{"id":"21be304f-d0fd-4975-8df4-7b5a9bbfd295","arxiv_id":"2501.15728","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper proposes FedAvg plus an exponential learning-rate decay driven by loss reduction and calls it a control system, but provides insufficient evidence for the claimed gains.","lead":"This paper claims that adding a control-theoretic learning-rate scheduler to personalized federated learning improves accuracy and convergence. The supporting experiments are a small simulated study with no code, data, or statistical detail, so the claim is not verifiable from the manuscript.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Undefined P and f_i plus an unspecified baseline make the claimed control-system gain non-identifiable and unreproducible.","rationale":"The reader's verdict of REJECT is supported: the central claim is an empirical improvement from a control mechanism, but the paper does not provide a fully specified algorithm or enough experimental detail to identify which component produces the gain. My stress-test focuses on the single most load-bearing gap: the personalization function P in Eq. (3) and the contribution functions f_i in Eq. (5) are placeholders, and the baseline configuration is unspecified. This makes the reported improvement non-identifiable and unreproducible, which is an internal completeness problem rather than a disagreement with consensus. The missing stability analysis is secondary; even with full stability guarantees, the claim would remain untestable without concrete P and f_i. The proposed concrete test would settle the concern by isolating the control update from personalization and from a trivial learning-rate decay. I found no separate objection beyond this under-specification, and no reason to adjust the reader's REJECT verdict.","tokens_in":5052,"tokens_out":3284,"duration_ms":30820,"concrete_test":"Fix the algorithm with concrete instantiations, e.g., P = one local epoch of fine-tuning on client data and f_i = local gradient norm, then run on a standard public benchmark (FEMNIST or CIFAR-10 with a Dirichlet non-IID split across 50 clients). Compare four conditions: (1) FedAvg with constant learning rate; (2) FedAvg with the exponential learning-rate schedule from Eq. (4); (3) FedAvg with personalization P but constant learning rate; (4) the full method with Eqs. (3)-(5). Report mean plus/minus standard deviation over at least 5 seeds. If condition (4) does not beat conditions (2) and (3) by more than the seed noise, the claimed control-system benefit is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table I: 88.7% accuracy and 0.30 loss with the control system versus 82.5% and 0.45 without it after 10 global rounds. For the control system to be the cause, the comparison must isolate the exponential learning-rate update in Eq. (4) from the personalization step P in Eq. (3) and the client reweighting f_i in Eq. (5). Yet P and f_i are never defined, the baseline's learning-rate schedule is not stated, and no dataset, model, or experimental protocol is given. Eq. (4) also leaves Delta-L undefined; depending on the sign convention, the update implements either a decaying or an accelerating learning rate, and Table II's monotone decline from 0.01 to 0.0062 is consistent with a simple hand-set decay schedule, which is known to improve final accuracy over a constant learning rate. If the baseline used a constant learning rate, the reported gain is expected and does not demonstrate a control mechanism. Table III's client-level gains further conflate the personalization function P with the control loop, since P itself is not specified. The manuscript even contains a placeholder figure caption ('Fig. 1. Our overfiew figure') and no actual figure, reinforcing that the method and experiments are incompletely reported. The essential problem is not the absence of convergence theory; it is that the proposed algorithm is under-specified, making the empirical claim unfalsifiable and not attributable to any concrete mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a framework that combines personalized federated learning with a control-system-inspired learning-rate update and client reweighting. The method is specified through Eqs. (1)-(5): local gradient descent, weighted global aggregation, a personalization function P, an exponential learning-rate update, and client weighting via functions f_i. The paper claims that the integrated system outperforms standard federated learning, reporting in Table I 88.7% accuracy and 0.30 loss versus 82.5% and 0.45 without the control system after 10 global rounds on a simulated 50-client non-IID setting. The manuscript concludes that the integration substantially improves performance, convergence speed, and client-level personalization.","tokens_in":5340,"tokens_out":2352,"duration_ms":23241,"significance":"If the reported gains were real and attributable to a clearly specified mechanism, the idea of using control-theoretic feedback to adjust federated learning hyperparameters would be a useful practical contribution, and the claimed 6.2 percentage-point accuracy improvement in a non-IID setting would merit attention. However, the manuscript does not define the core components of the proposed algorithm, does not describe the experimental setup, and provides only bare comparative numbers with no variability or statistical support. The central empirical claim is therefore not verifiable, and the methodological contribution as stated is not reproducible. The paper also makes no theoretical claims, so the entire contribution rests on the ungrounded experimental tables.","major_comments":[{"comment":"The personalization function P is never defined. The statement that P \"could be a simple linear transformation or a more complex function\" is not a specification. Since Table III attributes client-level improvements to personalization, the absence of any concrete definition means the reported client-level gains cannot be attributed to a concrete mechanism or reproduced by any reader.","section":"Eq. (3), Section III-B3"},{"comment":"The quantity ΔL in the exponential learning-rate update is undefined. It is not clear which loss (global or per-client), over which interval, or with what sign convention ΔL is computed. Depending on the sign of ΔL, the update implements either an exponentially decaying or an accelerating learning rate. Without this definition and without stating the baseline's learning-rate schedule, the accuracy gap in Table I cannot be identified as an effect of the control system rather than a generic decaying learning-rate schedule.","section":"Eq. (4), Section III-C"},{"comment":"The functions f_i are never defined. The text says they represent \"a function of the client's contribution to the model's improvement,\" but no concrete formula is given. Since the weighted aggregation in Eq. (2) depends on these weights, the comparison \"with control system\" versus \"without control system\" conflates the control loop with an arbitrary reweighting of clients, making the claimed control-system benefit non-identifiable.","section":"Eq. (5), Section III-C"},{"comment":"The experimental section gives no dataset description, no data-generation procedure, no model architecture, no optimizer details, no number of local epochs, no number of samples per client, and no measure of variability. The tables report single numbers with no error bars or statistical tests. The central claim that the integrated system \"substantially improves performance across all metrics\" is therefore unsupported by the reported experiments and cannot be independently checked.","section":"Section IV, Tables I-III"},{"comment":"The pseudocode does not implement the claimed personalization step. The main loop contains only local training, aggregation, and learning-rate update; the personalization function P from Eq. (3) never appears in the control flow. This discrepancy means the paper describes an algorithm that is not actually the one used to produce the results, further undermining the reproducibility of the empirical claims.","section":"Algorithm 1, lines 7-12"}],"minor_comments":[{"comment":"The placeholder caption \"Fig. 1. Our overfiew figure\" with no actual figure indicates that the system-architecture description in Section III-A is incomplete and not supported by the intended figure.","section":"Figure 1"},{"comment":"The reference list contains several entries that do not match the in-text citations; for example, reference [6] is cited as the source for \"Smith et al.\" and vertically partitioned data, but the entry is a malware-detection paper, and reference [3] is labeled as Hanzely et al. while the entry is a different paper on personalized federated learning.","section":"References"},{"comment":"The learning-rate sequence in Table II (0.01 down to 0.0062) is consistent with a pre-specified hand-set decaying schedule; the paper should clarify whether these values are the actual output of Eq. (4) or a hand-set schedule, and if so, what values of γ and ΔL produced them.","section":"Table II"},{"comment":"The author list names Michael Geller but the corresponding author email is michel.geller@go.olemiss.edu; this inconsistency and the presence of unrelated cited works suggest the manuscript has not been carefully checked before submission.","section":"Author information"}],"recommendation":"reject","confidential_remarks":"In addition to the technical problems listed in the report, the citation and reference mismatches are severe enough to raise concerns about the integrity of the manuscript's scholarly apparatus: several cited works do not correspond to the cited authors or content, and the reference list includes unrelated papers. Combined with a placeholder figure, an under-specified algorithm, and non-reproducible experiments, the manuscript is not suitable for publication in its current form. These issues are not fixable by routine revision; a rewrite with full methodological detail and a properly conducted evaluation would be needed, which is beyond the scope of a revision of this submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Mike — you can skip the formal review on this one. The core comparison in Table I is meaningless as reported. The control update is a learning-rate multiplier (Eq. 4) with an undefined Delta L, personalization P in Eq. 3 and client weighting f_i in Eq. 5 are placeholders, and no dataset, architecture, hyperparameters, or error bars are given. The baseline's learning-rate schedule is never stated; if it was constant, the improvement is just exponential decay, which is standard. The placeholder figure caption in Fig. 1 and the mismatched reference list (a cancer biosensor paper cited as security work, etc.) reinforce that the manuscript was not checked for basic coherence.\n\nWhat is actually new: essentially nothing. The algorithm is FedAvg with a hand-set decay schedule and an unspecified personalization step. There is no convergence claim, no stability analysis, no identifiable control-theoretic contribution beyond the word 'control.' The one thing the paper does is articulate a plausible motivation — that non-IID personalization can destabilize convergence and a feedback rule might help — but motivation alone isn't a result.\n\nSoft spots are everywhere and load-bearing. The experimental claim is the only evidence, and it is unreproducible. No code, no data, no random seeds, no variance. Even the learning-rate trajectory in Table II is a monotone decline that could be written by hand, so it does not demonstrate a feedback mechanism. The stress-test note's point about Eq. 4's sign convention is fair: Delta L is undefined, so the update could be decaying or accelerating depending on sign. The personalization function P is the actual source of client-level gains in Table III, but it is unspecified, so those numbers cannot be attributed to the control loop.\n\nI disagree with the reader's novelty score of 1 only in that the framing is not even a parameterized variant with claims — it is a sketch. The paper has no identifiable new result at all.\n\nConclusion: this doesn't deserve referee time. It would need a full rewrite with specified functions, a real baseline, experiments with variance, code/data, and a reference cleanup. If you want to use it in a reading group as an example of why minimal reporting standards matter, fine, but not for content.\n\nRecommendation: desk reject.","headline":"Under-specified algorithm, missing experimental protocol, and placeholder artifacts make the claimed control-system gain unverifiable; this is not ready for review.","tokens_in":5825,"tokens_out":1976,"would_cite":false,"duration_ms":17371,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A control-system feedback loop that adjusts the learning rate from loss feedback is claimed to raise personalized federated learning accuracy from 82.5% to 88.7% on non-IID data.","keywords":["personalized federated learning","control systems","non-IID data","adaptive learning rate","model personalization","federated averaging"],"falsifier":"Repeat the 50-client non-IID simulation with Eq. (4) replaced by a fixed learning-rate schedule matching Table II; if accuracy and loss still come out at 88.7% and 0.30, the control feedback is not the active ingredient.","tokens_in":4838,"feed_emoji":"🎛️","tokens_out":7558,"duration_ms":66186,"temperature":0.7,"pith_summary":"This paper proposes combining personalized federated learning with a control-system feedback loop that adjusts the global learning rate in response to the recent change in loss. The central claim is that this dynamic adjustment, together with client-level personalization, makes distributed training faster and more accurate when client data are non-IID. The reported evidence is Table I: after 10 global rounds with 50 clients, global accuracy rises from 82.5% to 88.7% and loss falls from 0.45 to 0.30. The intended consequence is that federated systems in heterogeneous, privacy-sensitive networks can adapt their own training dynamics in real time.","feed_headline":"Adaptive control loop lifts federated accuracy to 88.7%","feed_subtitle":"A dynamic learning-rate rule plus client personalization beats standard federated training on non-IID data after 10 rounds.","key_machinery":"The load-bearing object is the adaptive learning-rate controller: a feedback law of the form $\\eta^{(t+1)} = \\eta^{(t)} \\exp(-\\gamma \\Delta L^{(t)})$, with gain $\\gamma$ and loss-reduction signal $\\Delta L$. On each global round the server uses this rule to decay $\\eta$, and the clients then run local gradient descent at that rate, producing updates that are aggregated by the weighted sum in Eq. (2) and finally passed through the personalization function $P$ in Eq. (3). This loop is what the authors credit for faster convergence and the accuracy improvements; without the control block the same pipeline reduces to vanilla personalized federated averaging.","core_discovery":"The paper's central discovery, stated on its own terms, is that the integration of a control system with personalized federated learning improves performance across all measured metrics. The mechanism is a feedback loop: after aggregating client updates, the server computes the loss reduction $\\Delta L$ and shrinks the learning rate by an exponential factor in Eq. (4), while each client applies a personalization step $P$ to tailor the aggregated global parameters to its local data in Eq. (3). In the simulations, the control-enabled system reaches 88.7% accuracy and 0.30 loss after ten global rounds, versus 82.5% and 0.45 without it, and selected clients show accuracy gains of five to eight points after personalization.","pith_inferences":["Editorial extension: because Eq. (4) resembles a standard exponential-decay schedule, a head-to-head test against a preset sequence would clarify whether the feedback signal is the active ingredient.","Editorial extension: a concrete testable version of the framework is to instantiate the unspecified personalization function $P$ as a small number of local fine-tuning steps and set $f_i$ to the client's loss reduction, then compare that version with standard federated averaging.","Editorial extension: the client-level numbers in Table III are best read as feasibility evidence for a family of personalization mechanisms rather than for one specific procedure, since $P$ and $f_i$ are left undefined."],"forward_implications":["If the reported improvement is real, federated training on non-IID data can be accelerated by an adaptive learning-rate controller, reducing the number of communication rounds needed to reach a target accuracy.","The client-level personalization gains in Table III imply that a single global model can be adapted locally to data distributions that differ from the aggregate, improving utility for individual nodes.","The control-loop design should, in principle, extend to settings with changing client availability or network conditions, since the learning rate reacts to a real-time loss signal.","The framework preserves the privacy property of federated learning: raw data still never leaves the client, and only model parameters are aggregated."],"supporting_citations":[{"why":"Supplies the standard federated-averaging baseline that the proposed control system extends and is compared against.","marker":"[10]"},{"why":"Motivates the need for personalization in face of non-IID client data.","marker":"[11]"},{"why":"Direct antecedent for dynamically adjusting learning rates in federated learning.","marker":"[8]"},{"why":"Provides the control-theoretic feedback-systems methodology that Eq. (4) instantiates.","marker":"[12]"},{"why":"Cited to support the claim that feedback-based control can improve convergence of federated systems under limited communication.","marker":"[13]"}],"fun_headline_variants":["Control loop boosts federated accuracy to 88.7%","Personalized FL plus control hits 88.7% accuracy","Feedback-driven FL training beats standard by 6.2%","Adaptive control sharpens federated learning outcomes","Control-integrated PFL reaches 88.7% in 10 rounds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the unspecified personalization function and client-contribution functions can be chosen so that local adaptation improves accuracy without hurting the global model, and that the exponential learning-rate rule remains stable; the paper supplies no definitions, stability proof, or convergence analysis.","fun_headline_variants_meta":{"raw":{"variants":["Control loop boosts federated accuracy to 88.7%","Personalized FL plus control hits 88.7% accuracy","Feedback-driven FL training beats standard by 6.2%","Adaptive control sharpens federated learning outcomes","Control-integrated PFL reaches 88.7% in 10 rounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000132,"raw_usage":{"total_tokens":1120,"prompt_tokens":918,"completion_tokens":202,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":116}},"tokens_in":534,"tokens_out":202,"duration_ms":2289,"temperature":1.0,"reasoning_tokens":116,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:59:20.942368+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the 50-client non-IID simulation with Eq. (4) replaced by a fixed learning-rate schedule matching Table II; if accuracy and loss still come out at 88.7% and 0.30, the control feedback is not the active ingredient.","supporting_citations":[],"review_version":1}