Pith. sign in

REVIEW 5 major objections 5 minor 39 references

Out of the Past: An AI-Enabled Pipeline for Traffic Simulation from Noisy, Multimodal Detector Data and Stakeholder Feedback

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A three-stage pipeline builds a granular traffic simulation from noisy camera footage, loop-detector counts, and plain-language stakeholder feedback, as shown on a 36-intersection network in Strongsville, Ohio.

desk verdict A genuinely useful pipeline integration with a solid CV counting step, but the simulation-level accuracy claim is not independently supported. read the letter →

arxiv 2505.21349 v2 pith:WLHON2CF submitted 2025-05-27 cs.CE

Rex Chen , Karen Wu , John McCartney , Norman Sadeh , Fei Fang This is my paper
classification cs.CE
keywords trafficsimulationdemandmodelingvehiclecountingcomputervisionquadraticintegerprogramminglargelanguagemodelmultimodaldetectordatastakeholderfeedback
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

The paper tries to establish that a faithful, granular traffic simulation can be built directly from the noisy detector data a city already collects, without expensive manual origin-destination surveys. Its three-stage pipeline counts vehicles from camera footage with a tracking-based vision method, converts the resulting count intervals and loop-detector counts into vehicle routes by solving a quadratic integer program, and then lets a large language model agent fold a traffic engineer's plain-language comments into additional optimization constraints. The testbed is a 24-hour, 36-intersection network in Strongsville, Ohio, where the pipeline produced a simulation whose total volume was about 200,000 vehicles and whose approach counts matched the vision-based counts on average. If these results hold, municipalities could make or refresh network-scale traffic simulations from existing cameras and loop detectors plus stakeholder knowledge, instead of treating detector data as noise-free ground truth or rebuilding demand from household activity surveys.

What carries the argument

The load-bearing object is the quadratic integer program over route usages $r_t$, with a binary route-location incidence matrix $A$ and slack variables $s^{CV}_t$ and $s^{LD}_t$ absorbing mismatch between the route-induced counts and the detector counts. Scalar multipliers $\alpha^{CV}_{LB}, \alpha^{CV}_{UB}, \alpha^{LD}_{LB}, \alpha^{LD}_{UB}$ convert calibration error into feasible count intervals. Two objective terms encode realism heuristics: penalizing routes that start or end inside the network and penalizing abrupt changes between adjacent time segments. The LLM agent's role is to modify this same optimization program by adding constraints generated from natural-language feedback, so that the final simulation is still a solution to one count-matching program.

What would settle it

Held-out manual counting can settle it: pick counting locations that were not among the four calibration intersections, count a peak hour not used to set $\alpha^{CV}_{LB}=0.94$ and $\alpha^{CV}_{UB}=1.12$, and compare the final simulation's approach volumes with those manual counts. Systematic violations of the calibrated ratio interval would refute the central accuracy claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that demand modeling can be reframed as constrained count matching. For each 15-minute segment, the authors enumerate the plausible routes — shortest paths between counting locations, with most trips starting and ending at network fringes — and solve a quadratic integer program that chooses how many vehicles take each route so that the implied counts at every counting location fall within calibrated error bounds for both the computer-vision and loop-detector sources. The same optimization penalizes routes that begin or end inside the network and encourages smooth changes between adjacent segments. The authors report that the vision-based counting step corrects the systematic undercounting of the deployed camera detector, that the route-matching program can be solved in under half a second per segment, and that an LLM agent with a self-reflection check converted real stakeholder feedback into executable constraints the stakeholder accepted. The Strongsville simulation is the concrete demonstration that these three steps together capture a city's traffic patterns.

Load-bearing premise

The load-bearing premise is that the error bounds estimated from manual counts at four intersections during two one-hour windows generalize to all 36 intersections and all 96 fifteen-minute segments; if those bounds are miscalibrated, the optimized route counts need not reflect real traffic.

Editorial extensions

If this is right

  • Existing camera and loop-detector installations can serve as a traffic-survey source, because tracking-based counting corrects the deployed detectors' systematic undercounting.
  • A municipality with no origin-destination survey can still obtain a plausible route set by enumerating shortest paths between counting locations and solving the per-segment quadratic integer program.
  • Traffic-engineering knowledge enters the model without the stakeholder writing math: the LLM agent converts qualitative comments into executable constraints on route counts.
  • Because the route-matching program solves in under half a second per 15-minute segment, the demand-modeling step is cheap enough to re-run as part of an interactive simulation-building workflow.
  • Treating both detector sources as noisy, with calibrated error bounds, is safer than the common practice of treating detector data as ground truth.

Reading between the lines

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

  • Editorial extension: the same per-segment count-matching program could run on streaming counts, turning this offline calibration process into a near-real-time simulation updater.
  • Editorial extension: for another city, the sparse calibration recipe (manual counting at a few intersections over a few hours) would need to be repeated locally, since the reported bounds are specific to Strongsville's detectors and conditions.
  • Editorial extension: the route-enumeration assumptions of shortest-path routing and fringe-heavy origins could be checked against independent GPS probe data; failing those checks would motivate adding route classes to the optimization rather than abandoning the pipeline.
  • Editorial extension: an external benchmark for the LLM-generated constraints, scored without the same model's self-reflection, would give a deploying agency independent evidence that the refinement step is faithful to the feedback.
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

5 major / 5 minor

Summary. The paper presents an end-to-end pipeline for traffic-demand modeling from noisy, multimodal detector data, applied to a road network in Strongsville, Ohio. The three steps are: (1) vehicle counting from camera footage using YOLOv11 plus BoT-SORT tracking, (2) route generation via a quadratic integer program (QIP) that enforces consistency with computer-vision and loop-detector counts through multiplicative error bounds, and (3) refinement of the simulation by an LLM agent that converts natural-language stakeholder feedback into additional optimization constraints. The authors claim that the resulting simulation accurately captures the city's traffic patterns, and that the approach generalizes to other municipalities. The CV-counting step is tested against manual counts at four intersections, the optimization is shown to be fast on this network, and the LLM step is evaluated on synthetic and real stakeholder feedback. However, the central simulation-level accuracy claim rests on validation against the same counts that were used as optimization constraints, and several auxiliary claims are not fully supported by the reported evidence.

Significance. If the central claim were established, the paper would make a valuable practical contribution: municipalities could build or update traffic simulations from existing camera and loop detectors plus stakeholder feedback, without costly manual origin-destination surveys. The paper has notable strengths: the code and prompts are released, the CV method is evaluated in situ against manual counts, the QIP formulation is exact rather than heuristic, and the LLM refinement is evaluated on synthetic feedback with a known ground truth. These elements make the pipeline a promising proof of concept. Nevertheless, the abstract's assertion that the pipeline 'accurately captures the city's traffic patterns in a granular simulation' is not supported by the current evidence, because the simulation-level validation is substantially circular and no independent route-level or link-level accuracy metric is provided. The paper therefore requires additional validation and a tempering of claims before the central message can be accepted.

major comments (5)
  1. [§5.2, Eq. (1)] The validation in Figure 5 compares simulation counts with the same f_CV and f_LD counts that appear as hard constraints (1b)-(1e) of the QIP. Because the solver is required to keep Ar_t within [alpha_LB f, alpha_UB f], agreement with the detector counts is partly by construction. The paper needs an independent accuracy metric: for example, held-out intersections, manual turn-movement counts, travel-time comparisons, or link-level count validation. Without such evidence, the abstract's claim that the pipeline 'accurately captures the city's traffic patterns in a granular simulation' is unsupported.
  2. [§5.1, Table 1] The statement that 'for all 16 of the footage excerpts that we manually counted, our method had an error of less than 2 vehicles per minute' is contradicted by Table 1. At Intersection 33 (Royalton Rd & I-71 Ramps), eastbound, 17:00, the manual count is 2088 and the CV count is 1962, an absolute difference of 126 vehicles per hour, or 2.1 vehicles per minute. Please correct this claim or adjust the reported threshold.
  3. [§4.2, §5.1] The error bounds alpha_CV_LB = 0.94 and alpha_CV_UB = 1.12 are computed from 16 excerpts at four intersections during two one-hour windows, and then applied globally to all 36 intersections and all 96 fifteen-minute time segments in Section 4.2. No held-out manual counts are used to test whether these bounds actually bracket the true counts at other locations or times. If the bounds are miscalibrated, the QIP constraints will not contain the true counts, directly undermining the route-generation step. Please report a coverage test on held-out excerpts or otherwise justify the temporal and spatial transfer. In addition, the loop-detector bounds [0.02, 19.06] are so loose that they impose almost no restriction on the solution; the multimodality claim should be tempered accordingly.
  4. [§4.2] The route-generation step relies on two assumptions: all vehicles follow shortest paths, and traffic originates or terminates only at the fringes of the network. These assumptions are imposed rather than tested, and the objective function's nonfringe penalty further forces solutions toward them. Even if aggregate intersection counts match the detector data, route-level and link-level errors can cancel while counts still agree. Since no route-level, link-level, or travel-time validation is provided, the claim that the simulation is 'granular' in accuracy is not established.
  5. [§5.3] For the real stakeholder feedback, there is no ground truth for the correct simulation outcome. The LLM's semantic-correctness reflection only checks whether the generated constraint is consistent with the feedback text; it does not verify that the resulting counts are accurate in the real world. For the 8 pieces of feedback where the stakeholder indicated that the simulation needed improvement, no before/after quantitative comparison against independent data is reported. The stakeholder's final concurrence is a useful sanity check, but it is not evidence that the refined simulation is more accurate.
minor comments (5)
  1. [Eq. (1)] The matrix A is defined as {0,1}^{n×m} with A_ij = 1 if route i passes counting location j, but then Ar_t is dimensionally invalid (an n×m matrix times an n-vector). Please use A^T r_t or define A as an m×n matrix.
  2. [§2.1, §2.3] In Section 2.1, 'we do not consider activity modeling in this chapter' should refer to 'this paper'; in Section 2.3, 'to my knowledge' should be rephrased or attributed to the authors.
  3. [§4.3] The notation QP(0), r(0), and QP(k) is used without a formal definition. Please define the iterative sequence of optimization problems and route solutions explicitly.
  4. [§6] The conclusion states that the methodology 'adheres more faithfully to real-world traffic conditions than approaches used in past work,' but no comparative experiment against prior methods is presented. Please temper this claim or add a comparison.
  5. [Figure 5] Reporting only the mean and range of differences across counting locations is not sufficient to assess fit. Please include quantitative error metrics, such as mean absolute error, root mean square error, or counts of locations exceeding the alpha bounds.

Circularity Check

2 steps flagged · score 7.0 of 10

The paper's central accuracy validation is circular: the QIP-generated simulation is constrained to match the same CV and loop-detector counts that Figure 5 then uses as the accuracy benchmark, and the LLM refinement step verifies itself through self-reflection by the same model that wrote the constraints.

  1. fitted input called prediction [Section 4.2, QIP constraints (1b)-(1e); Section 5.2, Figure 5]
    "s.t.Ar t +s CV t ≥α CV LB fCV t (1b) Art +s CV t ≤α CV UB fCV t (1c) Art +s LD t ≥α LD LB fLD t (1d) Art +s LD t ≤α LD UB fLD t (1e) Across counting locations on average, the simulation was accurate to our computer vision counts f CV j , with no overflow or underflow. This can be attributed to the relatively narrow range of [α CV LB ,α CV UB]."

    The QIP's decision variable r_t is explicitly constrained so that Ar_t (the simulation's counts at the counting locations) lies between alpha_CV_LB * f_CV and alpha_CV_UB * f_CV, and likewise between the loop-detector bounds. Figure 5 plots exactly Ar_t - f_CV and Ar_t - f_LD, which are the slack variables that the objective (1a) minimizes under those bounds. An average near zero therefore only shows that the solver honored its own constraints; it is not independent evidence that the generated routes reflect real traffic. No route-level, turn-movement, travel-time, or held-out intersection validation is provided, and the alpha_CV bounds were calibrated from only four intersections and two one-hour windows before being applied to all 36 intersections and 96 time segments.

  2. other [Section 4.3 (semantic correctness verification) applied in Section 5.3 (real stakeholder feedback)]
    "To do so, we use the LLM agent to perform self-reflection (Shinn et al. 2023)."

    For real stakeholder feedback, the paper explicitly notes that 'there is no ground truth,' yet the semantic-correctness verdict is produced by the same Claude 3.7 Sonnet model that generated the modified QIP. The reported 100% semantic correctness rate from reflection is therefore a self-consistency check between the model's generated code and the model's own interpretation of that code, not an external measurement that the refined simulation matches real traffic. No independent quantitative follow-up is reported for the 8 stakeholder feedback items that indicated needed improvement, so the LLM refinement step's accuracy claim is self-referential.

full rationale

The load-bearing circularity is in the central simulation-accuracy claim. In QIP (1), constraints (1b)-(1e) force Ar_t, the simulated counts at counting locations, to lie inside intervals built from f_CV and f_LD, with the slack s_CV and s_LD minimized in the objective. Section 5.2 then validates the simulation by plotting differences from those same f_CV and f_LD values (Figure 5) and concludes that the simulation was accurate to the computer-vision counts. That agreement is by construction: it expresses constraint satisfaction, not independent confirmation. The only external benchmark in the paper is the manual-count comparison for the CV step (Section 5.1), which is legitimate but is used to calibrate the alpha bounds, not to validate the final simulated routes against manual counts. The LLM refinement step adds a second self-referential element: for real stakeholder feedback, where the paper states there is no ground truth, the semantic-correctness signal comes from self-reflection by the same model that wrote the constraints. There are no load-bearing self-citations or imported uniqueness theorems, and the CV counting component does have an independent manual-count check, so the circularity is partial rather than total. Nevertheless, the central claim that the pipeline 'accurately captures the city's traffic patterns' is supported mainly by by-construction agreement with optimization inputs, warranting a score of 7.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on a chain of fitted bounds and unvalidated modeling assumptions. The most consequential are the alpha error bounds, fit to 16 manually counted excerpts and then applied globally, and the shortest-path and fringe-origin assumptions that define the route set. No new physical entities are introduced.

free parameters (6)
  • alpha_CV_LB = 0.94
    Lower bound ratio between manual ground-truth counts and CV counts, computed from 16 manually counted excerpts; applied as a global constraint to all CV counting locations and time segments.
  • alpha_CV_UB = 1.12
    Upper bound ratio likewise fit from manual counts; used in constraint (1c).
  • alpha_LD_LB = 0.02
    Derived from the minimum ratio of CV to loop counts over overlapping locations; extremely loose, making the loop constraints nearly vacuous.
  • alpha_LD_UB = 19.06
    Derived from the maximum CV-to-loop ratio; likewise loose, so the loop detector constraints carry little information.
  • lambda (nonfringe and temporal penalties) = 10
    Chosen by hand to balance the objective terms; no sensitivity analysis is reported. The text says 'we set the hyperparameter lambda=10'.
  • stop-bar crossing threshold epsilon = not specified
    Used in the initial counting check |b_y - S_y| <= epsilon; the numerical value is not reported.
assumptions (6)
  • domain assumption Drivers follow shortest paths (Dijkstra) between origin and destination.
    Used in Section 4.2 to enumerate the full route set; route realism depends on it.
  • domain assumption Most traffic originates and terminates at the fringes of the network; interior endpoints are rare.
    Used to limit routes and to add the nonfringe penalty in objective (1a); stated to hold only if all major sources and sinks are modeled.
  • domain assumption Manual counts of 16 excerpts are error-free ground truth.
    Manual counting is treated as ground truth without inter-rater reliability or error bars; it sets the CV error bounds.
  • ad hoc to paper Error bounds from 16 excerpts generalize to all counting locations and time segments.
    The alpha bounds are applied globally to all 96 time segments and all intersections, though measured at only 4 intersections and 2 hours.
  • domain assumption The QIP objective's realism heuristics (fringe penalty, temporal smoothness) select realistic routes among many feasible solutions.
    The problem is underconstrained; the chosen solution depends on these heuristics, which are not validated.
  • domain assumption LLM self-reflection is a reliable check on LLM-generated constraints.
    Semantic correctness on real feedback is verified by the same model that wrote the code, plus stakeholder agreement; no independent ground truth exists for real feedback.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Out of the Past: An AI-Enabled Pipeline for Traffic Simulation from Noisy, Multimodal Detector Data and Stakeholder Feedback." pith.science (2026). https://pith.science/paper/WLHON2CF

@misc{pith2026250521349,
  author       = {Pith},
  title        = {Pith review of: Out of the Past: An AI-Enabled Pipeline for Traffic Simulation from Noisy, Multimodal Detector Data and Stakeholder Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WLHON2CF}},
  note         = {Machine review of arXiv:2505.21349}
}
read the original abstract

How can a traffic simulation be designed to faithfully reflect real-world traffic conditions? One crucial step is modeling the volume of traffic demand. But past demand modeling approaches have relied on unrealistic or suboptimal heuristics, and they have failed to adequately account for the effects of noisy and multimodal data on simulation outcomes. In this work, we integrate advances in AI to construct a three-step, end-to-end pipeline for systematically modeling traffic demand from detector data: computer vision for vehicle counting from noisy camera footage, combinatorial optimization for vehicle route generation from multimodal data, and large language models for iterative simulation refinement from natural language feedback. Using a road network from Strongsville, Ohio as a testbed, we show that our pipeline accurately captures the city's traffic patterns in a granular simulation. Beyond Strongsville, incorporating noise and multimodality makes our framework generalizable to municipalities with different levels of data and infrastructure availability.

Figures

Figures reproduced from arXiv: 2505.21349 by the authors.

Figure 1
Figure 1. Our pipeline for generating a traffic simulation from multimodal detector data. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Demonstration of our vehicle tracking method on camera detector footage from intersection 8 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Screenshot of traffic simulation of Strongsville, Ohio. Counts are available for 36 intersec [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Plot of counts from our vehicle tracking method (green) and AutoScope (red), as ratios relative [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Plot of difference between counts in the QIP-generated simulation, counts from our vehicle [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Evaluation of B = 10 optimization programs generated by Claude for K = 20 synthetically generated pieces of feedback. We used three criteria to evaluate the programs: syntactic correctness (perfect), feasibility, and semantic correctness (both high but not perfect). Se…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 19 canonical work pages

  1. [9]

    LLMScenario: Large Language Model Driven Scenario Generation

    “LLMScenario: Large Language Model Driven Scenario Generation”.IEEE Transactions on Systems, Man, and Cybernetics: Systems54(11):6581–6594 https://doi.org/10.1109/TSMC.2024.3392930. Chen, R., K. M. Carley, F. Fang, and N. Sadeh

    arXiv 2024

  2. [10]

    Purpose in the Machine: Do Traffic Simulators Produce Distributionally Equivalent Outcomes for Reinforcement Learning Applications?

    “Purpose in the Machine: Do Traffic Simulators Produce Distributionally Equivalent Outcomes for Reinforcement Learning Applications?”. InProceedings of the 2023 Winter Simulation Conference, WSC ’23, 1842–1853. San Antonio, USA: ACM https://doi.org/10.5555/3643142.3643294. Codeca, L., R. Frank, and T. Engel

    work page arXiv 2023

  3. [13]

    Prompt to transfer: sim-to-real transfer for traffic signal control with prompt learning

    “Prompt to transfer: sim-to-real transfer for traffic signal control with prompt learning”. InProceedings of the 38th AAAI Conference on Artificial Intelligence, AAAI ’24, 82–90. Vancouver, Canada: AAAI https://doi.org/10.1609/aaai.v38i1.27758. Diamond, S., and S. Boyd

    work page doi:10.1609/aaai.v38i1.27758

  4. [15]

    AutoReward: Closed-Loop Reward Design with Large Language Models for Autonomous Driving

    “AutoReward: Closed-Loop Reward Design with Large Language Models for Autonomous Driving”.IEEE Transactions on Intelligent VehiclesEarly Access:1–13 https://doi.org/ 10.1109/TIV .2024.3485964. Chen, Wu, McCartney, Sadeh and Fang He, B. Y ., Q. Jiang, H. Ma, and J. Ma

    arXiv 2024

  5. [16]

    Multi-Agent Multimodal Transportation Simulation for Mega-cities: Application of Los Angeles

    “Multi-Agent Multimodal Transportation Simulation for Mega-cities: Application of Los Angeles”.Procedia Computer Science238:736–741 https://doi.org/10.1016/j.procs.2024.06.085. Jocher, G., and J. Qiu

    work page doi:10.1016/j.procs.2024.06.085 2024

  6. [18]

    A Demand Modelling Pipeline for an Agent-Based Traffic Simulation of the City of Barcelona

    “A Demand Modelling Pipeline for an Agent-Based Traffic Simulation of the City of Barcelona”. InProceedings of the 2023 Winter Simulation Conference, WSC ’23, 1777–1782. San Antonio, USA: ACM https://doi.org/10.1109/WSC60868.2023.10407551. Li, S., T. Azfar, and R. Ke

    arXiv 2023

  7. [19]

    ChatSUMO: Large Language Model for Automating Traffic Scenario Generation in Simulation of Urban MObility

    “ChatSUMO: Large Language Model for Automating Traffic Scenario Generation in Simulation of Urban MObility”.IEEE Transactions on Intelligent VehiclesEarly Access:1–12 https://doi.org/10.1109/TIV .2024.3508471. Lobo, S., S. Neumeier, E. M. G. Fernandez, and C. Facchi

    arXiv 2024

  8. [21]

    Effects of fog, snow, and rain on video detection systems at intersections

    “Effects of fog, snow, and rain on video detection systems at intersections”. Transportation Letters2(1):1–12 https://doi.org/10.3328/TL.2010.02.01.1-12. Mei, H., X. Lei, L. Da, B. Shi, and H. Wei

    work page doi:10.3328/tl.2010.02.01.1-12 2010

Show all 39 references
  1. [22]

    Libsignal: an open library for traffic signal control

    “Libsignal: an open library for traffic signal control”.Machine Learning113:5235–5271 https://doi.org/10.1007/s10994-023-06412-y. Michalopoulos, P. G

    doi

  2. [24]

    Towards Real-World Deployment of Reinforcement Learning for Traffic Signal Control

    “Towards Real-World Deployment of Reinforcement Learning for Traffic Signal Control”. InProceedings of the 20th IEEE International Conference on Machine Learning and Applications, ICMLA ’21, 507–514. IEEE https://doi.org/10.1109/ICMLA52953.2021.00085. Qiu, A., P. A. Sathish, D...

  3. [25]

    Advanced Traffic Demand Generation in SUMO: ML-based Prediction of Flow Rate based on Real-world Measured Datasets

    “Advanced Traffic Demand Generation in SUMO: ML-based Prediction of Flow Rate based on Real-world Measured Datasets”. InProceedings of the 2024 IEEE 99th Vehicular Technology Conference, VTC ’24, 1–7. Singapore: IEEE https://doi.org/10.1109/VTC2024-Spring62846.2024.10683300. R...

    2024

  4. [26]

    Vehicular traffic simulation in the city of Turin from raw data

    “Vehicular traffic simulation in the city of Turin from raw data”.IEEE Transactions on Mobile Computing21(12):4656–4666 https://doi.org/10.1109/TMC.2021.3075985. Rhodes, A., D. M. Bullock, J. R. Sturdevant, and Z. T. Clark

    2021

  5. [28]

    Reflexion: Language Agents with Verbal Reinforcement Learning

    “Reflexion: Language Agents with Verbal Reinforcement Learning”. InProceedings of the 37th Conference on Neural Information Processing Systems, NeurIPS ’23, 8634–8652. New Orleans, USA: NeurIPS https://doi.org/10.5555/3666122.3666499. Sun, Y ., N. S. Pargoo, P. J. Jin, and J. Ortiz

  6. [29]

    Optimizing Autonomous Driving for Safety: A Human-Centric Approach with LLM-Enhanced RLHF

    “Optimizing Autonomous Driving for Safety: A Human-Centric Approach with LLM-Enhanced RLHF”. InCompanion of the 2024 ACM International Joint Conference on Pervasive and Ubiquitous Computing, UbiComp Companion ’24, 76–80. New York, USA: ACM https://doi.org/10.1145/3675094.36775...

    2024

  7. [30]

    Vehicle Counting using Computer Vision: A Survey

    “Vehicle Counting using Computer Vision: A Survey”. In Proceedings of the 2022 IEEE 7th International Conference for Convergence in Technology, I2CT ’22, 1–7. Pune, India: IEEE https://doi.org/10.1109/I2CT54291.2022.9824432. Uppoor, S., and M. Fiore

    2022

  8. [32]

    LLM-Assisted Light: Leveraging Large Language Model Capabilities for Human-Mimetic Traffic Signal Control in Complex Urban Environments

    “LLM-Assisted Light: Leveraging Large Language Model Capabilities for Human-Mimetic Traffic Signal Control in Complex Urban Environments”.arXiv preprint2403.08337:1–20. Wei, H., N. Xu, H. Zhang, G. Zheng, X. Zang, C. Chen,et al

    arXiv

  9. [34]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”. InProceedings of the 36th Conference on Neural Information Processing Systems, NeurIPS ’22, 24824–24837. New Orleans, USA: NeurIPS https://doi.org/10.5555/3600270.3602070. Weinblatt, H., E. Minge, and S. Petersen

  10. [36]

    Hierarchically and cooperatively learning traffic signal control

    “Hierarchically and cooperatively learning traffic signal control”. InProceedings of the 35th AAAI Conference on Artificial Intelligence, AAAI ’21, 669–677. Virtual: AAAI https: //doi.org/10.1609/aaai.v35i1.16147. Zhang, H., S. Feng, C. Liu, Y . Ding, Y . Zhu, Z. Zhou,et al

    doi

  11. [37]

    CityFlow: A Multi-Agent Reinforcement Learning Environment for Large Scale City Traffic Scenario

    “CityFlow: A Multi-Agent Reinforcement Learning Environment for Large Scale City Traffic Scenario”. InProceedings of the 2019 World Wide Web Conference, WWW ’19, 3620–3624. San Francisco, USA: ACM https://doi.org/10.1145/3308558.3314139. Zheng, G., Y . Xiong, X. Zang, J. Feng,...

    2019

  12. [38]

    Learning Phase Competition for Traffic Signal Control

    “Learning Phase Competition for Traffic Signal Control”. InProceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM ’19, 1963–1972. Beijing, China: ACM https://doi.org/10.1145/3357384.3357900. Zhuo, T. Y ., M. C. Vu, J. Chim, H. Hu, W....

    1963

  13. [40]

    Fine-Tuning Language Models from Human Preferences

    “Fine-Tuning Language Models from Human Preferences”.arXiv preprint1909.08593:1–26. AUTHOR BIOGRAPHIES REX CHEN(email rexc@cmu.edu) is a PhD graduate from the Societal Computing program of the Software and Societal Systems Department in the School of Computer Science at Carneg...

    2002 arXiv

  14. [1991]

    A real-time computer vision system for vehicle tracking and traffic surveillance

    “A real-time computer vision system for vehicle tracking and traffic surveillance”.IEEE Transactions on Vehicular Technology40(1):21–29 https://doi.org/10.1109/25.69968. Müller, A., V . Rangras, G. Schnittker, M. Waldmann, M. Friesen, T. Ferfers,et al

    doi

  15. [2005]

    Evaluation of Stop Bar Video Detection Accuracy at Signalized Intersections

    “Evaluation of Stop Bar Video Detection Accuracy at Signalized Intersections”. Technical Report FHW A/IN/JTRP-2005/28, Joint Transportation Research Program, Indiana Department of Transportation and Purdue University. Shinn, N., F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao

    2005

  16. [2010]

    Models, Traffic Models, Simulation, and Traffic Simulation

    “Models, Traffic Models, Simulation, and Traffic Simulation”. InFundamentals of Traffic Simulation, edited by J. Barceló, 1–62. Springer https://doi.org/10.1007/978-1-4419-6142-6_1. Behari, N., E. Zhang, Y . Zhao, A. Taneja, D. Nagaraj, and M. Tambe

    doi

  17. [2011]

    Non-Local Means Denoising

    “Non-Local Means Denoising”.Image Processing On Line1:208–212 https: //doi.org/10.5201/ipol.2011.bcm_nlm. Chang, C., S. Wang, J. Zhang, J. Ge, and L. Li

    2011 doi

  18. [2012]

    Identifying chronic splashover errors at freeway loop detectors

    “Identifying chronic splashover errors at freeway loop detectors”.Transportation Research Part C: Emerging Technologies24:141–156 https://doi.org/10.1016/j.trc.2012.02.005. Leon, J. F., F. Giancola, A. Boccolucci, and M. Neroni

    2012 doi

  19. [2013]

    Length-Based Vehicle Classification Schemes and Length Bin Boundaries

    “Length-Based Vehicle Classification Schemes and Length Bin Boundaries”. Transportation Research Record2339(1):19–29 https://doi.org/10.3141/2339-03. Chen, Wu, McCartney, Sadeh and Fang Xu, B., Y . Wang, Z. Wang, H. Jia, and Z. Lu

    doi

  20. [2014]

    Traffic Simulation for All: A Real World Traffic Scenario from the City of Bologna

    “Traffic Simulation for All: A Real World Traffic Scenario from the City of Bologna”. InProceedings of the 2014 SUMO Conference, SUMO ’14, 47–60. Berlin, Germany: Springer https://doi.org/10.1007/978-3-319-15024-6_4. Bochenina, K., A. Taleiko, and L. Ruotsalainen

    2014 doi

  21. [2015]

    Luxembourg SUMO Traffic (LuST) Scenario: 24 hours of mobility for vehicular networking research

    “Luxembourg SUMO Traffic (LuST) Scenario: 24 hours of mobility for vehicular networking research”. InProceedings of the 2015 Vehicular Networking Conference, VNC ’15, 1–8. Kyoto, Japan: IEEE https://doi.org/10.1109/VNC.2015.7385539. Codecá, L., and J. Härri

    2015

  22. [2016]

    CVXPY: A Python-embedded modeling language for convex optimization

    “CVXPY: A Python-embedded modeling language for convex optimization”.Journal of Machine Learning Research17(83):1–5 https://doi.org/10.5555/2946645.3007036. Euthenics, and TranSystems

  23. [2017]

    Towards multimodal mobility simulation of C-ITS: The Monaco SUMO traffic scenario

    “Towards multimodal mobility simulation of C-ITS: The Monaco SUMO traffic scenario”. In Proceedings of the 2017 Vehicular Networking Conference, VNC ’17, 97–100. Turin, Italy: IEEE https://doi.org/10.1109/ VNC.2017.8275627. Coifman, B., D. Beymer, P. McLauchlan, and J. Malik

    2017

  24. [2018]

    Microscopic Traffic Simulation using SUMO

    “Microscopic Traffic Simulation using SUMO”. InProceedings of the 21st International Conference on Intelligent Transportation Systems, ITSC ’18, 2575–2582. Piscataway: IEEE https://doi.org/10.1109/ITSC.2018.8569938. Austin, J., A. Odena, M. Nye, M. Bosma, H. Michalewski, D. Do...

    2018

  25. [2019]

    A Comparative Study of State-of-the-Art Deep Learning Algorithms for Vehicle Detection

    “A Comparative Study of State-of-the-Art Deep Learning Algorithms for Vehicle Detection”.IEEE Intelligent Transportation Systems Magazine11(2):82–95 https://doi.org/10.1109/MITS.2019. 2903518. Wang, M., A. Pang, Y . Kan, M.-O. Pun, C. S. Chen, and B. Huang

    2019 doi

  26. [2020]

    InTAS - The Ingolstadt Traffic Scenario for SUMO

    “InTAS - The Ingolstadt Traffic Scenario for SUMO”. In Proceedings of the 2020 SUMO User Conference, SUMO ’20, 73–92. Berlin, Germany: SUMO https://doi.org/10.52825/ scp.v1i.102. Medina, J., M. Chitturi, and R. Benekohal

    2020

  27. [2021]

    Fine-Tuning Language Models from Human Preferences

    “Fine-Tuning Language Models from Human Preferences”.arXiv preprint2108.07732:1–34 https://doi.org/10.48550/arXiv.1909.08593. Barceló, J

    1909 doi arXiv

  28. [2022]

    BoT-SORT: Robust Associations Multi-Pedestrian Tracking

    “BoT-SORT: Robust Associations Multi-Pedestrian Tracking”.arXiv preprint2206.14651:1–13 https://doi.org/10.48550/arXiv.2206.14651. Alvarez Lopez, P., M. Behrisch, L. Bieker-Walz, J. Erdmann, Y .-P. Flötteröd, R. Hilbrich,et al

    doi arXiv

  29. [2023]

    Simulation-Based Origin-Destination Matrix Reduction: A Case Study of Helsinki City Area

    “Simulation-Based Origin-Destination Matrix Reduction: A Case Study of Helsinki City Area”. InProceedings of the 2023 SUMO User Conference, SUMO ’23, 1–7. Berlin, Germany: SUMO https://doi.org/10.52825/scp.v4i.197. Buades, A., B. Coll, and J.-M. Morel

    2023 doi

  30. [2024]

    A Decision-Language Model (DLM) for Dynamic Restless Multi-Armed Bandit Tasks in Public Health

    “A Decision-Language Model (DLM) for Dynamic Restless Multi-Armed Bandit Tasks in Public Health”. InProceedings of the 38th Conference on Neural Information Processing Systems, NeurIPS ’24, 1–38. Vancouver, Canada: NeurIPS https://doi.org/10.48550/arXiv.2402.14807. Bieker, L.,...

    doi arXiv

  31. [2025]

    BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions

    “BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions”. InProceedings of the 2025 International Conference on Learning Representations, ICLR ’25, 1–55. Singapore: ICLR https://doi.org/10.48550/arXiv.2406.15877. Ziegler, D. M., N. Stie...

    2025 doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.