3 scoring function needs the final target's position p_tar to compute the goal-alignment score, and the paper does not state whether that position is the benchmark's ground-truth goal or something the model derives from the language instruction; if it is the ground truth, the selection step is partially informed by the evaluation label.

What would settle it

Re-run the exact pipeline on the TravelUAV TS set with the goal-alignment term computed from an instruction-only estimate of the target (or with S_tar removed entirely) and compare SR/OSR to the reported 24.96/47.39; if the margin over TravelUAV (22.94/44.92) disappears, the reported gains hinge on a leaked target coordinate.

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

If this is right

  • If the claim holds, any existing UAV-navigation VLM can be upgraded at inference time — same weights, same simulator — without a training run.
  • Because gains accrue along both the parallel (breadth) and serial (depth) axes, operators can trade compute against accuracy by dialing the number of candidates and refinement rounds.
  • The reported token-versus-SR curve implies there is no obvious plateau at the tested budget, suggesting larger inference budgets may yield further gains.
  • The smaller but positive gains on unseen objects and unseen maps indicate the mechanism generalizes beyond memorized scenes, though less strongly.
  • A safety-weighted selection, rather than raw model confidence, is what the authors credit for collision avoidance; if so, the framework is a safety intervention, not just an accuracy one.

Where Pith is reading between the lines

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

  • One testable extension is to estimate the target point p_tar from the language instruction and current view alone and re-run Stage 3; the paper never states whether p_tar is the simulator's ground-truth goal, so the goal-alignment term may be doing more of the work than the model's own reasoning.
  • The parallel-exploration step resembles sampling from the model's predictive distribution over waypoints; treating the candidate set as a Monte Carlo estimate of waypoint uncertainty could motivate adaptive N — more candidates when candidate dispersion is high.
  • The same Explore–Refine–Select loop could be ported to ground-vehicle or manipulator language-guided tasks where a frozen VLM currently emits one shot; the scoring function would need task-specific safety and progress terms.
  • The positive token–SR correlation hints at a compute-versus-training trade-off: for a fixed total budget, the authors' results suggest spending on test-time deliberation can rival fine-tuning, but the paper does not directly compare those two spending options.

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 proposes a training-free, three-stage 'Explore–Refine–Select' framework for test-time scaling of a frozen UAV vision-language navigation (VLN) policy. Stage 1 samples N candidate waypoints via independent inferences of the same frozen VLM; Stage 2 refines each candidate through M rounds of self-correction using a reflective prompt ('Wait, let's reconsider...'); Stage 3 scores the refined candidates with a weighted combination of a safety score (based on depth-image nearest-obstacle distance), a goal-alignment score (cosine similarity between the candidate direction and the direction to a 'global final target point' p_tar), and a forward-progress score (tanh-normalized distance), then selects the argmax. Experiments on the TravelUAV benchmark report consistent but small gains over the TravelUAV baseline (e.g., Table 1: TS SR 24.96 vs 22.94, OSR 47.39 vs 44.92, SPL 20.93 vs 19.50), together with analyses of parallel/serial scaling and token-budget vs. success-rate correlation.

Significance. The core idea—improving a frozen VLN policy at inference time through parallel sampling, iterative self-refinement, and heuristic selection—is timely and practically relevant, and the three-stage framework is simple to understand. If the reported gains were robust and the scoring function did not depend on privileged information, the paper would constitute a useful contribution to test-time scaling for embodied VLN. However, the central SOTA claim is currently not evaluable for two reasons: first, Eq. (8) uses p_tar, the 'global final target point,' without stating whether this is the benchmark's ground-truth goal coordinate; second, the reported improvements are small and come with no error bars or significance testing. The paper also provides no code or release plan, which further limits verification. These issues must be resolved before the claim can be assessed.

major comments (3)
  1. [§3.2.3, Eq. (8)] The Goal-Alignment Score S_tar is computed from the 'global final target point' p_tar, but the paper never states where p_tar comes from. In the TravelUAV benchmark, the ground-truth goal coordinate is exactly the label used to compute NE, SR, OSR, and SPL. If p_tar is read from the simulator's episode metadata, then Stage 3 re-ranks the refined candidates by cosine similarity to the true target, leaking the evaluation label into the selection. This alone could explain the Table 1 gains—even if the N candidate waypoints are no better than TravelUAV's single output, selecting the one pointing at p_tar would inflate SR/OSR/SPL. The paper must either state explicitly that p_tar is derived purely from the language instruction and visual input (and explain how), or report an ablation without S_tar and with p_tar estimated online. As written, the SOTA claim is not evaluable.
  2. [§4.1, Table 1] The paper reports no error bars, no confidence intervals, and no significance tests for any metric. The gains are small (SR +2.02 on TS, +1.28 on UO, +0.94 on UM; SPL +1.43, +1.13, +0.65). Moreover, the TravelUAV baseline is described as 'our replication values' with no details on replication fidelity, number of runs, or seed variance. With differences of this magnitude, the claim that the method 'achieves SOTA performance' is not statistically supported. The authors should provide multiple-seed results, report variance, and ideally run a paired significance test (e.g., per-episode bootstrap) on the TS/UO/UM splits.
  3. [§3.2.3, Eqs. (5) and (10); §4.1] The scoring weights w_obs=0.5, w_tar=0.3, w_prog=0.2 and the scaling alpha=10.0 are presented as fixed choices, but there is no sensitivity analysis, no ablation, and no description of a validation set used to select them. Because the scoring function is the mechanism that produces the reported improvements, the paper should show how performance varies with these hyperparameters and whether the chosen values were selected on the TS test set. Without this, it is unclear whether the reported gains are robust or the result of tuning to the evaluated splits.
minor comments (4)
  1. [Throughout] There are numerous typographical and formatting issues, e.g., 'UA V' spacing, 'typically relies' (Abstract), 'test-time scalingparadigm' (§1), and the stray 'i' in 'Scoretotal(c′ i), i' (§3.2.3). The paper needs a careful proofread.
  2. [§4.2 and Figure 4] Figure 4's caption says 'after 100 parallel reflections,' but the experimental configuration in Section 4 uses Par up to 3. This inconsistency should be corrected.
  3. [§4.4 and Figure 5] The 'positive correlation' between token count and SR is based on exactly five configurations with no error bars and no statistical test. Reporting a correlation coefficient with a confidence interval, or at least multiple runs per configuration, would strengthen the claim.
  4. [§3.2.3, Eq. (8)] The notation in Eq. (8) is inconsistent: p_c is defined as the candidate point but the candidate vector is labeled v_cand; similarly p_tar is 'global final target point' but v_tar is the direction. Clarify that p_c is the candidate and p_tar is the endpoint used to define the direction.

Circularity Check

1 steps flagged

Stage-3 selection uses the unstated-origin 'global final target point' p_tar (Eq. 8) to re-rank candidate waypoints; if p_tar is the benchmark ground-truth goal, the reported SOTA gains are partially forced by construction.

specific steps
  1. self definitional [Section 3.2.3 (Stage 3: Scoring & Selection), Eqs. (5), (8), (11)]
    "Score_total(c'_i) = w_obs · S_obs + w_tar · S_tar + w_prog · S_prog ... Goal-Alignment Score (S_tar). This score is used to evaluate whether the candidate points of orientation align with the user’s final goal. Let the UAV’s current position be p_curr, the candidate point be p_c, and the global final target point be p_tar. ... S_tar = 1/2( (v_cand · v_tar)/(||v_cand||_2 · ||v_tar||_2) + 1 ) ... w_best = arg max_{c'_i in C_refined} Score_total(c'_i)"

    The final waypoint w_best is selected by maximizing Score_total, which includes a 0.3-weighted Goal-Alignment term S_tar. S_tar measures cosine similarity between the candidate direction and the direction to p_tar, called the 'global final target point.' In the TravelUAV benchmark, the final target coordinate is exactly the label used to compute NE/SR/OSR/SPL. The paper never states that p_tar is derived from the language instruction and visual input; if it is read from the episode, then the candidate selection is re-ranked by the ground-truth answer itself. Selecting the candidate pointing closest to the true target can by itself inflate SR/OSR/SPL in Table 1, and TravelUAV does not use this signal. No ablation removing S_tar or disclosure of p_tar's origin is provided, so the central SOT

full rationale

The only significant circularity/leakage path is the Stage-3 scoring function. Parallel exploration and serial refinement are ordinary test-time scaling techniques, and no load-bearing self-citation chain or ansatz-smuggling is present. However, Eq. (8) defines the Goal-Alignment Score using p_tar, the 'global final target point,' and Eq. (11) selects the executed waypoint by maximizing a weighted sum that includes this score. The paper does not specify whether p_tar is obtained from the language instruction and egocentric visual input or from the simulator's ground-truth goal coordinate. In the TravelUAV evaluation protocol, the ground-truth goal coordinate is precisely the label used to compute the reported metrics. If p_tar is that label, then the method's final prediction is selected by comparing candidates against the answer, so the Table 1 improvements over TravelUAV are not independent evidence of improved navigation; the central claim becomes partially circular. If p_tar is predicted from language and vision, the concern would disappear, but then the origin must be stated and the same signal made available to baselines for a fair comparison. Because the manuscript leaves this ambiguous and provides no S_tar ablation, the headline result is not evaluable and the circularity burden is high.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The method introduces no new physical entities. The central load-bearing assumptions are the availability of the ground-truth target p_tar for scoring, the reliability of the depth-based safety proxy, and the diversity of stochastic sampling. All numerical hyperparameters (weights, alpha, N, M) are chosen empirically, and the safety prompt is adapted to observed test-set failures.

free parameters (5)
  • Scoring weights w_obs, w_tar, w_prog = 0.5, 0.3, 0.2
    Set by hand in §3.2.3 to prioritize safety; no sensitivity analysis or principled derivation, likely tuned to the test sets.
  • tanh scaling alpha = 10.0
    Chosen in §3.2.3 to control saturation of the forward-progress score; no justification from data or theory.
  • Parallel candidate count N = 3
    Selected based on Table 2, which tests N=1,2,3; the paper does not report N>3 or a principled selection criterion.
  • Serial refinement depth M = 2
    Selected based on Table 3, which tests M=1,2; no exploration of larger M or an adaptive budget.
  • Global safety instruction in prompt = Unspecified wording
    Added after analyzing failure cases (§3.2.2); effectively a test-set-driven prompt modification, a free text parameter of the method.
axioms (4)
  • domain assumption The global final target point p_tar is available for computing S_tar during evaluation.
    Stage 3 scoring uses p_tar in Eq. 8; if p_tar is the simulator's ground-truth goal, the selection leaks label information. The paper does not state where p_tar comes from or whether the baseline also has access to it.
  • domain assumption Minimum depth in the central ROI of each of K depth views is a reliable proxy for collision risk.
    S_obs = log(d_safe + 1) assumes the nearest obstacle distance from depth maps monotonically reflects safety; no validation of this proxy against actual collisions.
  • domain assumption Independent calls to the VLM with stochastic decoding produce diverse candidate waypoints.
    Stage 1 relies on sampling diversity; no temperature, top-p, or seed settings are reported, and no diversity measure is given.
  • domain assumption The TravelUAV baseline configuration used in the replication matches the original benchmark's reported configuration.
    Table 1 states TravelUAV values are 'our replication values'; without comparing to the original paper's numbers or releasing code, the baseline strength is unverified.

pith-pipeline@v1.3.0-alltime-deepseek · 12580 in / 11268 out tokens · 104479 ms · 2026-08-01T12:50:29.556758+00:00 · methodology

0 comments
read the original abstract

Test-time scaling offers a promising method to improve the inference performance of Vision-Language Models (VLMs) without additional training. Existing approaches to vision-language navigation (VLN) for Unmanned Aerial Vehicle (UAV) typically relies on a single inference pass, which can falter in complex environments by producing suboptimal or unsafe trajectories. In this paper, we explore a simple and effective approach to apply test-time scaling to VLN for UAV. We enhance navigation reasoning through an iterative refinement process that requires no extra model training, guiding the model to re-evaluate its initial navigation plan for better accuracy and safety. Our method first prompts the model to generate multiple parallel candidates and then performs a self-correction step, achieving deeper and more robust planning without changing the underlying model. To further strengthen decision-making, we design a multi-criteria scoring function to evaluate the refined candidates based on safety, goal alignment, and forward-progress. This simple yet powerful combination enables a frozen UAV navigation VLMs to self-correct and generate more accurate and reliable flight plans, achieving SOTA performance in this task.

Figures

Figures reproduced from arXiv: 2607.19288 by Ang Liu, Dongliang Xu, Feinan Cheng, Tianyu Wang, Wenli Nong, Yue Yao, Zhiheng Zhang.

Figure 1
Figure 1. Figure 1: Task definition for UAV navigation. This set of images shows the target recognition process for a UAV in a simulation environment. The red boxes represent negative samples, which are distracting objects whose color or shape might cause confusion (such as traffic cones and trash cans). The green boxes represent positive samples, which are fire hydrants that need to be located. Fan et al. [2], have explored … view at source ↗
Figure 2
Figure 2. Figure 2: The Multi-Step Pipeline for Trajectory Generation and Refinement. (Left) The VLMs processes multimodal inputs, including multi-view images and a language instruction, to produce a navigation output. This panel illustrates the potential outcomes of different planning qualities, ranging from mission failure (e.g., getting stuck or collision) to success. (Right) The core of the method is a three-stage process… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of the traditional prompt and self-reflective prompt workflows. (Top) shows the single-step, direct generation process of the traditional Prompt. (Bottom) shows the“reflection-optimization” process introduced by our self-reflective Prompt, where an initial plan is reevaluated to produce an optimized output. where c ′ i is a more considered, refined point generated after M rounds of self-correcti… view at source ↗
Figure 4
Figure 4. Figure 4: A scatter plot of the model’s first phase of self-reflection after 100 parallel reflections. The green dots represent 100 possible solutions, and the red stars represent the next pathpoints selected by a multi-criteria scoring function [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Correlation between the number of inference tokens and model accuracy (measured in SR). This figure shows the re￾lationship between the number of tokens generated by the model and SR under different test conditions. The figure uses a bar chart to represent the number of tokens generated during the genera￾tion phase (left vertical axis), and a broken line to represent the corresponding SR (right vertical ax… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of trajectories under different outcome categories. This figure visualizes the comparison between the 3D flight paths generated by model evaluation and the actual trajectories. Blue lines represent the ground truth, while red lines indicate the trajectories predicted and executed by the VLMs. The starting point (Start) is marked by a square, and the endpoint (End) is marked by a triangle. The… 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

38 extracted references · 16 linked inside Pith

  1. [1]

    Wang andet al

    X. Wang andet al. Towards realistic uav vision-language navigation: Platform, benchmark, and methodology.arXiv preprint arXiv:2410.07087, 2024. 1, 3, 8, 9

    Pith/arXiv arXiv 2024

  2. [2]

    Fan andet al

    Y . Fan andet al. Aerial vision-and-dialog navigation.arXiv preprint arXiv:2205.12219, 2022. 2

    Pith/arXiv arXiv 2022

  3. [3]

    Wei andet al

    J. Wei andet al. Chain-of-thought prompting elicits reason- ing in large language models. InAdvances in Neural Infor- mation Processing Systems, volume 35, pages 24824–24837,

  4. [4]

    Madaan andet al

    A. Madaan andet al. Self-refine: Iterative refinement with self-feedback. InAdvances in Neural Information Process- ing Systems, volume 36, 2023. 2

    2023

  5. [5]

    Yao andet al

    S. Yao andet al. Tree of thoughts: Deliberate prob- lem solving with large language models.arXiv preprint arXiv:2305.10601, 2023. 2

    Pith/arXiv arXiv 2023

  6. [6]

    Chang andet al

    Y . Chang andet al. A review of uav autonomous naviga- tion in gps-denied environments.Robotics and Autonomous Systems, 170:104533, 2023. 3

    2023

  7. [7]

    Pritzl andet al

    V . Pritzl andet al. Fusion of visual-inertial odometry with li- dar relative localization for cooperative guidance of a micro- scale aerial vehicle.arXiv preprint arXiv:2306.17544, 2023. 3

    arXiv 2023

  8. [8]

    Bedn ´aˇr andet al

    J. Bedn ´aˇr andet al. Deployment of reliable visual inertial odometry approaches for unmanned aerial vehicles in real- world environment. In2022 International Conference on Unmanned Aircraft Systems (ICUAS), pages 167–176. IEEE,

  9. [9]

    D ´ıez-Gonz´alez andet al

    J. D ´ıez-Gonz´alez andet al. Time-based uwb localization architectures analysis for uavs positioning in industry.Ad Hoc Networks, 157:103419, 2024. 3

    2024

  10. [10]

    Zhong andet al

    J. Zhong andet al. A safer vision-based autonomous plan- ning system for quadrotor uavs with dynamic obstacle trajec- tory prediction and its application with llms. InProceedings of the IEEE/CVF winter conference on applications of com- puter vision, pages 920–929, 2024. 3

    2024

  11. [11]

    Kaufmann andet al

    E. Kaufmann andet al. Champion-level drone racing using deep reinforcement learning.Nature, 620(7976):982–987,

  12. [12]

    Hanover andet al

    D. Hanover andet al. Autonomous drone racing: A survey. IEEE Transactions on Robotics, 40:3044–3067, 2024. 3

    2024

  13. [13]

    Oyinlola andet al

    S. Oyinlola andet al. Reinforcement learning for au- tonomous point-to-point uav navigation.arXiv preprint arXiv:2509.13943, 2025. 3

    arXiv 2025

  14. [14]

    Guo andet al

    J. Guo andet al. Autonomous uav last-mile delivery in urban environments: A survey on deep learning and rein- forcement learning solutions.Control Engineering Practice, 165:106491, 2025. 3

    2025

  15. [15]

    Jia andet al

    Z. Jia andet al. Drone-nerf: Efficient nerf based 3d scene re- construction for large-scale drone survey.Image and Vision Computing, 143:104920, 2024. 3

    2024

  16. [16]

    Liu andet al

    C. Liu andet al. Cooperative relative localization in mav swarms with ultra-wideband ranging.arXiv preprint arXiv:2405.18234, 2024. 3

    Pith/arXiv arXiv 2024

  17. [17]

    Pritzl andet al

    V . Pritzl andet al. Drones guiding drones: Cooperative navi- gation of a less-equipped micro aerial vehicle in cluttered en- vironments. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10597–10604. IEEE, 2024. 3

    2024

  18. [18]

    Wu andet al

    R. Wu andet al. Aeroduo: Aerial duo for uav-based vision and language navigation.arXiv preprint arXiv:2508.15232,

    Pith/arXiv arXiv

  19. [19]

    Liu andet al

    S. Liu andet al. Aerialvln: Vision-and-language naviga- tion for uavs. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15384–15394, 2023. 3, 4

    2023

  20. [20]

    Liu andet al

    H. Liu andet al. Visual instruction tuning.Advances in neu- ral information processing systems, 36:34892–34916, 2023. 3

    2023

  21. [21]

    Li andet al

    B. Li andet al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3

    Pith/arXiv arXiv 2024

  22. [22]

    Li andet al

    F. Li andet al. Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models.arXiv preprint arXiv:2407.07895, 2024. 3

    Pith/arXiv arXiv 2024

  23. [23]

    Chen andet al

    Z. Chen andet al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 3

    2024

  24. [24]

    Wang andet al

    P. Wang andet al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024. 3

    Pith/arXiv arXiv 2024

  25. [25]

    Bai andet al

    S. Bai andet al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025. 3

    Pith/arXiv arXiv 2025

  26. [26]

    Ye andet al

    X. Ye andet al. Atp-llava: Adaptive token pruning for large vision language models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 24972– 24982, 2025. 4

    2025

  27. [27]

    Li andet al

    W. Li andet al. Tokenpacker: Efficient visual projector for multimodal llm.International Journal of Computer Vision, pages 1–19, 2025. 4

    2025

  28. [28]

    Song andet al

    X. Song andet al. Towards long-horizon vision-language navigation: Platform, benchmark and method. InProceed- ings of the Computer Vision and Pattern Recognition Con- ference, pages 12078–12088, 2025. 4

    2025

  29. [29]

    Zhang andet al

    Y . Zhang andet al. Grounded vision-language navigation for uavs with open-vocabulary goal understanding.arXiv preprint arXiv:2506.10756, 2025. 4

    Pith/arXiv arXiv 2025

  30. [30]

    Su andet al

    Y . Su andet al. Learning fine-grained alignment for aerial vision-dialog navigation. InProceedings of the AAAI Con- ference on Artificial Intelligence, volume 39, pages 7060– 7068, 2025. 4

    2025

  31. [31]

    Cai andet al

    H. Cai andet al. Flightgpt: Towards generalizable and in- terpretable uav vision-and-language navigation with vision- language models.arXiv preprint arXiv:2505.12835, 2025. 4

    Pith/arXiv arXiv 2025

  32. [32]

    Muennighoff andet al

    N. Muennighoff andet al. s1: Simple test-time scaling.arXiv preprint arXiv:2501.19393, 2025. 4

    Pith/arXiv arXiv 2025

  33. [33]

    Song andet al

    X. Song andet al. Efficient evaluation methods for neural architecture search: A survey.arXiv preprint arXiv:2301.05919, 2023. 4

    Pith/arXiv arXiv 2023

  34. [34]

    Montello andet al

    F. Montello andet al. A survey on dynamic neural networks: from computer vision to multi-modal sensor fusion.arXiv preprint arXiv:2501.07451, 2025. 4

    Pith/arXiv arXiv 2025

  35. [35]

    Xiao and C

    Z. Xiao and C. G. Snoek. Beyond model adaptation at test time: A survey.arXiv preprint arXiv:2411.03687, 2024. 5

    Pith/arXiv arXiv 2024

  36. [36]

    Snell andet al

    C. Snell andet al. Scaling llm test-time compute optimally can be more effective than scaling model parameters, 2024. URL https://arxiv. org/abs/2408.03314, 20, 2024. 5

    Pith/arXiv arXiv 2024

  37. [37]

    Garcia-Aunon andet al

    P. Garcia-Aunon andet al. Behavior-based control for an aerial robotic swarm in surveillance missions.Sensors, 19(20):4584, 2019. 8

    2019

  38. [38]

    Yang andet al

    J. Yang andet al. Vision-language pre-training with triple contrastive learning.Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 15671–15680, 2022. 8

    2022