Understanding Driving Risks using Large Language Models: Toward Elderly Driver Assessment

Asuka Harada; Hitoshi Kanamori; Linjing Jiang; Nihan Karatas; Takahiro Tanaka; Yuki Yoshihara

REVIEW 3 major objections 5 minor 21 references

Understanding Driving Risks using Large Language Models: Toward Elderly Driver Assessment

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 LLM sizes up traffic scenes like a human driver—with the right prompts

desk verdict A modest, honest feasibility study showing prompt design moves LLM agreement on driving-scene labels, but the two-rater reference standard and pseudo-replicated statistics keep the 'human-like judgment' conclusion on thin ice. read the letter →

arxiv 2507.08367 v1 pith:VT6SWCQ7 submitted 2025-07-11 cs.CV cs.SYeess.SY

Yuki Yoshihara , Linjing Jiang , Nihan Karatas , Hitoshi Kanamori , Asuka Harada , Takahiro Tanaka This is my paper · ORCID

classification cs.CVcs.SYeess.SY

keywords multimodallargelanguagemodelselderlydriverassessmentdrivingrisktrafficsceneunderstandingpromptengineeringchain-of-thoughtpromptingdashcamimageanalysisintersectionvisibility

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 asks whether a multimodal large language model can make the contextual traffic judgments that matter for older drivers: judging traffic density, intersection visibility, and whether a visible stop sign actually applies to the ego vehicle. The authors tested ChatGPT-4o on 200 static dashcam images using zero-shot, few-shot chain-of-thought, and multi-shot chain-of-thought prompts, comparing its labels with two human raters. They report that prompt design changes performance substantially: recall for intersection visibility rose from 21.7% (zero-shot) to 57.0% (multi-shot), traffic-density agreement rose from 53.5% to 67.6%, and stop-sign judgments had high precision (up to 86.3%) but lower recall (about 76.7%). The paper's conclusion is that, with well-designed prompts, these models can approximate human-like scene interpretation for elderly-driver-relevant assessments, while still requiring human oversight and larger validation.

What carries the argument

The machinery is the prompting protocol, not a new model architecture. Each dashcam image is sent to ChatGPT-4o with a role, task description, expected output format, and a varying number of exemplars: zero-shot gives no example, few-shot chain-of-thought gives one or two labeled examples with explanations, and multi-shot chain-of-thought lists every category with a detailed explanation. The structured output (label plus explanation) is parsed programmatically, compared with human annotations using precision, recall, and F1-score, and run ten times per image to measure response stability. This design isolates prompt content as the experimental variable.

What would settle it

Build a new, larger set of dashcam frames where scene-level truth is fixed by expert consensus or direct measurement (counted vehicles, line-of-sight visibility proxies, stop-sign orientation), run the multi-shot prompt, and compare the model's labels with that reference; if the model's agreement with the objective reference is no better than chance, or if the recall gains from added examples disappear on held-out scenes, the claim that prompt-guided LLMs approximate human-like risk judgment fails.

Watch

Extended reading notes

Core claim

The central discovery is that a single vision-language model can be steered toward human-like relational traffic judgments by the quantity and structure of the examples in its prompt. On the stop-sign task the model is conservative: its positive predictions are reliable (precision 0.838–0.863) but it misses some signs (recall 0.749–0.767). On traffic density the model tracks human global trends but struggles to recall moderate and highest-density scenes (recall as low as 29% and 0%, respectively, under the best prompt). On intersection visibility, the task where the two human raters themselves disagreed most (kappa = 0.325), multi-shot prompting lifted recall from 21.7% to 57.0%, and model F1-scores (0.531–0.577) came close to the human-human F1-score (0.579). The authors present this as a preliminary feasibility demonstration, not a deployment-ready system, and emphasize that explanatory outputs accompanied the labels.

Load-bearing premise

The load-bearing premise is that the two human raters' labels are a trustworthy reference for what each traffic scene truly contains and means, even though the raters disagreed substantially on visibility (a standard agreement statistic of 0.325).

Editorial extensions

If this is right

Prompt design is decisive: agreement between model and human labels moves substantially with the number and structure of examples, so any evaluation of LLM-based scene risk labeling must specify the prompting strategy.
The model is currently conservative for safety-relevant positives: it rarely overclaims stop signs or dense traffic but does miss them, so it is a screening aid, not an autonomous assessor.
Explanations track predictions in the tested scenes, which makes the labels auditable and usable for driver feedback and training.
Repeated model runs flag borderline scenes: images where repeated outputs disagree tend to overlap with images where the human raters disagree, so output instability can be used to identify ambiguous cases.
Stop-sign recall is markedly lower than precision, implying that a diagnostic tool built on this model would need a separate handling mechanism for missed signs before it could support elderly-driver assessment.

Reading between the lines

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

Editorial extension: the ten-run repetition protocol could be converted into an explicit uncertainty signal—images with high output variability, especially those overlapping human disagreement, could be routed automatically to human experts for review.
Editorial extension: the high-precision/low-recall stop-sign pattern implies a tunable safety tradeoff; a deployment-oriented prompt could deliberately bias toward recall for missed signs at the cost of extra false alarms, a tradeoff the paper observed but did not test.
Editorial extension: since the two human raters disagreed most on visibility, the measured ceiling for the model may be set by label ambiguity rather than visual understanding; expert adjudication or objective sight-distance measures would separate model error from reference error.
Editorial extension: the step from one still frame per scene to a short video clip is a directly testable next move for resolving the structurally ambiguous scenes identified here, and could be evaluated with the same prompting protocol.

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.

Desk Editor's Note

read the letter

Quick take: this paper is a small empirical study of ChatGPT-4o on three structured dashcam classification tasks, and it does what it claims—shows that prompting strategy changes agreement with human labels. It is not a breakthrough, and the authors do not oversell it too badly, but the "approximate human-like judgments" conclusion rests on a weaker reference standard than the framing implies.

What is genuinely useful: the comparison of zero-shot, few-shot CoT, and multi-shot CoT on the same images, with repeated sampling to test stability. The finding that visibility recall climbs from 21.7% to 57.0% with multi-shot is a concrete measurement, even if the underlying effect of prompt engineering is well established. The trace of conservative stop-sign behavior—high precision, low recall—matches other diagnostic LLM applications and is worth reporting. The stability overlap (Venn diagrams) between human disagreement and model inconsistency is a nice analytic touch; it suggests both share difficulty with ambiguous scenes.

The soft spots are real but not disqualifying. First and most load-bearing: the ground truth is two anonymized raters described as "experienced cautious drivers," not validated experts. Inter-rater kappa for visibility is 0.325, which is low enough that "agreement with human raters" mostly means "agreement with these two individuals, or with their majority." The paper says this in Limitations, but the main conclusion still leans on the same weak reference. Second, the statistical tests inflate the sample: replicating the four hundred human annotations ten times gives four thousand paired entries and drives the F-test degrees of freedom to 11,997. The independent sample is 200 images; the repeated model runs are clustered. Those p-values are overconfident. Third, no data, code, or exact prompts are released, so the numbers cannot be independently checked.

I would not reject this out of hand. The descriptive results—agreement rates, confusion matrices, stability analysis—are internally consistent, and the authors are explicit about the preliminary nature of the work. Who gets value: researchers working on LLM-based driver assessment or dashboard-cam analytics, and anyone building structured classification prompts for visual safety tasks. A serious referee could push for corrected statistics, a more defensible reference standard (more raters or expert adjudication), and artifact release. As is, it is a plausible feasibility study, not a demonstrated capability.

Recommendation: send it to review, but require the statistical fix and a softened conclusion or a better ground truth.

Referee Report

3 major / 5 minor

Summary. This paper investigates whether ChatGPT-4o can perform human-like interpretation of static dashcam images for three tasks relevant to elderly driver assessment: traffic density, intersection visibility, and stop-sign relevance. The authors compare three prompting strategies (zero-shot, few-shot chain-of-thought, and multi-shot chain-of-thought) against annotations from two human raters on 200 images, using accuracy, precision, recall, F1-score, confusion matrices, and inter-rater agreement statistics. The paper reports that prompt design significantly affects agreement rates, with the largest gains for intersection visibility (21.7% to 57.0%), and concludes that, with well-designed prompts, multimodal LLMs can partially approximate human-like judgments in traffic scene interpretation.

Significance. If the results hold, the study provides a useful feasibility benchmark for using multimodal LLMs as supportive tools in elderly driver diagnostics, particularly because it goes beyond object detection to contextual relational judgments and includes interpretable explanations. The paper has concrete strengths: it systematically varies prompting strategies, reports class-level confusion matrices rather than aggregate accuracy alone, compares the model against each rater and their majority vote, and explicitly examines output stability across repeated runs. These design choices make the evaluation more transparent than many similar LLM feasibility studies. However, the significance is limited by the small, non-public dataset and, more importantly, by a reference standard composed of only two non-expert raters with low inter-rater agreement on one of the three tasks.

major comments (3)

[Section III-F] The statistical analysis pseudo-replicates each human annotation ten times, creating 4,000 paired entries from only 200 images and two raters. The logistic regression F-tests (e.g., F(2,11997)=109.21 in Section IV-B) treat these replicated entries as independent observations, which inflates the effective sample size and produces artificially small p-values and confidence intervals. Please reanalyze the data using a model that accounts for the nesting of multiple model runs within images and raters (e.g., a mixed-effects logistic regression with random intercepts for image and rater, or cluster-robust standard errors at the image level), or use nonparametric paired tests on image-level summaries. The reported prompt effects may be qualitatively robust, but the inferential statistics as presented are not valid.
[Sections III-E and IV-D] The reference standard for the central claim consists of only two raters described as experienced cautious drivers, with low inter-rater agreement on intersection visibility (Cohen's kappa = 0.325) and only moderate agreement on traffic density (0.688). Because the paper operationalizes "human-like judgment" as agreement with these two individuals, the low kappa undermines the claim that the model approximates human-like risk perception: the model's F1 parity with the raters on visibility could reflect matching one rater's idiosyncratic biases or shared ambiguity rather than valid traffic-scene understanding. The acknowledgment in Section V-A that "only two raters... may have introduced individual bias" is accurate but does not reduce the load. Please either (a) validate the reference standard with a larger, more diverse annotation panel including expert raters, or (b) restrict the human-like judgment claim to tasks with high inter-rater agreement (e.g., stop-sign presence, kappa = 0.927) and frame the visibility results as demonstrating prompt sensitivity rather than human-like accuracy.
[Sections IV-C and I (Abstract)] The class-level results do not support the broad conclusion that the model "can approximate human-like judgments" across all three tasks. For stop-sign presence, the model's macro-averaged F1 (0.783-0.803 in Table I) is substantially below the human-human F1 (0.964), and the recall for the "present" class is only 55% (Section IV-C), not the 76.7% cited in the abstract as "lower recall" (which is the macro-averaged recall across both classes). For traffic density, model F1 (0.419-0.445) is far below human-human agreement (0.663). Only for visibility does the model approach human-level F1 (0.531-0.577 vs. 0.579). The abstract and conclusion should be revised to state that the model reaches human-like performance only on visibility and shows clearly inferior but above-chance performance on the other two tasks, unless the authors provide a specific argument for why lower F1 still constitutes "approximation."

minor comments (5)

[Abstract] The phrase "recall for intersection visibility increasing from 21.7% (zero-shot) to 57.0% (multi-shot)" mislabels what Section IV-B reports as "agreement probability," which is exact-match accuracy, not recall. Also, "recognizing stop signs recognition" contains a typo.
[Section IV-C] The recall reported for the stop-sign "present" class is 55%, while the abstract cites 76.7% as the "lower recall"; the latter is the macro-averaged recall across present and not-present classes. Please clarify which metric is used in each place.
[Section III-C] The term "multi-shot CoT" is misleading because the prompt merely lists all possible categories with descriptions rather than providing multiple exemplar scenes. Consider renaming it "definitional prompting" or "category-description prompting" to avoid confusion with the standard few-shot paradigm.
[Figure 5] The dashed lines labeled "chance-level accuracy" need a definition. For multi-class tasks, chance-level accuracy depends on the marginal distribution of human labels, which is not uniform and may differ across tasks; please specify how the chance level was computed.
[Section IV-E] The Venn diagrams would benefit from a precise definition of "inconsistent" for the model outputs; for example, is a model output considered inconsistent if any of the ten repeated runs disagrees with the majority label? Please add the criterion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the model outputs are evaluated against independently collected human annotations, and the prompt comparisons are empirical rather than reducing to any fitted input.

full rationale

The paper's derivation chain is empirical and self-contained. ChatGPT-4o is prompted with zero-shot, few-shot, and multi-shot templates to label traffic density, intersection visibility, and stop-sign presence (Sec. III-C), and its outputs are compared with labels independently provided by two human raters (Sec. III-E). No parameter is fitted to the human labels and then reported as a prediction; the agreement rates, F1 scores, and logistic regressions are computed directly from raw model outputs and raw human annotations. The prompting strategies are manipulated inputs, not outputs derived from the reference standard. The low inter-rater kappa for visibility (0.325) and the acknowledged limitation of using only two raters (Sec. V-A) are validity concerns about the reference standard, but they do not make the model's predictions equivalent to the human labels by construction. Self-citations in the paper (e.g., [6]) are contextual and not load-bearing for the central empirical claim. Therefore, no circular step can be exhibited, and the circularity score is 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new mathematical entities or fitted models are introduced. The central claim rests on domain assumptions about the validity of human annotations, sample representativeness, and prompt design; these are only partially acknowledged in the limitations (Section V-A).

free parameters (1)

API sampling hyperparameters = temperature=0.2, top_p=0.5, frequency_penalty=0.3, presence_penalty=0.2
Chosen by hand in Section III-D to balance accuracy and consistency. They are not fitted to the target tasks, but they could influence the reported agreement rates and are not varied in the analysis.

assumptions (4)

domain assumption Human labels are a valid reference standard for driving risk judgments
Section III-E defines two human raters as the ground truth; the paper does not validate these labels against expert opinion or objective risk data, and inter-rater agreement is low for visibility (kappa=0.325).
domain assumption The 200-image sample is representative of the larger dashcam corpus and of elderly driving scenarios
Section III-B states random selection from a non-public 3,000-segment corpus but gives no scene-diversity statistics; generalizability is assumed.
domain assumption Ten repeated model runs adequately characterize output stability
Section IV-E uses 10 repeats per image to define model inconsistency, but no convergence or saturation analysis is reported.
domain assumption The few-shot and multi-shot exemplar prompts are unbiased operationalizations of the tasks
Section III-C describes prompt construction but reports no pilot testing; the chosen exemplars may steer the model toward particular labels.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Understanding Driving Risks using Large Language Models: Toward Elderly Driver Assessment." pith.science (2026). https://pith.science/paper/VT6SWCQ7

@misc{pith2026250708367,
  author       = {Pith},
  title        = {Pith review of: Understanding Driving Risks using Large Language Models: Toward Elderly Driver Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VT6SWCQ7}},
  note         = {Machine review of arXiv:2507.08367}
}

read the original abstract

This study investigates the potential of a multimodal large language model (LLM), specifically ChatGPT-4o, to perform human-like interpretations of traffic scenes using static dashcam images. Herein, we focus on three judgment tasks relevant to elderly driver assessments: evaluating traffic density, assessing intersection visibility, and recognizing stop signs recognition. These tasks require contextual reasoning rather than simple object detection. Using zero-shot, few-shot, and multi-shot prompting strategies, we evaluated the performance of the model with human annotations serving as the reference standard. Evaluation metrics included precision, recall, and F1-score. Results indicate that prompt design considerably affects performance, with recall for intersection visibility increasing from 21.7% (zero-shot) to 57.0% (multi-shot). For traffic density, agreement increased from 53.5% to 67.6%. In stop-sign detection, the model demonstrated high precision (up to 86.3%) but a lower recall (approximately 76.7%), indicating a conservative response tendency. Output stability analysis revealed that humans and the model faced difficulties interpreting structurally ambiguous scenes. However, the model's explanatory texts corresponded with its predictions, enhancing interpretability. These findings suggest that, with well-designed prompts, LLMs hold promise as supportive tools for scene-level driving risk assessments. Future studies should explore scalability using larger datasets, diverse annotators, and next-generation model architectures for elderly driver assessments.

Figures

Figures reproduced from arXiv: 2507.08367 by the authors.

**Figure 2.** Evaluation pipeline. outputs were programmatically integrated (5) and compared to human annotations (6). B. Dashcam Dataset We used image frames extracted from a dashcam video corpus collected in a previously approved naturalistic driving study. Although the dataset is not publicly available, it was carefully curated and anonymized to ensure privacy and reflect the real-world driving risks faced by elderly drivers. … view at source ↗

**Figure 3.** Prompts used for each classification tasks: Zero-shot (top), few-shot [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Representative dashcam images and example outputs using the LLM [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Agreement rates across three prompting strategies (Zero-shot, Few [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Distribution of categorical labels generated by the model (top row) [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Confusion matrices (top row), recall matrices (middle row), and [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: Venn diagrams showing the overlap between human labeling inconsis [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages

[1]

Traffic accident statistics open data,

National Police Agency of Japan, “Traffic accident statistics open data,” https://www.npa.go.jp/publications/statistics/koutsuu/opendata/index opendata.html, 2024, available from the National Police Agency of Japan

work page 2024
[2]

The long-term effects of active training strategies on improving older drivers’ scanning in intersections: a two-year follow- up to romoser and fisher (2009),

M. R. Romoser, “The long-term effects of active training strategies on improving older drivers’ scanning in intersections: a two-year follow- up to romoser and fisher (2009),” Human factors , vol. 55, no. 2, pp. 278–284, 2013

work page 2009
[3]

J. L. Bell, M. A. Taylor, G.-X. Chen, R. D. Kirk, and E. R. Leatherman, “Evaluation of an in-vehicle monitoring system (ivms) to reduce risky driving behaviors in commercial drivers: Comparison of in-cab warning lights and supervisory coaching with videos of driving behavior,” Journal of safety research , vol. 60, pp. 125–136, 2017

work page 2017
[4]

Are inter- ventions effective at improving driving in older drivers?: A systematic review,

H. Castellucci, G. Bravo, P. Arezes, and M. Lavalli `ere, “Are inter- ventions effective at improving driving in older drivers?: A systematic review,” BMC geriatrics, vol. 20, pp. 1–25, 2020

work page 2020
[5]

The use of monitor- ing and feedback devices in driving: An assessment of acceptability and its key determinants,

A. Picco, A. Stuiver, J. de Winter, and D. de Waard, “The use of monitor- ing and feedback devices in driving: An assessment of acceptability and its key determinants,” Transportation research part F: traffic psychology and behaviour, vol. 92, pp. 1–14, 2023

work page 2023
[6]

Study on driver agent based on analysis of driving instruction data—driver agent for encouraging safe driving behavior (1)—,

T. Tanaka, K. Fujikake, T. Yonekawa, M. Yamagishi, M. Inagami, F. Kinoshita, H. Aoki, and H. Kanamori, “Study on driver agent based on analysis of driving instruction data—driver agent for encouraging safe driving behavior (1)—,” IEICE TRANSACTIONS on Information and Systems, vol. 101, no. 5, pp. 1401–1409, 2018

work page 2018
[7]

How ai from automated driving systems can contribute to the assessment of human driving behavior,

T. Driessen, O. Siebinga, T. de Boer, D. Dodou, D. de Waard, and J. de Winter, “How ai from automated driving systems can contribute to the assessment of human driving behavior,” Robotics, vol. 13, no. 12, p. 169, 2024

work page 2024
[8]

Integrating visual large language model and reasoning chain for driver behavior analysis and risk assessment,

K. Zhang, S. Wang, N. Jia, L. Zhao, C. Han, and L. Li, “Integrating visual large language model and reasoning chain for driver behavior analysis and risk assessment,” Accident Analysis & Prevention, vol. 198, p. 107497, 2024

work page 2024

Show all 21 references

[9]

Exploring the potential of multi-modal ai for driving hazard prediction,

K. Charoenpitaks, V .-Q. Nguyen, M. Suganuma, M. Takahashi, R. Ni- ihara, and T. Okatani, “Exploring the potential of multi-modal ai for driving hazard prediction,” IEEE Transactions on Intelligent Vehicles , 2024

2024
[10]

Surrealdriver: Designing llm-powered generative driver agent framework based on human drivers’ driving-thinking data,

Y . Jin, R. Yang, Z. Yi, X. Shen, H. Peng, X. Liu, J. Qin, J. Li, J. Xie, P. Gao et al., “Surrealdriver: Designing llm-powered generative driver agent framework based on human drivers’ driving-thinking data,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and ...

2024
[11]

Drivegpt4: Interpretable end-to-end autonomous driving via large language model,

Z. Xu, Y . Zhang, E. Xie, Z. Zhao, Y . Guo, K.-Y . K. Wong, Z. Li, and H. Zhao, “Drivegpt4: Interpretable end-to-end autonomous driving via large language model,” IEEE Robotics and Automation Letters , 2024

2024
[12]

Chatbot and fatigued driver: Exploring the use of llm-based voice assistants for driving fatigue,

S. Huang, X. Zhao, D. Wei, X. Song, and Y . Sun, “Chatbot and fatigued driver: Exploring the use of llm-based voice assistants for driving fatigue,” in Extended Abstracts of the CHI Conference on Human Factors in Computing Systems , 2024, pp. 1–8

2024
[13]

Older driver perception-reaction time for intersection sight distance and object detection, volume i,

N. D. Lerner, R. W. Huey, H. W. McGee, A. Sullivan et al. , “Older driver perception-reaction time for intersection sight distance and object detection, volume i,” United States. Department of Transportation. Federal Highway Administration, Tech. Rep., 1995

1995
[14]

Urban and rural differences in older drivers’ failure to stop at stop signs,

L. Keay, S. Jasti, B. Munoz, K. A. Turano, C. A. Munro, D. D. Duncan, K. Baldwin, K. J. Bandeen-Roche, E. W. Gower, and S. K. West, “Urban and rural differences in older drivers’ failure to stop at stop signs,” Accident Analysis & Prevention , vol. 41, no. 5, pp. 995–1000, 2009

2009
[15]

Left turns by older drivers with vision impairment: A naturalistic driving study,

T. A. Swain, G. McGwin Jr, J. F. Antin, J. M. Wood, and C. Owsley, “Left turns by older drivers with vision impairment: A naturalistic driving study,” Innovation in aging , vol. 5, no. 3, p. igab026, 2021

2021
[16]

Large lan- guage models are zero-shot reasoners,

T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large lan- guage models are zero-shot reasoners,” Advances in neural information processing systems, vol. 35, pp. 22 199–22 213, 2022

2022
[17]

Language mod- els are few-shot learners,

T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod- els are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

1901
[18]

Nuplanqa: A large-scale dataset and benchmark for multi- view driving scene understanding in multi-modal large language mod- els,

S.-Y . Park, C. Cui, Y . Ma, A. Moradipari, R. Gupta, K. Han, and Z. Wang, “Nuplanqa: A large-scale dataset and benchmark for multi- view driving scene understanding in multi-modal large language mod- els,” arXiv preprint arXiv:2503.12772 , 2025

2025 arXiv
[19]

When technology tells you how you drive—-truck drivers’ attitudes towards feedback by technology,

M. Roetting, Y .-H. Huang, J. R. McDevitt, and D. Melton, “When technology tells you how you drive—-truck drivers’ attitudes towards feedback by technology,” Transportation Research Part F: Traffic Psy- chology and Behaviour , vol. 6, no. 4, pp. 275–287, 2003

2003
[20]

Exploring whether chatgpt-4 with image analysis capabilities can diagnose os- teosarcoma from x-ray images,

Y . Ren, Y . Guo, Q. He, Z. Cheng, Q. Huang, and L. Yang, “Exploring whether chatgpt-4 with image analysis capabilities can diagnose os- teosarcoma from x-ray images,” Experimental Hematology & Oncology, vol. 13, no. 1, p. 71, 2024

2024
[21]

Unlocking the potential of medical imaging with chatgpt’s intelligent diagnostics,

A. Alzahem, S. Latif, W. Boulila, and A. Koubaa, “Unlocking the potential of medical imaging with chatgpt’s intelligent diagnostics,” Procedia Computer Science , vol. 225, pp. 3583–3592, 2023

2023

Pith tools

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

[1] [1]

Traffic accident statistics open data,

National Police Agency of Japan, “Traffic accident statistics open data,” https://www.npa.go.jp/publications/statistics/koutsuu/opendata/index opendata.html, 2024, available from the National Police Agency of Japan

work page 2024

[2] [2]

The long-term effects of active training strategies on improving older drivers’ scanning in intersections: a two-year follow- up to romoser and fisher (2009),

M. R. Romoser, “The long-term effects of active training strategies on improving older drivers’ scanning in intersections: a two-year follow- up to romoser and fisher (2009),” Human factors , vol. 55, no. 2, pp. 278–284, 2013

work page 2009

[3] [3]

J. L. Bell, M. A. Taylor, G.-X. Chen, R. D. Kirk, and E. R. Leatherman, “Evaluation of an in-vehicle monitoring system (ivms) to reduce risky driving behaviors in commercial drivers: Comparison of in-cab warning lights and supervisory coaching with videos of driving behavior,” Journal of safety research , vol. 60, pp. 125–136, 2017

work page 2017

[4] [4]

Are inter- ventions effective at improving driving in older drivers?: A systematic review,

H. Castellucci, G. Bravo, P. Arezes, and M. Lavalli `ere, “Are inter- ventions effective at improving driving in older drivers?: A systematic review,” BMC geriatrics, vol. 20, pp. 1–25, 2020

work page 2020

[5] [5]

The use of monitor- ing and feedback devices in driving: An assessment of acceptability and its key determinants,

A. Picco, A. Stuiver, J. de Winter, and D. de Waard, “The use of monitor- ing and feedback devices in driving: An assessment of acceptability and its key determinants,” Transportation research part F: traffic psychology and behaviour, vol. 92, pp. 1–14, 2023

work page 2023

[6] [6]

Study on driver agent based on analysis of driving instruction data—driver agent for encouraging safe driving behavior (1)—,

T. Tanaka, K. Fujikake, T. Yonekawa, M. Yamagishi, M. Inagami, F. Kinoshita, H. Aoki, and H. Kanamori, “Study on driver agent based on analysis of driving instruction data—driver agent for encouraging safe driving behavior (1)—,” IEICE TRANSACTIONS on Information and Systems, vol. 101, no. 5, pp. 1401–1409, 2018

work page 2018

[7] [7]

How ai from automated driving systems can contribute to the assessment of human driving behavior,

T. Driessen, O. Siebinga, T. de Boer, D. Dodou, D. de Waard, and J. de Winter, “How ai from automated driving systems can contribute to the assessment of human driving behavior,” Robotics, vol. 13, no. 12, p. 169, 2024

work page 2024

[8] [8]

Integrating visual large language model and reasoning chain for driver behavior analysis and risk assessment,

K. Zhang, S. Wang, N. Jia, L. Zhao, C. Han, and L. Li, “Integrating visual large language model and reasoning chain for driver behavior analysis and risk assessment,” Accident Analysis & Prevention, vol. 198, p. 107497, 2024

work page 2024