{"id":"470af986-f6e5-40da-b6b1-2709caa0968d","arxiv_id":"2506.17929","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"ASTER couples a resource-aware spatio-temporal encoder with a multi-objective reinforcement learning agent to jointly predict incidents and dispatch emergency resources, claiming better outcomes than decoupled predict-then-optimize pipelines.","lead":"ASTER is a model that forecasts where urban incidents will happen and, in the same step, allocates a limited number of response resources to the most useful places. The paper tests it on traffic, crime, and emergency datasets from New York and Orange County, and reports better success rates and lower false alarms than nine baseline methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 contradicts the central 'consistently outperforms all baselines' claim: on XTraffic SR, STAEformer (0.34) beats ASTER (0.30), and ASTER is not best in several other metric–dataset cells, so the reported SOTA gains are not uniform and the 11.15% average is undefined.","rationale":"Reader's weakest assumption is about the simulator and reward coefficients; I do not rely on that. The most load-bearing issue is narrower and internal: the sentence in Section 4.2 making the universal SOTA claim is contradicted by Table 1 before any external validity question arises. Under the arrows printed in Table 1, ASTER is beatable on XTraffic SR (0.30 vs STAEformer 0.34, EARLIEST 0.31) and on at least six other cells. The paper's reported 11.15% average SR improvement is not reproducible from the table under the natural best-baseline-per-dataset definition (which gives about 14.5%, and includes a negative value for XTraffic). Since the paper's framing and abstract rest on 'state-of-the-art performance ... across six downstream metrics,' this overstatement is load-bearing. The confound identified in the reader's verdict—ASTER trains on the reward components that define the metrics, while baselines use a fixed rank-and-allocate rule—compounds the problem but is not needed to show the universal claim is unsupported. A scripted re-derivation from Table 1 is the minimal check; if the authors supply per-cell best counts and an explicit aggregation rule, the remaining claim can be re-evaluated. This does not change my overall rejection, so the verdict remains unchanged.","tokens_in":17177,"tokens_out":8432,"duration_ms":80685,"concrete_test":"Write a script that parses Table 1 and, for every dataset–metric cell, applies the stated arrow direction and counts whether ASTER is strictly best, tied, or worse; then recompute the claimed 11.15% average under explicitly stated definitions of 'second-best approach' (e.g., best baseline per dataset, best baseline per metric, averaged over datasets or over all 24 cells). If the per-cell count shows at least one baseline strictly better than ASTER, the 'consistently outperforms all' sentence must be revised; if no definition yields 11.15% for SR, the reported average must be re-derived or removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 makes the paper's strongest empirical claim: ASTER 'consistently outperforms all baseline methods across diverse datasets and evaluation metrics,' with an 'average 11.15% improvement in success rate over the second-best approaches.' The table that immediately follows does not support this. Under the arrow directions in Table 1, ASTER is not best in at least seven cells: XTraffic SR (STAEformer 0.34 vs ASTER 0.30; EARLIEST 0.31), NYC AD (STAEformer 120.65 vs ASTER 148.35), NYC AET (ESTGCN 6.76 vs ASTER 3.26), NYPD FAR (LSTM 0.02 vs ASTER 0.25), EMS FAR (ARIMA 0.04 vs ASTER 0.16), XTraffic AD (ESTGCN 26.31 vs ASTER 43.82), and XTraffic AET (STEMO 7.08 vs ASTER 3.35). Thus the universal wording of the headline result is internally contradicted by the paper's own reported numbers. Moreover, under the natural definition of 'second-best approaches' as the best non-ASTER SR per dataset, the per-dataset improvements are +28.6%, +20.7%, +12.0%, and −3.2%, which average +14.5%, not 11.15%; the reported figure requires an explicit aggregation convention that the paper does not provide. This matters because the entire contribution is positioned as state-of-the-art empirical performance; if the claim is only 'best on some metrics,' the revision needs to say exactly which cells and under what aggregation the 11.15% holds. The confound that ASTER trains on the same reward components used as metrics reinforces the concern: even the cells where ASTER wins may reflect direct reward optimization rather than a generic advantage of joint prediction and decision.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ASTER, a model for early spatio-temporal decision-making that jointly learns forecasting and resource allocation under dynamic resource constraints. ASTER combines a Resource-aware Spatio-Temporal interaction module (RaST) with a Preference-oriented decision agent (Poda) trained by multi-objective reinforcement learning. The authors evaluate on four real-world datasets (NYC, NYPD, EMS, XTraffic) with six downstream metrics and claim an average 11.15% improvement in success rate over the second-best baselines, as well as consistent superiority across all datasets and metrics. The manuscript also includes ablation studies, a sensitivity analysis, a case study, and appendices with dataset details, training protocol, and limitation discussion.","tokens_in":17583,"tokens_out":6265,"duration_ms":60468,"significance":"If the central claim were established, ASTER would be a useful contribution: it would demonstrate that jointly optimizing prediction and downstream allocation can outperform decoupled predict-then-allocate pipelines in a resource-constrained setting. The paper introduces a plausible modeling idea—resource-aware graph masking and preference-conditioned decision-making—and provides a code link, pseudo-code, and detailed appendices, which are strengths. However, the empirical evidence as presented does not support the headline claim. The evaluation is confounded because ASTER's training reward shares its components with the reported metrics, the claimed universal superiority is contradicted by the paper's own Table 1, and the reported 11.15% average cannot be reproduced from the table without an explicit aggregation convention. These issues are central to the paper's contribution and require substantial revision.","major_comments":[{"comment":"The claim that \"ASTER consistently outperforms all baseline methods across diverse datasets and evaluation metrics\" is contradicted by Table 1. For example, on XTraffic SR, STAEformer achieves 0.34 versus ASTER's 0.30; on NYC AD, STAEformer achieves 120.65 versus ASTER's 148.35; on NYC AET, ESTGCN achieves 6.76 versus ASTER's 3.26; on NYPD FAR, LSTM achieves 0.02 versus ASTER's 0.25; on EMS FAR, ARIMA achieves 0.04 versus ASTER's 0.16; on XTraffic AD, ESTGCN achieves 26.31 versus ASTER's 43.82; and on XTraffic AET, STEMO achieves 7.08 versus ASTER's 3.35. At least seven cells in the table show that ASTER is not the best. Furthermore, the reported \"average 11.15% improvement in success rate over the second-best approaches\" is not reproducible from the table. Using the natural definition of second-best as the best non-ASTER SR per dataset, the per-dataset improvements are roughly +28.6%, +20.7%, +12.0%, and (depending on whether the best or second-best baseline is used for XTraffic) either -11.8% or -3.2%, which average to about +12.4% or +14.5%, not 11.15%. The paper must state exactly which cells are considered wins, which aggregation is used, and why \"consistently outperforms\" is appropriate when multiple cells show otherwise.","section":"Section 4.2, Table 1"},{"comment":"The evaluation protocol is confounded: all baselines are coupled to a fixed rank-and-allocate policy based on forecasted intensities, while ASTER's Poda is trained to maximize a reward (Eq. 5) that is composed of the same quantities used in the reported metrics (accuracy reward, false alarm reward, distance reward, and temporal reward). This means ASTER is directly optimizing the measurement instrument, whereas the baselines are not. Without an ablation that separates the forecasting component from the allocation component—for example, applying Poda to baseline forecasts, or applying the fixed allocation policy to ASTER's predictions—the reported gains cannot be attributed to the joint forecasting–decision model rather than to the benefit of reward optimization itself. The paper should include such controlled comparisons or explicitly argue why the current comparison is informative despite this confound.","section":"Section 4.1 and Appendix C.2"},{"comment":"The multi-objective TD target in Eq. (6) is not well-defined: the vector y is set equal to r_t + gamma * max_{a,omega'} omega^T Q(s_{t+1}, a, omega'; theta), which adds a vector r_t to a scalar. The norm ||y - Q||_2^2 requires y and Q to have the same dimension. The correct vector target would be r_t + gamma * Q(s_{t+1}, a*, omega'; theta) where a* maximizes omega^T Q, or the formulation must be rewritten to make the dimensionality consistent. This is load-bearing because both losses LA and LB depend on y, and a dimension mismatch makes the training objective undefined as written.","section":"Section 3.3, Eq. (6)"},{"comment":"The hidden-preference inference procedure is nearly tautological for the tested setting: the agent infers omega by maximizing sum_t omega^T r_t, i.e., by selecting the objective that accumulated the largest total reward in a trajectory. When the true preferences are one-hot, any trajectory generated with that preference will tend to have the largest reward on that objective, so the inferred preference will trivially match. The appendix's claim that the agent \"successfully infers and selects preferences that closely match the hidden objectives\" therefore does not demonstrate genuine preference recovery or adaptation. To make this claim meaningful, the authors should test on non-one-hot preferences, hold out behavioral data for inference, or compare inferred preferences against a ground-truth preference in a setting where the argmax is not the only plausible estimator.","section":"Section 3.3 and Appendix C.4"}],"minor_comments":[{"comment":"The paper reports means for three trials but no standard deviations or significance tests; adding error bars or confidence intervals would help assess whether the differences are meaningful.","section":"Table 1"},{"comment":"In the Cost-Effectiveness Ratio definition, CER = (1/n) * sum_i r_i / c_i, but the cost c_i is never defined. Please specify how c_i is computed.","section":"Section 4.1, CER definition"},{"comment":"The text says \"All spatio-temporal models are implemented and trained under unified framework, from which we extract their optimal parameters\" but does not describe how optimal parameters are selected; please clarify the hyperparameter search procedure.","section":"Appendix C.2, baseline details"},{"comment":"There is a typo: \"Observe recoded values\" should be \"Observe recorded values\".","section":"Appendix B.2, Algorithm 1"},{"comment":"The table header uses \"GWNeT\" but the model is GraphWaveNet; please correct the capitalization.","section":"Table 2"},{"comment":"The resource mask M is defined as 1 if \"resource in both node i and node j are available,\" but it is unclear whether this is intended to be symmetric and how it interacts with the asymmetric affinity matrix A*; please clarify.","section":"Section 3.2, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an interesting idea, and the code release and detailed appendices are appreciated. However, the empirical evaluation as it stands does not support the stated SOTA claims: the universal superiority claim is internally contradicted by Table 1, and the reward-metric confound is a serious concern that would require new controlled experiments to address. The Eq. (6) dimensionality issue and the trivial hidden-preference test further weaken the technical contribution. I therefore recommend major revision rather than rejection, because the central research direction is plausible and the issues—while load-bearing—are potentially fixable with an honest re-framing of the claims, corrected equations, and additional evaluation protocols."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First: this paper is worth reading for the problem framing, not for its reported results. The core idea—jointly learning to predict and to allocate scarce resources in spatio-temporal settings, with a resource-aware graph mask and a preference-conditioned DQN—is a legitimate and underexplored direction. RaST's resource masking and the dual long-/short-term encoders with dynamic fusion are sensible architectural choices; Poda's scalarized multi-objective Q-learning is borrowed but competently applied.\n\nThe trouble is the evidence. Section 4.2 claims ASTER 'consistently outperforms all baselines,' but Table 1 contradicts that in at least seven dataset–metric cells: XTraffic SR (STAEformer 0.34 vs 0.30), NYC AD (120.65 vs 148.35), NYC AET (6.76 vs 3.26), NYPD FAR (0.02 vs 0.25), EMS FAR (0.04 vs 0.16), XTraffic AD (26.31 vs 43.82), and XTraffic AET (7.08 vs 3.35). The quoted '11.15% average improvement' doesn't match any standard aggregation: per-dataset SR gains over the best baseline average about +12.4%, and if XTraffic is compared to the second-best method (since ASTER is third), the average is +14.5%. Neither yields 11.15%. So the universal SOTA claim is unsupported by the paper's own numbers.\n\nThe evaluation is also confounded: ASTER's decision agent is trained on the same reward components used as metrics, while all baselines are coupled to a fixed rank-and-allocate policy. That makes the comparison between 'joint' and 'decoupled' unfair, and even the wins could reflect reward hacking rather than a general advantage. No error bars or significance tests are reported, and despite the abstract's claim about 'early prediction accuracy,' no raw forecasting metrics appear.\n\nThere are smaller technical problems worth noting: Eq. (6) mixes a scalar target ω^T Q into a vector regression, and the hidden-preference inference is essentially selecting the objective with the highest observed sum—the Appendix's 'recovery' of one-hot preferences is near-tautological. These are fixable but add noise.\n\nBottom line: the architecture may have genuine merit, but the paper as written overclaims and the evaluation needs a redesign. I'd recommend a reject-and-resubmit decision rather than a desk reject, and invite the authors to address the baseline fairness and metric aggregation issues. The problem is real; the current evidence isn't.","headline":"A well-motivated framework for coupling spatio-temporal prediction with resource allocation, but the SOTA claim is contradicted by the paper's own Table 1 and the evaluation is confounded.","tokens_in":18145,"tokens_out":5625,"would_cite":false,"duration_ms":48362,"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":"ASTER jointly optimizes early spatio-temporal prediction and constrained resource allocation, and reports an average 11.15 percent success-rate gain over the second-best decoupled baseline on four real-world datasets.","keywords":["spatio-temporal prediction","early decision making","dynamic resource allocation","multi-objective reinforcement learning","resource-aware graph learning","predict-then-optimize","emergency response","decision-focused learning"],"falsifier":"Re-run the comparison in a simulator where success requires the allocated resource to arrive within a fixed statutory response window rather than merely being assigned before the event, keeping the same data, baselines, resource levels, and evaluation protocol; the central claim fails if ASTER's reported 11.15 percent average success-rate advantage over the second-best approach disappears or reverses.","tokens_in":16931,"feed_emoji":"🚑","tokens_out":12262,"duration_ms":110738,"temperature":0.7,"pith_summary":"The paper argues that emergency-response and similar dynamic resource-allocation systems lose real-world utility when prediction is treated as a separate stage from allocation: forecasts are made first, then converted into actions, so errors and delays accumulate. To fix this, it proposes ASTER, a model that jointly learns to predict incidents and to allocate limited resources under dynamic availability, so the same representations drive both what will happen and what to do. The central claim, supported by experiments on four real-world datasets, is that this joint approach outperforms decoupled forecast-then-allocate pipelines, with an average 11.15 percent improvement in success rate over the second-best baseline and the best cost-effectiveness ratio on every dataset. A sympathetic reading is that coupling prediction with decision-making, rather than only improving forecast accuracy, is the route to actionable spatio-temporal intelligence.","feed_headline":"Joint prediction-and-dispatch model lifts success rate 11.15 percent","feed_subtitle":"ASTER couples forecasting and dispatch, outperforming decoupled baselines on four real-world datasets.","key_machinery":"Two coupled modules carry the argument. RaST's resource-aware graph learning layer multiplies an asymmetric learned affinity matrix $A^*$ by a binary resource mask $M$, zeroing out edges between busy nodes; its dynamic spatio-temporal convolution layers weight four parallel temporal filters by the resource ratio $\\gamma = S_t/S$, so scarce resources shorten the effective temporal horizon, and the state generator fuses long- and short-term encoder outputs with the same $\\gamma$ while predicting a node-specific horizon $k = \\gamma k_{\\max}$. Poda replaces downstream optimization post-processing with a multi-objective Q-network $Q_\\theta: \\mathcal{S}\\times\\mathcal{A}\\times\\Omega\\to\\mathbb{R}^d$; action selection uses the preference-scalarized value $\\omega^\\top Q(s,a,\\omega)$, and training anneals $\\lambda$ from 0 to 0.6 between vector regression and preference-scalarization losses. The same resource state therefore shapes representation, horizon choice, and reward, which is what couples prediction to feasible action.","core_discovery":"On the paper's own terms, the discovery is that early spatio-temporal prediction and downstream resource allocation should be optimized as one coupled system rather than cascaded. ASTER couples two components: RaST, a resource-aware encoder that masks the learned spatial adjacency matrix so message passing only occurs among nodes with idle resources, adapts temporal receptive fields to resource abundance, and fuses long- and short-term encodings by a resource ratio; and Poda, a preference-conditioned multi-objective Q-network $Q_\\theta(s,a,\\omega)$ that outputs per-objective values and selects actions by $\\omega^\\top Q$, with hidden preferences inferred from observed trajectories. Under a common evaluation where all baselines feed a fixed rank-and-allocate policy in a resource-constrained simulator, the paper reports an average 11.15 percent improvement in success rate over the second-best approach and the best cost-effectiveness ratio on all four datasets, which it takes as evidence that joint optimization with resource awareness is the effective route to actionable spatio-temporal intelligence.","pith_inferences":["The paper leaves implicit that the resource-adaptive horizon $k=\\gamma k_{\\max}$ can be read as an explicit early-decision policy; comparing it with optimal-stopping rules from early time-series classification would test whether the learned horizons match information-theoretic stopping boundaries.","Because the training protocol alternates encoder updates and agent updates, one can swap in a different spatio-temporal encoder and retrain Poda on top; if the reported gains transfer across encoders, the joint-coupling mechanism itself—not RaST's specific convolutions—is the active ingredient.","The same resource-mask-plus-preference-agent design applies to fleet repositioning, wildfire resource staging, or vaccine allocation; the paper's experiments cover only crime, accidents, EMS dispatch, and traffic incidents."],"forward_implications":["Joint training is the active ingredient the paper identifies: coupling the encoder and the decision agent in one objective outperforms feeding the same allocation policy with separately trained forecasters.","Resource abundance becomes a planning signal: ASTER extends its horizon when resources are plentiful and contracts to short-term patterns when resources are scarce, a behavior the case studies show preserves resources over time.","One preference-conditioned Q-network can cover many stakeholder trade-offs without retraining per preference, and hidden preferences can be recovered from observed behavior when no explicit priority is given.","The paper's six-metric evaluation (success rate, false alarm rate, average distance, average early time, resource utilization, cost-effectiveness) offers a common yardstick for early-decision systems that pure forecast error cannot provide."],"supporting_citations":[{"why":"STEMO is the strongest early-forecasting baseline; it supplies the multi-objective-reinforcement-learning comparison ASTER must beat.","marker":"[8]"},{"why":"STAEformer is a strong spatio-temporal forecasting baseline whose rank-and-allocate behavior ASTER contrasts in the case study.","marker":"[31]"},{"why":"GraphWaveNet is a spatio-temporal forecasting baseline used in the main comparison and the resource-level robustness table.","marker":"[15]"},{"why":"EARLIEST provides the adaptive-halting early-classification baseline that anchors the early-decision comparison.","marker":"[49]"},{"why":"NYC traffic accident records form one of the four benchmark datasets for the central performance claim.","marker":"[43]"},{"why":"NYPD crime complaint records form the second benchmark dataset and host the ablation study.","marker":"[44]"},{"why":"EMS dispatch records form the third benchmark dataset and host the resource-level robustness and hidden-preference experiments.","marker":"[45]"},{"why":"XTraffic is the fourth benchmark dataset, testing the method on traffic incident prediction and allocation.","marker":"[46]"},{"why":"The generalized multi-objective RL algorithm motivates treating the preference vector as a latent variable inferred from trajectories.","marker":"[38]"},{"why":"Decision-focused learning results justify the paper's resource-constrained evaluation protocol rather than unconstrained optimization.","marker":"[10]"}],"fun_headline_variants":["Coupling prediction and dispatch lifts success 11.15%","Adaptive spatio-temporal decisions boost allocation success 11.15%","Coupling forecasting and resource allocation boosts success 11.15%","Adaptive model for early resource decisions: 11.15% higher success","Resource-aware coupling beats decoupled pipelines by 11.15%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's own limitation note (Appendix D.1) concedes that real-world constraints are more complex and dynamically evolving than the simulator models, so the load-bearing premise is that the simulated success rule (a resource assigned before an incident counts as a success) and the reward weights in Equation (5) capture true stakeholder trade-offs, and that the fixed rank-and-allocate policy is a fair decoupled baseline; if either is off, the reported 11.15 percent gain measures the simulator's reward, not real-world utility.","fun_headline_variants_meta":{"raw":{"variants":["Coupling prediction and dispatch lifts success 11.15%","Adaptive spatio-temporal decisions boost allocation success 11.15%","Coupling forecasting and resource allocation boosts success 11.15%","Adaptive model for early resource decisions: 11.15% higher success","Resource-aware coupling beats decoupled pipelines by 11.15%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00078,"raw_usage":{"total_tokens":3464,"prompt_tokens":982,"completion_tokens":2482,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":2389}},"tokens_in":598,"tokens_out":2482,"duration_ms":19936,"temperature":1.0,"reasoning_tokens":2389,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:57:50.203423+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison in a simulator where success requires the allocated resource to arrive within a fixed statutory response window rather than merely being assigned before the event, keeping the same data, baselines, resource levels, and evaluation protocol; the central claim fails if ASTER's reported 11.15 percent average success-rate advantage over the second-best approach disappears or reverses.","supporting_citations":[{"cited_title":"Stemo: Early spatio-temporal forecasting with multi-objective reinforcement learning","cited_arxiv_id":null,"evidence_quote":"STEMO is the strongest early-forecasting baseline; it supplies the multi-objective-reinforcement-learning comparison ASTER must beat."},{"cited_title":"Spatio-temporal adaptive embedding makes vanilla transformer sota for traffic forecasting","cited_arxiv_id":null,"evidence_quote":"STAEformer is a strong spatio-temporal forecasting baseline whose rank-and-allocate behavior ASTER contrasts in the case study."},{"cited_title":"Adaptive-halting policy network for early classification","cited_arxiv_id":null,"evidence_quote":"EARLIEST provides the adaptive-halting early-classification baseline that anchors the early-decision comparison."},{"cited_title":"https://www.kaggle.com/datasets/mysarahmadbhat/nyc-traffic-accidents","cited_arxiv_id":null,"evidence_quote":"NYC traffic accident records form one of the four benchmark datasets for the central performance claim."},{"cited_title":"https://data.cityofnewyork.us/Public-Safety/NYPD-Complaint-Data-Historic/qgea-i56i","cited_arxiv_id":null,"evidence_quote":"NYPD crime complaint records form the second benchmark dataset and host the ablation study."},{"cited_title":"https://data.cityofnewyork.us/Public-Safety/EMS-Incident-Dispatch-Data/76xm-jjuj","cited_arxiv_id":null,"evidence_quote":"EMS dispatch records form the third benchmark dataset and host the resource-level robustness and hidden-preference experiments."},{"cited_title":"Decision-focused learning without decision-making: Learning locally optimized decision losses.Advances in Neural Information Processing Systems, 35:1320–1332, 2022","cited_arxiv_id":null,"evidence_quote":"Decision-focused learning results justify the paper's resource-constrained evaluation protocol rather than unconstrained optimization."}],"review_version":2}