pith. sign in

arxiv: 2606.31157 · v1 · pith:64ANXH6Lnew · submitted 2026-06-30 · 💻 cs.CV

Rethinking Foundation Model Collaboration: Enhancing Specialized Models through Proxy Task Reasoning

Hongyi Lin , Yang Liu , Jinhua Zhao , Xiaobo Qu This is my paper

Pith reviewed 2026-07-01 06:36 UTC · model grok-4.3

classification 💻 cs.CV
keywords foundation modelsproxy task reasoningspecialized modelsobject detectionsemantic segmentationtrajectory predictionmodel collaborationvision-language models
0
0 comments X

The pith

Foundation models enhance specialized vision models by performing proxy selection or verification on reconstructed hypotheses rather than direct structured prediction.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that foundation models should collaborate with specialized models through task decomposition instead of replacing them on structured prediction problems where geometry and numerics matter. It introduces the FAT framework that splits the work into specialist hypothesis generation, information-space reconstruction into multimodal candidates, and a bounded proxy task executed by the foundation model. Experiments instantiate this as ProxySelect using a vision-language model and report consistent gains over specialist baselines plus lower cost than direct foundation-model regression on 2D detection, 3D detection, trajectory prediction, and semantic segmentation. A sympathetic reader would care because the approach preserves the strengths of each model type while avoiding the capability mismatch that arises when foundation models are forced into precise estimation roles.

Core claim

FAT decomposes structured prediction into specialist prediction of geometrically and physically valid hypotheses, information-space reconstruction that produces multimodal candidates, and foundation-model proxy reasoning such as selection or verification. The ProxySelect instantiation demonstrates that this decomposition improves specialized baselines across four vision tasks and substantially outperforms direct foundation-model regression while using less compute, supporting the principle that specialized models keep task-specific structure and foundation models refine it through contextual proxy reasoning.

What carries the argument

ProxySelect, the instantiation of FAT in which a vision-language model executes a bounded proxy task of selection or verification over multimodal candidates reconstructed from specialist outputs.

Load-bearing premise

The information-space reconstruction step yields multimodal candidates on which the foundation model's bounded proxy task stays effective and does not introduce errors that outweigh the specialist's initial hypotheses.

What would settle it

Running ProxySelect on any of the four evaluated tasks and observing no accuracy gain or higher error than the pure specialist baseline would falsify the claim that the proxy-reasoning collaboration improves performance.

Figures

Figures reproduced from arXiv: 2606.31157 by Hongyi Lin, Jinhua Zhao, Xiaobo Qu, Yang Liu.

Figure 1
Figure 1. Figure 1: Overview of the FAT framework and its ProxySelect instantiation. A task-specific model first generates structurally valid hypotheses. These hypotheses are reconstructed into a foundation-model-readable information space through overlays, projections, structured descriptions, or numbered regions. The foundation model then solves a bounded proxy task, such as selection, verification, comparison, ranking, or … view at source ↗
Figure 4
Figure 4. Figure 4: Trajectory prediction. Panel (a) compares direct VLM regression with HPNet against the annotated future, illustrating endpoint deviation and continuity errors in direct regression. Panel (b) shows proxy selection over HPNet modes and perturbed candidates in map context. The gray reference path is used only as an inference-time reference and is distinct from the annotated future used for supervision and eva… view at source ↗
Figure 5
Figure 5. Figure 5: Cityscapes semantic segmentation. Panels (a) and (b) show two representative examples. Each example includes the ground-truth car mask, Mask2Former predictions, numbered FastSAM partitions, VLM-confirmed regions, and the fused result. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
read the original abstract

Foundation models are increasingly integrated into embodied intelligence systems, but directly assigning them structured prediction tasks requires precise geometric and numerical estimation, where specialized models often remain stronger. This capability mismatch raises a key question: should foundation models replace task-specific predictors, or should they collaborate through tasks better aligned with their strengths? We propose FAT, a foundation-model-augmented task-specific reasoning framework that treats collaboration as task decomposition rather than model replacement. FAT decomposes structured prediction into specialist prediction, information-space reconstruction, and foundation-model proxy reasoning. The specialist generates geometrically and physically valid hypotheses in the native output space, while the foundation model performs a bounded proxy task, such as selection or verification, over reconstructed multimodal candidates. We instantiate this principle as ProxySelect with a vision--language model. Across 2D object detection, 3D object detection, trajectory prediction, and semantic segmentation, ProxySelect consistently improves specialized baselines and substantially outperforms direct foundation-model regression at lower computational cost. These results suggest a general collaboration principle: specialized models preserve task-specific structure, while foundation models refine their hypotheses through contextual proxy reasoning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes the FAT framework for foundation-model collaboration with specialized models on structured prediction tasks. It decomposes the problem into specialist hypothesis generation in native output space, information-space reconstruction to produce multimodal candidates, and a bounded proxy task (selection or verification) performed by a foundation model. The ProxySelect instantiation is evaluated on 2D object detection, 3D object detection, trajectory prediction, and semantic segmentation, with the claim that it consistently improves specialized baselines while outperforming direct foundation-model regression at lower computational cost.

Significance. If the empirical results and the reconstruction step hold, the work supplies concrete evidence for a task-decomposition collaboration principle that exploits the respective strengths of specialists (geometric/physical validity) and foundation models (contextual reasoning), rather than model replacement. The multi-task scope is a strength. However, the central claim cannot be assessed without verification that reconstruction preserves sufficient fidelity and diversity.

major comments (2)
  1. [Abstract / method description] Abstract and method description: the claim that ProxySelect improves upon specialists requires that the information-space reconstruction step produces multimodal candidates on which the foundation model's bounded proxy task remains effective. No quantitative check (e.g., reconstruction error versus specialist residual error, mode-collapse metrics, or ablation removing the reconstruction stage) is supplied to confirm this assumption does not introduce new errors that outweigh the gains.
  2. [Abstract] Abstract: performance claims are stated without implementation details, error bars, dataset splits, ablation tables, or statistical significance tests, so the reported improvements across the four tasks cannot be verified from the manuscript text.
minor comments (1)
  1. The expansion of the acronym FAT is given only in the abstract; a dedicated section or footnote would improve clarity for readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our work. We address each major comment below and outline planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract / method description] Abstract and method description: the claim that ProxySelect improves upon specialists requires that the information-space reconstruction step produces multimodal candidates on which the foundation model's bounded proxy task remains effective. No quantitative check (e.g., reconstruction error versus specialist residual error, mode-collapse metrics, or ablation removing the reconstruction stage) is supplied to confirm this assumption does not introduce new errors that outweigh the gains.

    Authors: We agree that quantitative validation of the reconstruction step is valuable for supporting the core claim. The manuscript provides qualitative examples of multimodal candidate generation, but we will add a dedicated ablation subsection that reports reconstruction fidelity metrics (e.g., error relative to specialist residuals), mode diversity statistics, and performance with the reconstruction stage removed. This will confirm that the step enhances rather than degrades the proxy task effectiveness. revision: yes

  2. Referee: [Abstract] Abstract: performance claims are stated without implementation details, error bars, dataset splits, ablation tables, or statistical significance tests, so the reported improvements across the four tasks cannot be verified from the manuscript text.

    Authors: The abstract is intentionally concise; full details including dataset splits, error bars, ablation tables, and significance tests appear in Sections 4–5 and Tables 1–4 of the main text. To improve accessibility, we will expand the abstract with a short clause referencing the evaluation protocol and point readers to the detailed results. The claims remain verifiable from the complete manuscript. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical framework with externally validated gains

full rationale

The paper describes an empirical collaboration framework (FAT/ProxySelect) that decomposes structured prediction into specialist hypotheses, information-space reconstruction, and bounded FM proxy tasks (selection/verification). Performance gains are reported across four tasks via direct comparison to baselines and direct FM regression, with no equations, fitted parameters renamed as predictions, or self-citation chains invoked to justify the decomposition. The central claims rest on experimental outcomes rather than any derivation that reduces to its own inputs by construction; the method is presented as a testable principle without load-bearing self-referential steps.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework rests on domain assumptions about relative model strengths rather than new mathematical axioms or fitted parameters.

axioms (2)
  • domain assumption Specialized models remain stronger than foundation models at precise geometric and numerical estimation.
    Explicitly stated as the capability mismatch motivating the work.
  • domain assumption Foundation models can reliably execute bounded proxy tasks such as selection or verification once candidates are reconstructed.
    Central premise of the FAT decomposition.

pith-pipeline@v0.9.1-grok · 5722 in / 1150 out tokens · 34788 ms · 2026-07-01T06:36:48.343607+00:00 · methodology

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

49 extracted references · 20 canonical work pages · 3 internal anchors

  1. [1]

    Envisioning the Future of Transportation: Inspiration of ChatGPT and Large Models

    X. Qu, H. Lin, and Y. Liu. “Envisioning the Future of Transportation: Inspiration of ChatGPT and Large Models”. In:Communications in TransportationResearch3(2023),p.100103.doi: 10.1016/j.commtr. 2023.100103

    work page doi:10.1016/j.commtr 2023

  2. [2]

    HarnessingMultimodal Large Language Models for Traffic Knowledge Graph Generation and Decision-Making

    S.Kuang,Y.Liu,X.Wang,X.Wu,andY.Wei.“HarnessingMultimodal Large Language Models for Traffic Knowledge Graph Generation and Decision-Making”. In:Communications in Transportation Research4.4 (2024), p. 100146.doi:10.1016/j.commtr.2024.100146

    work page doi:10.1016/j.commtr.2024.100146 2024

  3. [3]

    Multimodal Large-Language Model Empowering Next-Generation Autonomous Driving Systems

    Z. Hu, M. Xu, and Q. Cheng. “Multimodal Large-Language Model Empowering Next-Generation Autonomous Driving Systems”. In:Jour- nal of Intelligent and Connected Vehicles8.2 (2025), p. 9210059.doi: 10.26599/JICV.2025.9210059

    work page doi:10.26599/jicv.2025.9210059 2025

  4. [4]

    Vision language models in autonomous driving: A survey and outlook

    X.Zhou,M.Liu,E.Yurtsever,B.L.Zagar,W.Zimmer,H.Cao,andA.C. Knoll. “Vision language models in autonomous driving: A survey and outlook”. In:IEEE Transactions on Intelligent Vehicles(2024), pp. 1–20. doi:10.1109/TIV.2024.3402136

    work page doi:10.1109/tiv.2024.3402136 2024

  5. [5]

    DriveVLM: The convergence of autonomous driving and largevision–languagemodels

    X. Tian, J. Gu, B. Li, Y. Liu, Y. Wang, Z. Zhao, K. Zhan, P. Jia, X. Lang, and H. Zhao. “DriveVLM: The convergence of autonomous driving and largevision–languagemodels”.In:Proceedingsofthe8thConferenceon Robot Learning. Vol. 270. Proceedings of Machine Learning Research. PMLR, 2025, pp. 4698–4726

    2025

  6. [6]

    Embodied large language models enable robots to complete complex tasks in unpredictable environments

    R. Mon-Williams, G. Li, R. Long, W. Du, and C. G. Lucas. “Embodied large language models enable robots to complete complex tasks in unpredictable environments”. In:Nature Machine Intelligence7 (2025), pp. 592–601.doi:10.1038/s42256-025-01005-x

    work page doi:10.1038/s42256-025-01005-x 2025

  7. [7]

    Zhang, K

    Z. Zhang, K. Zheng, Z. Chen, J. Jang, Y. Li, S. Han, C. Wang, M. Ding, D. Fox, and H. Yao. “GRAPE: Generalizing robot policy via preference alignment”. In:arXiv preprint arXiv:2411.19309(2024). arXiv:2411.19309 [cs.RO]

    work page arXiv 2024

  8. [8]

    VLM-AD:End-to-endautonomousdriving through vision–language model supervision

    Y. Xu, Y. Hu, Z. Zhang, G. P. Meyer, S. K. Mustikovela, S. Srinivasa, E.M.Wolff,andX.Huang.“VLM-AD:End-to-endautonomousdriving through vision–language model supervision”. In:Proceedings of the 9th Conference on Robot Learning. Vol. 305. Proceedings of Machine Learning Research. PMLR, 2025, pp. 3778–3803

    2025

  9. [9]

    AEFusion: An Attention-Based EnsembleLearningApproachforBEVFusionPerceptioninAutonomous Modular Buses

    H. Lin, S. Ming, Y. Liu, and X. Qu. “AEFusion: An Attention-Based EnsembleLearningApproachforBEVFusionPerceptioninAutonomous Modular Buses”. In:IEEE Transactions on Intelligent Vehicles10.5 (2025), pp. 3468–3480.doi:10.1109/TIV.2024.3454288

    work page doi:10.1109/tiv.2024.3454288 2025

  10. [10]

    A High-Precision Calibration and Evaluation Method Based on Binocular Cameras and LiDAR for Intelligent Vehicles

    H. Lin, Y. Liu, L. Wang, and X. Qu. “A High-Precision Calibration and Evaluation Method Based on Binocular Cameras and LiDAR for Intelligent Vehicles”. In:IEEE Transactions on Vehicular Technology 74.5 (2025), pp. 7404–7415.doi:10.1109/TVT.2025.3530479

    work page doi:10.1109/tvt.2025.3530479 2025

  11. [11]

    Few- Shot Learning for Novel Object Detection in Autonomous Driving

    Y. Zhuang, P. Liu, H. Yang, K. Zhang, Y. Wang, and Z. Pu. “Few- Shot Learning for Novel Object Detection in Autonomous Driving”. In: Communications in Transportation Research5.3 (2025), p. 100194.doi: 10.1016/j.commtr.2025.100194

    work page doi:10.1016/j.commtr.2025.100194 2025

  12. [12]

    Use of cumulants to quantify uncertainties in the HBT measurements of the homogeneity regions

    P.Liu,H.Lin,Y.Zhao,Y.Liu,andX.Qu.“DesEAD:EnhancingEnd-to- EndAutonomousDrivingwithSceneDescriptions”.In:2025IEEE28th International Conference on Intelligent Transportation Systems. 2025, pp. 2906–2912.doi:10.1109/ITSC60802.2025.11423106

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1109/itsc60802.2025.11423106 2025

  13. [13]

    The Dawn of LMMs: Preliminary Explorations with GPT-4V(ision)

    Z. Yang et al. “The dawn of LMMs: Preliminary explorations with GPT-4V(ision)”. In:arXiv preprint arXiv:2309.17421(2023). arXiv: 2309.17421 [cs.CV]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  14. [14]

    Qwen2.5-VL Technical Report

    S. Bai et al. “Qwen2.5-VL technical report”. In:arXiv preprint arXiv:2502.13923(2025). arXiv:2502.13923 [cs.CV]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    Grounding DINO: Marrying DINO with grounded pre- training for open-set object detection

    S. Liu et al. “Grounding DINO: Marrying DINO with grounded pre- training for open-set object detection”. In:Computer Vision – ECCV

  16. [16]

    Vol. 15118. Lecture Notes in Computer Science. Springer, 2024, pp. 38–55.doi:10.1007/978-3-031-72970-6_3

    work page doi:10.1007/978-3-031-72970-6_3 2024

  17. [17]

    Pix2Seq: A Language Modeling Framework for Object Detection

    T. Chen, S. Saxena, L. Li, D. J. Fleet, and G. Hinton. “Pix2Seq: A Language Modeling Framework for Object Detection”. In:International Conference on Learning Representations. 2022. 8

    2022

  18. [18]

    LISA: ReasoningSegmentationviaLargeLanguageModel

    X. Lai, Z. Tian, Y. Chen, Y. Li, Y. Yuan, S. Liu, and J. Jia. “LISA: ReasoningSegmentationviaLargeLanguageModel”.In:Proceedingsof the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024

    2024

  19. [19]

    MultiPath: Multiple Probabilistic Anchor Trajectory Hypotheses for Behavior Prediction

    Y. Chai, B. Sapp, M. Bansal, and D. Anguelov. “MultiPath: Multiple Probabilistic Anchor Trajectory Hypotheses for Behavior Prediction”. In:Proceedings of the Conference on Robot Learning. 2019

    2019

  20. [20]

    CoverNet: Multimodal Behavior Prediction Using Trajectory Sets

    J. Philion, A. Kar, S. Fidler, and M. Behl. “CoverNet: Multimodal Behavior Prediction Using Trajectory Sets”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020

    2020

  21. [21]

    HPNet: Dynamic trajectory forecasting with historical prediction attention

    X. Tang, M. Kan, S. Shan, Z. Ji, J. Bai, and X. Chen. “HPNet: Dynamic trajectory forecasting with historical prediction attention”. In:Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 15261–15270

    2024

  22. [22]

    HowGenerativeAdversarialNetworks Promote the Development of Intelligent Transportation Systems: A Survey

    H.Lin,Y.Liu,S.Li,andX.Qu.“HowGenerativeAdversarialNetworks Promote the Development of Intelligent Transportation Systems: A Survey”. In:IEEE/CAA Journal of Automatica Sinica10.9 (2023), pp. 1781–1796.doi:10.1109/JAS.2023.123744

    work page doi:10.1109/jas.2023.123744 2023

  23. [23]

    A Dynamic Prompting and Scenario Generation Method for Autonomous Driving Perception via Large-Model Optimization

    S. Zhang, H. Lin, M. Wang, B. Wei, Y. Liu, and X. Qu. “A Dynamic Prompting and Scenario Generation Method for Autonomous Driving Perception via Large-Model Optimization”. In:Transportation Research Part C: Emerging Technologies188 (2026), p. 105672.doi:10.1016/ j.trc.2026.105672

    work page arXiv 2026

  24. [24]

    Risk-Controllable Multi-View Diffusion for Driving Scenario Generation

    H. Lin, W. Shi, H. Huang, D. Zhuang, S. Zhang, Y. Liu, X. Qu, and J. Zhao. “Risk-Controllable Multi-View Diffusion for Driving Scenario Generation”.In:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition Workshops. June 2026, pp. 5169–5178

    2026

  25. [25]

    Training language models to follow instructions with human feedback

    L. Ouyang et al. “Training language models to follow instructions with human feedback”. In:Advances in Neural Information Processing Systems. Vol. 35. 2022, pp. 27730–27744

    2022

  26. [26]

    A meta- analysis on the reliability of comparative judgement

    S. Verhavert, R. Bouwer, V. Donche, and S. De Maeyer. “A meta- analysis on the reliability of comparative judgement”. In:Assessment in Education: Principles, Policy & Practice26.5 (2019), pp. 541–562.doi: 10.1080/0969594X.2019.1602027

    work page doi:10.1080/0969594x.2019.1602027 2019

  27. [27]

    LoRA: Low-Rank Adaptation of Large Language Models

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen. “LoRA: Low-Rank Adaptation of Large Language Models”. In:International Conference on Learning Representations. 2022

    2022

  28. [28]

    DETRs beat YOLOs on real-time object detection

    Y. Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y. Liu, and J. Chen. “DETRs beat YOLOs on real-time object detection”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 16965–16974

    2024

  29. [29]

    3D-MOOD: Lifting 2D to 3D for monocular open- set object detection

    Y.-H.Yang,L.Piccinelli,M.Segu,S.Li,R.Huang,Y.Fu,M.Pollefeys,H. Blum, and Z. Bauer. “3D-MOOD: Lifting 2D to 3D for monocular open- set object detection”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision. 2025, pp. 7429–7439

    2025

  30. [30]

    Masked- attention mask transformer for universal image segmentation

    B.Cheng,I.Misra,A.G.Schwing,A.Kirillov,andR.Girdhar.“Masked- attention mask transformer for universal image segmentation”. In:Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022, pp. 1290–1299

    2022

  31. [31]

    Fast segment anything,

    X. Zhao, W. Ding, Y. An, Y. Du, T. Yu, M. Li, M. Tang, and J. Wang. “Fast Segment Anything”. In:arXiv preprint arXiv:2306.12156(2023). arXiv:2306.12156 [cs.CV]

    work page arXiv 2023

  32. [32]

    In: Eu- ropean Conference on Computer Vision (2020),https://doi.org/10.1007/978- 3-030-58452-8_241

    T.-Y. Lin et al. “Microsoft COCO: Common objects in context”. In: Computer Vision – ECCV 2014. Vol. 8693. Lecture Notes in Computer Science. Springer, 2014, pp. 740–755.doi:10.1007/978- 3- 319- 10602-1_48

    work page doi:10.1007/978- 2014

  33. [33]

    Vision meets robotics: The KITTI dataset

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. “Vision meets robotics: The KITTI dataset”. In:The International Journal of Robotics Research 32.11 (2013), pp. 1231–1237.doi:10.1177/0278364913491297

    work page doi:10.1177/0278364913491297 2013

  34. [34]

    Argoverse: 3D tracking and forecasting with rich maps

    M.-F. Chang et al. “Argoverse: 3D tracking and forecasting with rich maps”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019, pp. 8748–8757

    2019

  35. [35]

    The Cityscapes dataset for semantic urban scene understanding

    M.Cordts,M.Omran,S.Ramos,T.Rehfeld,M.Enzweiler,R.Benenson, U. Franke, S. Roth, and B. Schiele. “The Cityscapes dataset for semantic urban scene understanding”. In:Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016, pp. 3213–3223

    2016

  36. [36]

    Toward Human-in-the-Loop AI: Enhancing Deep Reinforcement Learning via Real-Time Human GuidanceforAutonomousDriving

    J. Wu, Z. Huang, Z. Hu, and C. Lv. “Toward Human-in-the-Loop AI: Enhancing Deep Reinforcement Learning via Real-Time Human GuidanceforAutonomousDriving”.In:Engineering21.2(2023),pp.75– 91.doi:10.1016/j.eng.2022.05.017. 9 Appendix Thisappendixreportsthecompletetask-specificmetricsandabsoluteinferencecostsomittedfromthecompactmain-textcomparison, followed b...

    work page doi:10.1016/j.eng.2022.05.017 2023

  37. [37]

    Analyze the position of each candidate (horizontal and vertical alignment)

  38. [38]

    Analyze the size and aspect ratio of each candidate

  39. [39]

    Evaluate the overlap (IoU) with the target object

  40. [40]

    Verify the class accuracy of each candidate

  41. [41]

    Select the most suitable candidate based on comprehensive analysis Finally answer: Candidate X is the best choice. 3D object detection – selection prompt Analyze the following 3D bounding box candidates and select the most accurate one through step−by−step analysis of each candidate's characteristics: Candidate 0: class xx, location x y z, dimensions w h ...

  42. [42]

    Analyze the position characteristics of each candidate

  43. [43]

    Analyze the size characteristics of each candidate

  44. [44]

    Analyze the orientation characteristics of each candidate

  45. [45]

    11 Trajectory prediction – selection prompt You are analyzing a trajectory prediction scenario

    Based on the above analysis, select the most suitable candidate Finally answer: Candidate X is the best choice. 11 Trajectory prediction – selection prompt You are analyzing a trajectory prediction scenario. The image shows: −Green solid line with dots: vehicle historical trajectory −Gray lines: road lane boundaries and centerlines −Gray line: reference t...

  46. [46]

    Visual Continuity: Does it smoothly continue from the historical path?

  47. [47]

    Endpoint Accuracy: Does it reach a reasonable destination based on the reference?

  48. [48]

    Driving Realism: Would a human driver naturally follow this path?

  49. [49]

    Candidate 3

    Road Compliance: Does it follow the road structure shown in gray? Select the candidate that best balances these criteria. Respond with ONLY the number (e.g., "Candidate 3" or "3"). Semantic segmentation – selection prompt Query category: <TARGET_CATEGORY> The image shows N candidate masks outlined in different colors and numbered 1−N. Select all masks tha...