Pith. sign in

REVIEW 3 major objections 6 minor 37 references

Pairwise risk orderings learned from crash structure beat surrogate safety scores for earlier, more precise collision warnings.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Pairwise ordinal learning from temporal, event-contrast, and physics comparisons yields collision-relevant risk scores that improve high-recall warning precision and lead time over surrogate baselines on 100-Car and SHRP2.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Clean ordinal formulation for risk scoring from event structure; DRL gains on 100-Car/SHRP2 look real, with one load-bearing temporal assumption that is only partly stress-tested. the 3 major comments →

arxiv 2607.11128 v1 pith:T2LBI6JU submitted 2026-07-13 cs.RO cs.LG

Comparison-Based Ordinal Learning for Proactive Driving Risk Assessment

classification cs.RO cs.LG
keywords driving risk assessmentcollision-related riskordinal risk learningpairwise comparisonsproactive collision warningsurrogate safety measuresnaturalistic driving data
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Proactive driving risk assessment is hard because real collisions are rare and no one can label every frame with a true risk number. Most existing methods therefore fall back on kinematic proxies, trajectory predictors, or statistical anomaly scores that only loosely track actual crash danger. This paper argues that crash and near-crash sequences already contain usable ordinal information: later moments in a conflict are usually riskier than earlier ones, conflict frames are riskier than normal driving, and simple physical perturbations that tighten gaps or raise closing speeds are riskier than the unperturbed states. By turning those three sources into pairwise comparisons and training a scalar risk score with a Bradley–Terry ranking loss plus collision-endpoint anchoring, the method learns collision-relevant ordering without calibrated labels. On 100-Car and out-of-distribution SHRP2 crash data the resulting scores raise high-recall discrimination, warning precision, and lead time over standard surrogates, whether the score is learned from scratch or used to realign existing indicators.

Core claim

Event-structured pairwise supervision drawn from temporal progression inside safety-critical sequences, contrast between conflict and normal events, and physics-based counterfactual perturbations is sufficient to train collision-relevant risk scores that improve high-recall discrimination, warning precision, and precision-constrained lead time relative to representative kinematic and distribution-based surrogates on both in-distribution 100-Car and out-of-distribution SHRP2 crash data.

What carries the argument

Comparison-based ordinal risk learning: multi-source pairwise comparisons fed into a source-weighted Bradley–Terry ranking loss regularized by collision-endpoint anchoring, then applied to three parameterizations (direct neural risk score, single-surrogate alignment, multi-surrogate weighted fusion).

Load-bearing premise

Inside each recorded crash or near-crash, risk is assumed to rise (or at least not systematically fall) as the impact or critical endpoint approaches, so later frames can be treated as strictly riskier than earlier ones within a bounded time window.

What would settle it

If, on a held-out set of naturalistic crashes whose internal risk ordering is independently verified (for example by expert annotation or high-fidelity reconstruction), the learned score fails to rank later pre-impact frames above earlier ones and consequently yields no improvement in high-recall warning metrics over the same surrogates, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Automated-vehicle planners can substitute the learned ordinal scores for hand-crafted TTC-style proxies when issuing early collision warnings.
  • Existing surrogate models can be fine-tuned or re-weighted under the same pairwise objective without redesigning their architectures.
  • Scarce crash logs become more valuable because each sequence yields many temporal and counterfactual training pairs rather than a single binary label.
  • Cross-dataset transfer of risk ordering becomes feasible even when absolute risk scales differ between recording systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same construction of temporal, contrastive and counterfactual pairs could supply weak supervision for risk scoring in multi-agent robotics or pedestrian–vehicle interaction domains that also lack dense labels.
  • If the temporal-monotonicity assumption is relaxed with soft or bidirectional constraints, the framework might still work for near-misses that de-escalate after a peak.
  • Learned multi-surrogate weights could serve as a diagnostic of which classical indicators remain informative under modern traffic mixes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a comparison-based ordinal learning framework for proactive driving risk assessment that learns a scalar risk score from pairwise supervision rather than calibrated frame-level risk labels. Pairwise comparisons are constructed from three event-structured sources—temporal progression within crash/near-crash sequences, cross-sequence contrast between conflict and normal driving, and physics-based counterfactual perturbations—and optimized under a source-weighted Bradley–Terry objective with collision-endpoint anchoring (Eqs. 8–11). Three parameterizations are instantiated: direct neural risk-score learning (DRL), single-surrogate alignment of GSSM (SSRA), and multi-surrogate alignment of TAdv/EI/GSSM (MSRA). Evaluation on 100-Car (5-fold event-level CV) and held-out SHRP2 crashes uses high-recall warning metrics (A_ROC_α, P_PRC_α, mTTI_90%_p). DRL substantially outperforms kinematic and distributional baselines in-distribution and out-of-distribution; SSRA yields modest gains over GSSM; MSRA achieves high precision with interpretable weights. Source ablations and simulation-augmentation analyses are also reported.

Significance. If the results hold, the work offers a practical weak-supervision route for collision-relevant risk scoring when dense numerical labels are unavailable—an important and recurring bottleneck in naturalistic driving safety research. Framing risk assessment as ordinal comparison learning, and showing that the same objective can train a score from scratch or align existing surrogates, is a clear conceptual contribution relative to pure kinematic SSMs, prediction-then-risk pipelines, and distributional anomaly scores. Strengths include multi-source comparison construction with explicit weighting, collision-endpoint anchoring, three complementary instantiations, event-level CV, a true OOD crash set (SHRP2), operationally motivated high-recall metrics, and source/simulation ablations. These elements make the paper useful both as a learning method and as a calibration tool for existing surrogate indicators.

major comments (3)
  1. Section 3.2 (Temporal comparisons) and Algorithm 1 (lines 5–12) assert o_t2 ≻ o_t1 whenever t1 < t2 ≤ t_impact inside a bounded gap. This monotonicity assumption is load-bearing for fine-grained within-event supervision and for the long mTTI_90%_p gains claimed for DRL. Figure 2 Case 2 itself shows non-monotonic mid-sequence dynamics (score drop when relative speed falls), so a non-negligible fraction of temporal pairs may reverse true instantaneous risk. The paper should (i) explicitly state and bound this assumption, (ii) quantify how often surrogate or learned scores are non-monotonic within conflict sequences, and (iii) report a sensitivity check (e.g., soft labels, random temporal flips, or gap-restricted ablations) so that lead-time improvements cannot be attributed mainly to forced later-is-riskier labeling.
  2. Tables 1–2 report full in-distribution and OOD results for DRL and unaligned baselines, but SSRA (Table 3) and MSRA (Table 4) are evaluated only on 100-Car. The abstract and contributions claim the framework improves warning metrics “across both in-distribution and out-of-distribution evaluations” for the proposed instantiations. Either provide SHRP2 results for SSRA and MSRA under the same protocol, or narrow the OOD claim to DRL and state that aligned-surrogate generalization remains untested.
  3. Section 4.4 states that experiments report means and standard deviations over multiple random seeds under 5-fold event-level CV, yet Tables 1–4 show only point estimates with no ±std, confidence intervals, or significance tests. For SSRA especially, the gains are small (e.g., A_ROC_90% 82.82→83.18, mTTI 3.49→3.54 s in Table 3) and cannot be interpreted without variability. Please add fold/seed dispersion (and, where appropriate, paired tests) so that modest alignment gains are distinguishable from noise.
minor comments (6)
  1. In MSRA (Eq. 15 / Table 4), α_k is described as automatically accounting for surrogate risk direction, yet TAdv receives a negative weight. Clarify how sign(α_k) and w_k interact after sigmoid normalization, and whether negative w_k is expected or an optimization artifact.
  2. Hyperparameters β_temporal/cross/physics, λ_impact/λ_upper, r*, u, and [Δ_min, Δ_max] are free parameters (Section 3.3) but receive little sensitivity analysis beyond source-type ablations. A short appendix table on β/λ sensitivity would strengthen reproducibility.
  3. Related Work §2.2 discusses supervised accident-anticipation methods (e.g., RiskProp) but they are not included as baselines. A brief note on modality mismatch (video vs. kinematic features) or why they are out of scope would help readers place the comparison set.
  4. Figure 2 is useful but axis scales, threshold τ, and impact markers are hard to read in the text rendering; ensure consistent time axes and annotate t_impact and first-alert times for each method.
  5. Minor prose issues: duplicated “as follows::” in §4.3; occasional missing spaces after commas in compound adjectives; arXiv-style line breaks in the abstract/intro that should be cleaned for journal production.
  6. Algorithm 1 Step 4 constructs physics pairs for every frame and every variable then randomly subsamples to M; state the perturbation magnitudes used for each variable in 𝒱 so the physics source is fully reproducible.

Circularity Check

0 steps flagged

No circularity: pairwise ordinal targets and evaluation metrics are constructed from external event structure (impact times, crash/normal labels, physics directions) independent of the learned score R_φ.

full rationale

The paper does not claim a first-principles derivation of numerical risk; it constructs pairwise supervision from observed event structure (temporal order within conflict sequences ending at recorded t_impact, cross-sequence conflict-vs-normal labels, and directional physics perturbations) and optimizes a Bradley–Terry ranking loss plus impact-window anchoring (Eqs. 3–11, Algorithm 1). These inputs exist prior to and independently of R_φ; the learned score is then evaluated on held-out warning metrics (A_ROC_90%, P_PRC_90%, mTTI_90%_p) that use the same external endpoints only as ground-truth labels, which is ordinary supervised evaluation rather than a reduction of the claim to its inputs by construction. No parameter is fitted to a quantity and then re-presented as a prediction of that quantity; no uniqueness theorem or load-bearing premise is imported solely via overlapping-author citation; GSSM and kinematic baselines are external surrogates that are either left unaligned or fine-tuned under the same external ordinal objective. The temporal monotonicity assumption is a modeling choice that may inject label noise, but it is not circularity under the stated criteria. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on a small set of domain assumptions about how risk orders inside crash sequences and under physics perturbations, plus a handful of free weighting and anchoring constants that are chosen or tuned rather than derived. No new physical entities are postulated; the risk score is simply a learned scalar function of existing kinematic features.

free parameters (4)
  • source-level coefficients β_temporal, β_cross, β_physics
    Hand-chosen or tuned relative importance of the three comparison sources; they directly scale the pairwise loss terms.
  • anchoring weights λ_impact, λ_upper and reference values r*, u
    Balance the regression and upper-bound penalties that fix the absolute scale of R_φ; chosen to keep impact states near a high reference.
  • temporal gap bounds [Δ_min, Δ_max] and pair count M
    Control which frame pairs are admitted and how many comparisons are sampled per source; affect the training distribution.
  • MSRA affine-sigmoid parameters {w_k, α_k, β_k}
    Learned combination weights and per-surrogate scalings that determine the final multi-surrogate score.
axioms (5)
  • domain assumption Within a conflict sequence culminating in crash or near-crash, risk generally increases as the impact moment is approached, justifying o_t2 ≻ o_t1 for t1 < t2 ≤ t_impact.
    Stated in Section 3.2 (Temporal comparisons) and used to generate the majority of fine-grained supervision.
  • domain assumption A conflict-event observation is strictly riskier than a normal-driving observation (cross-sequence comparisons).
    Section 3.2; supplies the global calibration signal between dangerous and benign regimes.
  • domain assumption Local physics perturbations that decrease headway, increase closing speed, reduce lateral clearance or increase vehicle size directionally raise collision risk.
    Section 3.2 (Physics-based counterfactual comparisons); supplies shape constraints independent of event outcomes.
  • standard math Bradley–Terry model correctly maps score differences into pairwise ordering probabilities (Eq. 3).
    Standard preference-learning assumption used for the ranking loss.
  • domain assumption True risk V(o) is monotone non-decreasing in each chosen surrogate when others are held fixed (MSRA).
    Section 3.4.3; required for positive-weighted combinations to preserve ordering direction.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Comparison-Based Ordinal Learning for Proactive Driving Risk Assessment." pith.science (2026). https://pith.science/paper/T2LBI6JU

@misc{pith2026260711128,
  author       = {Pith},
  title        = {Pith review of: Comparison-Based Ordinal Learning for Proactive Driving Risk Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T2LBI6JU}},
  note         = {Machine review of arXiv:2607.11128}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Real-time driving risk assessment provides an essential basis for proactive safety by identifying and quantifying the danger of ongoing road interactions before adverse outcomes occur. However, due to the scarcity of collision data and frame-level risk labels, existing driving risk assessment methods often rely on surrogate objectives, which may imperfectly align with true collision risk and not faithfully reflect the relative danger of driving interaction. This paper proposes a comparison-based ordinal risk learning framework that learns collision-relevant risk scores from pairwise supervision in driving data, directly modeling relative risk ordering without requiring numerical frame-level risk labels. We derive pairwise comparisons from three sources of event-structured driving data for such ordinal risk learning: temporal progression within safety-critical sequences, event-level contrast between dangerous and normal interactions, and physics-based counterfactual perturbations. On this basis, instantiations with three risk-scoring function parameterizations are implemented, including directly learning risk scores from comparison data, and aligning existing single or multiple surrogate-based risk models. The proposed framework is evaluated on the 100-Car and SHRP2 naturalistic driving datasets using a proactive collision warning task. Results show that the proposed framework improves high-recall risk discrimination, warning precision, and warning lead time over representative surrogate-based baselines across both in-distribution and out-of-distribution evaluations. These results suggest that the proposed framework can contribute to proactive safety research by providing more reliable risk assessment for automated driving systems and safety-critical driving interactions.

Figures

Figures reproduced from arXiv: 2607.11128 by Bo Leng, Chongfeng Wei, Lu Xiong, Weiqi Zhang, Xinrui Zhang, Yi Zhong, Zhuoren Li.

Figure 1
Figure 1. Figure 1: The overall framework of our proposed comparison-based ordinal risk learning. where 𝑜 (𝑖) 𝑤 ≻ 𝑜(𝑖) 𝑙 denotes that observation 𝑜 (𝑖) 𝑤 is judged more dangerous than 𝑜 (𝑖) 𝑙 . We construct three comparison sources, each exploiting a different structural property of driving data. Temporal comparisons. Within a single conflict sequence that culminates in a crash or near-crash, risk generally in￾creases as the … view at source ↗
Figure 2
Figure 2. Figure 2: Risk score trajectories in three representative scenarios. s, and EI and TTC2D perform similarly. DRL, by contrast, maintains mTTI90% 𝑝 = 5.96 s, the longest warning lead time across both datasets. These results confirm that the ordinal supervision encoded in the comparison data provides a robust training signal: a model guided solely by collision￾anchored pairwise constraints generalises across recording … view at source ↗
Figure 3
Figure 3. Figure 3: Individual informativeness of pairwise supervision sources for DRL under different training event fractions. This trend suggests that their local monotonic constraints are useful for establishing the directional effects of risk￾relevant variables, but are insufficient by themselves to cap￾ture full event-level risk progression. Simulation-derived comparisons perform poorly when used alone, indicating that … view at source ↗
Figure 4
Figure 4. Figure 4: DRL performance aggregated by subset size over all combinations of three real-data comparison sources and one simulation-derived augmentation source [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of incrementally adding MetaDrive simulation comparison pairs on DRL performance, evaluated at small and large real-data budgets. 5. Conclusion This paper proposed a comparison-based ordinal learn￾ing framework for proactive driving risk assessment. The key observation is that naturalistic driving data, despite the scarcity of collision events, contain rich ordinal informa￾tion. Temporal progression… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references

  1. [1]

    Hybrid Action-Based Reinforcement Learning for Multiobjective Compatible Autonomous Driving , year=

    Jin, Guizhe and Li, Zhuoren and Leng, Bo and Han, Wei and Xiong, Lu and Sun, Chen , journal=. Hybrid Action-Based Reinforcement Learning for Multiobjective Compatible Autonomous Driving , year=

  2. [2]

    Safety-Enhanced Deep Reinforcement Learning for Autonomous Driving: Dare to Make Mistakes to Learn Better and Faster , year=

    Li, Zhuoren and Leng, Bo and Xiong, Lu and Eichberger, Arno and Huang, Chao and Hu, Jia , journal=. Safety-Enhanced Deep Reinforcement Learning for Autonomous Driving: Dare to Make Mistakes to Learn Better and Faster , year=

  3. [3]

    Accident Analysis & Prevention , volume=

    A review of surrogate safety measures and their applications in connected and automated vehicles safety modeling , author=. Accident Analysis & Prevention , volume=

  4. [4]

    2023 , publisher =

    Global Status Report on Road Safety 2023 , author =. 2023 , publisher =

  5. [5]

    and Klauer, Sheila G

    Dingus, Thomas A. and Klauer, Sheila G. and Neale, Vicki L. and Petersen, Andy and Lee, Suzanne E. and Sudweeks, Jeremy and Perez, Miguel A. and Hankey, Jonathan and Ramsey, David and Gupta, Santosh and Bucher, Christian and Doerzaph, Zachary R. and Jermeland, Julie and Knipling, Ronald R. , title =

  6. [6]

    Accident Analysis & Prevention , volume =

    Evaluation of traffic safety, based on micro-level behavioural data: Theoretical framework and first implementation , author =. Accident Analysis & Prevention , volume =

  7. [7]

    Transportation Research Part C: Emerging Technologies , volume =

    Emergency Index (EI): A two-dimensional surrogate safety measure considering vehicles’ interaction depth , author =. Transportation Research Part C: Emerging Technologies , volume =

  8. [8]

    Analytic Methods in Accident Research , volume=

    A unified probabilistic approach to traffic conflict detection , author=. Analytic Methods in Accident Research , volume=

  9. [9]

    Nature Machine Intelligence , volume =

    Learning collision risk proactively from naturalistic driving data at scale , author =. Nature Machine Intelligence , volume =

  10. [10]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , address =

    Uncertainty-Aware Probabilistic Risk Quantification of SOTIF for Autonomous Vehicles , author =. 2025 IEEE International Conference on Robotics and Automation (ICRA) , address =

  11. [11]

    Transportation Research Part E: Logistics and Transportation Review , volume =

    RiskNet: interaction-aware risk forecasting for autonomous driving in long-tail scenarios , author =. Transportation Research Part E: Logistics and Transportation Review , volume =. 2026 , issn =

  12. [12]

    Probabilistic Risk Metric for Highway Driving Leveraging Multi-Modal Trajectory Predictions , year=

    Wang, Xinwei and Alonso-Mora, Javier and Wang, Meng , journal=. Probabilistic Risk Metric for Highway Driving Leveraging Multi-Modal Trajectory Predictions , year=

  13. [13]

    Accident Analysis & Prevention , volume=

    Dynamic risk assessment for autonomous vehicles from spatio-temporal probabilistic occupancy heatmaps , author=. Accident Analysis & Prevention , volume=

  14. [14]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Zou, Yiyang and Zhao, Tianhao and Xiao, Peilun and Jin, Hongyu and Qi, Longyu and Li, Yuxuan and Liang, Liyin and Qian, Yifeng and Lai, Chunbo and Lin, Yutian and Li, Zhihui and Wu, Yu , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2026 , pages =

  15. [15]

    Transportation Research Part C: Emerging Technologies , volume=

    Predicted occupancy map-based online safety verification and resilient motion planning for autonomous driving , author=. Transportation Research Part C: Emerging Technologies , volume=

  16. [16]

    arXiv preprint arXiv:2601.20367 , year=

    Unsupervised Anomaly Detection in Multi-Agent Trajectory Prediction via Transformer-Based Models , author=. arXiv preprint arXiv:2601.20367 , year=

  17. [17]

    , title =

    Bradley, Ralph Allan and Terry, Milton E. , title =. Biometrika , volume =

  18. [18]

    2024 , month = apr, version =

    Jiao, Yiru , title =. 2024 , month = apr, version =

  19. [19]

    and Lee, Suzanne E

    Antin, Jonathan F. and Lee, Suzanne E. and Perez, Miguel A. and Dingus, Thomas A. and Hankey, Jonathan M. and Brach, Alison , title =. Safety Science , volume =

  20. [20]

    and Perez, Miguel A

    Hankey, Jonathan M. and Perez, Miguel A. and McClafferty, Julie A. , title =

  21. [21]

    Bird’s eye view trajectory reconstruction of naturalistic crashes and near-crashes in the SHRP2 NDS (Public Version) , year =

    Jiao, Yiru and Calvert, Simeon and Costa, Rufina , publisher =. Bird’s eye view trajectory reconstruction of naturalistic crashes and near-crashes in the SHRP2 NDS (Public Version) , year =

  22. [22]

    Hayward , title =

    John C. Hayward , title =. Highway Research Record , volume =

  23. [23]

    Accident Analysis & Prevention , volume =

    Modeling driver’s evasive behavior during safety–critical lane changes: Two-dimensional time-to-collision and deep reinforcement learning , author =. Accident Analysis & Prevention , volume =

  24. [24]

    2025 , issn =

    Interactive Risk (IR): An omnidirectional safety metric of CAVs based on multimodal trajectory prediction and driving risk field , journal =. 2025 , issn =

  25. [25]

    2025 , url =

    Shuguang Wang and Qian Zhou and Kui Wu and Dapeng Wu and Wei-Bin Lee and Jianping Wang , booktitle =. 2025 , url =

  26. [26]

    ACCV , publisher =

    Anticipating accidents in dashcam videos , author =. ACCV , publisher =

  27. [27]

    IROS , publisher =

    Unsupervised traffic accident detection in first-person videos , author =. IROS , publisher =

  28. [28]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Dota: Unsupervised detection of traffic anomaly in driving videos , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

  29. [29]

    ACMMM , pages =

    Uncertainty-based traffic accident anticipation with spatio-temporal relational learning , author =. ACMMM , pages =

  30. [30]

    ICCV , pages =

    Drive: Deep reinforced accident anticipation with visual explanation , author =. ICCV , pages =

  31. [31]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year =

    Accident Anticipation via Temporal Occurrence Prediction , author =. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year =

  32. [32]

    2022 IEEE Intelligent Vehicles Symposium (IV) , address =

    Learning to Predict Collision Risk from Simulated Video Data , author =. 2022 IEEE Intelligent Vehicles Symposium (IV) , address =

  33. [33]

    Accident Analysis & Prevention , volume =

    A physics-informed attention model for integrated driving risk assessment , author =. Accident Analysis & Prevention , volume =

  34. [34]

    Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , pages =

    Thorsten Joachims , title =. Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , pages =. 2002 , publisher =

  35. [35]

    Christiano and Jan Leike and Tom B

    Paul F. Christiano and Jan Leike and Tom B. Brown and Miljan Martic and Shane Legg and Dario Amodei , title =. Advances in Neural Information Processing Systems , volume =

  36. [36]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Li, Quanyi and Peng, Zhenghao and Feng, Lan and Zhang, Qihang and Xue, Zhenghai and Zhou, Bolei , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  37. [37]

    Physical Review E , volume =

    Treiber, Martin and Hennecke, Ansgar and Helbing, Dirk , title =. Physical Review E , volume =

This paper was first reviewed by grok-4.5 on July 14, 2026.