pith. machine review for the scientific record. sign in

arxiv: 2604.10436 · v1 · submitted 2026-04-12 · 💻 cs.CV

Recognition: unknown

SignReasoner: Compositional Reasoning for Complex Traffic Sign Understanding via Functional Structure Units

Ruibin Wang, Xinhai Zhao, Zhenyu Lin

Pith reviewed 2026-05-10 15:55 UTC · model grok-4.3

classification 💻 cs.CV
keywords traffic sign understandingcompositional reasoningvision language modelsfunctional structure unitsautonomous drivinggeneralizationFSU-Reasoning
0
0 comments X

The pith

Decomposing traffic signs into minimal functional blocks lets general VLMs reason about novel sign layouts without any model changes.

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

Vision language models often fail on traffic signs that combine symbols, languages, and layouts in unfamiliar ways, which creates risks for autonomous driving. The paper shows that breaking signs down into small Functional Structure Units such as Direction or Lane lets models learn the underlying rules of how these parts combine. A two-stage post-training process first distills accurate captions and unit labels, then optimizes reasoning using tree-edit-distance rewards. This produces strong gains on a new benchmark while using little extra data and leaving the original model architecture untouched. The result is more reliable understanding of complex signs that current systems miss.

Core claim

Shifting from whole-sign instance modeling to function-based decomposition into minimal Functional Structure Units enables general VLMs to acquire compositional reasoning for traffic signs. The method defines an FSU-Reasoning task and applies a two-stage pipeline of Iterative Caption-FSU Distillation followed by FSU-GRPO optimization that uses Tree Edit Distance as the reward signal, yielding state-of-the-art accuracy and data efficiency on the TrafficSignEval benchmark with no architectural modifications.

What carries the argument

Functional Structure Unit (FSU): a minimal core functional block such as Direction, Notice, or Lane that represents the sign's meaning through flexible composition rather than fixed whole-sign templates.

Load-bearing premise

That any traffic sign can be broken down into a small fixed vocabulary of functional blocks whose combinations fully determine the sign's meaning and allow reliable generalization to unseen layouts.

What would settle it

Create a test set of traffic signs that use entirely new combinations of the defined functional units, then check whether the model's accuracy on these signs falls to the level of unmodified baseline VLMs.

Figures

Figures reproduced from arXiv: 2604.10436 by Ruibin Wang, Xinhai Zhao, Zhenyu Lin.

Figure 1
Figure 1. Figure 1: Visualizations of complex traffic sign understanding of VLMs. Existing VLMs, including Qwen2.5-VL-72B [2], InternVL3- 78B [42], and Gemini 2.5-Pro, demonstrate a significant inability to correctly comprehend traffic signs. Their failures primarily stem from misidentifying the number or type of directional arrows, or failing to accurately associate the U-Turn prohibition symbol with the correct lane. In con… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of Functional Structure Units (FSUs) for modeling traffic signs. Four kinds of FSUs are shown, including Notice, Direction, Construction and Lane. Notably, a traffic sign can either be decomposed into multiple FSUs with same or different functions. diversions. The four categories are designed to be highly orthogonal in function and can be viewed in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the Hierarchical Key-Value Schema. The “Sign-1,2,3” tags correspond to images in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the SignReasoner overall pipeline. The FSU Modeling (Sect. 3.1) decomposes the traffic sign multiple into FSUs, which would be concatenated with captions generated from base VLMs for Iterative Caption-FSU Distillation (Sect. 3.2). Finally, the distilled models would go through the FSU-GRPO Training (Sect. 3.3) to further boost reasoning capacity. Comparing with string similarity, TED offers… view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparison between SignReasoner and base VLMs. Caption-FSU Iterative SignReasoner Format Step Qwen2.5-VL-7B Pangu-MM-7B 0 74.87 78.26 ✓ 1 77.95 81.03 ✓ 2 83.08 83.59 ✓ 3 82.37 83.06 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualizations of Traffic Sign Understanding in SignReasoner (Pangu-MM-7B). HUAWEI CONFIDENTIAL 内部资料 注意保密 Caption: 这张交通牌是黄色的，分为左右两部分。左侧包含限速40和施工的交通标志 。右侧是前方右转东海道的道路信息，包括道南段（深盐路口-中港三路）封闭施工，车 辆需提前择道行驶。\n\n左侧限速40的标志是圆形的，红色边框，白色背景，下方有一 个施工标志，施工标志是三角形的，黑色背景，白色施工图案。右侧道路信息是文字描 述，无其他图案。\n\n交通牌的空间排布是左侧限速和施工信息，右侧道路信息。限速 和施工信息提醒驾驶员注意前方道路施工，最高速度40。道路信息告知驾驶员前方右转 东海道，具体路段封闭施工，车辆需提前择道行驶。 FSU: {'交通牌': '是', '电子牌': '否', '遮挡… view at source ↗
Figure 7
Figure 7. Figure 7: Visualizations of Traffic Sign Understanding in SignReasoner (Pangu-MM-7B). 7.3. Evaluation Details The specific process of automated evaluation in the TrafficSignEval-Structure protocol is illustrated in Fig.8. The input consists of the Predicted and True FSU Dictio￾naries. First, the matching score, Score1, for the top-level keys is computed. This score is a weighted sum where pg is the binary match indi… view at source ↗
Figure 8
Figure 8. Figure 8: Flowchart of automatic evaluation algorithm. 8. More Visualizations In this section, we provide more comprehensive visualized results (Fig.6-7) in both Chinese and English to demonstrate the powerful capabilities of SignReasoner [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
read the original abstract

Accurate semantic understanding of complex traffic signs-including those with intricate layouts, multi-lingual text, and composite symbols-is critical for autonomous driving safety. Current models, both specialized small ones and large Vision Language Models (VLMs), suffer from a significant bottleneck: a lack of compositional generalization, leading to failure when encountering novel sign configurations. To overcome this, we propose SignReasoner, a novel paradigm that transforms general VLMs into expert traffic sign reasoners. Our core innovation is Functional Structure Unit (FSU), which shifts from common instance-based modeling to flexible function-based decomposition. By breaking down complex signs into minimal, core functional blocks (e.g., Direction, Notice, Lane), our model learns the underlying structural grammar, enabling robust generalization to unseen compositions. We define this decomposition as the FSU-Reasoning task and introduce a two-stage VLM post-training pipeline to maximize performance: Iterative Caption-FSU Distillation that enhances the model's accuracy in both FSU-reasoning and caption generation; FSU-GRPO that uses Tree Edit Distance (TED) to compute FSU differences as the rewards in GRPO algorithm, boosting reasoning abilities. Experiments on the newly proposed FSU-Reasoning benchmark, TrafficSignEval, show that SignReasoner achieves new SOTA with remarkable data efficiency and no architectural modification, significantly improving the traffic sign understanding in various VLMs.

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 / 2 minor

Summary. The paper proposes SignReasoner, a post-training paradigm that converts general VLMs into traffic-sign reasoners by decomposing complex signs into Functional Structure Units (FSUs) such as Direction, Notice, and Lane. It introduces the FSU-Reasoning task and a two-stage pipeline—Iterative Caption-FSU Distillation followed by FSU-GRPO using Tree Edit Distance (TED) rewards—to improve both caption accuracy and compositional reasoning. Experiments on the new TrafficSignEval benchmark claim new SOTA results with high data efficiency and no architectural changes, attributing gains to learning an underlying structural grammar that enables generalization to unseen sign compositions.

Significance. If the FSU decomposition and training pipeline demonstrably enable productive recombination of unseen FSU combinations, the work would offer a practical, architecture-agnostic route to stronger compositional visual reasoning in safety-critical domains. The emphasis on data efficiency and the introduction of a structured benchmark could influence post-training methods for other structured visual tasks beyond traffic signs.

major comments (2)
  1. [Experiments / TrafficSignEval benchmark] The central compositional-generalization claim (abstract and §1) rests on FSUs capturing a reusable structural grammar. However, the TrafficSignEval benchmark description and experimental splits do not appear to isolate recombination of novel FSU co-occurrences; without a held-out set of unseen FSU combinations or an ablation that replaces FSU labels with generic tags while preserving layout, it remains possible that gains arise from improved general captioning rather than grammar learning.
  2. [FSU-GRPO subsection] §3.2 (FSU-GRPO): The TED reward is defined on FSU trees, yet the manuscript does not specify the exact tree-construction procedure from the FSU decomposition or provide validation that TED correlates with human judgments of sign similarity. This makes it difficult to assess whether the reward truly incentivizes compositional structure rather than surface-level caption similarity.
minor comments (2)
  1. [Abstract / §1] The abstract and introduction use the term 'FSU-Reasoning task' without a concise formal definition or example of input/output format; adding a short illustrative figure or table early would improve readability.
  2. [Results tables] Table captions and axis labels in the experimental results should explicitly state whether reported metrics are on seen versus unseen FSU combinations to allow direct assessment of the generalization claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments on our work. We address the major comments point by point below, and we will make revisions to improve the clarity and rigor of the manuscript as outlined.

read point-by-point responses

  1. Referee: [Experiments / TrafficSignEval benchmark] The central compositional-generalization claim (abstract and §1) rests on FSUs capturing a reusable structural grammar. However, the TrafficSignEval benchmark description and experimental splits do not appear to isolate recombination of novel FSU co-occurrences; without a held-out set of unseen FSU combinations or an ablation that replaces FSU labels with generic tags while preserving layout, it remains possible that gains arise from improved general captioning rather than grammar learning.

    Authors: We agree that demonstrating isolation of novel FSU recombinations is essential to support the claim of learning a reusable structural grammar. Although the TrafficSignEval benchmark was designed to include diverse compositions, the manuscript's description of the experimental splits may not have sufficiently highlighted the held-out novel combinations. In the revised version, we will expand the benchmark section to explicitly describe the construction of splits that hold out specific FSU co-occurrences for testing generalization. We will also include the suggested ablation, where FSU labels are replaced with generic tags while preserving the visual layout, to verify that the performance improvements are attributable to compositional structure learning rather than enhanced general captioning capabilities. revision: yes

  2. Referee: [FSU-GRPO subsection] §3.2 (FSU-GRPO): The TED reward is defined on FSU trees, yet the manuscript does not specify the exact tree-construction procedure from the FSU decomposition or provide validation that TED correlates with human judgments of sign similarity. This makes it difficult to assess whether the reward truly incentivizes compositional structure rather than surface-level caption similarity.

    Authors: The referee is correct that additional details are needed for full reproducibility and to confirm the reward's focus on structure. The current manuscript introduces the use of Tree Edit Distance on FSU trees but omits the precise mapping from FSU decompositions to tree structures and any empirical validation against human judgments. We will revise §3.2 to provide the exact tree-construction procedure, including how individual FSUs (such as Direction, Notice, and Lane) and their attributes are organized into a hierarchical tree. We will also report a validation study correlating TED scores with human similarity assessments on a sample of traffic signs. These additions will help demonstrate that the GRPO reward effectively promotes compositional reasoning. revision: yes

Circularity Check

0 steps flagged

No circularity: method proposal and benchmark evaluation are self-contained

full rationale

The paper introduces FSU decomposition and a two-stage post-training pipeline (Iterative Caption-FSU Distillation + FSU-GRPO with TED reward) as a novel paradigm for improving VLMs on traffic sign understanding. No mathematical derivation chain, equations, or first-principles predictions appear; performance claims rest on empirical results from the new TrafficSignEval benchmark rather than any quantity that reduces to its own inputs by construction. The compositional generalization assumption is an empirical hypothesis tested via experiments, not a self-referential definition or fitted parameter renamed as a prediction. Self-citations are absent from the provided text, and the central claims do not rely on load-bearing uniqueness theorems or ansatzes imported from prior author work.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated.

pith-pipeline@v0.9.0 · 5552 in / 938 out tokens · 31546 ms · 2026-05-10T15:55:15.553829+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 · 19 canonical work pages · 14 internal anchors

  1. [1]

    Qwen Technical Report

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report.arXiv preprint arXiv:2309.16609, 2023. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025. 2, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Driving by the rules: A benchmark for integrating traffic sign regulations into vectorized hd map

    Xinyuan Chang, Maixuan Xue, Xinran Liu, Zheng Pan, and Xing Wei. Driving by the rules: A benchmark for integrating traffic sign regulations into vectorized hd map. InProceed- ings of the Computer Vision and Pattern Recognition Con- ference, pages 6823–6833, 2025. 1, 2, 3

    2025

  4. [4]

    Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 3

    2024

  5. [5]

    Yolo-world: Real-time open-vocabulary object detection

    Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xing- gang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16901–16911, 2024. 1

    2024

  6. [6]

    Fast r-cnn

    Ross Girshick. Fast r-cnn. InProceedings of the IEEE inter- national conference on computer vision, pages 1440–1448,

  7. [7]

    A survey of deep learning techniques for autonomous driving.Journal of Field Robotics, 37(3):362– 386, 2020

    Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving.Journal of Field Robotics, 37(3):362– 386, 2020. 1

    2020

  8. [8]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 3, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Learning to understand traffic signs

    Yunfei Guo, Wei Feng, Fei Yin, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Learning to understand traffic signs. InPro- ceedings of the 29th ACM International Conference on Mul- timedia, pages 2076–2084, 2021. 2

    2076

  10. [10]

    Visual traffic knowledge graph generation from scene images

    Yunfei Guo, Fei Yin, Xiao-hui Li, Xudong Yan, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Visual traffic knowledge graph generation from scene images. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 21604–21613, 2023. 2

    2023

  11. [11]

    Sign- parser: An end-to-end framework for traffic sign understand- ing.International Journal of Computer Vision, 132(3):805– 821, 2024

    Yunfei Guo, Wei Feng, Fei Yin, and Cheng-Lin Liu. Sign- parser: An end-to-end framework for traffic sign understand- ing.International Journal of Computer Vision, 132(3):805– 821, 2024. 1, 2, 6

    2024

  12. [12]

    Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749,

    work page internal anchor Pith review arXiv

  13. [13]

    Vlm-rl: A unified vision language models and reinforcement learning framework for safe autonomous driv- ing.arXiv preprint arXiv:2412.15544, 2024

    Zilin Huang, Zihao Sheng, Yansong Qu, Junwei You, and Sikai Chen. Vlm-rl: A unified vision language models and reinforcement learning framework for safe autonomous driv- ing.arXiv preprint arXiv:2412.15544, 2024. 3

    work page arXiv 2024

  14. [14]

    Al Sallab, Senthil Yogamani, and Patrick P´erez

    B Ravi Kiran, Ibrahim Sobh, Victor Talpaert, Patrick Man- nion, Ahmad A. Al Sallab, Senthil Yogamani, and Patrick P´erez. Deep reinforcement learning for autonomous driving: A survey.IEEE Transactions on Intelligent Transportation Systems, 23(6):4909–4926, 2022. 1

    2022

  15. [15]

    arXiv preprint arXiv:2502.21321

    Komal Kumar, Tajamul Ashraf, Omkar Thawakar, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, Phillip HS Torr, Fahad Shahbaz Khan, and Salman Khan. Llm post-training: A deep dive into reasoning large language models.arXiv preprint arXiv:2502.21321, 2025. 3

    work page arXiv 2025

  16. [16]

    Seg-zero: Reasoning-chain guided segmentation via cognitive reinforcement.arXiv preprint arXiv:2503.06520, 2025

    Yuqi Liu, Bohao Peng, Zhisheng Zhong, Zihao Yue, Fanbin Lu, Bei Yu, and Jiaya Jia. Seg-zero: Reasoning-chain guided segmentation via cognitive reinforcement.arXiv preprint arXiv:2503.06520, 2025. 3

    work page arXiv 2025

  17. [17]

    Understanding R1-Zero-Like Training: A Critical Perspective

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective.arXiv preprint arXiv:2503.20783, 2025. 3, 5

    work page internal anchor Pith review arXiv 2025

  18. [18]

    Visual-RFT: Visual Reinforcement Fine-Tuning

    Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual- rft: Visual reinforcement fine-tuning.arXiv preprint arXiv:2503.01785, 2025. 3

    work page internal anchor Pith review arXiv 2025

  19. [19]

    MM-Eureka: Exploring the Frontiers of Multimodal Reasoning with Rule-based Reinforcement Learning

    Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Tiancheng Han, Botian Shi, Wenhai Wang, Junjun He, et al. Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforce- ment learning.arXiv preprint arXiv:2503.07365, 2025. 3

    work page Pith review arXiv 2025

  20. [20]

    Medvlm-r1: Incentivizing medical reasoning ca- pability of vision-language models (vlms) via reinforcement learning

    Jiazhen Pan, Che Liu, Junde Wu, Fenglin Liu, Jiayuan Zhu, Hongwei Bran Li, Chen Chen, Cheng Ouyang, and Daniel Rueckert. Medvlm-r1: Incentivizing medical reasoning ca- pability of vision-language models (vlms) via reinforcement learning. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 337–

  21. [21]

    Efficient computa- tion of the tree edit distance.ACM Transactions on Database Systems (TODS), 40(1):1–40, 2015

    Mateusz Pawlik and Nikolaus Augsten. Efficient computa- tion of the tree edit distance.ACM Transactions on Database Systems (TODS), 40(1):1–40, 2015. 5

    2015

  22. [22]

    You only look once: Unified, real-time object de- tection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object de- tection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016. 1

    2016

  23. [23]

    Faster r-cnn: Towards real-time object detection with region proposal networks.IEEE transactions on pattern analysis and machine intelligence, 39(6):1137–1149, 2016

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks.IEEE transactions on pattern analysis and machine intelligence, 39(6):1137–1149, 2016. 1

    2016

  24. [24]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 2, 3, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  25. [25]

    Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition.Neural net- works, 32:323–332, 2012. 6

    2012

  26. [26]

    Llamav-o1: Rethinking step-by-step vi- sual reasoning in llms

    Omkar Thawakar, Dinura Dissanayake, Ketan Pravin More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Il- muz Zaman Mohammed Zumri, Jean Lahoud, Rao Muham- mad Anwer, et al. Llamav-o1: Rethinking step-by-step vi- sual reasoning in llms. InFindings of the Association for Computational Linguistics: ACL 2025, pages 24290–24315,

    2025

  27. [27]

    LLaMA: Open and Efficient Foundation Language Models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  28. [28]

    Improved yolov5 network for real-time multi-scale traffic sign detection.Neural Computing and Applications, 35(10): 7853–7865, 2023

    Junfan Wang, Yi Chen, Zhekang Dong, and Mingyu Gao. Improved yolov5 network for real-time multi-scale traffic sign detection.Neural Computing and Applications, 35(10): 7853–7865, 2023. 1

    2023

  29. [29]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reason- ing in language models.arXiv preprint arXiv:2203.11171,

    work page internal anchor Pith review Pith/arXiv arXiv

  31. [31]

    Chain-of-thought prompting elicits reasoning in large lan- guage models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large lan- guage models.Advances in neural information processing systems, 35:24824–24837, 2022. 3, 5

    2022

  32. [32]

    Llava-cot: Let vision language models reason step-by-step

    Guowei Xu, Peng Jin, Ziang Wu, Hao Li, Yibing Song, Lichao Sun, and Li Yuan. Llava-cot: Let vision language models reason step-by-step. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2087– 2098, 2025. 3

    2087

  33. [33]

    Traffic sign interpretation via natural language description.IEEE Transactions on Intelligent Transportation Systems, 2024

    Chuang Yang, Kai Zhuang, Mulin Chen, Haozhao Ma, Xu Han, Tao Han, Changxing Guo, Han Han, Bingxuan Zhao, and Qi Wang. Traffic sign interpretation via natural language description.IEEE Transactions on Intelligent Transportation Systems, 2024. 1, 3

    2024

  34. [34]

    Signeye: Traffic sign interpretation from vehicle first-person view

    Chuang Yang, Xu Han, Tao Han, Yuejiao Su, Junyu Gao, Hongyuan Zhang, Yi Wang, and Lap-Pui Chau. Signeye: Traffic sign interpretation from vehicle first-person view. IEEE Transactions on Intelligent Transportation Systems,

  35. [35]

    Towards real-time traffic sign detection and classification

    Yi Yang, Hengliang Luo, Huarong Xu, and Fuchao Wu. Towards real-time traffic sign detection and classification. IEEE Transactions on Intelligent transportation systems, 17 (7):2022–2031, 2015. 6

    2022

  36. [36]

    R1-onevision: Advancing gen- eralized multimodal reasoning through cross-modal formal- ization.arXiv preprint arXiv:2503.10615, 2025

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. R1-onevision: Advancing gen- eralized multimodal reasoning through cross-modal formal- ization.arXiv preprint arXiv:2503.10615, 2025. 3

    work page arXiv 2025

  37. [37]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xi- aochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, et al. Dapo: An open-source llm reinforce- ment learning system at scale, 2025.URL https://arxiv. org/abs/2503.14476, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  38. [38]

    A survey of autonomous driving: Com- mon practices and emerging technologies.IEEE Access, 8: 58443–58469, 2020

    Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Com- mon practices and emerging technologies.IEEE Access, 8: 58443–58469, 2020. 1

    2020

  39. [39]

    DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection.arXiv preprint arXiv:2203.03605, 2022. 1

    work page internal anchor Pith review arXiv 2022

  40. [40]

    Simple fast algorithms for the editing distance between trees and related problems

    Kaizhong Zhang and Dennis Shasha. Simple fast algorithms for the editing distance between trees and related problems. SIAM journal on computing, 18(6):1245–1262, 1989. 5

    1989

  41. [41]

    Lili Zhang, Kang Yang, Yucheng Han, Jing Li, Wei Wei, Hongxin Tan, Pei Yu, Ke Zhang, and Xudong Yang. Tsd- detr: A lightweight real-time detection transformer of traffic sign detection for long-range perception of autonomous driv- ing.Engineering Applications of Artificial Intelligence, 139: 109536, 2025. 1

    2025

  42. [42]

    InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shen- glong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025. 2, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  43. [43]

    Deformable DETR: Deformable Transformers for End-to-End Object Detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable trans- formers for end-to-end object detection.arXiv preprint arXiv:2010.04159, 2020. 1

    work page internal anchor Pith review arXiv 2010

  44. [44]

    Traffic-sign detection and classifica- tion in the wild

    Zhe Zhu, Dun Liang, Songhai Zhang, Xiaolei Huang, Baoli Li, and Shimin Hu. Traffic-sign detection and classifica- tion in the wild. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2110–2118,

  45. [45]

    1, 6 SignReasoner: Compositional Reasoning for Complex Traffic Sign Understanding via Functional Structure Units Supplementary Material

  46. [46]

    SignReasoner Implementations We elaborate on several implementation details of SignRea- soner in this section, including the uniquely designed keys for each FSU (Sect.6.1), the specific prompt and response format employed to instruct the VLM for both Captioning and FSU-Reasoning (Sect.6.2), and the algorithmic imple- mentation of Tree Edited Distance (Sec...

  47. [47]

    Training and Evaluation This section details the specifics of the training and evaluation datasets (Sect.7.1), the training particulars (Sect.7.2)—including data construction, hyper-parameters, and hardware—and the comprehensive evaluation method- ology on the Structure protocols (Sect.7.3). 7.1. Datasets Training Data.We primarily utilize two types of tr...

  48. [48]

    Traffic Sign

    Mingle Road (Mingle Rd) is on the left side of the sign, indicating a left turn. 3. Yangtaishan Road (Yangtaishan Rd) is on the right side of the sign, indicating a right turn. The color and shape of the traffic sign are designed to be simple and clear, using the white crossroad pattern to clearly indicate the options for going straight ahead or turning l...

  49. [49]

    More Visualizations In this section, we provide more comprehensive visualized results (Fig.6-7) in both Chinese and English to demonstrate the powerful capabilities of SignReasoner. Algorithm 2Tree Edit Distance Require:Tree nodesn 1,n 2 Ensure:Minimum edit cost 1:FunctionHELPER(n 1, n2): 2:cost←0 3:ifISLEAF(n 1)andISLEAF(n 2)then 4:# Case 1: Both are lea...