{"id":"00f4c1e5-fa97-43de-afeb-20a8c6671a3e","arxiv_id":"2501.09934","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HEART schedules multiple federated learning tasks across vehicles and edge servers using hybrid synchronous-asynchronous aggregation and a two-stage PSO-GA plus greedy optimizer, reducing total training time in simulations.","lead":"A new scheduler, HEART, assigns several machine learning tasks to moving vehicles and edge servers in hierarchical federated learning, then orders each vehicle's tasks to cut total training time. If the simulations hold, vehicle fleets could run multiple models such as object detection and traffic prediction together with lower latency.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section V's claim that HEART finishes all tasks fastest is contradicted by Fig. 4(d): TSGD completes at 103.4 s, beating HEART's 110.2 s; the text uses 170.7 s, which is TSGD's Driver-Yawning time, not its overall completion.","rationale":"I read the paper in good faith: the central contribution is a two-stage scheduler (PSO-GA plus greedy ranking) that should reduce idle time and balance multi-task training in VEC-HFL. The claimed payoff is that HEART finishes all tasks faster in wall-clock time than four baselines. That payoff is the load-bearing empirical claim. It fails on the paper's own reported data: Fig. 4(d) shows TSGD completing its last task at 103.4 s, whereas HEART completes at 110.2 s. The Section V text substitutes 170.7 s for TSGD, a number that belongs to TSGD's Driver-Yawning curve in Fig. 4(c), and 215.8 s for TSSO, a number that belongs to TSSO's MNIST curve in Fig. 4(b). Thus the paper compares HEART's overall time against per-task times of baselines, producing a false impression of superiority. Even if all other idealizations (e.g., zero downlink delay) were accepted, this internal contradiction invalidates the strongest claim. Since the central claim is contradicted by the paper's own evidence, the verdict should move from conditional to reject. I do not raise this as a matter of consensus; it is a direct inconsistency within the manuscript.","tokens_in":25576,"tokens_out":5705,"duration_ms":51974,"concrete_test":"Re-read Fig. 4(d) and identify which legend entry is highlighted in red, indicating the last task to complete for each method. If TSGD's red entry is 103.4 s, compute the maximum completion time across the four subplots for each method and compare with the Section V text. Then re-run the simulation with the stated settings, recording per-task and overall completion times in a table; if TSGD's overall time remains below HEART's, the headline claim must be withdrawn.","verdict_should_be":"REJECT","load_bearing_attack":"The strongest claim (Section V) asserts HEART has the lowest wall-clock completion of all tasks: 110.2 s versus 191.9 s (TSPSO), 154.1 s (TSGA), 215.8 s (TSSO), and 170.7 s (TSGD). These comparison numbers are not all overall completion times. Fig. 4(d) reports the last-task completion for each method: TSPSO 191.9 s, TSGA 154.1 s, TSSO 130.1 s, TSGD 103.4 s. The 215.8 s for TSSO is its MNIST completion in Fig. 4(b), and the 170.7 s for TSGD is its Driver-Yawning completion in Fig. 4(c). Using the paper's own Fig. 4(d), TSGD finishes all four tasks at 103.4 s, which is 6.8 s before HEART. Therefore the central superiority claim is false as stated and cannot support the abstract's and conclusion's claims of reduced training time. This is an internal inconsistency, independent of idealizations such as the downlink-delay assumption.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers the problem of training multiple ML models simultaneously in a vehicle-edge-cloud hierarchical federated learning (VEC-HFL) system. It proposes a hybrid synchronous-asynchronous aggregation rule, and a two-stage method, HEART, that first assigns tasks to vehicles using a combined PSO-GA heuristic and then orders the assigned tasks on each vehicle with a greedy algorithm that maximizes an 'aggregate score' of task-overlap and upload-time. The paper claims, based on simulations on real datasets, that HEART reduces global training latency while balancing task execution compared to several two-stage baselines. The central empirical claim is that HEART finishes all tasks faster than the baselines in wall-clock time.","tokens_in":25863,"tokens_out":3201,"duration_ms":32090,"significance":"The problem of multi-model training in VEC-HFL is timely and the paper addresses a genuine gap in the literature. The proposed hybrid aggregation rule is a plausible design choice, and the two-stage decomposition is practical. If the reported performance advantages were correct, the paper would provide a useful heuristic with reasonable complexity. However, the paper's headline claim of wall-clock superiority is contradicted by its own experimental data, and several load-bearing theoretical/modeling points are not adequately supported. The contribution is best seen as a design study with a heuristic method; the current empirical basis does not substantiate the claimed advantage.","major_comments":[{"comment":"The claim that HEART achieves the fastest finish of all tasks is contradicted by the paper's own data. Fig. 4(d) reports the wall-clock completion time for the 20 Newsgroups task, which is the last task to finish for each method: TSGD completes at 103.4 s, while HEART completes at 110.2 s. The text compares HEART (110.2 s) against TSGD (170.7 s), but 170.7 s is the Driver-Yawning completion time from Fig. 4(c), not the overall completion time. Thus, the central superiority claim in the abstract and Section V is false as stated and cannot support the conclusion that HEART reduces training time.","section":"Section V, Fig. 4(d)"},{"comment":"The NP-hardness of problem P is asserted without proof. The statement 'P belongs to a sub-category of integer programming (IP) problems which are known to be NP-hard' is insufficient, since many integer programs are polynomial-time solvable. Because the NP-hardness claim is used to justify the heuristic approach, the authors should provide a formal reduction from a known NP-hard problem (e.g., Partition or a scheduling problem) or qualify the claim as a conjecture.","section":"Section IV-C"},{"comment":"The fitness function in Eq. (19) appears to have a sign inconsistency with the text. The penalty for imbalance is written as -(ξ3 Σ|ψ_j−χ| + Σ ρ_j |ψ_j−χ|), i.e., with a negative sign, so larger ρ_j makes the fitness more negative for imbalance. However, the text states that ρ_j 'increases the reward for assigning tasks with longer training durations,' which suggests ρ_j should positively weight tasks with longer training times. Either the signs in Eq. (19) are wrong, or the description of ρ_j is misleading; this needs clarification.","section":"Equation (19)"},{"comment":"The assumption that downlink communication delays (ES-to-vehicle and CS-to-ES) are negligible is load-bearing for the timing model. Equations (8)-(15) and the case (i)/(ii) classification depend on when a vehicle receives the next edge model; a nonzero downlink delay would add to the inactive time in case (ii) and could change the optimal task ordering. The authors should quantify typical downlink delays or provide a sensitivity analysis to show that the conclusions are robust to this assumption.","section":"Section III-A, footnote 1"}],"minor_comments":[{"comment":"In Section V, the phrase 'notable lower than' should be 'notably lower than'.","section":"Section V"},{"comment":"The caption of Fig. 3 labels the right subplot tasks as 'C1-C9' but the left subplot uses 'C1-C4'; this inconsistency should be resolved.","section":"Figure 3 caption"},{"comment":"The notation 'for j ∈{ x⟨j⟩m,n;[g] = 1} in turn do' is unclear; it should specify the set of assigned tasks for vehicle n under ES m.","section":"Algorithm 1, line 8"},{"comment":"The variable j' is used as the task preceding j in the vehicle's sequence, but the notation does not clearly define the relationship; a short explanation or a change of notation would improve readability.","section":"Equation (9)"}],"recommendation":"major_revision","confidential_remarks":"The central claim of the paper is directly contradicted by its own Fig. 4(d). If the authors cannot produce corrected experiments or a revised claim that is consistent with their data, the paper may not be salvageable. The other issues (NP-hardness proof, Eq. (19), downlink delay) also need to be addressed. The paper's self-citation [15] is contextual and not a concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the stress-test's central complaint doesn't survive contact with the paper. The Section V comparison uses the red-highlighted per-method last-task completion times, not the 20 Newsgroups completion for every baseline. TSGD's 170.7s in the comparison is its Driver-Yawning completion from Fig 4(c); TSSO's 215.8s is its MNIST time from Fig 4(b). Read as the paper defines it, HEART's 110.2s is lower than TSPSO's 191.9s, TSGA's 154.1s, TSSO's 215.8s, and TSGD's 170.7s. So the headline superiority claim is not internally contradicted.\n\nWhat is actually new: the hybrid synchronous-asynchronous aggregation rule — per-task edge-synchronous, cloud-asynchronous — is a sensible combination for multi-model VEC-HFL, and the two-stage PSO-GA plus greedy ranking is a reasonable heuristic. The balance constraint on task assignment is a practical addition, and the related-work table is useful. The experiments use four real datasets and compare against four baselines, more than many scheduling papers do.\n\nSoft spots, in order. First, the NP-hardness claim in Section IV-C is asserted without a reduction; either prove it or soften the claim. Second, Eq. (19) has a sign issue: the rho_j term as written rewards imbalance rather than penalizing it. Likely a typo, but it needs fixing because the fitness drives Stage 1. Third, the headline numbers come from a single simulation campaign: no code, no error bars, and several hyperparameters (xi_3 through xi_7, rho_j, pi_min/max, chi) are unreported. That limits confidence in the reported 17-37% gains. Fourth, the downlink-delay assumption is an idealization, but the paper acknowledges it and it is common in this literature.\n\nOverall, this is a solid systems paper with presentation flaws. It is not a theory contribution, but it is a reasonable engineering solution to a real problem. A serious editor should send it to peer review; reviewers should ask for a clarification of Eq. (19), a proof or softening of the NP-hard claim, and a reproducibility statement.","headline":"Stress-test's contradiction dissolves on a close read; the paper is a competent heuristic contribution with a few presentation flaws that deserve peer review.","tokens_in":26414,"tokens_out":5731,"would_cite":true,"duration_ms":49740,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that jointly optimizing task-to-vehicle assignment and per-vehicle training order under a hybrid synchronous-asynchronous aggregation rule makes multi-model federated training over moving vehicles finish faster than four…","keywords":["hierarchical federated learning","vehicle-edge-cloud","multi-model training","task scheduling","hybrid synchronous-asynchronous aggregation","particle swarm optimization","genetic algorithm","greedy task ranking"],"falsifier":"Run the same four-task, five-edge-server simulation with a nonzero downlink delay model, for example a finite broadcast data rate derived from a standard vehicular communication protocol, and check whether HEART's 110.2-second completion time still beats the 191.9-second TSPSO result; if a baseline improves more than HEART or the optimal task order shifts, the paper's central timing claim is undermined.","tokens_in":25371,"feed_emoji":"🚗","tokens_out":4865,"duration_ms":46361,"temperature":0.7,"pith_summary":"The paper addresses a setting in which vehicles simultaneously train several machine learning models in a hierarchical vehicle-edge-cloud federated learning system. It claims that the combination of a hybrid aggregation rule and a two-stage scheduler, HEART, minimizes the wall-clock time until every task reaches its target accuracy while keeping training effort balanced across tasks. The first stage assigns tasks to vehicles with a hybrid particle swarm and genetic algorithm; the second stage greedily orders each vehicle's assigned tasks by an aggregate score that rewards overlap with other vehicles and short upload times. If the claimed simulation results hold, balanced multi-task scheduling can cut total training time by roughly 20 to 40 percent compared with the tested baselines, without sacrificing individual-task convergence.","feed_headline":"Two-stage scheduler cuts multi-task vehicle FL time to 110 seconds","feed_subtitle":"A hybrid PSO-GA assignment plus greedy task ranking beats four baselines on wall-clock time.","key_machinery":"The load-bearing mechanism is the hybrid synchronous-asynchronous aggregation rule combined with the two-stage scheduler HEART. At each edge server, aggregation for a task waits for all assigned vehicles' local models; at the cloud, global aggregation proceeds once a subset of edge models arrives, which the paper argues limits model obsolescence while keeping data utilization high. Stage 1 uses an improved PSO with dynamically decreasing inertia weight and a GA crossover-mutation step, guided by a fitness function that penalizes imbalanced task assignment and rewards assigning longer-training tasks. Stage 2 defines an aggregate-score, a weighted sum of an overlap score counting how many vehicles train the same task in the same rank position and an upload score inversely related to vehicle-to-edge upload time, and greedily builds each vehicle's training sequence to maximize it.","core_discovery":"The paper's central claim is that the min-max global training latency problem in VEC-HFL with multiple simultaneous ML tasks is NP-hard, and that its proposed HEART method reaches a near-optimal solution with low complexity. HEART combines an improved PSO-GA for balanced task-to-vehicle assignment with a greedy algorithm that ranks each vehicle's tasks to maximize an aggregate-score of inter-vehicle training overlap and upload efficiency. Under a hybrid rule that aggregates synchronously at edge servers and asynchronously at the cloud, the paper reports that all four tasks finish in 110.2 seconds, compared with 191.9 seconds for TSPSO, 154.1 seconds for TSGA, 215.8 seconds for TSSO, and 170.7 seconds for TSGD in the tested configuration.","pith_inferences":["Editorial inference: if downlink broadcast latency is added to the model, the optimal task order may change because vehicles would spend time receiving the next edge model; the greedy aggregate-score could be extended to include a downlink-aware term.","Editorial inference: the aggregate-score idea is not specific to vehicles and could be transferred to other hierarchical federated systems with mobile or heterogeneous clients, such as drones or smartphones, where upload windows vary over time.","Editorial inference: a testable extension is to combine HEART's assignment with stale-model control or adaptive client selection, since the paper's hybrid rule does not explicitly bound staleness at the cloud aggregation.","Editorial inference: the authors' own numbers imply that comparing HEART against a downlink-aware scheduler would isolate how much of the reported 110.2-second gain comes from task ordering versus from the zero-downlink-delay assumption."],"forward_implications":["Balancing task assignments across vehicles can shorten the overall wall-clock time even when no single task converges fastest, because all tasks finish at similar times instead of one task being prioritized at the expense of others.","The low-complexity greedy ranking step can be executed in near real time, which matters for vehicles that move between edge servers and have limited dwell time.","The hybrid synchronous-asynchronous aggregation rule reduces non-task-training time compared with a fully synchronous rule, with reported reductions of about 15.6 percent for 25 vehicles and 10 percent for 50 vehicles.","HEART's reported reductions in average and longest non-task-training time relative to TSGA, TSPSO, and TSSO suggest that task-order optimization mainly saves time by shrinking idle and upload gaps at the edge.","In the tested settings, increasing the number of vehicles from 25 to 50 preserves or improves HEART's time savings over the baselines, indicating the method scales with more participating vehicles."],"supporting_citations":[{"why":"Supplies the precedent for ignoring downlink broadcast delay and the multi-task FL IoV setting that this paper extends.","marker":"[17]"},{"why":"Provides the hierarchical staleness-aware FL method and the 4-layer CNN used for the MNIST task.","marker":"[25]"},{"why":"Provides the HFEL cost model and parameter settings for hierarchical edge learning used in the simulations.","marker":"[28]"},{"why":"Supplies the vehicle-to-edge channel and data-rate model used in the timing formulas.","marker":"[29]"},{"why":"One of the GA-based scheduling methods that HEART builds on and compares against as TSGA.","marker":"[30]"},{"why":"One of the PSO methods that HEART builds on and compares against as TSPSO.","marker":"[31]"},{"why":"Provides a UAV two-tier hierarchical federated learning method used for comparison with synchronous aggregation.","marker":"[32]"},{"why":"Defines the VGG16 network used for the CIFAR-10 task.","marker":"[37]"},{"why":"Defines the ResNet-18 network used for the Driver Yawning task.","marker":"[38]"}],"fun_headline_variants":["HEART solves NP-hard multi-model FL with hybrid PSO-GA","Multi-task vehicle FL: HEART hits 110s via hybrid scheduler","Greedy ranking plus PSO-GA halve VEC-HFL training latency","Hybrid sync-async aggregation trims fleet FL to 110 seconds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The timing model assumes edge servers and the cloud can broadcast updated models to vehicles with zero delay; if those downlink transmissions take real time, every completion time is underestimated and the optimized training order may no longer be optimal.","fun_headline_variants_meta":{"raw":{"variants":["HEART solves NP-hard multi-model FL with hybrid PSO-GA","Multi-task vehicle FL: HEART hits 110s via hybrid scheduler","Greedy ranking plus PSO-GA halve VEC-HFL training latency","Hybrid sync-async aggregation trims fleet FL to 110 seconds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000284,"raw_usage":{"total_tokens":1709,"prompt_tokens":1011,"completion_tokens":698,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":616}},"tokens_in":627,"tokens_out":698,"duration_ms":7734,"temperature":1.0,"reasoning_tokens":616,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:31:09.201554+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same four-task, five-edge-server simulation with a nonzero downlink delay model, for example a finite broadcast data rate derived from a standard vehicular communication protocol, and check whether HEART's 110.2-second completion time still beats the 191.9-second TSPSO result; if a baseline improves more than HEART or the optimal task order shifts, the paper's central timing claim is undermined.","supporting_citations":[{"cited_title":"Matching Game for Multi-Task Federated Learning in Internet of Vehicles,","cited_arxiv_id":null,"evidence_quote":"Supplies the precedent for ignoring downlink broadcast delay and the multi-task FL IoV setting that this paper extends."},{"cited_title":"HiFlash: Communication-Efficient Hierarchical Federated Learning With Adaptive Staleness Control and Heterogeneity- Aware Client-Edge Association,","cited_arxiv_id":null,"evidence_quote":"Provides the hierarchical staleness-aware FL method and the 4-layer CNN used for the MNIST task."},{"cited_title":"HFEL: Joint edge association and resource allocation for cost-efficient hierar- chical federated edge learning,","cited_arxiv_id":null,"evidence_quote":"Provides the HFEL cost model and parameter settings for hierarchical edge learning used in the simulations."},{"cited_title":"Joint Optimization of Platoon Control and Resource Scheduling in Cooperative Vehicle-Infrastructure System,","cited_arxiv_id":null,"evidence_quote":"Supplies the vehicle-to-edge channel and data-rate model used in the timing formulas."},{"cited_title":"Joint Optimiza- tion of Completion Ratio and Latency of Offloaded Tasks With Multiple Priority Levels in 5G Edge,","cited_arxiv_id":null,"evidence_quote":"One of the GA-based scheduling methods that HEART builds on and compares against as TSGA."},{"cited_title":"Dynamic Stochastic Reorientation Particle Swarm Optimization for Adaptive Latent Factor Analysis in High-Dimensional Sparse Matrices,","cited_arxiv_id":null,"evidence_quote":"One of the PSO methods that HEART builds on and compares against as TSPSO."},{"cited_title":"UA V Swarm-Assisted Two-Tier Hierarchical Federated Learn- ing,","cited_arxiv_id":null,"evidence_quote":"Provides a UAV two-tier hierarchical federated learning method used for comparison with synchronous aggregation."}],"review_version":1}