{"id":"4dcd88a6-6735-4bd2-86fd-65458dfcb341","arxiv_id":"2501.13461","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"I2XTraj, a roadside-unit model, jointly predicts all vehicles' trajectories at signalized intersections by encoding traffic signals and maneuver priors, and outperforms several baselines on V2X-Seq and SinD, though not on every reported setting.","lead":"This paper proposes I2XTraj, an infrastructure-based model that predicts the future trajectories of all vehicles at signalized intersections using traffic light states and maneuver priors, then transmits those predictions to vehicles. The authors report large accuracy gains over prior methods on two real-world datasets, though some headline figures are selective.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SinD results contradict the claimed 15% improvement: on the 12-12 setting I2XTraj has minADE 0.09 vs KI-GAN 0.05, so the headline performance claim is not supported by the paper's own table.","rationale":"The reader identified the maneuver-interval partition as the weakest load-bearing premise, and I agree it is under-validated. However, the more immediate load-bearing concern is the empirical claim itself: the SinD table contradicts the abstract's 15% improvement, and the V2X-Seq multi-agent comparison has only one baseline. Since the paper's central contribution is the comparative performance, and the reported tables do not support the headline numbers as stated, the paper should not be accepted without either corrected claims or additional experiments. I do not recommend rejection because the method may still offer value if the comparisons are clarified and the claims are narrowed. The conditional verdict is appropriate, with required revisions including: (1) reconcile the SinD numbers with the abstract, (2) broaden the V2X-Seq multi-agent baselines or explicitly limit the claim to the HPNet comparison, and (3) explain or remove the under-specified D-based interval prior. I therefore keep the reader's CONDITIONAL verdict, and my agreement is partial because my primary concern is the performance evidence rather than the geometric prior itself.","tokens_in":16464,"tokens_out":6634,"duration_ms":57641,"concrete_test":"Re-run the SinD evaluation for both I2XTraj and KI-GAN under identical data splits, mode counts, and metrics, reporting per-seed minADE and minFDE for both 12-12 and 12-18. If the 12-12 minADE remains 0.09 vs 0.05, the abstract's 'over 15% improvement on SinD' claim is false and must be revised to specify the exact settings and baselines where an improvement holds. Additionally, inspect the implementation to verify whether the D-based interval boundaries from Section IV-B are used in computing P_gt or in initializing mode anchors; if they are not used anywhere, the claimed 'driving strategy awareness' mechanism is not actually exercising the stated prior.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim, stated in the abstract and contributions, is that I2XTraj 'outperforms existing state-of-the-art methods by more than 30% on V2XSeq and 15% on SinD.' Table III directly contradicts this for SinD: in the 12-12 setting, KI-GAN achieves minADE 0.05 and minFDE 0.12, while I2XTraj achieves minADE 0.09 and minFDE 0.12. I2XTraj is therefore 80% worse on the primary minADE metric and only ties on minFDE. In the 12-18 setting, I2XTraj improves over KI-GAN by 9% in minADE (0.10 vs 0.11) and 19% in minFDE (0.21 vs 0.26), so the claimed uniform '>15%' improvement is not consistently supported. Furthermore, the '>30% on V2XSeq' claim rests on a single multi-agent baseline (HPNet) in Table II; Table I shows improvements ranging from 5.9% to 47% depending on baseline and metric. Because the quantitative comparison is the central evidence for the contribution, the paper's headline results as stated are not established. The interval-based maneuver prior in Section IV-B is also under-specified: the D-dependent intervals are not visibly connected to the mode-query equations (8)-(13), so it is unclear whether the prior actually shapes the predictions or merely provides an untested label source for L_cls.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Infrastructure-to-Everything (I2X) scheme in which roadside units predict the joint future trajectories of all vehicles at signalized intersections and transmit these predictions unidirectionally to subscribing vehicles. The accompanying model, I2XTraj, combines a continuous traffic-signal encoding, a driving-strategy-awareness module built from interval-based maneuver priors, and a spatial-temporal-mode attention network. The method is evaluated on two real-world datasets, V2X-Seq and SinD, in single-infrastructure and online collaborative settings, including robustness experiments against data loss and communication delay. The paper claims that I2XTraj outperforms state-of-the-art methods by more than 30% on V2X-Seq and more than 15% on SinD.","tokens_in":16786,"tokens_out":7299,"duration_ms":58440,"significance":"If the claims are supported, the I2X prediction scheme is a practical contribution: it decouples infrastructure-side prediction from vehicle-side models, uses traffic signal information and maneuver priors that are naturally available to infrastructure, and reports a lightweight model (3.18M parameters, 112 ms inference). The ablations show that each proposed component contributes positively, and the robustness experiments on data loss and delay address an important deployment concern. However, the headline quantitative claims are not fully supported by the paper's own tables. In particular, the SinD 12-12 result in Table III contradicts the abstract's '>15%' claim, and the V2X-Seq '>30%' claim is not uniformly supported by the multi-agent comparison in Table II. These issues are load-bearing because the paper's central contribution is the claimed empirical superiority.","major_comments":[{"comment":"The claimed 'more than 15%' improvement on SinD is not supported by Table III. In the 12-12 setting, I2XTraj's minADE is 0.09 against KI-GAN's 0.05, which is an 80% worse minADE, and its minFDE ties KI-GAN at 0.12. In the 12-18 setting, the improvement over KI-GAN is 9.09% in minADE (0.10 vs 0.11) and 19.23% in minFDE (0.21 vs 0.26), so even the better setting does not meet a uniform 15% threshold on both metrics. The abstract and contribution should state per-setting, per-metric improvements and identify the baseline for each percentage.","section":"Abstract, Contribution (3), and Table III"},{"comment":"The five-interval maneuver partition is internally inconsistent. The right-turn interval is printed as [\\sqrt{2}D/2, D/2), whose lower endpoint exceeds its upper endpoint for any positive D; the U-turn interval [1, \\sqrt{2}D/2) overlaps the left-turn interval [D/2, D) for typical intersection diameters; and the straight interval [\\sqrt{2}D, \\infty) leaves the range [D, \\sqrt{2}D) unassigned. The subsequent K-mode formula {[0,1), [1, D/(K-2)), ..., [(k-3)D/(K-2), D), [D,\\infty)} does not preserve the five-interval semantics and omits the \\sqrt{2}D/2 boundary. Please correct the intervals and specify exactly how the interval memberships produce the labels used in L_cls (Eq. 21) and the mode queries in Eqs. (8)-(13).","section":"Section IV-B"},{"comment":"The 'more than 30% on V2X-Seq' claim is not uniformly supported. Table II, the only multi-agent V2X-Seq comparison, shows improvements over HPNet of 23.53% in minJointADE, 30.8% in minJointFDE, and 19.05% in minJointMR; only one metric exceeds 30%. In the single-agent Table I, improvements over the strongest collaborative baselines are much smaller, for example 5.9% in minADE and 4.5% in minFDE relative to Co-MTP. The authors should identify the exact baseline and metric for the 30% figure or weaken the claim.","section":"Section V-B and Tables I-II"},{"comment":"The off-map loss is claimed in Section V-B a to be responsible for the MR improvements, but no ablation isolates it. Table IV varies CSI, MAP, and DSA only, and Eq. (22) always includes L_off-map. Please add an ablation without the off-map term, or explicitly retract the attribution.","section":"Section IV-D and Table IV"}],"minor_comments":[{"comment":"The text lists 'twelve state-of-the-art models,' but the enumeration contains ten V2X-Seq baselines and five SinD baselines; the count should be reconciled.","section":"Section V-A c"},{"comment":"The rows use only check marks without row labels or a legend, so it is not clear which components are present or absent in each row; add a configuration column or a caption that defines each row.","section":"Table IV"},{"comment":"The notation \\sqrt{2}D/2 is ambiguous: it should be written as \\frac{\\sqrt{2}}{2}D to avoid confusion with \\sqrt{2D}/2.","section":"Section IV-B"},{"comment":"The implementation statement 'each dynamic graph attention consists of four layers of multi-head attention' is ambiguous; please specify whether each GATv2 operation has four attention heads or four stacked layers.","section":"Section V-A d"},{"comment":"The sentence 'I2XTraj framework also performances improved prediction accuracy' is ungrammatical and should be rephrased.","section":"Section V-B b"},{"comment":"After the quantitative claims are corrected, the percentages in the abstract and the conclusion should be synchronized with the revised per-setting, per-metric results.","section":"Abstract and Section VI"}],"recommendation":"major_revision","confidential_remarks":"The manuscript describes a useful system and the ablations suggest the components help, but the quantitative headline claims need substantive correction. I recommend major revision rather than rejection because the core method and experiments are largely sound; however, the authors must recompute and report all percentages relative to the appropriate baselines and settings, and clarify the maneuver-interval construction. If the claims are corrected to the actual table values, the contribution is more modest but still publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core idea is genuinely useful: an RSU predicts joint trajectories for all vehicles and broadcasts them unidirectionally to any subscribing vehicle, without requiring the vehicle to send data back. That is a clean framing, and it is new relative to the coupled VIC pipelines they compare against. Second, the headline numbers are not supported by their own tables. On SinD 12-12, I2XTraj gets minADE 0.09 while KI-GAN gets 0.05; that is 80% worse, not 15% better. The 15% claim only survives in the 12-18 setting, and there it is 9–19% depending on metric. The abstract and conclusion repeat the over-15% claim, so this is a real overstatement, not a rounding issue.\n\nWhat the paper does well: the continuous traffic light encoding is a sensible adaptation, the ablations show each component contributes, and the robustness experiments on data loss and communication delay are the kind of practical evaluation this line of work often skips. The I2X scheme itself is a worthwhile contribution even if the model is assembled from GATv2 and query-based decoding.\n\nThe soft spots, in order of importance. First, the performance claims need to be rewritten to match the tables. Second, the multi-agent comparison on V2X-Seq uses only HPNet as baseline; that is too thin to support “over 30%.” Third, the maneuver-interval prior in Section IV-B is ad hoc: boundaries like [√2D/2, D/2) are defined from a single intersection diameter D, with no validation across T-shaped or multi-branch intersections, and the connection between these intervals and the mode queries in Eqs. (8)–(13) is not made explicit. Fourth, there is no code, no seeds, no error bars, and no data split details, so the robustness results cannot be independently checked.\n\nThe central mechanism is not circular and the paper is coherent on its own terms; it just overclaims. I think a serious editor should send this to review, but the revision should be major: fix the comparisons, add baselines, clarify or remove the D-based interval prior, and release code or at least detailed experimental setup. I would not cite it in its current form, but I would watch the revised version.","headline":"A promising infrastructure-side prediction design with overstated headline results: the SinD table itself undercuts the 15% claim, but the I2X framing and robustness work merit a serious referee.","tokens_in":17330,"tokens_out":1429,"would_cite":false,"duration_ms":14100,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A roadside unit can jointly predict every vehicle's future trajectory at a signalized intersection using traffic-light timing and maneuver priors, outperforming vehicle-centric and cooperative baselines by 30% on real-world V2I data.","keywords":["trajectory prediction","signalized intersections","Infrastructure-to-Everything","multi-agent prediction","traffic signal encoding","maneuver strategy prior","V2X communication","joint multi-modal trajectory prediction"],"falsifier":"Train I2XTraj on a T-shaped or multi-branch signalized intersection where a single-diameter radial partition does not match the lane geometry, and check whether per-mode probability calibration and minJointFDE degrade compared with an intersection-agnostic mode clustering; a correct geometric prior should also keep the predicted mode distribution nearly unchanged when $D$ is varied by a small amount.","tokens_in":16210,"feed_emoji":"🚦","tokens_out":8896,"duration_ms":68697,"temperature":0.7,"pith_summary":"This paper argues that the right place to predict what vehicles will do at a signalized intersection is the roadside, not the vehicle. It proposes an Infrastructure-to-Everything (I2X) scheme in which roadside units forecast the joint future trajectories of every vehicle in the intersection and broadcast those predictions unidirectionally to any subscribing autonomous vehicle. The accompanying model, I2XTraj, feeds real-time traffic light states, prior maneuver-strategy knowledge, and multi-agent interactions into a single attention-based network. On two real-world signalized-intersection datasets the model reports over 30% and 15% gains over existing baselines, and it degrades only gradually under data loss and communication delay. If the claims hold, infrastructure can provide a common prediction service that any autonomous driving system can use without tight coupling to specific on-board models.","feed_headline":"Roadside AI predicts all vehicles' paths at once","feed_subtitle":"Using traffic-light timing and maneuver priors, it beats prior models by 30% on real-world intersection data.","key_machinery":"Two mechanisms carry the argument. The continuous signal-informed mechanism encodes each traffic light's remaining time with a nonlinear function $PE_{t,\\ell} = \\sin\\left(\\frac{t_{\\text{remain}}}{T_\\Omega} \\cdot \\left(\\frac{1}{3}\\right)^d\\right)$, making the model sensitive to signal transitions and usable across intersection layouts. The driving strategy awareness mechanism partitions feasible maneuvers by radial distance from a reference point using the intersection diameter $D$: $[0,1)$ for stop, $[\\sqrt{2}D/2, D/2)$ for right turn, $[D/2, D)$ for left turn, $[1, \\sqrt{2}D/2)$ for U-turn, and $[\\sqrt{2}D, \\infty)$ for straight, generalized to $K$ modes as $\\{[0,1), [1, D/(K-2)), \\ldots, [(k-3)D/(K-2), D), [D,\\infty)\\}$. These intervals become mode queries through dynamic graph attention and $K$ independent decoders, producing joint trajectory proposals that an off-map loss keeps on drivable areas.","core_discovery":"I2XTraj's central discovery is that the infrastructure can jointly predict all vehicles' future trajectories at signalized intersections by making the prediction problem knowledge-driven: a continuous encoding of remaining traffic-light times guides trajectory proposals; a driving strategy awareness mechanism maps each vehicle onto a small set of maneuver intervals derived from intersection diameter $D$ and geometry; and a spatial-temporal-mode attention network refines joint proposals across agents, time steps, and modes. The paper reports that this outperforms existing methods by more than 30% on V2X-Seq and 15% on SinD, remains accurate when half the historical data is dropped, and keeps single-agent errors nearly stable up to 400ms latency.","pith_inferences":["If the radial-distance partition generalizes, the same strategy-aware design could extend to roundabouts, highway ramps, or other constrained maneuver sets, though the paper only demonstrates signalized intersections.","A direct test of whether the geometric prior is essential would replace the $D$-based intervals with data-driven maneuver clustering and compare; if gains persist, the prior could be learned rather than assumed.","The I2X scheme's value for real traffic depends on how subscribing planners consume joint multi-modal scenes; the paper evaluates prediction quality but not the closed-loop safety outcome.","The same subscription architecture could be extended to pedestrians and cyclists by adding their state encodings and maneuver priors, since the model already handles a small shared space qualitatively."],"forward_implications":["Infrastructure can offer joint multi-agent trajectory prediction as a public subscription service, so heterogeneous autonomous vehicles do not need a unified on-board prediction model or shared data with the roadside.","Traffic-signal encoding should let predictors anticipate phase transitions, including early 'jump-start' accelerations and vehicles rushing a yellow light, rather than treating signals as static context.","Strategy-aware mode generation is what gives the model its robustness to missing data, because maneuver priors stay stable when vehicle-state observations are dropped.","Joint multi-agent predictions degrade gracefully with communication delay up to roughly 400ms, suggesting roadside or edge deployment is feasible for real-time use.","The combination of strategy priors and an off-map loss sharply reduces miss rate compared to baselines, which matters for downstream planning and safety."],"supporting_citations":[{"why":"Supplies the real-world V2I dataset with 28 intersections and traffic-light data on which the main 30% improvement claim is demonstrated.","marker":"[19]"},{"why":"Supplies the drone-captured signalized-intersection dataset used to test generalizability and the 15% improvement claim.","marker":"[40]"},{"why":"Provides the strong unsignalized-intersection baseline that I2XTraj beats on V2X-Seq and SinD, and supplies the relative spatio-temporal edge encoding used in the graph.","marker":"[34]"},{"why":"Supplies the factorized joint multi-agent prediction baseline and the joint regression loss formulation adapted here.","marker":"[5]"},{"why":"Provides the lane-graph representation used to extract topological map relationships in the scene encoder.","marker":"[36]"},{"why":"Defines the dynamic graph attention used throughout the strategy-aware and spatial-temporal-mode modules.","marker":"[38]"},{"why":"Contributes the object-query concept that the mode-specific decoders borrow to generate adaptive trajectory anchor queries.","marker":"[37]"},{"why":"Motivates the multi-iteration attention across spatial, temporal, and mode dimensions.","marker":"[35]"},{"why":"Supplies a strong knowledge-informed baseline on SinD that I2XTraj outperforms.","marker":"[29]"},{"why":"Provides a cooperative trajectory representation baseline on V2X-Seq against which I2XTraj's gains are measured.","marker":"[6]"}],"fun_headline_variants":["Roadside AI predicts all vehicles, beats vehicle-centric by 30%","Infrastructure beats onboard sensing: 30% better trajectory forecasts","Traffic-light awareness boosts multi-agent path prediction by 30%","I2X: joint prediction from roadside units, 30% over SOTA","Signalized intersections: infrastructure predicts all paths better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every vehicle's feasible maneuver at an intersection can be read off from a small set of radial-distance intervals defined by a single intersection diameter $D$, so distance alone separates stop, turn, and straight maneuvers; if $D$ is poorly defined for T-shaped or multi-branch intersections, the strategy-mode assignments and probabilities are mis-specified.","fun_headline_variants_meta":{"raw":{"variants":["Roadside AI predicts all vehicles, beats vehicle-centric by 30%","Infrastructure beats onboard sensing: 30% better trajectory forecasts","Traffic-light awareness boosts multi-agent path prediction by 30%","I2X: joint prediction from roadside units, 30% over SOTA","Signalized intersections: infrastructure predicts all paths better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00036,"raw_usage":{"total_tokens":1960,"prompt_tokens":975,"completion_tokens":985,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":895}},"tokens_in":591,"tokens_out":985,"duration_ms":9048,"temperature":1.0,"reasoning_tokens":895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:55:32.092482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train I2XTraj on a T-shaped or multi-branch signalized intersection where a single-diameter radial partition does not match the lane geometry, and check whether per-mode probability calibration and minJointFDE degrade compared with an intersection-agnostic mode clustering; a correct geometric prior should also keep the predicted mode distribution nearly unchanged when $D$ is varied by a small amount.","supporting_citations":[{"cited_title":"V2x-seq: A large-scale sequential dataset for vehicle- infrastructure cooperative perception and forecasting,","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world V2I dataset with 28 intersections and traffic-light data on which the main 30% improvement claim is demonstrated."},{"cited_title":"Sind: A drone dataset at signalized intersection in china,","cited_arxiv_id":null,"evidence_quote":"Supplies the drone-captured signalized-intersection dataset used to test generalizability and the 15% improvement claim."},{"cited_title":"Hpnet: Dynamic trajectory forecasting with historical prediction attention,","cited_arxiv_id":null,"evidence_quote":"Provides the strong unsignalized-intersection baseline that I2XTraj beats on V2X-Seq and SinD, and supplies the relative spatio-temporal edge encoding used in the graph."},{"cited_title":"Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic inter- action graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the factorized joint multi-agent prediction baseline and the joint regression loss formulation adapted here."},{"cited_title":"Learning lane graph representations for motion forecasting,","cited_arxiv_id":null,"evidence_quote":"Provides the lane-graph representation used to extract topological map relationships in the scene encoder."},{"cited_title":"How attentive are graph attention networks?","cited_arxiv_id":null,"evidence_quote":"Defines the dynamic graph attention used throughout the strategy-aware and spatial-temporal-mode modules."},{"cited_title":"End-to-end object detection with transformers,","cited_arxiv_id":null,"evidence_quote":"Contributes the object-query concept that the mode-specific decoders borrow to generate adaptive trajectory anchor queries."},{"cited_title":"Query-centric trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"Motivates the multi-iteration attention across spatial, temporal, and mode dimensions."},{"cited_title":"Ki- gan: Knowledge-informed generative adversarial networks for enhanced multi-vehicle trajectory forecasting at signalized intersections,","cited_arxiv_id":null,"evidence_quote":"Supplies a strong knowledge-informed baseline on SinD that I2XTraj outperforms."},{"cited_title":"Learning cooperative trajectory representations for motion forecasting,","cited_arxiv_id":null,"evidence_quote":"Provides a cooperative trajectory representation baseline on V2X-Seq against which I2XTraj's gains are measured."}],"review_version":1}