{"id":"f8102057-9671-473b-b78d-3408f0346950","arxiv_id":"2607.00027","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"K-means clustering on 1219 Argoverse 2 deceleration events identifies four stable modes with bootstrap ARI 0.897, and an early HistGradientBoosting classifier reaches macro-F1 0.758 with modest scene-context gains.","lead":"This paper extracts over 1200 sustained deceleration events from Argoverse 2 urban driving data, clusters them into four stable kinematic modes via K-means, and trains a classifier to predict mode from the first second of motion. Scene context adds limited value, with only pair age showing a medium effect.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"K-means on 19D kinematic vectors may yield artifacts of feature engineering or event selection rather than distinct behavioral modes","rationale":"The reader's weakest_assumption directly isolates the same methodological hinge that the abstract's central claim rests on. No stronger internal inconsistency or external-consensus violation appears from the provided material; the concern is therefore already correctly located and the UNVERDICTED verdict does not require adjustment.","tokens_in":1898,"tokens_out":334,"duration_ms":16349,"concrete_test":"Recompute the four-mode solution after (i) replacing K-means with GMM on the identical 19D vectors and (ii) ablating the three highest-variance kinematic features; if adjusted Rand index between the new and original partitions falls below 0.6 or the mode proportions shift by >15 percentage points, the stability claim is sensitive to the exact pipeline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result (four modes with bootstrap ARI 0.897) depends on the claim that K-means applied to the chosen 19 kinematic features on the 1,219 extracted events recovers interpretable, stable modes. This is vulnerable because (a) no justification is given for K-means over alternatives (GMM, hierarchical), (b) the 19 features and the precise definition of 'sustained deceleration events' are fixed without reported ablation, and (c) bootstrap ARI only tests resampling stability, not sensitivity to those upstream choices. The reported regime dependence at low speed (ARI 0.166) already hints that the recovered partition is not uniformly robust.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper extracts 1,219 sustained deceleration events from 234 Argoverse 2 urban driving logs, encodes them as 19-dimensional kinematic feature vectors, applies K-means clustering with bootstrap stability analysis to identify four modes (anticipatory soft 62.8%, reactive closing 30.6%, brake-like jerk 4.8%, outlier 1.8%) with ARI 0.897, quantifies scene-context modulation (pair age epsilon^2=0.085 dominant), and trains a HistGradientBoosting early-event classifier achieving macro-F1=0.758 at 1.0 s (scene context adds +0.059 F1). Modes are reported as regime-invariant at medium speeds (ARI 0.817) but dependent at low speeds (ARI 0.166).","tokens_in":2067,"tokens_out":686,"duration_ms":18897,"significance":"If the recovered modes are shown to be robust to upstream choices rather than artifacts, the work supplies a data-driven taxonomy for urban deceleration with direct utility for AV motion planning and car-following models. Credit is due for leveraging a public multi-agent dataset, reporting bootstrap ARI across 50 resamples, and providing an early-window classifier that isolates jerk as the dominant predictive signal.","major_comments":[{"comment":"The central claim of four stable, interpretable modes with bootstrap ARI 0.897 rests on K-means applied to a fixed 19-dimensional kinematic feature set extracted from 1,219 events; however, the manuscript provides no justification for K-means over alternatives (GMM, hierarchical) and reports no ablation on the feature definitions, the precise criteria for 'sustained deceleration events,' or the number of clusters. This leaves the partition vulnerable to being an artifact of feature engineering or event selection rather than distinct behavioral regimes.","section":"Methods (event extraction and clustering)"},{"comment":"The regime analysis reports ARI=0.166 at low speed versus 0.817 at medium speed; this sensitivity directly undermines the headline assertion of 'stable modes' and 'regime-invariant' structure in medium-speed driving without additional controls or re-clustering experiments that isolate speed as a covariate.","section":"Results (regime dependence)"},{"comment":"The classifier performance (macro-F1=0.758 at 1.0 s) and scene-context contribution (+0.059 F1) are presented without reported sensitivity to the kinematic feature definitions or to the clustering-derived labels themselves; because the labels are the output of the unablated K-means step, downstream predictive claims inherit the same upstream uncertainty.","section":"Results (early-event classifier)"}],"minor_comments":[{"comment":"The abstract and methods should explicitly list or tabulate the 19 kinematic features and the numerical thresholds used to define 'sustained deceleration events' to support reproducibility.","section":"Abstract/Methods"},{"comment":"Figure captions and the bootstrap procedure description would benefit from stating the exact K-means hyperparameters (initialization, distance metric, max iterations) used for the reported ARI values.","section":"Methods/Figures"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive critique, which identifies key areas where our methodological choices require stronger justification. We address each major comment below with clarifications drawn from the manuscript and indicate revisions where the points are well-taken.","responses":[{"response":"K-means was selected for its scalability on 1219 samples in 19D space and for producing directly interpretable centroids that align with kinematic intuition (e.g., jerk magnitude). The bootstrap ARI of 0.897 across 50 resamples supplies quantitative evidence against pure artifact. We nevertheless agree that explicit rationale and sensitivity checks are absent. In revision we will (i) state the rationale and cite comparable trajectory-clustering studies, (ii) report the elbow/silhouette analysis used to select k=4, and (iii) add a limitations paragraph on feature and event-definition choices. Full ablations on every upstream decision exceed the scope of the current study but will be flagged as future work.","revision_made":"partial","referee_comment":"[Methods (event extraction and clustering)] The central claim of four stable, interpretable modes with bootstrap ARI 0.897 rests on K-means applied to a fixed 19-dimensional kinematic feature set extracted from 1,219 events; however, the manuscript provides no justification for K-means over alternatives (GMM, hierarchical) and reports no ablation on the feature definitions, the precise criteria for 'sustained deceleration events,' or the number of clusters. This leaves the partition vulnerable to being an artifact of feature engineering or event selection rather than distinct behavioral regimes."},{"response":"The manuscript already reports both ARI values and explicitly qualifies the claim: modes are regime-invariant only at medium speeds. The low-speed dependence is presented as a substantive finding rather than a flaw. To avoid any overstatement we will revise the abstract, results, and discussion to foreground the speed qualification and remove any unqualified use of 'regime-invariant.' Additional controlled re-clustering experiments that hold speed fixed are not performed in the current work; we will note this as a limitation and direction for follow-up.","revision_made":"yes","referee_comment":"[Results (regime dependence)] The regime analysis reports ARI=0.166 at low speed versus 0.817 at medium speed; this sensitivity directly undermines the headline assertion of 'stable modes' and 'regime-invariant' structure in medium-speed driving without additional controls or re-clustering experiments that isolate speed as a covariate."},{"response":"We acknowledge that the classifier inherits label uncertainty from the upstream K-means step. The bootstrap stability metric offers partial mitigation by showing label consistency under resampling, yet we agree that direct sensitivity of the F1 score to feature definitions or alternative labelings is unreported. In revision we will add a short limitations subsection stating this dependency and noting that the reported scene-context gain is conditional on the current pipeline. No new classifier ablations will be added at this stage.","revision_made":"partial","referee_comment":"[Results (early-event classifier)] The classifier performance (macro-F1=0.758 at 1.0 s) and scene-context contribution (+0.059 F1) are presented without reported sensitivity to the kinematic feature definitions or to the clustering-derived labels themselves; because the labels are the output of the unablated K-means step, downstream predictive claims inherit the same upstream uncertainty."}],"tokens_in":1627,"tokens_out":727,"duration_ms":23077,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this paper pulls 1,219 sustained deceleration events from Argoverse 2, runs K-means on 19 kinematic features, and reports four modes that hold up under bootstrap resampling with ARI 0.897. It also shows a gradient boosting classifier reaching macro-F1 0.758 from the first second of data, with a small lift from scene variables. That specific quantification on this dataset is new.\n\nWhat works is the use of a public multi-agent trajectory set and the stability check across resamples. Pair age comes out as the only context variable with a medium effect, and the low-speed regime dependence is called out plainly. The early-window prediction result is straightforward to interpret for behavior modeling.\n\nThe soft spot is the clustering step itself. No ablation on the 19 features, no comparison to GMM or other methods, and the event extraction rules are not detailed enough to judge selection bias. The low ARI at low speeds already flags that the partition is not uniform, so the modes could shift with different upstream choices. The abstract does not show sensitivity tests on those decisions.\n\nThis is for researchers working on car-following models or AV simulation who need empirical mode counts from real urban data. It is not a methods paper and does not claim broader theory.\n\nI would send it to peer review. The data source is solid and the numbers are reported clearly enough that referees can ask for the missing checks.","headline":"Four-mode taxonomy from K-means on Argoverse 2 deceleration events is new empirical work but rests on untested feature and event choices.","tokens_in":2560,"tokens_out":369,"would_cite":false,"duration_ms":12430,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Urban deceleration events cluster into four stable kinematic modes identifiable from the first second of motion.","keywords":["urban deceleration","behavior modes","K-means clustering","Argoverse 2","kinematic classifier","car following","scene context","early prediction"],"falsifier":"Re-clustering the same 1,219 events using a different algorithm such as Gaussian mixture models or altering the feature vector to exclude jerk components, then finding low agreement with the original four-mode partition, would falsify the stability of the reported modes.","tokens_in":2783,"feed_emoji":"🚗","tokens_out":702,"duration_ms":18982,"temperature":0.7,"pith_summary":"The paper extracts over a thousand sustained deceleration events from urban driving logs in the Argoverse 2 dataset. It encodes each as a 19-dimensional vector of kinematic features and applies K-means clustering to reveal four modes that prove stable under bootstrap resampling. An early classifier predicts which mode an event belongs to using only the initial 1.0 seconds of data, achieving a macro-F1 score of 0.758, with scene context providing a modest boost. Pair age emerges as the primary scene variable influencing mode distribution, while most other context factors show little effect.","feed_headline":"Four modes organize urban deceleration from early kinematics","feed_subtitle":"Argoverse 2 data yields stable anticipatory and reactive types with 0.758 F1 prediction at 1 second using pair age as key modulator.","key_machinery":"K-means clustering applied to 19-dimensional kinematic feature vectors extracted from sustained deceleration events, validated by bootstrap stability analysis, and used to train an early-event HistGradientBoosting classifier.","core_discovery":"Four stable behavioral modes structure urban deceleration: anticipatory soft braking at 62.8 percent, reactive closing at 30.6 percent, brake-like jerk at 4.8 percent, and an outlier category at 1.8 percent. These modes are recovered via K-means on 19 kinematic features with a bootstrap Adjusted Rand Index of 0.897. The modes remain consistent in medium-speed driving but shift in low-speed regimes. An early kinematic classifier reaches macro-F1 of 0.758 at one second, and pair age is the dominant contextual modulator.","pith_inferences":["These modes could serve as building blocks for more human-like car-following controllers in autonomous vehicles.","Low-speed dependence suggests the need for speed-regime-specific models in dense traffic.","Early detection of mode could enable proactive safety interventions in shared road scenarios.","The dominance of early jerk features points to jerk as a key signal for intent inference."],"forward_implications":["Mode membership can be predicted reliably from the first second of an event's kinematics.","The four modes hold across medium-speed driving conditions but vary at low speeds.","Scene context adds only a small improvement to classification accuracy beyond kinematics.","Pair age shows a medium effect on mode distribution while geometry and vulnerable road user proximity show negligible effects."],"fun_headline_variants":["Urban deceleration has four stable kinematic modes","Early classifier achieves 0.758 F1 on deceleration modes","Argoverse 2 data clusters deceleration into four modes","Pair age is main modulator of deceleration behavior modes"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That applying K-means to the selected 19-dimensional kinematic features on the extracted events will identify distinct and interpretable behavioral modes instead of artifacts of the feature design or clustering choices.","fun_headline_variants_meta":{"raw":{"variants":["Urban deceleration has four stable kinematic modes","Early classifier achieves 0.758 F1 on deceleration modes","Argoverse 2 data clusters deceleration into four modes","Pair age is main modulator of deceleration behavior modes"]},"model":"grok-4.3","cost_usd":0.006854,"raw_usage":{"total_tokens":3232,"prompt_tokens":766,"num_sources_used":0,"completion_tokens":60,"cost_in_usd_ticks":68537000,"prompt_tokens_details":{"text_tokens":766,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2406,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":766,"tokens_out":60,"duration_ms":20404,"temperature":1.0,"reasoning_tokens":2406,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-02T21:44:19.248589+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Re-clustering the same 1,219 events using a different algorithm such as Gaussian mixture models or altering the feature vector to exclude jerk components, then finding low agreement with the original four-mode partition, would falsify the stability of the reported modes.","supporting_citations":[],"review_version":1}