Pith. sign in

REVIEW 3 major objections 5 minor 30 references

Chat2SPaT: A Large Language Model Based Tool for Automating Traffic Signal Control Plan Management

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Chat2SPaT converts semi-structured, ambiguous traffic signal plan descriptions into exact phase-and-timing results, with reported accuracies above 94% on English and Chinese test cases for the best model choices.

desk verdict A genuinely useful pipeline and a new bilingual benchmark, but the reported accuracies are in-sample numbers because the test set was used to tune prompts and scripts. read the letter →

arxiv 2507.05283 v1 pith:JX6QAJPF submitted 2025-07-04 cs.AI cs.CL

Yue Wang , Miao Zhou , Guijing Huang , Rui Zhuo , Chao Yi , Zhenliang Ma This is my paper
classification cs.AIcs.CL
keywords trafficsignalcontrolphaseandtimingSPaTlargelanguagemodelspromptengineeringplanmanagementintelligenttransportationsystemschat-basedtool
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that a large language model, guided by carefully written prompts and backed by rule-based scripts, can replace the tedious manual entry of traffic signal plans: a traffic engineer describes a plan in words, and the system outputs a second-by-second signal timing plan ready for controller software. It reports that on 306 bilingual plan descriptions, the pipeline matches expert-coded ground truth in over 91% of cases overall, with the best models exceeding 94% on each language. The point is not to replace traffic engineering with AI reasoning, but to have the LLM do what it is good at — understanding language — while scripts do arithmetic and safety checks.

What carries the argument

The central object is the two-step Chat2SPaT workflow. In step one, curated prompts instruct the LLM to output a JSON object with three parts: a phase sequence result (stage- or ring-based structure of major phases), a phase attribute result (key-value records for overlapped phases, standalone phases, permissive or prohibited left turns, pedestrian phases, late start and early cut-off), and a cycle length result. In step two, Python scripts locate every phase in the cycle: major phases are placed by stage or ring arithmetic, overlapped phases inherit their parent phase's split, standalone phases use explicit start and end times, connected occurrences are merged with depth-first search, and a conflict matrix validates the assembled plan. The scripts deliberately absorb numerical calculation and traffic-control nuance so the LLM only performs language understanding.

What would settle it

Take a set of real controller timing sheets from a city's traffic operations center, have engineers who did not write the prompts describe each plan in their own words, run Chat2SPaT with the released code and prompts against the controller's actual second-by-second output, and compare exact match rates; if the rate falls well below the reported 86–94%, the pipeline's external validity fails.

Watch

Extended reading notes

Core claim

Chat2SPaT's central claim is that converting natural-language plan descriptions to executable signal phase and timing results can be decomposed into an LLM skill and a script skill: the LLM rewrites the user's description into a JSON structure of phase sequence, phase attributes, and cycle length; deterministic Python scripts then compute start and end times, merge connected phase occurrences, apply overlapping rules, and validate conflicts. The paper reports that this division of labor yields exact second-by-second plan matches on 91.83% of 306 cases overall with ChatGPT-4o, 94.12% on English cases, and 94.55% on Chinese cases with Qwen2.5-72B-Instruct, where correctness requires every phase and every second's color code to equal expert-coded ground truth.

Load-bearing premise

The test set was written by the authors and coded by experts using the same movement diagram that appears in the prompts; the claim that the tool works in practice assumes real engineers describe plans in similar words for an intersection that has all the same signal heads.

Editorial extensions

If this is right

  • Traffic engineers can create and edit signal plans conversationally, with iterative edits within the same chat session, reducing manual parameter input for time-of-day and day-of-week plan changes.
  • The same pipeline can output stage-based or ring-based plan structures, so it can feed both dual-ring controllers and stage-based traffic control software.
  • Because scripts handle calculations, smaller or weaker LLMs can still produce usable plans; accuracy stays above 90% with a 72B-parameter model, indicating the tool is feasible where large proprietary models are unavailable.
  • The open test dataset of 306 bilingual plan descriptions provides a first benchmark for measuring LLM understanding of traffic signal plan language.
  • The plan validation step catches conflicting movements and short pedestrian walk intervals before deployment, warning users to revise unsafe plans.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the approach transfers to real controller interfaces, it could turn high-level policy statements such as 'give the main street more green in the afternoon peak' into concrete timing plan variants, and the same prompt-attribute schema could be extended to actuated control parameters like detector calls and max extensions.
  • The reported accuracy is measured against the authors' own coding scheme; a natural next test is whether the pipeline survives intersection-specific movement codings and varied signal-head layouts without prompt changes.
  • One could build a version that asks the user clarifying questions when ambiguity is detected, rather than guessing, which would likely cut the residual semantic comprehension errors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. Chat2SPaT is an LLM-based pipeline that converts natural-language descriptions of pre-timed traffic signal plans into executable signal phase and timing (SPaT) plans. The pipeline uses curated prompts to elicit a JSON structure containing phase sequence, phase attributes, and cycle length, then applies Python scripts for data cleansing, phase timing calculation, occurrence merging, overlap handling, and plan validation. The authors contribute an open dataset of 306 bilingual (English/Chinese) plan descriptions with second-by-second ground-truth color codes and evaluate four LLMs (ChatGPT-4o, Qwen2.5-72B-Instruct, Deepseek-V3, Qwen3-32B), reporting overall accuracies between 86.06% and 91.83% and language-specific accuracies up to 94.55%. The paper claims that Chat2SPaT is the first benchmark for LLM understanding of traffic signal control plan descriptions and that it can reduce manual plan-management workload in practice.

Significance. The strongest parts of the paper are the strict second-by-second color-code evaluation, the script-based assembly that offloads arithmetic and domain nuances from the LLM, and the open release of code, prompts, and test data. If the reported accuracies hold out of sample, the tool would be a practically useful building block for traffic signal plan management. However, the evaluation protocol has three load-bearing gaps: prompts and fault-tolerance functions appear to have been tuned on the same 306 cases used for testing, the aggregation of three repeated runs per case is unspecified, and the ground truth was constructed by the authors under the same movement-coding scheme used in the prompts. These gaps prevent the headline accuracy from being interpreted as an unbiased estimate of real-world performance.

major comments (3)
  1. [Sections V-B, V-C] The reported accuracies in Table II are in-sample estimates because prompts, fault-tolerance functions, and the temperature setting were tuned on the same 306 cases used for evaluation. Section V-C states that 'prompts and fault tolerance functions are designed and iteratively refined' on the basis of observed LLM errors, and Section V-B reports temperature 0.7 chosen 'after trial-and-error testing.' With no development/test split described, Table II cannot be interpreted as measuring generalization to new plan descriptions. The authors should fix a validation split before measuring accuracy, report results on a held-out subset that was never used to modify prompts or scripts, and describe the refinement protocol.
  2. [Section V-B, Table II] The paper does not specify how the three repeated runs per case are aggregated into the case-level accuracy reported in Table II. Since correctness is defined as exact equality of phases and second-by-second color codes (Section V-A), the choice of 'any run correct,' 'majority correct,' or 'all runs correct' can materially change the percentages. The authors should state the aggregation rule, report per-run accuracy, and provide confidence intervals or variance estimates across runs and across cases.
  3. [Section V-A] The test set and ground truth are author-constructed: 306 descriptions written by the authors and coded by 'TSC experts' using the same movement phase-coding scheme that appears in the prompts (Fig. 1 and Section III-A). No inter-rater reliability, independent second coding, or external corpus of real practitioner descriptions is provided. This limits external validity: real-world descriptions may involve intersection-specific signal-head configurations, agency-specific terminology, and different ambiguity patterns. The authors should present evidence that the descriptions resemble real-world usage, for example by collecting cases from practitioners or by reporting coding instructions and agreement measures.
minor comments (5)
  1. [Abstract and Table II] The abstract's claim of 'accuracy of over 94% for both English and Chinese cases' is only true when different models are used for the two languages (ChatGPT-4o for English, Qwen2.5-72B-Instruct for Chinese); as written, it implies a single model surpasses 94% in both languages.
  2. [Section V-B] There is a typo in the model name: 'Qwen2.5B-72B-Instruct' should read 'Qwen2.5-72B-Instruct.'
  3. [Equation (1)] The typesetting of Equation (1) is garbled; the condition 'if T_start^S + S = C' and the modulo expression are hard to parse. Please rewrite the equation with unambiguous notation.
  4. [Section V-A and Table II] Please report the denominators for the English/Chinese subsets and for the creating/editing subsets (188 creating and 118 editing cases are mentioned in Section V-A but Table II gives only percentages).
  5. [Section IV-A] There is a grammatical error in 'assign the default the value of attribute parentPhase'; it should be 'assign the default value of attribute parentPhase.'

Circularity Check

1 steps flagged · score 6.0 of 10

Reported accuracies are in-sample: prompts, fault-tolerance functions, and temperature were iteratively refined on the same 306-case test set used for the headline numbers.

  1. fitted input called prediction [Section V-B, V-C; Table II]
    "In LLM outputs, it is difficult to distinguish user specified values from LLMs' overthinking results. Therefore, prompts and fault tolerance functions are designed and iteratively refined, to reduce the chance and impact of LLMs' typical overthinking errors. [...] To account for the randomness in LLMs' reasoning, the plan result is generated and evaluated three times for each case, with a temperature of 0.7 after trial-and-error testing with different parameters."

    Table II's accuracies are computed on the same 306 cases used to develop the pipeline. The paper states that after finding 'various erroneous results in LLM outputs' in the experiments, 'prompts and fault tolerance functions are designed and iteratively refined,' and that temperature 0.7 was chosen 'after trial-and-error testing with different parameters.' No held-out validation split is reported; the dataset is only described as 306 'created' plan descriptions. Therefore the reported 91.83%/94.12%/94.55% figures are in-sample estimates of how well the prompts/scripts/temperature fit the test set, not independent predictions for new plan descriptions. The headline empirical claim reduces, in part, to a report of the tuning process.

full rationale

No definitional circularity is present: the final evaluation compares assembled plans to second-by-second color sequences coded by TSC experts, not to the JSON schema the LLM is prompted to emit, so the comparison is not tautological. Self-citations (e.g., reference [2]) are not load-bearing. The central issue is evaluation leakage: the paper's own text says prompts and fault tolerance functions were iteratively refined after observing errors in experiments, and temperature was chosen by trial-and-error, all without reporting a held-out split. Thus the central empirical claim — Chat2SPaT's high accuracy on the 306-case benchmark — is an in-sample result and should not be read as a validated out-of-sample performance figure. This is a fitted-input-called-prediction pattern rather than a mathematical derivation, so the score is 6 rather than higher.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are postulated. The main load-bearing choices are the data representation, the JSON schema, the split-versus-green-time convention, and the representativeness of the self-created benchmark. The only hand-tuned numeric parameter is the LLM temperature. These assumptions are reasonable for a first tool, but they limit how far the reported accuracy can be extrapolated.

free parameters (1)
  • LLM sampling temperature = 0.7
    Set after trial-and-error testing (Section V-B). Reported accuracies are specific to this choice, and no sensitivity analysis is provided.
assumptions (5)
  • domain assumption Representing a complete signal control plan as per-second color codes for all movements of the hypothetical intersection in Fig. 1 is sufficient for plan correctness.
    Section V-A defines ground truth as second-by-second color codes; the entire evaluation depends on this encoding being the right target.
  • domain assumption The 306 test descriptions and their expert-coded ground truth are correct and representative of real traffic engineering plan descriptions.
    Section V-A introduces the dataset; external validity of the accuracy numbers rests on this assumption.
  • ad hoc to paper The phase sequence, phase attribute, and cycle length JSON schema captures every nuance needed to assemble an executable pre-timed plan.
    Section III defines this schema; scripts fill in defaults such as ignored U-turns and permissive right-turns, which are conventions adopted by the paper.
  • domain assumption Inter-green intervals belong to the split of the associated phase.
    Section IV-B, Eq. (2): this stipulation is used to convert green time to split duration; it is a simplification, not a universal standard.
  • domain assumption All 16 vehicular and 12 pedestrian movements in Fig. 1 are available at the test intersection.
    Section III-A assumes this to build the test framework; real intersections must restrict the schema to actual signal groups.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Chat2SPaT: A Large Language Model Based Tool for Automating Traffic Signal Control Plan Management." pith.science (2026). https://pith.science/paper/JX6QAJPF

@misc{pith2026250705283,
  author       = {Pith},
  title        = {Pith review of: Chat2SPaT: A Large Language Model Based Tool for Automating Traffic Signal Control Plan Management},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JX6QAJPF}},
  note         = {Machine review of arXiv:2507.05283}
}
read the original abstract

Pre-timed traffic signal control, commonly used for operating signalized intersections and coordinated arterials, requires tedious manual work for signaling plan creating and updating. When the time-of-day or day-of-week plans are utilized, one intersection is often associated with multiple plans, leading to further repetitive manual plan parameter inputting. To enable a user-friendly traffic signal control plan management process, this study proposes Chat2SPaT, a method to convert users' semi-structured and ambiguous descriptions on the signal control plan to exact signal phase and timing (SPaT) results, which could further be transformed into structured stage-based or ring-based plans to interact with intelligent transportation system (ITS) software and traffic signal controllers. With curated prompts, Chat2SPaT first leverages large language models' (LLMs) capability of understanding users' plan descriptions and reformulate the plan as a combination of phase sequence and phase attribute results in the json format. Based on LLM outputs, python scripts are designed to locate phases in a cycle, address nuances of traffic signal control, and finally assemble the complete traffic signal control plan. Within a chat, the pipeline can be utilized iteratively to conduct further plan editing. Experiments show that Chat2SPaT can generate plans with an accuracy of over 94% for both English and Chinese cases, using a test dataset with over 300 plan descriptions. As the first benchmark for evaluating LLMs' capability of understanding traffic signal control plan descriptions, Chat2SPaT provides an easy-to-use plan management pipeline for traffic practitioners and researchers, serving as a potential new building block for a more accurate and versatile application of LLMs in the field of ITS. The source codes, prompts and test dataset are openly accessible at https://github.com/yuewangits/Chat2SPaT.

Figures

Figures reproduced from arXiv: 2507.05283 by the authors.

Figure 2
Figure 2. Various description styles for phase sequence. Based on users’ description, LLMs are instructed to record the plan structure in the phase sequence result. The phase sequence result is written as a list of stage or ring objects, with key ‘stageStyle’ or ‘ringStyle’ to label the structure style. For each ‘stageStyle’ (‘ringStyle’) object, the value is a nested list, with phases in the same stage (ring) recorded sequen… view at source ↗
Figure 3
Figure 3. Example plan description and LLM outputs. B. Left Turn Movements There are five common options for the left-turn phasing: permissive only, protected only, protected-permissive, split phasing, and prohibited [15]. To serve these phasing options, in the prompts, a left-turn phase is assumed to be protected when mentioned by the user, and could be specified as permissive or prohibited with attributes ‘isPermissive’ and… view at source ↗
Figure 4
Figure 4. Signal color sequence of each phase [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 24 canonical work pages

  1. [1]

    The study proposes Chat2SPaT, an LLM powered tool to enable automated chat based TSC plan management. This innovation allows for TSC plan creating and editing based on user descriptions, relieving traffic engineers of tedious manual SPaT parameter inputting and boosting efficiency for traffic agencies

    work page

  2. [2]

    Through extensive tests, the results show that the proposed methodology could understand a variety of description styles, and accurately generate and validate TSC plans

    An open -access test dataset with over 300 plan descriptions in both English and Chinese is created, covering common TSC plan schemes and special phase treatments. Through extensive tests, the results show that the proposed methodology could understand a variety of description styles, and accurately generate and validate TSC plans

    work page

  3. [3]

    NBT runs till the end of the cycle

    In Chat2SPaT, most traffic engineering nuances and numerical calculations are handled by TSC specific scripts, which minimizes the requirement for LLMs’ capability in complex mathematical reasoning. Moreover, fault tolerance functions are developed, targeting at typical errors in LLM outputs. These features make Chat2SPaT robust against LLM size, increasi...

    work page

  4. [4]

    Large Language Models for Intelligent Transportation: A Review of the State of the Art and Challenges,

    S. Wandelt, C. Zheng, S. Wang, Y. Liu, and X. Sun, “Large Language Models for Intelligent Transportation: A Review of the State of the Art and Challenges,” Applied Sciences, vol. 14, no. 17, Aug. 2024, Art. no. 7455

    work page 2024

  5. [5]

    TransGPT: Multi -modal generative pre-trained transformer for transportation,

    P. Wang, X. Wei, F. Hu, and W. Han, “TransGPT: Multi -modal generative pre-trained transformer for transportation,” in IEEE CLNLP, Yinchuan, China, Jul. 2024, pp. 96-100

    work page 2024

  6. [6]

    Formatting errors. As introduced in Section III, LLMs are instructed to record the plan structure as a list of json objects labeled as stages or rings, with phases in the same stage or ring written in nested lists. Yet, LLMs may get confused by the usage of json objects and lists, and sometimes omit the label of stage. Common formatting errors could be ad...

    work page

  7. [7]

    LLMs are used to convert user descriptions to keywords, with which scripts are designed to generate simulation networks

    to enable simulation scenario generation through chats. LLMs are used to convert user descriptions to keywords, with which scripts are designed to generate simulation networks. T 2 Strategic level . Leveraging LLMs’ skills of reasoning and task planning, traffic management frameworks are developed, combining LLMs and transportation functions. Prompts are ...

    work page

  8. [8]

    With a tendency to overthink [21] , LLMs sometimes decide to derive redundant results by mathematical calculation, based on available information in the description

    Overthinking errors. With a tendency to overthink [21] , LLMs sometimes decide to derive redundant results by mathematical calculation, based on available information in the description. However, as LLMs tend to fail in such tasks requiring professional domain knowledge and precise numerical calculations [6] , LLMs’ calculation results could potentially b...

    work page

Show all 30 references
  1. [9]

    Conforming to LLM’s scaling law [22] , the performance of Chat2SPaT declines when using Qwen3-32B, with an overall accuracy 3~6% lower than the larger models

    Semantic comprehension errors. Conforming to LLM’s scaling law [22] , the performance of Chat2SPaT declines when using Qwen3-32B, with an overall accuracy 3~6% lower than the larger models. In particular, for English cases, the accuracy drops by 12.3%, compared with ChatGPT -4...

  2. [10]

    and [13]. In other words, TSC is not a purely text and reasoning based task by nature, and replacing traditional TSC algorithms solely with language models at their current stage may not be the best option. Instead, using LLMs to aid traffic engineers in TSC strategically seem...

  3. [11]

    Exploring Traffic Signal Control: A Comprehensive Survey on Reinforcement Learning Techniques,

    K. ThamaraiSelvi, A.R. Bohra, P.S. Sunkara, B. Sunku, and B. NityaJignesh, “Exploring Traffic Signal Control: A Comprehensive Survey on Reinforcement Learning Techniques,” in IEEE ICIDeA, Bhubaneswar, India, Feb. 21, 2025, pp. 1-6

    2025

  4. [12]

    Traffic Signal Phase and Timing Estimation Using Trajectory Data from Radar Vision Integrated Camera,

    W. Zhou, Y. Wang, M. Liu, T. Liu, P. Zhang and Z. Ma, “Traffic Signal Phase and Timing Estimation Using Trajectory Data from Radar Vision Integrated Camera,” in IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 11, pp. 18279-18291, Aug. 2024

    2024

  5. [13]

    Advancing ITS Applications with LLMs: A Survey on Traffic Management, Transportation Safety, and Autonomous Driving,

    D. Zhang, H. Zheng, W. Yue, and X. Wang , “Advancing ITS Applications with LLMs: A Survey on Traffic Management, Transportation Safety, and Autonomous Driving,” in IJCRS , Halifax, Canada, May. 17, 2024, pp. 295-309

    2024

  6. [14]

    CityBench: Evaluating the capabilities of large language models for urban tasks,

    J. Feng et al. “CityBench: Evaluating the capabilities of large language models for urban tasks,” arXiv preprint arXiv:2406.13945v3, May 2025

    2025 arXiv

  7. [15]

    ChatSUMO: Large language model for automating traffic scenario generation in simulation of urban mobility,

    S. Li, T. Azfar, and R. Ke, “ChatSUMO: Large language model for automating traffic scenario generation in simulation of urban mobility,” IEEE Trans. Intell. Veh., early access, doi: 10.1109/TIV.2024.3508471, Aug. 2024

    2024

  8. [16]

    TrafficGPT: Viewing, processing and interacting with traffic foundation models,

    S. Zhang, et al., “TrafficGPT: Viewing, processing and interacting with traffic foundation models,” Transp. Policy , vol. 150, pp. 95 -105, May 2024

    2024

  9. [17]

    Open -ti: Open traffic intelligence with augmented language model,

    L. Da et al. “Open -ti: Open traffic intelligence with augmented language model,” Int. J. Mach. Lean. Cybern., vol. 15, no. 10, pp. 4761-4786, Oct. 2024

    2024

  10. [18]

    LLM-assisted light: Leveraging large language model capabilities for human-mimetic traffic signal control in complex urban environments,

    M. Wang, A. Pang, Y. Kan, M.O. Pun, C.S. Chen, and B. Huang, “LLM-assisted light: Leveraging large language model capabilities for human-mimetic traffic signal control in complex urban environments,” arXiv preprint arXiv:2403.08337, Mar. 2024

    2024 arXiv

  11. [19]

    Large language model -assisted arterial traffic signal control,

    Y. Tang, X. Dai, and Y. Lv, “Large language model -assisted arterial traffic signal control,” IEEE J. Radio Freq. Identif., vol. 8, pp. 322-326, Apr. 2024

    2024

  12. [20]

    Large language model-powered digital traffic engineers: The framework and case studies,

    X. Dai, Y. Tang, Y. Chen, Y. Zhang and Y. Lv, “Large language model-powered digital traffic engineers: The framework and case studies,” IEEE J. Radio Freq. Identif., vol. 8, pp. 780-787, Aug. 2024

    2024

  13. [21]

    LLMLight: Large language models as traffic signal control agents,

    S. Lai, Z. Xu, W. Zhang, H. Liu, and H. Xiong, “LLMLight: Large language models as traffic signal control agents,” arXiv preprint arXiv:2312.16044, Dec. 2023

    2023 arXiv

  14. [22]

    Development of conflict techniques for left-turn and cross-traffic at protected left-turn signalized intersections,

    J. Oh, E. Kim, M. Kim, and S. Choo, “ Development of conflict techniques for left-turn and cross-traffic at protected left-turn signalized intersections,” Safety Sci., vol. 48, no. 4, pp. 460-468, 2010

    2010

  15. [23]

    Traffic Signal Timing Manual

    United States Department of Transportation - Federal Highway Admini stration. Traffic Signal Timing Manual. (2021, Apr. 27). [Online]. Avail able: https://ops.fhwa.dot.gov/publications/fhwahop08024/chapter4.htm

    2021

  16. [24]

    Safety effectiveness of flashing yellow arrow: evaluation of 222 signalized intersections in North Carolina,

    C. L. Simpson and S. A. Troy, “Safety effectiveness of flashing yellow arrow: evaluation of 222 signalized intersections in North Carolina,” Transp. Res. Rec., vol. 2492, no. 1, pp. 46-56, Jan. 2015

    2015

  17. [25]

    Traffic signal phasing concepts and definitions

    Transportation Research Board and National Academies of Sciences, Engineering, and Medicine , “Traffic signal phasing concepts and definitions”, in Decision- Making Guide for Traffic Signal Phasing , Washington D.C., USA: NAP, 2020, pp. 10-11

    2020

  18. [26]

    Delay estimation and signal timing design techniques for multi -stage pedestrian crossings and two -stage bicycle left turns

    P. Furth and Y. Wang, “Delay estimation and signal timing design techniques for multi -stage pedestrian crossings and two -stage bicycle left turns”, in Transportation Research Board Annual Meeting. 2015, Washington D.C

    2015

  19. [27]

    Traffic Advisory Leaflet

    Department for Transport (UK). Traffic Advisory Leaflet. (2006, Mar.). [Online]. Available: https://www.tsrgd.co.uk/pdf/tal/2006/tal -1- 11 06_4.pdf

    2006

  20. [28]

    PTV Vissim and PTV Viswalk Help

    PTV VISSIM. PTV Vissim and PTV Viswalk Help . (2023, Jan.). [Onlin e]. Available: https://cgi.ptvgroup.com/vision -help/VISSIM_2022_EN G/Content/11_Auswertungen/AuswertungSignalzeitenplan.htm?Highli ght=signal%20times%20table

    2023

  21. [29]

    Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models,

    S. Yang et al. “Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models,” arXiv preprint arXiv:2503.16419, Mar. 2025

    2025 arXiv

  22. [30]

    Scaling Laws for Neural Language Models,

    K. Jared et al. “Scaling Laws for Neural Language Models,” arXiv preprint arXiv:2001.08361, Jan. 2020

    2001 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.