{"id":"24988613-3539-41d0-bde9-8b5f374e120f","arxiv_id":"2603.18540","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"GAPSL adds leader-gradient selection and direction-alignment regularization to parallel split learning, and its testbed runs show consistent accuracy and convergence-time gains over PSL, SFL, EPSL, and vanilla SL.","lead":"GAPSL is a new training method for split learning that tries to fix divergence caused by uneven data across edge devices by picking out the most consistent device gradients and pulling every other gradient toward them. The practical payoff would be faster, more stable training of models on phones and sensors without the communication-heavy aggregation step of federated learning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Footnote 6's detached-gradient implementation makes the GDA regularizer's gradient identically zero, so Eq. (7) cannot drive the reported alignment gains.","rationale":"The reader's weakest-assumption analysis correctly targets the load-bearing flaw. The paper is clearly written and the PSL problem is well motivated, but the central claimed contribution—gradient direction alignment via Eq. (7)—is internally inconsistent with the implementation note in Footnote 6. If g_i and g_lead are detached, the regularizer has zero derivative; therefore it cannot align gradients or improve convergence as claimed. This is not a matter of disagreement with the community consensus but an internal inconsistency between the stated algorithm and the stated implementation. The proposed concrete test would settle the issue by checking whether the regularizer has any measurable effect once thresholding is controlled. If the effect is zero, the reported GDA gains must be attributed to LGI plus adaptive client filtering, not to the regularization mechanism. Secondary concerns such as missing error bars, unspecified η in Eq. (6), and absence of code/data further weaken the empirical support, but the primary issue is sufficient to support the reader's rejection.","tokens_in":21197,"tokens_out":5317,"duration_ms":58253,"concrete_test":"Re-run the GDA ablation on CIFAR-10/VGG-16 non-IID with λ=0 and with λ=5×10^-4, holding LGI, thresholding, seeds, and all other hyperparameters fixed. If the two runs produce identical parameter updates and identical test-accuracy curves, the direction-aware regularizer contributes no gradient, confirming that Footnote 6 makes it a no-op. If the curves differ, inspect the backward graph to identify the nonzero path for ∂λ(1−cosθ)/∂w and update the implementation description accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C3 introduces the regularized loss L̃_i = L_i + λ(1−cos θ_lead,i) as the core mechanism by which GDA aligns client gradients with the leader gradient. Footnote 6 states that the implementation treats g_i and g_lead as detached constants from the computational graph, explicitly to avoid double backpropagation. Under that implementation, θ_lead,i is a function only of detached tensors, so ∂λ(1−cos θ_lead,i)/∂w = 0 for every trainable parameter w. The direction-aware regularization term therefore contributes no gradient and cannot change any parameter update. What remains in GDA is the adaptive angular threshold of Eq. (6), which filters clients out of the summed loss; that is a client-selection mechanism, not the claimed direction-alignment force. Consequently, the ablation in Fig. 21 cannot be explained by the published loss: if the implementation truly detaches, then keeping thresholding and LGI fixed, the λ(1−cosθ) term is inert; if LGI+GDA and LGI+Non-GDA differ, the implemented gradient path must be different from the one described. The paper needs to either document a real differentiable path for Eq. (7) or withdraw the regularization claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GAPSL, a parallel split learning (PSL) framework designed to mitigate the training divergence caused by non-IID client data. GAPSL has two components: leader gradient identification (LGI), which computes pairwise angular deviations between client gradients, adaptively selects the top-K most directionally consistent gradients, and averages them into a leader gradient; and gradient direction alignment (GDA), which computes each client's angular deviation to the leader gradient, filters out clients whose deviation exceeds an adaptive threshold based on mean−η·std, and adds a λ(1−cosθ) regularization term to the losses of the remaining clients. The method is evaluated on a Jetson-based testbed with CIFAR-10/VGG-16 and CIFAR-100/ViT-Base under IID and non-IID settings. The paper reports consistent improvements in accuracy and convergence time over PSL, SFL, EPSL, and vanilla SL, including non-IID gains of up to 16.0 points over PSL on CIFAR-10.","tokens_in":21399,"tokens_out":4995,"duration_ms":57715,"significance":"If the method worked as described, it would be a valuable contribution to aggregation-free split learning: it removes client-side model aggregation while addressing gradient conflict, and the hardware testbed evaluation with two model families is more realistic than simulation-only studies. The empirical motivation in Section II is clearly presented, and the ablation studies are a useful attempt to isolate LGI and GDA. However, the central mechanism is undermined by the paper's own Footnote 6, and several hyperparameters are selected on the same datasets used for evaluation, without independent validation or variance reporting. As written, the reported results do not establish the claimed direction-alignment effect.","major_comments":[{"comment":"The central mechanism is not present in the evaluated system. Footnote 6 states that the implementation treats g_i and g_lead as detached constants from the computational graph. Under that implementation, θ_lead,i depends only on detached tensors, so ∂λ(1−cos θ_lead,i)/∂w = 0 for every trainable parameter w. The regularization term contributes exactly zero gradient, and Eq. (7) reduces to L_i plus a constant. Therefore the ablation in Fig. 21 (LGI+GDA vs. LGI+Non-GDA) cannot be explained by the proposed GDA; the only active component is the adaptive threshold of Eq. (6), which is client selection, not alignment. The authors must either document the actual differentiable path for Eq. (7) or withdraw the alignment claim.","section":"§III-C3, Eq. (7) and Footnote 6"},{"comment":"The threshold θ_th = max(min(μ − ην, π/2), 0) is presented as a 'robust confidence boundary' without any distributional justification. For skewed or multi-modal angular deviations, mean−η·std is not necessarily a meaningful outlier cutoff, and the clamps at 0 and π/2 can discard all clients or admit all clients in ways that are not analyzed. Moreover, η is not reported in Section IV-A4 and is not ablated in Section V-C. Since this threshold is the only remaining part of GDA that can affect training once Eq. (7) is inert, its unspecified handling is a concrete gap.","section":"§III-C2, Eq. (6)"},{"comment":"The hyperparameter protocol is under-specified. λ is set to 5×10−4 for VGG-16 and 3×10−5 for ViT-Base, K_min/K_max to 20%/80%, and η is never stated. These values appear to be chosen on CIFAR-10/CIFAR-100, and Tables I–III report single runs without standard deviations or significance tests. Without a described model-selection procedure (hold-out validation, grid ranges, repeated seeds), the claimed consistent gains may reflect tuning to the test sets rather than a robust algorithmic advantage.","section":"§IV-A4 and §V"},{"comment":"The adaptive selection ratio is justified only by the empirical observations in Fig. 10, which come from the same CIFAR-10/VGG-16 setup. The linear t/T term and the multiplicative coupling with the normalized dispersion term are ad hoc; no theoretical derivation or comparison with alternative functional forms is given. This weakens the claim that LGI's adaptivity is principled rather than curve-fit to the observed datasets.","section":"§III-B2, Eq. (2)"}],"minor_comments":[{"comment":"Typo: 'deivce' should be 'device'. Also, the sentence 'PSL incurs the same communication overhead as standard PSL' is tautological and should be rephrased.","section":"§III-A"},{"comment":"Duplicate wording: 'The gradient selection ratio is is bounded...' Also, the regularization coefficient λ is defined twice in the same paragraph.","section":"§IV-A4"},{"comment":"The legend appears to omit the LGI+GDA curve, making it impossible to identify which line corresponds to the full method. Please fix the figure.","section":"Fig. 20(b)"},{"comment":"The symbol S is used both as the set of client devices and as the number of devices (|S|), which is slightly confusing. Consider using a different symbol for the set, such as C.","section":"§III-B1, Eq. (1)"}],"recommendation":"reject","confidential_remarks":"The Footnote 6 issue is decisive: the paper's own description of the implementation makes the GDA regularizer mathematically inert, so the central claimed mechanism does not exist in the evaluated system. Although the testbed and experimental scope are substantial, the main result does not support the paper's conclusions. The authors might fix this by reimplementing the regularizer with a true differentiable path, but that would require re-running all experiments and would effectively constitute a new study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jan — quick take on GAPSL. The problem is genuinely important: aggregation-free PSL collapses under non-IID data, and a coordination mechanism that doesn't add communication is worth having. The paper's empirical motivation is solid — the comparison of SFL vs PSL on CIFAR-10 with VGG-16 shows the divergence clearly, and the consistency-score analysis (Fig. 5-11) is a nice piece of diagnostic work. LGI, the leader-gradient selection mechanism, is well designed and plausibly explains part of the gains; the adaptive ratio based on dispersion and training progress is reasonable, even if it relies on hand-set bounds.\n\nThe soft spot is not in the framing but in the math. Section III-C3 and Algorithm 2 define the GDA regularized loss as L_i + λ(1−cosθ_lead,i). Footnote 6 then says the implementation treats g_i and g_lead as detached constants to avoid double backprop. That means θ_lead,i is a constant scalar w.r.t. the model parameters, so the regularization term contributes exactly zero gradient. The entire alignment force described in the abstract and Section III-C simply does not exist in the evaluated system. What remains is the adaptive angular threshold of Eq. (6), which is pure client selection. So Fig. 21's ablation, which attributes accuracy gains to GDA, is unexplained by the stated mechanism. Either the authors actually implemented a differentiable path (e.g., they did not detach, or they used a straight-through estimator) — in which case they need to describe it — or the GDA claims should be withdrawn and the paper reframed as an LGI-plus-thresholding method. This is a load-bearing issue, not a stylistic quibble.\n\nSecondary concerns: Eq. (6) is a mean-minus-η·std outlier rule with no distributional justification, and η is never specified in the experiments. Hyperparameters λ and K_min/K_max are tuned on the same datasets used for evaluation, with no separate validation protocol. No code or data is released, and there are no error bars or seeds reported anywhere. These are all fixable, but they compound the uncertainty.\n\nIf the mechanism can be properly documented and the claims rescaled, the paper has value for the split-learning community. As it stands, I wouldn't cite the GDA result, but I would send it to a careful referee because the problem is real and the LGI design deserves scrutiny. It's a good reading-group piece — the footnote is a textbook example of why detaching gradients in a regularizer kills it.","headline":"Real problem, solid LGI story, but the GDA regularizer as implemented is mathematically inert — the core alignment claim needs a fix before the paper can be taken at face value.","tokens_in":21981,"tokens_out":3891,"would_cite":false,"duration_ms":37543,"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":"GAPSL argues that the convergence collapse of aggregation-free parallel split learning under non-IID client data is caused by inter-device gradient directional inconsistency, and that a dynamically selected leader gradient plus a cosine ali","keywords":["parallel split learning","gradient directional inconsistency","data heterogeneity","leader gradient identification","gradient direction alignment","non-IID data","edge computing","distributed machine learning"],"falsifier":"Inspect the evaluated implementation's backward pass and compute the gradient of λ(1−cos θ_lead,i) with respect to the server-side model parameters while g_i and g_lead are detached; if that gradient is numerically zero, the GDA term changes no update, and any accuracy difference between LGI+GDA and LGI+Non-GDA must come from the filtering or averaging strategy rather than from the claimed alignment penalty.","tokens_in":1416,"feed_emoji":"🧭","tokens_out":1838,"duration_ms":89037,"temperature":0.7,"pith_summary":"The paper proposes GAPSL, a parallel split learning framework that keeps the communication and deployment advantages of aggregation-free PSL while fixing its tendency to diverge when client data is unevenly distributed. The central idea is to have the server identify a leader gradient from the client gradients that point most consistently in the same direction, then pull each client's update toward that leader with a cosine-based penalty. If the mechanism works as claimed, aggregation-free split learning can remain stable under severe data heterogeneity and even beat split federated learning in accuracy and convergence speed. The paper reports non-IID gains of up to 16.0 accuracy points over PSL on CIFAR-10 and 13.8 points on CIFAR-100.","feed_headline":"A leader gradient keeps split learning stable on uneven data","feed_subtitle":"Selects directionally consistent updates and aligns the rest, beating other split learning by up to 16 accuracy points.","key_machinery":"Two coupled server-side mechanisms. Leader Gradient Identification (LGI) computes pairwise angular deviations arccos(⟨g_i,g_j⟩/(‖g_i‖‖g_j‖)), averages them into a per-client consistency score, selects the top K_t% of clients with K_t = K_min + (t/T)·(relative stability score)·(K_max−K_min), and averages their server-side gradients into g_lead. Gradient Direction Alignment (GDA) then computes each client's angular deviation θ_lead,i from g_lead, builds an adaptive threshold max(min(μ−ησ, π/2), 0) from the mean and standard deviation of those deviations, filters out clients beyond the threshold, and adds λ(1−cos θ_lead,i) to the surviving clients' losses. The cosine penalty is intended to vani","core_discovery":"GAPSL's central claim is that the divergence of aggregation-free PSL under heterogeneous data is a gradient-direction problem, not a fundamental limitation of split learning. The server computes each client's gradient on its own segment, scores each client by the average angular deviation from every other client's gradient, selects a top fraction whose size adapts to dispersion and training round, and averages the survivors into a 'leader gradient' that approximates the global descent direction. Then, for clients within an adaptive angular threshold of the leader, the loss is augmented with λ(1−cos θ_lead,i), a cosine penalty that vanishes when aligned. Evaluated on CIFAR-10/VGG-16 and CIFAR","pith_inferences":["If the footnote to Eq. (7) is taken literally, that g_i and g_lead are detached constants, then the GDA penalty contributes zero gradient to the update, so the measured GAPSL-vs-PSL gains would be explained by LGI's selective averaging and threshold filtering rather than by the direction-aware alignment mechanism described in Section III-C.","A direct test would be to implement the regularizer without detachment, using true double backpropagation; if accuracy rises, the original first-order implementation was leaving the paper's own mechanism unused, and if it does not, the stability comes from client selection, not alignment.","The adaptive threshold mean−η·std assumes the angular-deviation distribution is roughly symmetric and bell-shaped; a rank-based or percentile threshold would be a natural alternative if the distribution is skewed, as is often the case under extreme class imbalance."],"forward_implications":["If GAPSL is correct, aggregation-free PSL can remain stable under severe non-IID data, preserving PSL's lower communication and deployment costs relative to SFL.","The same server-side coordination applies without any client-side model sharing, so the privacy properties of split learning are retained.","Reported non-IID accuracy gains over PSL are up to 16.0 points on CIFAR-10 and 13.8 points on CIFAR-100, with faster convergence on both datasets and models.","The adaptive selection ratio and angular threshold remove the need to tune a fixed selection fraction or threshold for each task.","The method maintains accuracy better than PSL as the number of devices grows to 40 under non-IID data, suggesting scalability to larger edge systems."],"fun_headline_variants":["Gradient alignment tames split learning on skewed data","How leaders stop split learning divergence on uneven data","GAPSL: align gradients, dodge split-learning divergence","Leader gradients anchor split learning across uneven clients","Beat heterogeneous split learning drift with leader gradients"],"cache_read_input_tokens":23168,"weakest_assumption_plain":"The load-bearing premise is the paper's footnote to Eq. (7), that treating the client and leader gradient vectors as detached constants preserves the GDA regularizer's alignment effect; if a detached constant is used, the penalty has zero gradient, so the described alignment mechanism would not exist in the evaluated system and any gains would have to come from LGI selection and threshold filtering alone.","fun_headline_variants_meta":{"raw":{"variants":["Gradient alignment tames split learning on skewed data","How leaders stop split learning divergence on uneven data","GAPSL: align gradients, dodge split-learning divergence","Leader gradients anchor split learning across uneven clients","Beat heterogeneous split learning drift with leader gradients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000529,"raw_usage":{"total_tokens":2393,"prompt_tokens":758,"completion_tokens":1635,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":1562}},"tokens_in":502,"tokens_out":1635,"duration_ms":13474,"temperature":1.0,"reasoning_tokens":1562,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T05:45:39.454590+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the evaluated implementation's backward pass and compute the gradient of λ(1−cos θ_lead,i) with respect to the server-side model parameters while g_i and g_lead are detached; if that gradient is numerically zero, the GDA term changes no update, and any accuracy difference between LGI+GDA and LGI+Non-GDA must come from the filtering or averaging strategy rather than from the claimed alignment penalty.","supporting_citations":[],"review_version":1}