{"id":"8bc13f61-5213-4991-b20d-f17393204129","arxiv_id":"2608.07730","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FedSceneX uses a single calibrated price per client to jointly pick participation, precision, and placement, achieving the highest Macro-F1 within a 20-hour budget on nuScenes while keeping all four sensing modalities.","lead":"Federated learning typically counts progress in communication rounds, but a round can take 3 to 10 hours depending on which sensors participate. FedSceneX prices each client with one number to choose who uploads, at what precision, and where, reaching the best accuracy within a 20-hour budget on the nuScenes benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The VHP derivation replaces the makespan objective with total work and the deployed weights are calibrated constants, so the claimed one-number price is not established.","rationale":"The reader's weakest assumption about undisclosed calibrated constants is valid and important; I partially agree with it. My concern widens the same weak spot: even if the constants were reported, the derivation in Section IV-C optimizes a surrogate objective (total client work) rather than the active-time makespan that Eq. (2) and the evaluation use. The Graham list-scheduling bound cited there does not justify replacing T by sum_k L_k when selecting clients, because equal total work can yield different makespans depending on how the work balances across the two edge servers. The paper's own text admits kappa is calibrated and the weights appear as constants, directly contradicting the assertion that Eq. (4) is 'not a weighted sum chosen by hand.' Since both the theoretical derivation and the deployed constants are load-bearing for the central claim that one utility prices participation, precision, and placement, the current manuscript does not support that claim as written. The empirical comparison is suggestive and honestly reports the crossover, but the combination of a flawed derivation, missing constants, and single-run baselines prevents acceptance. I would not reject the underlying idea outright: a revised version that either proves a bounded approximation of makespan by total work or reframes VHP as an explicit heuristic, together with reported or swept constants and multi-seed baselines, could warrant conditional acceptance. For the paper as submitted, the central derivation is not established, so the verdict should move from conditional to reject.","tokens_in":13979,"tokens_out":10765,"duration_ms":110576,"concrete_test":"Run a replication on nuScenes in which VHP is implemented exactly as specified: compute the Dinkelbach root kappa of Eq. (3) and the dual multiplier lambda each round, and measure the round cost as the true makespan under least-loaded assignment, not as sum_k L_k. If the true makespan and exact kappa/lambda do not reproduce the reported 3.31 +/- 0.13 h/round and the >=0.6486 Macro-F1 at 20 h, the reported advantage comes from the calibrated surrogate, not from VHP. A smaller analytical sanity check: construct a two-client, two-server instance (two 1 h clients with V=2 versus one 2 h client with V=3) and check whether the VHP threshold selects the solution with lower true V/T; if it does, Proposition 1's makespan substitution is not a bounded approximation.","verdict_should_be":"REJECT","load_bearing_attack":"The claimed one-number price is not derived from the stated problem. In Eq. (2), T is the active time, i.e., the makespan over edge servers. Section IV-C replaces T by \"the total time the round buys,\" citing Graham's list-scheduling bound. That bound applies to a fixed client set; it does not make optimizing V / sum_k L_k equivalent to optimizing V / makespan. With two servers, two clients of 1 h on separate servers have makespan 1 h and total work 2 h, while one client of 2 h has the same total work but makespan 2 h; the surrogate and the real objective can disagree by a factor of 2 on the same total work. Per-client subtraction of eta L_k is therefore not the shadow price of critical-path time. Independently, Proposition 1 introduces alpha, gamma, mu as weights on P, G, F without deriving them from V, and the paper itself says kappa is \"calibrated once\" and the weights \"appear as constants\" (Section IV-C). Eq. (4) is thus a hand-weighted linear score, not a closed-form price obtained from the Lagrangian. The empirical 20-hour lead may be a property of the specific calibrated constants on nuScenes rather than of the pricing mechanism. This directly undermines the central claim that one utility replaces hand-tuned thresholds.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FedSceneX is an orchestration layer for same-scene multimodal federated edge learning. It selects clients, quantizes their updates, and places them on edge servers so as to maximize learning value per active hour, subject to an uplink budget and a modality-coverage floor. The claimed technical novelty, Value-per-Hour Pricing (VHP), converts the fractional program via Dinkelbach and dualization into a closed-form per-client utility whose weights are resource shadow prices. Experiments on the nuScenes benchmark with fifteen clients and twelve baselines report a per-round cost of 3.31 ± 0.13 active hours (baselines 4.85–9.78 h), highest Macro-F1 within a 20-hour budget, and a crossover beyond 45 hours; the paper also reports a modality-inference privacy leak from the server-visible scalars.","tokens_in":14255,"tokens_out":6280,"duration_ms":62181,"significance":"The paper addresses a real and under-appreciated issue (rounds are not equal work) and proposes a concrete budget-aware objective with coverage constraints. The empirical evaluation is broad (12 baselines, full nuScenes, five seeds for the proposed method) and unusually honest about the crossover region and the privacy cost of orchestration. If the central derivation were sound and the calibrated constants transparently reported, the paper would be a useful contribution to client selection and resource allocation in federated edge learning. As it stands, however, the derivation of the utility is not what is implemented, the makespan surrogate is not equivalent to the stated objective, and the headline comparison lacks baseline variance, so the core claims are not yet established.","major_comments":[{"comment":"The derivation presented for the utility U_k is not what the implementation uses. In Eq. (4) the positive terms αP+γG+µF are asserted as 'expected worth' without any derivation from V(Γ) in Eq. (2); the proposition therefore defines V rather than deriving a price from it. The paper then states that κ is 'calibrated once' and 'the weights appear as constants' (Section IV-C). Consequently the implementation is a hand-weighted linear score with five constants (plus κ, λ, ρ) whose values are never reported. This contradicts the abstract's claim of a closed-form price and leaves the central claim that one utility replaces hand-tuned thresholds unsupported. Please either solve/update the Dinkelbach and dual parameters per round and report them, or present Eq. (4) as a learned/calibrated score and provide the calibration procedure, values, and a sensitivity analysis.","section":"Section IV-C, Eq. (4), Proposition 1"},{"comment":"The reduction from T to total work via Graham's list-scheduling bound is invalid as stated. Graham's bound bounds makespan of a given assignment by a factor times the optimal makespan; it does not imply that the ratio V / sum_k L_k is within a constant of V / makespan. With two servers and two one-hour jobs on separate servers, sum L = 2 h and makespan = 1 h, while a single two-hour job also has sum L = 2 h but makespan = 2 h; the surrogate and the true objective disagree by a factor of 2 on equal total work. Therefore the term η L_k in Eq. (4) is a price on total work, not a shadow price of the round's critical path, and the claim that η is 'the worth of an hour' is not established. This is load-bearing because the entire VHP decomposition depends on separability of T.","section":"Section IV-B/C, Eq. (2)/(3)"},{"comment":"All twelve baselines are reported as single runs while FedSceneX has five seeds. At the 20-hour budget the weakest FedSceneX seed is 0.6486 vs. the strongest baseline 0.6458; the margin 0.0028 is almost 70% of FedSceneX's own between-seed standard deviation at 20 h (0.0040). Without baseline variance it is impossible to know whether the claimed lead across random seeds is significant. Similarly the crossover at 42.5–45 h and the Table V ablation comparisons are based on single-run baselines/variants. Please provide multiple seeds for the main baselines (or at least error bars on the budget curves) and statistically test the 20-hour claim.","section":"Section V-B, Table IV, Fig. 3(c)"}],"minor_comments":[{"comment":"The symbol b_t^k is used both for the precision/rate variable in the contract (Eq. (5)) and for the uplink rate allocation in Eq. (6); rename one of them to avoid confusion.","section":"Section IV-D, Eqs. (5)–(6)"},{"comment":"Table III lists κ, λ, ρ, and D^* but no values are reported anywhere; without these values (or the calibration protocol) the method is not reproducible.","section":"Section IV-C, Table III"},{"comment":"FedSceneX grants local epochs n_t^k in proportion to utility, while the baselines are said to use the same local budget; justify that this does not bias the per-round time comparison.","section":"Section V-A"},{"comment":"The modality-inference attackers are described only as 'logistic regression' and 'random forest'; specify the exact scalar features, normalization, and hyperparameters used so the measurement can be reproduced.","section":"Section V-E"},{"comment":"In Eq. (6), if all U_j^+ are zero the denominator is zero; state the fallback allocation for this edge case.","section":"Section IV-D, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical claims are potentially interesting, but the central theoretical contribution is currently not what is implemented. The authors should be asked to either run the announced Dinkelbach/dual procedure or reframe VHP as a calibrated score with reported constants and sensitivity analysis. Also, baseline variance is essential for the headline 20-hour claim. The heavy self-citation in the references (e.g., Refs. [4–6], [18–21], [25–37], [45–51]) is noticeable and should be trimmed during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper's empirical core is real and worth engaging with. The observation that round-based evaluation misleads when round durations vary by 3x is correct, and the authors build a system that selects clients, sets precision, and assigns servers based on one utility. They report a crossover: FedSceneX leads from 10 to 45 hours, then conventional methods overtake. That honesty is rare. The ablation shows each contract axis matters, and they report the privacy leak from telemetry with a 1.00 modality-inference accuracy — good to see that made explicit.\n\nNow the soft spots. The largest is that the VHP derivation does not match the implementation. Proposition 1 'derives' a client utility with weights alpha, gamma, mu, delta, eta, but alpha/gamma/mu are introduced without derivation from V, and the paper states that kappa is 'calibrated once' and the weights 'appear as constants.' No values are reported. So the claimed closed-form price is not what is actually deployed; what is deployed is a hand-weighted linear score. The stress-test point about the makespan surrogate also lands: replacing T (makespan) with total work via the list-scheduling bound is not valid for the optimization, since the bound applies to a fixed client set, not as an equivalence for the objective. The two-client example shows a factor-of-two disagreement. This directly weakens the 'one number replaces hand-tuned thresholds' claim.\n\nThe empirical comparison is also uneven: FedSceneX reports mean±std over 5 seeds, but the twelve baselines are single runs. That makes the 20-hour lead less certain. And the related-work description of FedMFS seems inaccurate; FedMFS already does selective modality communication, so the 'first' claim needs qualification.\n\nThese are addressable. Report the calibrated weights, run baselines with multiple seeds, and reframe VHP as a heuristic derived from a relaxation rather than the exact solution to (2). The empirical finding and the crossover are still valuable.\n\nWho gets value: anyone working on time-aware orchestration for multimodal federated edge learning. It deserves a serious referee; the theory needs major revision, but the problem identification and empirical protocol are worth the time.","headline":"A genuinely useful empirical study on time-based evaluation for multimodal FL, but the theoretical pricing story is not established; the deployed utility is hand-calibrated.","tokens_in":14769,"tokens_out":2265,"would_cite":false,"duration_ms":22959,"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":"One price per client decides who uploads, at what precision, and where, cutting federated round time to 3.31 hours against 4.85 to 9.78 for twelve baselines.","keywords":["federated learning","multimodal learning","edge orchestration","client selection","value-per-hour pricing","time-to-target","resource allocation","autonomous driving"],"falsifier":"A direct test is to apply the same calibrated utility weights to a different multimodal edge-learning benchmark without any retuning; if FedSceneX no longer completes rounds in roughly 3.3 active hours or no longer leads the same baseline set within a twenty-hour budget, the claimed time-to-target advantage is a fitted result rather than a property of the method.","tokens_in":13798,"feed_emoji":"⏱","tokens_out":9588,"duration_ms":80421,"temperature":0.7,"pith_summary":"FedSceneX claims that the right unit of progress for federated learning at the sensing edge is the active hour, not the communication round, because a round costs what its clients cost and on identical hardware the twelve baselines need 4.85 to 9.78 hours per round while FedSceneX needs 3.31. The paper argues that round composition should be the decision variable, and that one closed-form price per client can decide which clients upload, at what precision, and on which edge server. Counting by hours rather than rounds changes the ranking, and the paper reports that within a twenty-hour budget FedSceneX reaches the highest accuracy across all seeds while keeping all four sensing modalities in every round. The advantage is bounded: after about forty-five hours conventional methods overtake it, because the coarseness that makes rounds cheap also limits late-stage refinement.","feed_headline":"One per-client price cuts federated round time by up to 3x","feed_subtitle":"FedSceneX beats twelve baselines within a twenty-hour budget on a driving benchmark, then conventional methods take over.","key_machinery":"The central object is the per-client utility $U_t^k = \\alpha P_t^k + \\gamma G_t^k + \\mu F_t^k - \\delta X_t^k - \\eta L_t^k$, produced by applying a parametric transform to the fractional objective $\\max V/T$ and dualizing the uplink constraint, with $\\delta$ equal to the shadow price of a byte and $\\eta$ equal to the worth of an hour. This single scalar carries the whole orchestration: its sign is the participation threshold, its magnitude sets uplink rate and quantization level by reverse water-filling (giving larger updates more bits until every reconstruction reaches a common distortion floor), and it orders clients for least-loaded edge-server placement. The derivation also uses list scheduling to replace the makespan by a separable total with a bounded factor, so the coupled program collapses into one price per client.","core_discovery":"The central claim is that round composition, not compression, is the lever that determines how quickly a same-scene multimodal federation reaches a usable model, because the network cost of a round is only 1.99 seconds out of 3.52 hours while local training dominates. FedSceneX therefore maximizes learning value per active hour rather than per round. Value-per-Hour Pricing (VHP) transforms the fractional objective via a parametric transform of fractional programming and dualizes the uplink budget, producing a per-client scalar whose negative weights are the shadow price of a byte and the worth of an hour. That one scalar is read three ways: a positivity threshold selects clients subject to a modality-coverage floor, a reverse water-filling rule sets quantization to a common distortion floor, and a least-loaded rule places updates on edge servers with a bounded makespan loss. On the full fifteen-client benchmark with twelve baselines, a FedSceneX round costs 3.31 active hours, every seed leads within a twenty-hour budget, and the crossover happens between forty-five and fifty hours.","pith_inferences":["Because workload-pricing telemetry reveals modality with perfect accuracy, any orchestrator that prices clients by their workload inherits a privacy leak; this exposure should be budgeted as a cost independently of FedSceneX.","Updating the value-of-an-hour parameter by the parametric root-finding each round, instead of holding it constant, is a natural extension the paper does not test; it could either recover the late-stage accuracy or erase the speed advantage.","The reported crossover at forty-five hours is likely specific to the dataset and encoder choices; on tasks with stronger fine-grained refinement needs or higher quantization floors, the crossover should shift earlier.","Since network cost is about 0.016% of round time, communication-compression methods are roughly orthogonal to the bottleneck; the larger gains for same-scene multimodal edges lie in scheduling local compute, not in shrinking payloads."],"forward_implications":["Round-based convergence curves misreport progress by a factor of 1.5 to 3 on identical hardware, so evaluations should be drawn against cumulative active time or time-to-target instead of round index.","A deployment with a fixed-hour budget should price clients by one utility rather than tune separate selection, compression, and placement thresholds.","All four sensing modalities can be held in every round while uploading about 2.9 MB per round, less than baselines that drop a modality.","The speed advantage is bounded: for training budgets beyond about forty-five hours, conventional heterogeneity-aware optimizers catch up and pass FedSceneX."],"supporting_citations":[{"why":"Supplies the parametric transform that converts the value-per-hour ratio into a sequence of linear programs, the first step of VHP.","marker":"[52]"},{"why":"Supplies the list-scheduling bound that lets the round makespan be replaced by a separable total within a bounded factor.","marker":"[53]"},{"why":"Defines the complete multimodal driving benchmark and the fifteen-client sensor platform used in all evaluations.","marker":"[48]"},{"why":"The closest prior art, scoring clients by a utility that mixes statistical value with systems cost; the baseline FedSceneX extends to precision and placement.","marker":"[38]"},{"why":"The base federated averaging rule that FedSceneX leaves unchanged and a primary baseline.","marker":"[1]"},{"why":"A deadline-based client-selection baseline that also keeps all four modalities, providing the cost comparison.","marker":"[24]"},{"why":"A quantization and periodic-averaging baseline that anchors the communication-efficiency comparison.","marker":"[22]"},{"why":"The strongest late-stage baseline in the crossover comparison beyond forty-five hours.","marker":"[16]"}],"fun_headline_variants":["Round composition, not compression, is the real lever in federated learning","Per-client price cuts federated round time by up to 3x","One price per client selects, quantizes, and schedules federated edge updates","FedSceneX wins within 20 hours; conventional methods overtake later","Maximize learning per active hour, not per round"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported advantage rests on the assumption that the utility weights and the value-of-an-hour parameter, calibrated once rather than re-derived each round, are representative enough to preserve the speed and accuracy lead on other same-scene multimodal platforms and datasets.","fun_headline_variants_meta":{"raw":{"variants":["Round composition, not compression, is the real lever in federated learning","Per-client price cuts federated round time by up to 3x","One price per client selects, quantizes, and schedules federated edge updates","FedSceneX wins within 20 hours; conventional methods overtake later","Maximize learning per active hour, not per round"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000422,"raw_usage":{"total_tokens":2209,"prompt_tokens":1028,"completion_tokens":1181,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":1088}},"tokens_in":644,"tokens_out":1181,"duration_ms":10519,"temperature":1.0,"reasoning_tokens":1088,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:21:12.319452+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to apply the same calibrated utility weights to a different multimodal edge-learning benchmark without any retuning; if FedSceneX no longer completes rounds in roughly 3.3 active hours or no longer leads the same baseline set within a twenty-hour budget, the claimed time-to-target advantage is a fitted result rather than a property of the method.","supporting_citations":[{"cited_title":"On Nonlinear Fractional Programming,","cited_arxiv_id":null,"evidence_quote":"Supplies the parametric transform that converts the value-per-hour ratio into a sequence of linear programs, the first step of VHP."},{"cited_title":"Bounds for Certain Multiprocessing Anomalies,","cited_arxiv_id":null,"evidence_quote":"Supplies the list-scheduling bound that lets the round makespan be replaced by a separable total within a bounded factor."},{"cited_title":"nuScenes: A Multimodal Dataset for Autonomous Driving,","cited_arxiv_id":null,"evidence_quote":"Defines the complete multimodal driving benchmark and the fifteen-client sensor platform used in all evaluations."},{"cited_title":"Oort: Efficient Federated Learning via Guided Participant Selection,","cited_arxiv_id":null,"evidence_quote":"The closest prior art, scoring clients by a utility that mixes statistical value with systems cost; the baseline FedSceneX extends to precision and placement."},{"cited_title":"Communication-Efficient Learning of Deep Networks from Decentralized Data,","cited_arxiv_id":null,"evidence_quote":"The base federated averaging rule that FedSceneX leaves unchanged and a primary baseline."},{"cited_title":"Client Selection for Federated Learning with Heterogeneous Resources in Mobile Edge,","cited_arxiv_id":null,"evidence_quote":"A deadline-based client-selection baseline that also keeps all four modalities, providing the cost comparison."},{"cited_title":"FedPAQ: A Communication-Efficient Federated Learning Method with Periodic Averaging and Quantization,","cited_arxiv_id":null,"evidence_quote":"A quantization and periodic-averaging baseline that anchors the communication-efficiency comparison."},{"cited_title":"FedBABU: Toward Enhanced Represen- tation for Federated Image Classification,","cited_arxiv_id":null,"evidence_quote":"The strongest late-stage baseline in the crossover comparison beyond forty-five hours."}],"review_version":1}