{"id":"0f8ff26d-89a0-46a4-8974-e8435aacef75","arxiv_id":"2506.09769","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Load-aware Tram-FL greedily schedules the training node and per-label data amount each round to maximize samples trained per second, and simulations report faster convergence than three baselines.","lead":"This paper extends a model-circulation federated learning method with a scheduler that picks which node trains and how much data it uses, based on each node's current computing and network load. The reported simulations show faster training to a target accuracy than simpler scheduling rules on MNIST and CIFAR-10.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I contradicts the claim of 'consistently lowest' training time: Random beats Load-aware Tram-FL in the MNIST 3-node lower-accuracy setting (12.46 s vs 15.69 s).","rationale":"The reader's picked weakest assumption (perfect load prediction) is a real limitation, but it is explicitly scoped and affects only practical applicability. The more load-bearing concern is internal: the paper's own Table I gives a case where the proposed method is slower than Random, directly contradicting the strongest claim. Other elements of the derivation appear sound; the O3 endpoint argument is valid because the linear-fractional objective has constant derivative sign. The lack of error bars and the absence of code make the numerical support unverifiable, but the concrete contradiction is enough to require revision. Because the reader already returned CONDITIONAL with this same contradiction noted in the rationale, the verdict stays unchanged; the condition should explicitly require correcting the consistency claim and reporting full results with error bars.","tokens_in":9733,"tokens_out":4527,"duration_ms":51475,"concrete_test":"Re-run the MNIST 3-node, lower-accuracy condition exactly as described in Section III-A with at least 10 seeds, recording T_total to reach the 70% target for Load-aware Tram-FL and Random, and report mean and standard deviation. If Random's mean time is below Load-aware's, the 'consistently lowest' claim must be withdrawn or explicitly qualified. If the original numbers are not reproducible, the experimental support is unverifiable without code and seed statistics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B asserts that Load-aware Tram-FL 'consistently achieves the lowest training time in both MNIST and CIFAR-10 tasks.' The paper's own Table I contains a direct counterexample: in the MNIST 3-node scenario with the lower target accuracy, Random reports 12.46 s while Load-aware Tram-FL reports 15.69 s. Since the central claim is supported entirely by Table I, this single cell falsifies the claim as stated. The later percentage reductions are computed only for scenarios where baselines converged, so they do not establish consistency either. The perfect-load-prediction assumption in Section II-A is a genuine limitation that bounds the practical gains, but the more basic problem is internal: under the paper's own experimental assumptions, the headline claim is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Load-aware Tram-FL, a scheduling extension of the Tram-FL decentralized federated learning protocol. In each round, the node holding the model collects predicted computational and communication load information from all nodes, then greedily selects a training node and per-label data usage proportions by solving a sequence of optimization problems P1–P4. The objective is to maximize the ratio of total training samples to total training time, subject to a constraint that keeps per-label cumulative sample counts balanced (variance below a threshold V). The authors report simulations on MNIST and CIFAR-10 with 3, 5, 10, and uneven 5-node configurations, comparing against Random, Time-first, and Variance-first baselines, and claim that the proposed method consistently achieves the lowest training time.","tokens_in":9733,"tokens_out":5619,"duration_ms":64556,"significance":"If the approach works as claimed, it addresses a real practical problem in model-circulation decentralized FL: avoiding nodes and communication links that are temporarily heavily loaded. The problem formulation is clear, the decomposition into node-wise subproblems is a reasonable relaxation of a difficult global scheduling problem, and the method is described with enough detail to be reimplemented. The paper also deserves credit for disclosing the ideal-prediction assumption and for making the simulation metric (accumulated time from Eqs. (2)–(5)) explicit. However, the experimental evidence, as presented in the paper's own Table I, does not support the headline claim of consistent superiority, which limits the significance of the reported gains.","major_comments":[{"comment":"The statement that Load-aware Tram-FL 'consistently achieves the lowest training time in both MNIST and CIFAR-10 tasks' is directly contradicted by Table I. In the MNIST 3-node scenario at the 70% accuracy target, Random achieves 12.46 s while Load-aware Tram-FL requires 15.69 s. This is not an omitted baseline or a missing cell; it is a reported result in the paper's own main comparison table. The central claim must be revised to acknowledge that the proposed method is not uniformly best, or the authors must explain why this cell should be excluded (e.g., due to run-to-run variance, in which case repetitions and error bars are needed).","section":"Section III-B / Table I"},{"comment":"The percentage reductions highlighted in Section III-B, especially for the 5-node (uneven) scenarios, are computed only from the subset of baseline runs that actually reached the target accuracy. Many entries in Table I are dashes (e.g., Random in the MNIST 5-node and 10-node lower-accuracy rows; Time-first in several higher-accuracy CIFAR-10 rows), meaning the baseline did not converge within the simulation budget. The claim that Load-aware Tram-FL 'consistently' reduces training time is therefore not established by the table. The authors should report how non-converged runs are handled, provide the number of independent repetitions, and avoid percentage comparisons that implicitly ignore failed runs.","section":"Section III-B / Table I"},{"comment":"The entire scheduling procedure depends on the assumption, stated in Section II-A, that each node can accurately predict its own computation and communication load for the current round. In the simulation, the true sampled resource availability is fed directly into the scheduler, so the reported times represent an upper bound on achievable gains under realistic prediction error. The authors acknowledge this as future work, but the abstract and conclusion currently present the results without this scope restriction. The limitation should be stated in the abstract or conclusion so that readers do not overgeneralize the reported reductions.","section":"Section II-A / Algorithm 1"},{"comment":"The monotonicity argument for O3 does not fully account for the discontinuity of T_idle at S=1. For S<1, T_idle=H; for S>=1, T_idle=0. The derivation of the monotonicity condition treats T_idle as constant, so the statement that 'O3 is either monotonically increasing or decreasing' is only valid on [1, S★] (and separately on [0,1] with the idle term H). Comparing only O3(0) and O3(S★) could, in the decreasing case, skip a small-data update (S=1 or slightly larger) that is preferable to skipping the round when H is large. The authors should clarify why the S in [0,1] range can be ignored, or modify the comparison to include the S=1 endpoint.","section":"Section II-C / Eqs. (8)-(9)"}],"minor_comments":[{"comment":"The caption of Fig. 2 says '(CIFAR10)', but the same label-distribution setup is used for both MNIST and CIFAR-10 experiments. Please clarify whether the figure applies to both datasets or only to CIFAR-10.","section":"Section III-A / Fig. 2"},{"comment":"The '+1' in the denominator of O1 is introduced to avoid division by zero, but the notation is easy to misread as adding one second. A brief parenthetical explanation would improve clarity.","section":"Section II-C / Eq. (1a)"},{"comment":"Table I would be much easier to read if the lower- and higher-accuracy columns were visually separated (e.g., by a vertical rule or explicit subheadings), since the plain-text rendering conflates values like '15.6927.05' into a single number.","section":"Section III-B / Table I"},{"comment":"The variance constraint in P4 is a quadratic inequality in the variables x_{i',c}, and the text refers to a solver (COBYQA) without explaining how the problem is encoded or whether the feasible set is convex. A short remark on the solvability of P4 would strengthen the presentation.","section":"Section II-C / P4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a machine learning / distributed systems venue, and the core scheduling formulation is reasonable. The decisive issue is that the paper's own Table I contradicts the central 'consistently lowest training time' claim. This is a fixable problem if the authors reword the claim, report repetitions and variance, and clearly separate scenarios where the proposed method wins from those where it does not. The perfect-prediction assumption should also be moved from a buried sentence into the abstract or conclusion. I do not see a need for rejection, but the current version overstates the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The core scheduling formulation is a real, if small, extension of Tram-FL: P1–P4, the node-wise decomposition, and the monotonicity argument that reduces the data-amount choice to comparing O3(0) and O3(S*) are new relative to the fixed cyclic order in the original paper. The algebra in Eqs. (8)–(9) checks out, and the variance constraint is a sensible way to keep label usage balanced. I believe the greedy policy does what the paper says under perfect load information.\n\nThe problem is the evaluation, not the math. The abstract and Section III-B say Load-aware Tram-FL 'consistently achieves the lowest training time,' but Table I has a direct counterexample: MNIST 3 nodes, 70% target, Random 12.46 s vs Load-aware 15.69 s. That is the paper's own number. The later percentage savings are computed only for settings where baselines converged, so they don't support 'consistently' either. There are no seeds or error bars, so I can't tell if the difference is noise. The baselines are also handicapped: Random and Time-first fix the data usage ratio at a small constant, so they can't trade data volume for speed the way the proposed method does. Variance-first is the fair comparison, and in several CIFAR-10 cells it is close or better at the higher accuracy target.\n\nThe perfect-load-prediction assumption (Section II-A) is stated honestly and deferred to future work, but it means the reported gains are an upper bound. If predictions are wrong, the scheduler can pick a loaded node or enforce the variance constraint against stale counts. Also, the idle-time constant H and a few other experimental parameters are not reported, which makes reproduction harder than it should be.\n\nNone of this sinks the core idea. The formulation is coherent, the derivation is sound, and the limitation is acknowledged. It's an incremental contribution to a narrow subfield, not a breakthrough. With a corrected consistency claim, error bars, a flexible-data-usage baseline, and the missing constants reported, it could be a modest but useful paper.\n\nI'd send it to peer review, expecting major revision. It's not a desk reject. The reader's conditional verdict is about right; the stress-test point about Table I is exactly the thing that needs fixing.","headline":"Useful formulation, overclaimed evaluation: the paper's own Table I contradicts its 'consistently lowest' claim.","tokens_in":10423,"tokens_out":2785,"would_cite":true,"duration_ms":27060,"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":"This paper claims that a greedy, load-aware scheduler for the single-model circulation protocol Tram-FL reaches target accuracy in less simulated wall-clock time than random, time-first, and variance-first baselines under heterogeneous…","keywords":["Decentralized federated learning","Model circulation","Training scheduling","Load awareness","Client selection","Non-IID data","Latency minimization","Greedy optimization"],"falsifier":"Run the same MNIST and CIFAR-10 benchmark but generate each round's available resources after the scheduler commits, so that the scheduler's predictions are noisy (e.g., predicted availability from a past distribution while actual availability is resampled). If Load-aware Tram-FL no longer consistently beats Time-first in total time to target accuracy, the claimed advantage depends on the perfect-prediction assumption rather than on the scheduling rule itself.","tokens_in":9351,"feed_emoji":"⏱️","tokens_out":4977,"duration_ms":46887,"temperature":0.7,"pith_summary":"Decentralized federated learning typically has each node maintain its own model, which costs communication and suffers under non-IID data. Tram-FL instead circulates a single global model among nodes, but it ignores differences in computation and bandwidth. This paper proposes Load-aware Tram-FL, which each round greedily decides which node should train and how many samples per label it should use, by solving a decomposed optimization problem that maximizes the ratio of samples trained to time elapsed subject to a label-variance constraint. The authors report that in simulations on MNIST and CIFAR-10, this scheduling consistently reaches the target accuracy in less time than three baseline schedulers, with the largest gains under uneven data distributions. The practical interest is that a purely local, per-round scheduling rule could make decentralized FL usable on real heterogeneous clusters.","feed_headline":"Load-aware scheduling cuts decentralized FL training time up to 88%","feed_subtitle":"Greedy per-round node and data selection beats three baselines in MNIST and CIFAR-10 simulations.","key_machinery":"The load-bearing object is the round-level objective O3 (Eq. 8), the ratio of cumulative training samples to cumulative time. After rewriting, O3 = (r/M)·(S_round + S_prev)/(S_round + (r/M)(T_comm+T_idle+T_prev+1)); its derivative has a constant sign, so for each candidate node the maximum is attained either at S_round = 0 or at the largest feasible S_round, S*. The largest feasible sample count S* comes from solving P4, a quadratic-constrained problem that can be handled by the COBYQA solver, and the node with the highest O3 at its endpoint is chosen. The variance constraint (Eq. 1b/10b) is what forces balanced per-label usage across rounds.","core_discovery":"The central claim is that the training-scheduling problem for Tram-FL—which node trains next and how much data to use—can be formulated as a global optimization (P1), decomposed into per-node subproblems (P3, P4), and solved greedily each round because the per-round objective O3 is monotone in the amount of data trained, so only the endpoints O3(0) and O3(S*) need be compared. With this rule, the selected node is the one whose (predicted) computation-plus-communication time per trained sample is best, while the variance constraint keeps per-label usage balanced. The paper demonstrates that this 'Load-aware Tram-FL' achieves the lowest total training time to target accuracy in all tested MNIST and CIFAR-10 scenarios, with particularly large reductions in the 5-node uneven case.","pith_inferences":["If load predictions are imperfect, the greedy endpoint rule could be extended to a stochastic or bandit-style scheduler that treats predicted load as a prior and updates it from observed round durations; the paper leaves this open.","The monotone-ratio trick is generic: any system that alternates work and communication and wants to maximize throughput per unit time could use the same endpoint comparison, e.g., edge inference pipelines with heterogeneous devices.","The variance constraint acts as a fairness regularizer; tuning V beyond the fixed value may trade off convergence speed against per-class coverage in ways the paper does not explore.","The simulation models bandwidth as stable within a round; if the network is bursty, the constant-bandwidth assumption could be tested by time-varying link capacities."],"forward_implications":["With accurate load predictions, per-round greedy scheduling yields lower training time than fixed cyclic order, random selection, or time-only selection in the simulated scenarios.","Because the variance constraint is enforced while scheduling, the data per label used cumulatively becomes more balanced in non-IID settings, which is presented as the reason for faster convergence.","The method requires only one node to train at a time and skips rounds when no node is suitable, making it compatible with the single-model circulation protocol.","The comparison to Time-first shows that minimizing time alone causes repeated selection of low-cost nodes and overfitting; the variance constraint prevents this collapse.","The gains are largest in uneven node-data scenarios, where the scheduler can preferentially use nodes with richer label coverage when their load is low."],"supporting_citations":[{"why":"Defines the Tram-FL protocol—the single global model circulated among nodes—that Load-aware Tram-FL extends with load-aware scheduling.","marker":"[5]"},{"why":"COBYQA is the optimization solver used to solve P4 for the maximum feasible data amount under the variance constraint.","marker":"[10]"},{"why":"Cited to justify the ideal-scenario assumption that each node can accurately predict its own current load.","marker":"[8]"},{"why":"Also cited to support the load-prediction assumption underlying the scheduling decisions.","marker":"[9]"},{"why":"MNIST is one of the two benchmark datasets used to evaluate the method against baselines.","marker":"[11]"},{"why":"CIFAR-10 is the second benchmark dataset used in the experimental evaluation.","marker":"[12]"}],"fun_headline_variants":["Load-aware scheduling cuts decentralized FL training time 88%","New scheduling shaves 88% off decentralized federated learning time","Load-aware scheduling accelerates decentralized FL training by 88%","Greedy load-aware scheduling cuts decentralized FL training time 88%","Load-aware scheduling speeds up decentralized FL training 88%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheduling decision rests on the assumption that each node can perfectly predict its own computation and communication load for the upcoming round; the simulator feeds exactly those predicted values into the scheduler.","fun_headline_variants_meta":{"raw":{"variants":["Load-aware scheduling cuts decentralized FL training time 88%","New scheduling shaves 88% off decentralized federated learning time","Load-aware scheduling accelerates decentralized FL training by 88%","Greedy load-aware scheduling cuts decentralized FL training time 88%","Load-aware scheduling speeds up decentralized FL training 88%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000682,"raw_usage":{"total_tokens":3038,"prompt_tokens":829,"completion_tokens":2209,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":2123}},"tokens_in":445,"tokens_out":2209,"duration_ms":16436,"temperature":1.0,"reasoning_tokens":2123,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:42:35.295270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same MNIST and CIFAR-10 benchmark but generate each round's available resources after the scheduler commits, so that the scheduler's predictions are noisy (e.g., predicted availability from a past distribution while actual availability is resampled). If Load-aware Tram-FL no longer consistently beats Time-first in total time to target accuracy, the claimed advantage depends on the perfect-prediction assumption rather than on the scheduling rule itself.","supporting_citations":[{"cited_title":"Tram-fl: Routing-based model training for decentralized federated learning,","cited_arxiv_id":null,"evidence_quote":"Defines the Tram-FL protocol—the single global model circulated among nodes—that Load-aware Tram-FL extends with load-aware scheduling."},{"cited_title":"Model-based derivative-free optimization methods and software,","cited_arxiv_id":null,"evidence_quote":"COBYQA is the optimization solver used to solve P4 for the maximum feasible data amount under the variance constraint."},{"cited_title":"Forecasting short-term data center network traffic load with convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Cited to justify the ideal-scenario assumption that each node can accurately predict its own current load."},{"cited_title":"Runtime power al- location based on multi-gpu utilization in gamess,","cited_arxiv_id":null,"evidence_quote":"Also cited to support the load-prediction assumption underlying the scheduling decisions."},{"cited_title":"Gradient-based learning applied to document recognition,","cited_arxiv_id":null,"evidence_quote":"MNIST is one of the two benchmark datasets used to evaluate the method against baselines."}],"review_version":1}