{"id":"69b2b08b-089d-462a-bcd3-bb337e39cb8b","arxiv_id":"2607.14176","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An edge-aware graph neural network trained on MILP-optimal labels selects NR-V2X relay links with ~96% agreement and <5 ms inference, while pruned MILP keeps optimality below 30 ms for most instances.","lead":"Vehicles in dense cities often lose direct connection to roadside stations. This paper trains a graph neural network to pick relay routes as well as a slow exact optimizer does, but in milliseconds.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GINE outputs are never checked for feasibility; reported connectivity gains may be an artifact of a relaxed reachability metric.","rationale":"The reader's weakest_assumption identified the exact same load-bearing concern: GINE outputs are treated as feasible without any constraint enforcement, and the end-to-end metric counts only reachability, not feasibility. I agree this is the most critical issue because it directly threatens the practical validity of the GINE-only relay selection, which is the paper's main contribution. The link-level accuracy and latency numbers are measured independently and remain valid, but the connectivity gains—the key evidence of real-world usefulness—may be artifacts. The paper itself admits the limitation in the introduction and in future work, so the concern is not speculative. The proposed test would settle the issue by quantifying how much of the reported gain survives a feasible repair. A CONDITIONAL verdict remains appropriate: the paper should not be rejected outright, but the authors must either enforce feasibility (e.g., via post-processing) or clearly reposition the claims to acknowledge that the learned edge activations are only a heuristic input to a subsequent feasible-solution construction. The reader's verdict already reflects this conditionality, so I recommend no change.","tokens_in":8792,"tokens_out":4061,"duration_ms":38067,"concrete_test":"Sample a random subset of validation instances. For each, extract the GINE-activated edge set at τ=0.5 and check: (1) each CAV has at most one outgoing edge; (2) the induced subgraph over CAV–CAV edges is acyclic; (3) the flow-conservation constraints are satisfiable. If violations exist, repair the configuration by selecting, for each CAV with multiple outgoing links, the single best outgoing link (e.g., highest-capacity or the one that preserves connectivity), then recompute the fraction of CAVs reaching an RSU. Compare this repaired connectivity to the reported 9.2%/12% gains and to the 1-hop MILP baseline. If the repaired gains vanish or fall below the baseline, the headline claim is overstated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a GINE supervised by a MILP oracle enables real-time, optimal-grade relay selection. The load-bearing condition is that the GINE's edge activations constitute feasible relay configurations. However, the paper explicitly concedes in Section I that 'the proposed model does not explicitly enforce routing or flow-conservation constraints at inference time.' Section VI-C evaluates connectivity as mere reachability over activated edges, ignoring the MILP's single-outgoing-link (Eq. 4), flow-conservation (Eq. 2), and loop-free (Eq. 5) constraints. A CAV with two outgoing links or a cycle in the activated subgraph is counted as connected, even though no feasible single-path routing exists. Consequently, the reported 9.2% (four RSUs) and 12% (two RSUs) connectivity gains over the 1-hop MILP baseline may be inflated by infeasible edge sets that artificially extend reachability. The F1/accuracy metrics measure edge-level agreement with the oracle, not the quality of the resulting configuration, so a high F1 does not rescue the feasibility gap. Section VII's mention of 'constraint-aware post-processing' as future work further confirms that the current pipeline lacks feasibility guarantees.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a learning-to-optimize framework for multi-hop relay selection in NR-V2X vehicular networks. Each snapshot is modeled as a directed graph with node and edge features; an offline MILP oracle provides binary edge-activation labels, and a GINE is trained for edge-level relay activation. A hybrid GP-MILP variant uses GINE predictions to prune the MILP search space. Experiments on an OSM-SUMO-GEMV2 dataset report link-level accuracy 0.9589, F1 0.9544, connectivity gains up to 12% over a 1-hop MILP baseline, GINE inference below 5 ms for all instances, and GP-MILP below 30 ms for 98.14% of instances while preserving MILP objective values.","tokens_in":9091,"tokens_out":2805,"duration_ms":31767,"significance":"If the results hold, the work demonstrates a practical route to real-time relay selection with near-optimal quality, which would be valuable for low-latency NR-V2X. Strengths include the use of a large realistic simulation pipeline, the explicit adoption of an edge-aware GINE architecture, the hybrid pruning idea, and an honest disclosure of the model's limitations in the text. The main weakness is that the GINE output is not validated as a feasible relay configuration: the connectivity metric counts reachability through activated edges without enforcing the actual routing constraints of the problem. This gap is acknowledged in Section I and deferred to future work, but it is load-bearing for the central claim of 'relay selection' quality. The paper should therefore be revised to either enforce constraints at inference, report feasibility rates, or reframe the claims.","major_comments":[{"comment":"The end-to-end connectivity metric counts a vehicle as connected if it can reach an RSU over activated directed edges, but it does not require that the activated edge set satisfy the MILP's single-outgoing-link, flow-conservation, or loop-free constraints. A configuration with two outgoing links from one CAV or a cycle is treated as connected even though no valid single-path uplink route exists. The paper explicitly concedes in Section I that routing/flow constraints are not enforced at inference, and Section VII lists 'constraint-aware post-processing' as future work. No feasibility rate of GINE outputs is reported. Consequently, the reported 9.2% and 12% connectivity gains over the 1-hop MILP baseline may be inflated by infeasible edge sets. Please report the fraction of GINE outputs that satisfy Eqs. (2), (4), (5), and recompute connectivity on a feasible projection or after post-proc","section":"Section VI-C; Eq. (2), (4), (5)"},{"comment":"The MILP oracle is described as providing 'optimal' relay configurations, but the experimental setup imposes a per-instance time limit of 11 s. If any instances reach this limit, the ground-truth labels are not necessarily optimal, and the GINE's 'close agreement' is relative to potentially suboptimal labels. Similarly, the GP-MILP claim of 'preserving MILP-equivalent solutions (same objective value)' is only meaningful if the base MILP solves are proven optimal. Please report the percentage of instances solved to proven optimality within the time limit, the distribution of relative MILP gaps (e.g., objective bound gap) for time-limited instances, and whether GP-MILP changes those gaps.","section":"Section VI-A / Section IV"},{"comment":"All link-level metrics (accuracy, precision, recall, F1) are reported as point estimates on a single validation split, with no confidence intervals or multiple-seed training runs. The threshold sweep is performed on the same validation set, which can overestimate performance if thresholds are selected on the evaluation set. In addition, the connectivity gains in Section VI-C (9.2%, 12%) are reported without variance estimates or significance tests. Please add error bars (e.g., 95% CI over multiple seeds or bootstrapping), and ideally report results on a held-out test set that is not used for threshold selection.","section":"Table I / Section VI-B"}],"minor_comments":[{"comment":"Typo: 'iactive' should be 'inactive'. Throughout, the spacing around 'CA V' is inconsistent (e.g., 'CA Vs' vs. 'CAVs').","section":"Section VI-A"},{"comment":"Typo: 'stategy' should be 'strategy' in the sentence '...optimisation stategy effectively compatible...'.","section":"Section VI-D"},{"comment":"The table reports only percentages meeting latency thresholds; please also report mean/median and standard deviation of inference and solver runtimes, since the claim 'within 5 ms' is based on the mean over 10 runs per instance, which may mask tail latency. Figure 2 clips runtimes above 9 s; consider showing the full distribution or a log-scale inset.","section":"Table II and Figure 2"},{"comment":"The class weight α is computed from training statistics; please specify exactly when this is computed (prior to training) and report the value, since it directly affects the loss in Eq. (9).","section":"Section V-D"},{"comment":"Reference [23] is duplicated as [22] and [23] in the bibliography (both point to the GEMV2 model). Also, please check if the dataset and simulation code will be released, as the paper currently provides no link to code or data.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a machine-learning-for-systems venue, but the central claim requires a feasibility check of the GINE outputs. The authors are transparent about the limitation, which is commendable, but the current evaluation does not establish that the learned relay configurations are valid routing decisions. I recommend major revision rather than rejection because the issue is fixable: enforce constraints via post-processing (or at least report feasibility rates) and adjust the connectivity metric accordingly. The lack of error bars and the potential non-optimality of MILP labels are secondary but should be addressed in the same revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know three things about this one: it is a genuine learning-to-optimize application to multi-hop relay selection in NR-V2X, the paper is honest about its own main limitation, and the central evaluation metric has a hole that the authors acknowledge but do not fix.\n\nWhat is actually new: framing relay selection as edge-level GINE classification supervised by a MILP oracle, plus the GP-MILP pruning hybrid. The architecture itself is standard GINE + BCE loss, but the task framing is new in the cited literature, and the GP-MILP result — 98% of instances under 30 ms with the same objective value as full MILP — is a concrete, practical claim. The dataset is large (449,500 snapshots) and built from a realistic OSM-SUMO-GEMV2 pipeline. Link-level accuracy 0.9589 / F1 0.9544 is directly measured and, absent code, at least clearly specified.\n\nThe soft spots, in order of real softness. First and most important: the paper explicitly says the GINE does not enforce routing or flow-conservation constraints at inference. The end-to-end connectivity metric then counts a vehicle as connected if it can reach an RSU over any activated edge set, including sets with cycles or multiple outgoing links. That means the reported 9.2% and 12% connectivity gains over the 1-hop MILP baseline may be inflated by infeasible configurations. This is a load-bearing weakness, not a cosmetic one, and the stress-test note is right to flag it. The authors do disclose it up front and list constraint-aware post-processing as future work, so the paper is not hiding anything, but the headline gains should be treated as optimistic until feasibility is checked or enforced. Second: no error bars or confidence intervals on any metric, and no code or data release, which makes the 96% agreement hard to verify independently. Third: GP-MILP's \"MILP-equivalent\" claim is empirical on the evaluated dataset; there is no proof that pruning preserves optimality, and the threshold choice (0.15) is sensible but not analyzed. Fourth: the only baseline is the 1-hop MILP; no comparison to other learned or heuristic relay-selection methods.\n\nAll that said, the link-level agreement is a real measurement, the paper is clearly written, and the limitations are stated rather than buried. This is not circular: the GINE is trained on MILP labels and evaluated on agreement with those labels, which is standard L2O practice, not a logical circle. I would send this to a serious referee. It deserves review and probably a major revision that addresses the feasibility gap and adds uncertainty quantification. The core idea is solid enough to build on.\n\nFor your reading group, worth a maybe; for my own citation, not this year.","headline":"A solid, honest L2O engineering paper with a real task and a real feasibility gap: the learned edge activations are never checked against the MILP's routing constraints, so the connectivity-gain numbers are probably optimistic but the core link-level result is measured and worth reviewing.","tokens_in":9581,"tokens_out":1795,"would_cite":false,"duration_ms":23335,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a graph isomorphism network trained on exact mixed-integer solutions can select multi-hop V2X relay links with about 95% link-level F1 and all inference runs under 5 ms, bringing optimal-grade relay selection into rea","keywords":["NR-V2X","relay selection","graph neural networks","GINE","learning to optimize","mixed-integer linear programming","low-latency vehicular communication","edge classification"],"falsifier":"Take a validation snapshot where the MILP uses a multi-hop path and compare, edge by edge, whether the GINE's activated links form a directed forest with one outgoing link per vehicle. If a substantial fraction of high-confidence predictions contain cycles or multiple outgoing links yet still count as 'connected' under the paper's reachability metric, the reported F1 and connectivity gains overstate usable relay schedules. The same test could be rerun with a constraint-aware decoder to measure the gap.","tokens_in":8723,"feed_emoji":"📡","tokens_out":4352,"duration_ms":42181,"temperature":0.7,"pith_summary":"The paper tries to show that relay selection in NR-V2X vehicular networks, normally an NP-hard optimization solved by mixed-integer linear programming (MILP) with runtimes that scale poorly with graph size, can be delegated to a graph neural network trained on MILP-generated examples. The model, a Graph Isomorphism Network with Edge Features (GINE), classifies each directed link as active or inactive in a single forward pass. The paper reports close agreement with MILP decisions at the link level (F1 around 0.954) and consistent connectivity gains over a one-hop baseline (9-12%). It also shows that a hybrid GP-MILP scheme, where the network prunes the MILP search space, preserves the same objective value while keeping 98% of solver runs under 30 ms. A sympathetic reader should take away that learned surrogate optimization can plausibly fit within NR-V2X latency budgets, provided the model outputs are treated as feasible relay configurations.","feed_headline":"Learned relay selection matches exact optimizer in milliseconds","feed_subtitle":"A graph network trained on MILP solutions picks multi-hop uplinks with 95% link F1 and ~5 ms inference on urban V2X graphs.","key_machinery":"The central object is the GINE encoder with edge features: each directed link's radio capacity is projected through an MLP and injected into the message-passing updates, so the model sees not only topology but link quality. A classification head concatenates source and destination node embeddings with the edge feature and outputs an activation probability for each edge. Training uses MILP-optimal binary activation matrices as ground truth with class-weighted binary cross-entropy loss. For the hybrid GP-MILP, the same network's low-threshold predictions prune edges from the graph before MILP solving, shrinking the search space while retaining oracle-active links with high recall.","core_discovery":"The central claim is that an edge-aware GINE, supervised offline by a MILP oracle that maximizes the number of vehicles with a valid loop-free uplink path, learns to reproduce optimal relay activations at the edge level: accuracy 0.9589 and F1 0.9544 on validation. The learned model yields end-to-end connectivity gains of 9.2% with four RSUs and 12% with two RSUs over a 1-hop MILP baseline, with all tested instances completing inference in under 5 ms. The companion GP-MILP hybrid, which uses GINE predictions to prune the graph given to the MILP solver, attains the same objective values as the full MILP while completing 98.14% of instances in under 30 ms. As the paper states in Section I, the","pith_inferences":["If GINE outputs were passed through a lightweight feasibility repair that enforces single-outgoing-link and loop-free constraints, the reported accuracy numbers would translate more directly into deployable relay schedules; the paper's future-work note on constraint-aware post-processing points in this direction.","The threshold sweep suggests a family of operating points: a system designer could choose recall-oriented thresholds when preserving connectivity matters more than minimizing relay count, or precision-oriented thresholds to limit interference.","A natural extension is temporal: feeding consecutive GINE outputs could detect unstable relay flips, and the paper's future-work mention of temporal graph sequences signals that the authors intend this direction.","The pruning result hints that learned models can serve not only as replacements for optimization but as accelerators of exact solvers, a pattern transferable to other NP-hard network optimizations with expensive feasibility constraints."],"forward_implications":["If GINE accurately approximates MILP decisions, relay selection can shift from an online solver to a trained forward pass, removing exponential worst-case solve times from the decision loop.","The GP-MILP result implies that MILP-equivalent optimization can run within 30 ms for the vast majority of dense urban snapshots, meeting tighter latency budgets than unpruned MILP.","The 9-12% connectivity gains over a 1-hop baseline quantify the real-world value of coordinated multi-hop relaying in infrastructure-sparse urban deployments.","Because inference complexity scales linearly with the number of edges for a fixed network depth, the approach offers predictable latency as graph size grows."],"fun_headline_variants":["Graph net picks V2X relays like exact solver in under 5 ms","ML relay selection matches MILP optimality in milliseconds","GP-MILP: same optimal relays, solver under 30 ms","Learned relays improve V2X uplink connectivity by 12%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the neural model's unconstrained edge activations can be treated as feasible relay configurations, since the paper counts a vehicle as connected if it reaches an RSU over activated edges without checking the MILP's single-outgoing-link, flow-conservation, and loop-free constraints.","fun_headline_variants_meta":{"raw":{"variants":["Graph net picks V2X relays like exact solver in under 5 ms","ML relay selection matches MILP optimality in milliseconds","GP-MILP: same optimal relays, solver under 30 ms","Learned relays improve V2X uplink connectivity by 12%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001116,"raw_usage":{"total_tokens":4574,"prompt_tokens":926,"completion_tokens":3648,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":3572}},"tokens_in":670,"tokens_out":3648,"duration_ms":27431,"temperature":1.0,"reasoning_tokens":3572,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T03:41:14.085668+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a validation snapshot where the MILP uses a multi-hop path and compare, edge by edge, whether the GINE's activated links form a directed forest with one outgoing link per vehicle. If a substantial fraction of high-confidence predictions contain cycles or multiple outgoing links yet still count as 'connected' under the paper's reachability metric, the reported F1 and connectivity gains overstate usable relay schedules. The same test could be rerun with a constraint-aware decoder to measure the gap.","supporting_citations":[],"review_version":1}