{"id":"cad73c97-e8c0-44c0-828a-422d797b5882","arxiv_id":"2506.08426","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HASFL jointly optimizes per-device batch sizes and neural network split points to reduce training latency in heterogeneous split federated learning, guided by a new convergence bound.","lead":"HASFL is a framework that lets each edge device in a split federated learning system use its own batch size and model split point, chosen to balance per-round latency against convergence. The paper derives a convergence bound and uses it to pick these settings, reporting faster simulated training on CIFAR benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central optimization minimizes Corollary 1's lower bound on R as if it were the exact round count, and the paper provides no tightness check; the claimed latency minimum may hold only for the surrogate objective.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: Corollary 1's bound is converted into an exact round-count predictor without any tightness or validation. This is the critical link from the theoretical convergence bound to the optimization problem and to the empirical latency claims. I agree that the paper is otherwise largely coherent: the convergence proof is a standard smoothness/variance argument, the latency model is explicit, and the ablation studies support the qualitative effects of batch size and split point. However, the quantitative claim that HASFL minimizes training latency to a target accuracy depends on the unverified equality in Section VI. The proposed concrete test is feasible because it only requires running the same training procedure with fixed configurations and comparing predicted versus observed rounds. If the test supports rank preservation, the concern is resolved; if not, the optimization objective needs to be justified through an empirical tightness calibration or a different surrogate. Since the reader already conditioned acceptance on addressing this issue, my independent stress-test does not change the verdict, hence UNCHANGED.","tokens_in":26033,"tokens_out":7071,"duration_ms":93552,"concrete_test":"Run Algorithm 1 on CIFAR-10 with VGG-16 for a small set of fixed configurations, e.g., (b, Lc) in {(8,8), (16,8), (16,3), (32,3)}. For each configuration, measure the observed number of rounds R_obs needed to reach the convergence criterion in Eq. (26), using an empirical proxy for the running average squared gradient norm (e.g., full-batch gradient norm on a validation subset) at several epsilon values. Compare R_obs with R_bound from Eq. (27), using the paper's method for estimating beta, sigma_j^2, and G_j^2. If the rank ordering of configurations by R_bound differs from the rank ordering by R_obs, or if the ratio R_bound/R_obs varies by more than 50% across configurations, the equality assumption in Section VI fails and HASFL's optimized (b, mu) is not guaranteed to minimize actual training latency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's main claim is that jointly optimizing per-device batch sizes b and split points mu via problem P'' minimizes the wall-clock latency to reach a target accuracy. The bridge from the convergence analysis to this optimization is made in Section VI, where the authors state that 'the objective function is minimized if and only if (27) holds as an equality.' However, (27) is derived from Corollary 1, which is only a lower bound on the number of rounds R needed to satisfy the average squared gradient norm condition. Corollary 1 says R >= R_bound(epsilon, b, mu); it does not say R = R_bound. The subsequent objective Theta(b, mu) in Eq. (43) replaces R by this lower bound, so HASFL minimizes a surrogate of the true latency. For the surrogate to be valid, the bound must be tight enough that the ranking of configurations by R_bound matches the ranking by actual R. The paper does not provide any such tightness analysis: no experiment compares predicted R from (27) with observed rounds, and the evaluation in Figs. 5-6 measures wall-clock time to a test-accuracy plateau rather than the average gradient norm epsilon that appears in Corollary 1. Because the bound contains worst-case constants (beta, G_j^2, sigma_j^2) and uses multiple Cauchy-Schwarz and Jensen steps in Lemma 1 and Theorem 1, there is no a priori reason to expect equality or even rank preservation. If the bound is loose in a configuration-dependent way, the BCD solution can minimize the surrogate while increasing actual training latency. This is the load-bearing soft spot of the paper's central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HASFL, a heterogeneity-aware split federated learning framework that adaptively controls per-device batch sizes and model split points to minimize the training latency required to reach a target convergence accuracy. The authors derive a convergence upper bound for SFL with heterogeneous batch sizes and cut layers (Lemma 1 and Theorem 1), then use the lower bound on the required number of training rounds obtained from this bound (Corollary 1, Eq. (27)) to formulate a latency minimization problem. The problem is decomposed into batch-size and model-splitting subproblems, solved alternately by a block-coordinate descent algorithm (Algorithm 2). Experiments on CIFAR-10 and CIFAR-100 with VGG-16 and ResNet-18 report faster convergence and higher accuracy than several benchmarks.","tokens_in":26387,"tokens_out":4271,"duration_ms":56719,"significance":"If the optimization-to-latency bridge were valid, the paper would make a useful contribution to resource-efficient split federated learning: it provides the first convergence bound for SFL with jointly varied batch sizes and cut layers, and the proposed system-level optimization is evaluated across a broad set of scenarios. The theoretical derivations in Lemma 1 and Theorem 1 follow standard smoothness and bounded-variance machinery and appear algebraically sound. The experimental study is extensive and consistently shows improvements over the chosen benchmarks. However, the central claim that the proposed optimization minimizes actual training latency is not established, because the optimization is built on an unexamined equality assumption that replaces the convergence bound with an exact round-count prediction; the evaluation does not validate this assumption. The significance of the paper therefore hinges on a load-bearing point that currently lacks support.","major_comments":[{"comment":"The paper converts a lower bound into an exact equality. Corollary 1 establishes only that R >= R_bound(epsilon, b, mu) for the average squared gradient norm to be at most epsilon; it does not establish that the actual number of rounds equals R_bound. The statement in Section VI that 'the objective function is minimized if and only if (27) holds as an equality' is therefore unsupported. Since Eqs. (42)-(45) and the entire subsequent optimization minimize Theta(b, mu) = R_bound * [T_S + T_A/I], the claimed latency optimality is for a surrogate objective. The manuscript provides no tightness analysis, no comparison of predicted R_bound against observed rounds, and no argument that the ranking of configurations by R_bound matches the ranking by actual R. This is load-bearing for the paper's main claim, and the issue cannot be dismissed as a minor technicality because the bound contains worst-case constants and is derived through multiple relaxations in Lemma 1 and Theorem 1.","section":"Section VI, Eqs. (27) and (43)"},{"comment":"Proposition 1 claims 'The optimal BS decision' is given by Eq. (48), obtained by solving the first-order condition with the Newton-Jacobi method. The proof, however, only demonstrates that for each i', holding all other batch sizes fixed, the objective is initially decreasing and then increasing in b_i' (Eqs. (49)-(50)). This establishes coordinate-wise unimodality, not joint convexity or global optimality of the simultaneous root. The optimal solution under the coupling constraints C4, R3, and R4 may lie on the boundary even when the unconstrained root does not; the correction using kappa_i is heuristic. The remark's claim that exhaustive search over 3^N combinations identifies the global optimum is also inconsistent with the algorithm actually deployed, which solves the nonlinear system and applies a one-time correction. Consequently, the guarantee that P1 is solved exactly in the BCD algorithm is not proven.","section":"Section VI, Proposition 1"},{"comment":"The empirical evaluation does not test the key theoretical bridge. Corollary 1 concerns the average squared gradient norm condition (Eq. (26)), but the convergence criterion in the experiments is a test-accuracy plateau (e.g., 'test accuracy increases by less than 0.02 percent across five consecutive training rounds'). Therefore, the measured speedups in Figs. 5-6 cannot validate the assertion that Eq. (27) accurately predicts the number of rounds required to reach epsilon. The paper should report the actual versus predicted round counts, or measure the gradient-norm criterion directly, to support the equality assumption that underlies the optimization.","section":"Section VII-B, Figs. 5-6"}],"minor_comments":[{"comment":"There is a typo in the related-work discussion: 'incentive mechanisms to to balance the training load' should read 'incentive mechanisms to balance the training load'.","section":"Section II"},{"comment":"The approximation ceil(R/I) ≈ R/I is used without an error bound. Since the optimization can, in principle, select small R values, the relative error of this approximation should be quantified or justified.","section":"Section VI, Eq. (40)"},{"comment":"The remark states that the optimal solution can be obtained by exhaustive search over 3^N combinations, but the described algorithm solves the nonlinear system and performs a correction step. These two statements describe different procedures, and the paper should clarify which one is implemented and what guarantees each provides.","section":"Section VI, Proposition 1 remark"},{"comment":"Algorithm 2 terminates when the objective change is below epsilon, but no convergence analysis of the BCD scheme is provided. This is acceptable for a heuristic, but the text should avoid implying that the final solution is globally optimal.","section":"Algorithm 2 and Section VI"},{"comment":"The x-axis in Figs. 2(a) and 3(a) is labeled 'Training Rounds' while the captions say 'epochs'; the terminology should be made consistent.","section":"Figures 2 and 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's incremental contribution over the authors' prior works [21] and [23] is described only briefly in Section II; the editor may wish to verify that the novel element (jointly optimizing batch sizes and split points with a convergence bound) is sufficiently differentiated. Additionally, the equality assumption in Section VI is the main scientific risk; if the authors cannot provide tightness evidence or empirical validation of predicted versus actual round counts, the paper's claims should be reframed as heuristic latency reduction rather than latency-optimality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the useful news: this is the first SFL convergence bound I know that lets each device have its own batch size while also allowing cut-layer heterogeneity. The proof is standard smoothness/variance machinery and the algebra in Theorem 1 checks out. The latency model in Section V is careful, and the BCD/Dinkelbach/Newton-Jacobi stack is appropriate. If the optimization truly minimizes wall-clock training time, the 4-10x speedups in the simulations would be meaningful.\n\nThe soft spot is exactly where the reader and stress-test put it. Corollary 1 gives a lower bound on rounds, R >= ..., and Section VI converts that into an equality: \"the objective function is minimized if and only if (27) holds as an equality.\" That is a tightness assumption, and the paper never checks it. There is no experiment comparing predicted rounds from (27) with observed rounds, and the evaluation measures time to a test-accuracy plateau, which is a different stopping criterion than the average gradient norm in the bound. If the bound is loose in a configuration-dependent way, the optimized batch sizes and split points can minimize the surrogate while making actual latency worse. I read this as a load-bearing gap, not a proof error.\n\nMinor issues: the benchmarks are internal ablations, not published SOTA methods, so the \"state-of-the-art\" claim is overstated. No code is released, which makes the tightness check hard to do yourself. Proposition 1's \"global optimum\" claim is stronger than what the proof shows: the function is unimodal per coordinate, but the joint system is solved with Newton-Jacobi and then discretized, so the result is a well-justified heuristic, not a proven global optimum. The free constants (beta, sigma^2, G^2) are estimated following [24], but there is no sensitivity analysis, which matters because the optimization sits on top of those estimates.\n\nThese are addressable. The paper deserves a serious referee. I would ask for a tightness section - predicted vs. actual round counts under the bound, plus a check that the BCD solution on the surrogate does not inflate actual latency - before accepting. The convergence bound alone is a real contribution.","headline":"Solid first convergence bound for SFL with per-device batch sizes, but the latency-minimization claim rests on an unverified tightness assumption.","tokens_in":26896,"tokens_out":2001,"would_cite":false,"duration_ms":24462,"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 per-device batch sizes and model split points, chosen from a new convergence bound, minimize the training time of split federated learning on heterogeneous edge devices.","keywords":["federated learning","split federated learning","batch size","model splitting","mobile edge computing","straggler effect","convergence bound","latency optimization"],"falsifier":"Record, on the paper's CIFAR-10/VGG-16 setting, the round at which the average squared gradient norm first drops below a target $\\varepsilon$ and compare it with the value $R$ predicted by Corollary 1 using the same estimated parameters; a ratio far from 1 across different batch-size assignments would show the bound is too loose for the equality assumption to hold.","tokens_in":25844,"feed_emoji":"⚡","tokens_out":11853,"duration_ms":112634,"temperature":0.7,"pith_summary":"This paper claims that the straggler bottleneck in split federated learning can be overcome by giving each edge device its own batch size and its own model split point, and that the right choices can be computed from a convergence bound rather than tuned by hand. The authors derive an upper bound on the average squared gradient norm of the global loss that depends explicitly on every device's batch size and on the depth of the client-side cut layer. They then treat this bound as an exact predictor of the number of training rounds needed to reach a target accuracy, turning latency minimization into a joint, tractable optimization over batch sizes and split points. If the bound is tight enough, the resulting schedules cut training time by at least 4x and improve accuracy by about 1% on standard image benchmarks.","feed_headline":"HASFL cuts split-federated training time at least 4x","feed_subtitle":"Choosing batch sizes and cut layers from a convergence bound tames the straggler problem in split federated learning.","key_machinery":"The load-bearing object is the convergence bound of Theorem 1 together with its inversion in Corollary 1. The bound decomposes the optimization error into three terms: an initial-optimization-gap term $\\frac{2\\vartheta}{\\gamma R}$ that shrinks with more rounds, a mini-batch variance term $\\frac{\\beta\\gamma}{N^2}\\sum_i\\sum_j \\sigma_j^2/b_i$ that shrinks when any device raises its batch size, and a staleness term $4\\beta^2\\gamma^2I^2\\sum_{j=1}^{L_c}G_j^2$ that appears only when client-side sub-models are aggregated every $I>1$ rounds and grows with the client-side split depth $L_c$. Because the variance term depends on the sum of $\\sigma_j^2/b_i$ over devices, the bound predicts that batch sizes compensate across devices: a fast device can carry a large batch while a slow device uses a small one without changing the worst-case convergence guarantee. This bound converts the system design problem into the transformed program $P''$, solved by alternating a closed-form batch-size update (Proposition 1, using the Newton-Jacobi method) with a Dinkelbach-based split-point update.","core_discovery":"Under $\\beta$-smoothness of the local losses and bounded per-layer gradient variance and second moments, the paper proves (Theorem 1) that the average squared gradient norm after $R$ rounds satisfies $\\frac{1}{R}\\sum_{t=1}^R \\mathbb{E}\\|\\nabla f(w^{t-1})\\|^2 \\le \\frac{2\\vartheta}{\\gamma R} + \\frac{\\beta\\gamma}{N^2}\\sum_{i=1}^N \\sum_{j=1}^L \\frac{\\sigma_j^2}{b_i} + \\mathbf{1}_{\\{I>1\\}} 4\\beta^2 \\gamma^2 I^2 \\sum_{j=1}^{L_c} G_j^2$. Inverting this bound (Corollary 1) gives a lower bound on the rounds $R$ needed to reach accuracy $\\varepsilon$, and the paper asserts that the optimal schedule is obtained when this lower bound holds with equality. Substituting that expression for $R$ into a per-round latency model that includes client forward/backward passes, activation upload/download, and the periodic client-side model aggregation yields an explicit latency objective, which is then minimized over per-device batch sizes $b_i$ and binary split-point variables $\\mu_{i,j}$. The paper claims this is the first convergence bound for SFL with both varied batch sizes and cut layers, and that the resulting heterogeneity-aware schedule is what enables the reported speedups.","pith_inferences":["The paper never verifies its equality assumption: a reader who logs predicted versus actual rounds to reach $\\varepsilon$ on the paper's own testbed could check whether the optimized schedule is truly latency-optimal or only optimal for the surrogate.","The per-layer variance assumption ($\\sigma_j^2/b$ per layer, summed over layers) is a strong structural condition; if modern layers with normalization or adaptive optimizers violate it, the batch-size compensation rule could mis-rank devices.","Because the server-side common sub-model is assumed to synchronize every round for free, the analysis applies to a single edge server; a hierarchical or multi-hop version of SFL would need a modified staleness term.","A direct stress test is to compare HASFL's end-to-end measured latency against a brute-force grid search over batch sizes and split points on a small model; the size of any gap quantifies how much is lost by treating the upper bound as exact."],"forward_implications":["If the bound is tight, the optimal schedule assigns faster devices larger batch sizes and slower devices smaller ones, with each device's cut layer chosen to balance smashed-data communication cost against aggregation frequency.","For aggregation interval $I>1$, the bound predicts that shallower client-side splits accelerate convergence, so the optimizer sends more layers to the server even when that increases activation traffic, stopping where the latency trade-off balances.","Because the batch-size rule is nearly closed-form, the schedule can be recomputed cheaply each aggregation round, letting HASFL adapt to drifting device speeds and channel rates during training.","The experiments report at least 4x faster convergence and about 1% higher accuracy than benchmarks that randomize batch sizes, split points, or both."],"supporting_citations":[{"why":"Supplies FedAvg, the synchronous aggregation protocol that SFL inherits and whose straggler bottleneck this paper targets.","marker":"[5]"},{"why":"Introduces split learning, the layer-wise partitioning scheme that SFL builds on.","marker":"[12]"},{"why":"Defines SplitFed, the SFL framework this paper extends with adaptive batch sizes and split points.","marker":"[16]"},{"why":"Provides the parallel split-learning wireless system and latency model used in the problem formulation.","marker":"[20]"},{"why":"Gives the prior convergence analysis of adaptive SFL with model splitting, which this paper extends to varied batch sizes.","marker":"[21]"},{"why":"Gives the earlier hierarchical SFL convergence analysis with model splitting, extended here to per-device batch sizes.","marker":"[23]"},{"why":"Supplies the method for estimating β, σ_j², G_j² and the insight that shallower splits raise aggregation frequency.","marker":"[24]"},{"why":"Provides a convergence bound for FL with adaptive mini-batch size and aggregation frequency, conceptually adapted to SFL's client–server split.","marker":"[27]"},{"why":"Provides the parallel split-learning latency model with FLOPs-based computing costs used in the per-round latency formulas.","marker":"[31]"},{"why":"Supplies Dinkelbach's fractional-programming algorithm used to solve the split-point subproblem.","marker":"[46]"}],"fun_headline_variants":["HASFL: 4x faster split FL via adaptive scheduling","Split FL speedup: HASFL tunes batches and splits","HASFL tames stragglers with convergence-bound tuning","Fourfold faster split FL with HASFL's adaptive control","HASFL: adaptive batch and split control for straggler-free FL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes the convergence upper bound is tight enough that the round count it predicts is the round count really needed, so minimizing the bound is the same as minimizing training time; if the bound is loose, the optimized batch sizes and split points can lower the surrogate without lowering real latency.","fun_headline_variants_meta":{"raw":{"variants":["HASFL: 4x faster split FL via adaptive scheduling","Split FL speedup: HASFL tunes batches and splits","HASFL tames stragglers with convergence-bound tuning","Fourfold faster split FL with HASFL's adaptive control","HASFL: adaptive batch and split control for straggler-free FL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3235,"prompt_tokens":996,"completion_tokens":2239,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":2151}},"tokens_in":612,"tokens_out":2239,"duration_ms":21057,"temperature":1.0,"reasoning_tokens":2151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:13:09.811419+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record, on the paper's CIFAR-10/VGG-16 setting, the round at which the average squared gradient norm first drops below a target $\\varepsilon$ and compare it with the value $R$ predicted by Corollary 1 using the same estimated parameters; a ratio far from 1 across different batch-size assignments would show the bound is too loose for the equality assumption to hold.","supporting_citations":[{"cited_title":"Communication-efficient Learning of Deep Networks From Decentral- ized Data,","cited_arxiv_id":null,"evidence_quote":"Supplies FedAvg, the synchronous aggregation protocol that SFL inherits and whose straggler bottleneck this paper targets."},{"cited_title":"Splitfed: When Federated Learning Meets Split Learning,","cited_arxiv_id":null,"evidence_quote":"Defines SplitFed, the SFL framework this paper extends with adaptive batch sizes and split points."},{"cited_title":"Split learning over Wireless Networks: Parallel Design and Resource Management,","cited_arxiv_id":null,"evidence_quote":"Provides the parallel split-learning wireless system and latency model used in the problem formulation."},{"cited_title":"Hi- erarchical Split Federated Learning: Convergence Analysis and System Optimization,","cited_arxiv_id":null,"evidence_quote":"Gives the earlier hierarchical SFL convergence analysis with model splitting, extended here to per-device batch sizes."},{"cited_title":"Adaptive Federated Learning in Resource Constrained Edge Computing Systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the method for estimating β, σ_j², G_j² and the insight that shallower splits raise aggregation frequency."},{"cited_title":"DYNAMITE: Dynamic Interplay of Mini-batch Size and Aggregation Frequency For Federated Learning with Static and Streaming Datasets,","cited_arxiv_id":null,"evidence_quote":"Provides a convergence bound for FL with adaptive mini-batch size and aggregation frequency, conceptually adapted to SFL's client–server split."},{"cited_title":"Efficient Parallel Split Learning over Resource-constrained Wireless Edge Networks,","cited_arxiv_id":null,"evidence_quote":"Provides the parallel split-learning latency model with FLOPs-based computing costs used in the per-round latency formulas."},{"cited_title":"On Nonlinear Fractional Programming,","cited_arxiv_id":null,"evidence_quote":"Supplies Dinkelbach's fractional-programming algorithm used to solve the split-point subproblem."}],"review_version":1}