{"id":"490107a9-ed94-46a8-8aa0-35b6743e6fba","arxiv_id":"2608.09208","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FedA2L adapts per-layer learning rates from local weight-divergence and aggregation-stability signals, accelerating convergence in decentralized federated learning without extra communication.","lead":"FedA2L gives each layer of a model its own learning rate inside decentralized federated learning, using local signals about how much a layer moves during training and how much neighbors correct it. The method claims up to 4.94x faster convergence and fewer communication rounds, which matters when communication, not compute, is the bottleneck.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The layer-wise attribution of FedA2L's speedup is untested: its nonlinear fusion can raise the mean LR above the base schedule, so the gains may be a global LR-scale effect rather than per-layer allocation.","rationale":"The reader's weakest assumption (hyperparameter transferability) is valid and well-supported by Section 5.6: R_warm=10 on ResNet-18 raises rounds to target from 175 to 462, xi=0.3 causes non-convergence, and TinyImageNet uses a special xi=0.05. However, I find a more fundamental, untested assumption in the central mechanism. FedA2L's contribution is explicitly layer-wise adaptation, yet no experiment removes layer-wise variation while keeping the average LR scale. The construction in Eqs. 5-8 combines Z-scores through exp and tanh; this nonlinear mapping is not centered at unity for typical distributions of omega and delta, so the method may systematically operate at an effective global LR above eta0*gamma_r. If so, the observed speedups over vanilla (constant eta0) and over schedulers (decaying from eta0) could be explained by a simple global LR warm-up/boost, not by the layer-wise allocation that the paper motivates through foundational vs specialized layers. The ablations and Fig. 4 show that different layers receive different LRs and that both sigma and zeta contribute, but these do not disprove the global-scale confound. The convergence theorem (Theorem 1) does not help: its Assumption A5 is asserted rather than proved, and even if accepted it only shows FedA2L preserves the standard O(1/sqrt(R)) order, not that layer-wise adaptation causes faster convergence. I credit the paper for releasing code, testing six DFL algorithms, and reporting zero extra communication overhead; those are real strengths. But the central interpretive claim needs the proposed global-mean ablation before the mechanism can be accepted as layer-wise. Since the reader's verdict is already CONDITIONAL and this additional condition fits within that category, I recommend keeping the verdict unchanged.","tokens_in":27748,"tokens_out":8449,"duration_ms":85463,"concrete_test":"Run FedA2L on representative Table 2 configurations (e.g., ResNet-18/CIFAR-100/FedAvg and ResNet-34/TinyImageNet/FedAWA) with the per-layer multiplier m_{i,l}^r replaced by its layer-mean at each node, bar{m}_i^r = (1/L) sum_l m_{i,l}^r, applied uniformly to all layers while keeping all other components (warm-up, Z-score windows, decay) identical. If this global-mean variant retains most of FedA2L's round reduction versus vanilla and scheduler baselines, the reported gains do not require layer-wise differentiation. Additionally, report the evolution of bar{m}^r for FedA2L; if it exceeds 1 substantially during the first 100 rounds, the comparison against fixed-LR baselines is confounded by a global LR-scale effect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"FedA2L's central claim is that per-layer LR allocation drives the convergence gains, but the evidence does not isolate this mechanism. In Eqs. 7-8, the multiplier m_{i,l}^r = 1 + tanh(log(beta*e^{omega} + (1-beta)*e^{delta})) is applied per layer. Because the log of a weighted sum of exponentials is right-skewed even when the Z-scores omega and delta are centered near zero, the layer-average of m can exceed 1 in early rounds; combined with the cap near 2, this acts as a global LR boost relative to the vanilla eta0 and to schedulers that decay from eta0. The paper never reports the mean multiplier trajectory nor compares against a 'global-mean' version that removes layer-wise variation while preserving the average scale. The ablations in Table 6 and Fig. 5 only vary beta (sigma vs zeta) and hyperparameters; both arms retain layer-wise variation, so they cannot distinguish layer-wise allocation from a common scale effect. If the speedup survives with a single global multiplier applied uniformly to all layers, the central contribution (layer-wise adaptation) is not the cause of the reported 4.94x and 59% improvements. This challenge is more load-bearing than the transferability concern because it questions the mechanism on the very benchmarks where the headline numbers were obtained.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedA2L, a server-free, layer-wise learning-rate adaptation method for decentralized federated learning. At each node, FedA2L derives two per-layer metrics from the three intra-round model states: weight divergence (local update intensity) and aggregation stability (network consensus), normalizes them with Z-scores over a sliding window, fuses them into a score, and modulates the base learning rate with a bounded factor and a global polynomial decay. The method adds no communication overhead and is integrated into existing DFL protocols. The paper reports extensive experiments across six DFL algorithms, three image architectures, five datasets, wall-clock convergence, robustness to heterogeneity, scalability, topologies, and time-series forecasting, together with a non-convex convergence theorem. The main claimed results are up to 4.94x faster convergence than vanilla DFL and up to 59% fewer communication rounds than the best scheduler-based baselines. The authors also release code.","tokens_in":27980,"tokens_out":8972,"duration_ms":96142,"significance":"If the central claim holds, FedA2L is a practically attractive contribution: it is local-only, communication-free, easy to plug into existing DFL pipelines, and the empirical campaign is unusually broad for a DFL learning-rate method. The breadth of the evaluation (six base algorithms, three architectures, five datasets, wall-clock and round-based metrics, multiple topologies) is a genuine strength, as is the explicit release of code. The convergence result, if fully established, would show that the modification does not degrade the standard decentralized SGD rate. However, the significance is currently conditional on three issues: the evidence does not isolate layer-wise adaptation from a global learning-rate scale effect; several key hyperparameters are tuned on the same CIFAR-100 benchmark used for headline evaluation; and the convergence analysis rests on an unproved slow-variation assumption plus sketched proof steps. These are load-bearing for the paper's central claims, so the contribution is promising but not yet fully supported.","major_comments":[{"comment":"The paper does not isolate the claimed layer-wise mechanism from a global learning-rate scale effect. The multiplier m_{i,l}^r = 1+tanh(log(beta e^{omega} + (1-beta)e^{delta})) is not mean-zero over layers: by weighted AM-GM, log(beta e^a + (1-beta)e^b) >= beta a + (1-beta)b, and the Z-scores are centered, so the fusion score is systematically right-skewed and the layer-averaged multiplier can exceed 1, especially in early rounds. The paper never reports the mean multiplier trajectory and never compares against a control that applies a single common multiplier to all layers while preserving the average scale. The ablations in Table 6 and Fig. 5 vary beta and other hyperparameters, but both arms retain layer-wise variation, so they cannot distinguish per-layer allocation from a common scale boost. Please add a control experiment with a single global multiplier mbar^r = (1/L) sum_l m_{i,l}^r applied uniformly, and report the trajectory of the mean multiplier; if the speedup persists in that control, the central 'layer-wise' mechanism claim needs to be substantially revised.","section":"§3.4.3, Eqs. (7)-(8); §5.6"},{"comment":"The default hyperparameters are selected on CIFAR-100 and then used in the main evaluation that includes CIFAR-100, so the headline numbers are not an independent test of the method's robustness. Table 6 shows that the defaults are consequential: R_warm=10 on ResNet-18 raises rounds to target from 175 to 462, and xi=0.3 causes non-convergence. Table 7 then fixes R_warm and xi per architecture based on that sensitivity analysis, and even adds a post-hoc adjustment (xi=0.05 for TinyImageNet) after seeing that dataset. This is tuning on the evaluation benchmark. Please provide a holdout evaluation protocol: select defaults on one benchmark (e.g., CIFAR-100), then evaluate on the remaining datasets with those defaults fixed, or use nested validation, and report the sensitivity of the headline claims to that choice.","section":"§5.6.2, Table 7; §5.1"},{"comment":"The convergence analysis relies on an unproved slow-variation property. Assumption A5 requires |m^{r+1}-m^r| <= Delta_m, and the text asserts that Z-score normalization over a finite window, together with bounded tanh, 'ensures' this. Boundedness of m in (0,2) does not by itself bound round-to-round changes: Z-scores can jump when the sliding-window standard deviation is near epsilon, and the paper provides no lemma controlling Delta_m. Since Delta_m enters Eq. (15) through the term C3 L eta_max^2 Delta_m^2/phi^2, the claimed O(1/sqrt R) rate depends on Delta_m being a constant independent of R. In addition, the proof of Theorem 1 is only a sketch: the disagreement recursion for D^r with layer-wise heterogeneous step sizes is asserted, not derived, and the strongly convex result in §4.2 is stated without proof. The theoretical section needs a formal lemma establishing A5, or an alternative argument that does not depend on unproved slow variation, plus full proof details for the stated theorems.","section":"§4, Assumption A5 and Theorem 1"},{"comment":"The statistical reporting is inconsistent with the stated protocol. Section 5.1.3 says all results are mean values across five independent runs, but Tables 4, 5, and 6 report single numbers without standard deviations or confidence intervals. Table 2 itself shows large baseline variability (e.g., FedProx CAWR on CNN/CIFAR-10: 148 ± 64.78 rounds), so unquantified comparisons in the robustness, scalability, topology, and time-series tables do not support the corresponding claims at the stated level of rigor. Please report mean ± std (or CIs) for all tables, at least for the FedA2L rows and the best baseline in each setting.","section":"§5.1.3 with Tables 4, 5, 6"}],"minor_comments":[{"comment":"The pseudocode does not explicitly initialize the learning-rate vector before the warm-up rounds; please state that eta_i^r = eta0 for all layers during r <= R_warm.","section":"Algorithm 1"},{"comment":"The phrase 'When the fusion score lambda=1 (typical behavior)' is presented as a fact, but the paper does not demonstrate that lambda=1 is typical; given the skew of the fusion score discussed in the major comments, this statement should be either justified empirically or rephrased.","section":"§3.4.3"},{"comment":"The table header lists target accuracies (40%, 45%, 42%, 40%, 33%, 45%) but the column groups are not labeled in the displayed text; please add explicit column headers for Heterogeneity, Scalability, and Topology to improve readability.","section":"Table 4"},{"comment":"The right-axis curves are described as 'linearly rescaled to [0,1] for visualization,' but the raw scales of the metrics are not reported; please provide the actual ranges or a representative example so the reader can assess the magnitude of the divergence signals.","section":"§5.6.1, Fig. 4"},{"comment":"The constants C1, C2, C3 are stated to be independent of R but their dependencies on L, phi, and the multiplier bounds are not specified; please state them explicitly or give their definitions in the proof.","section":"§4.1, Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the empirical breadth is a real strength. My main concern is the mechanism attribution: if the reported speedups are reproduced by a single global multiplier control, the paper's central 'layer-wise' contribution is not supported. The tuning-on-CIFAR-100 issue and the missing error bars are also serious but fixable. I would not require perfection, but the global-multiplier control experiment is, in my view, necessary before the paper can be accepted. No concerns about citation practices or novelty disclosure beyond what is stated in the open comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"On the whole, this paper is a legitimate empirical contribution. It introduces a genuinely new trick: per-round, per-layer learning-rate adjustment for decentralized FL, computed entirely from local state transitions (base, trained, aggregated), with zero added communication. The three-state formalization is clean, and the dual metrics (weight divergence and aggregation stability) are reasonable signals. The empirical sweep is unusually wide: six DFL algorithms, three architectures, five datasets, plus robustness checks on heterogeneity, node count, and topology. The code is released, which makes the results checkable. If the reported reductions hold, this is a practical plug-in for DFL practitioners.\n\nThe soft spots are real but mostly fixable. Tables 4 and 5 omit error bars even though the protocol is five seeds, so we can't see variance in the robustness and time-series results. The headline hyperparameters (Table 7) were selected on CIFAR-100 and evaluated on CIFAR-100, and the sensitivity table shows sharp dependence (R_warm=10 on ResNet-18: 175 to 462 rounds; xi=0.3: non-convergence). So the \"no tuning needed\" claim is overstated; the paper itself acknowledges R_warm and xi depend on architecture. That's a caveat, not a fatal flaw.\n\nThe stress-test note is the more substantive concern. FedA2L's story is that per-layer allocation is what drives the gains, but the evidence never isolates that. The multiplier m = 1+tanh(log(weighted combination of exponentials)) is not mean-preserving; early in training the layer-average can sit above 1, effectively boosting the global LR relative to vanilla and to the schedulers. The paper never reports the mean multiplier trajectory, and the ablations (beta=0 vs beta=1) keep layer-wise variation in both arms, so they can't distinguish \"useful per-layer allocation\" from \"same average scale, applied everywhere.\" A simple control — replace the per-layer multiplier with a single global multiplier that has the same mean, and compare — would settle this. Without it, the 4.94x speedup is real but the attribution to \"layer-wise\" is unproven.\n\nThe convergence theorem is a sketch, and A5 (bounded, slowly varying multipliers) is asserted to follow from Z-score normalization, not shown. That part could be tightened, but it's not the main event.\n\nFor a DFL researcher, this is a useful paper to know about, and the code is a plus. It deserves a serious referee. I'd want the error bars, the global-mean ablation, and a clearer statement that the defaults come from tuning on the evaluation benchmark. If those land, this is a solid subfield-level contribution.","headline":"Useful plug-in LR adaptation for DFL with a clean empirical sweep, but the layer-wise attribution is untested and hyperparameters are tuned on the evaluation benchmark.","tokens_in":28569,"tokens_out":2773,"would_cite":true,"duration_ms":76075,"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":"FedA2L claims that assigning each layer its own learning rate, derived from locally observable divergence signals, reliably accelerates convergence in decentralized federated learning — up to 4.94x faster than vanilla DFL and 59% fewer…","keywords":["Decentralized federated learning","Layer-wise learning rate adaptation","Data heterogeneity","Network consensus","Convergence acceleration","Communication efficiency","Internet of Things (IoT)","Edge Computing"],"falsifier":"Run FedA2L with the paper's default hyperparameters on a held-out architecture/dataset pair (for instance, a small transformer on CIFAR-100 with $\\alpha=0.1$) and compare rounds-to-target against the best scheduler baseline; if FedA2L fails to converge or loses the speedup, the transferability premise is refuted. A sharper version sweeps $R_{\\mathrm{warm}}$ over $\\{10,40\\}$ on that pair to see whether the bound reported in Table 6 (175 to 462 rounds on ResNet-18) reproduces on a second architecture.","tokens_in":27474,"feed_emoji":"⚡","tokens_out":15999,"duration_ms":115914,"temperature":0.7,"pith_summary":"This paper tries to establish that a uniform learning rate across all neural-network layers is a primary brake on convergence when data is distributed unevenly across nodes in decentralized federated learning. It proposes FedA2L, which gives every layer its own learning rate each round, computed from two signals that are already present in the local training loop: how much the layer's weights moved during local training, and how stable the layer remained when neighbors' models were averaged in. FedA2L claims to reach target accuracy up to 4.94x faster than vanilla DFL and to use up to 59% fewer communication rounds than scheduler-based baselines, across six DFL algorithms and several architectures and datasets. The method adds no communication overhead and no central coordination, and the paper provides convergence theorems showing the usual decentralized-SGD rates are preserved. If the claim holds, a cheap per-layer rate adjustment could make large-scale edge and IoT learning markedly less communication-hungry.","feed_headline":"Adaptive per-layer rates cut federated learning rounds by up to 59%","feed_subtitle":"In decentralized federated learning, per-layer rates from local divergence signals cut rounds to target accuracy.","key_machinery":"The central object is the three-state decomposition of a DFL round: base state $\\theta^{B,r}_{i,l}$, trained state $\\theta^{T,r}_{i,l}$, and aggregated state $\\theta^{A,r}_{i,l}$, with the aggregated state becoming the next round's base state. From these, FedA2L computes two per-layer metrics — weight divergence $\\sigma = \\|\\theta^T - \\theta^B\\|_2 / (\\|\\theta^B\\|_2 + \\epsilon)$ and aggregation stability $\\zeta$, the fraction of parameters whose change under aggregation is below a threshold $\\tau$ — then normalizes them with Z-scores over a window of $\\rho$ rounds. The normalized signals are combined into a fusion score $\\lambda = \\beta e^{\\omega} + (1-\\beta) e^{\\delta}$, which is mapped to a learning-rate multiplier via $1 + \\tanh(\\log \\lambda)$ with global decay factor $\\gamma_r = (1+\\xi r)^{-1/2}$. This bounded mapping, which keeps every layer's effective rate in $(0, 2\\eta_0\\gamma_r)$ and re-centers persistent divergence back toward the global schedule, is what carries the argument: it converts locally observable optimization dynamics into per-layer step sizes without any extra messages.","core_discovery":"FedA2L's central claim is that the divergence between the three model states of a DFL round — base, trained, and aggregated — is enough information to assign useful per-layer learning rates. Weight divergence $\\sigma_{i,l}^r$ measures how strongly local training moved layer $l$ at node $i$, and aggregation stability $\\zeta_{i,l}^r$ measures how many of the layer's parameters survived neighbor aggregation nearly unchanged. After Z-score normalization over a sliding window, these two signals are fused into a bounded multiplier $1+\\tanh(\\log \\lambda_{i,l}^r)$ that scales the base learning rate for that layer, with a global decay factor keeping rates on a diminishing schedule. The paper reports that this mechanism reaches target accuracy in fewer rounds than vanilla, scheduler-based, and static-search baselines in the large majority of configurations tested, with the biggest gains on deep architectures, severe non-IID splits, and sparse topologies where many baselines fail to converge. The theory section proves that bounded, slowly varying layer-wise multipliers do not change the $\\mathcal{O}(1/\\sqrt{R})$ non-convex rate or the linear rate in the strongly convex case; they only affect constants.","pith_inferences":["Extension: the same three-state divergence signals could also drive layer-wise aggregation weights or per-layer communication compression, since they already summarize local drift and consensus mismatch without sending extra messages.","Extension: because Section 5.6 singles out $R_{\\mathrm{warm}}$ and $\\xi$ as the sensitive knobs, a short calibration phase that estimates them per architecture could make FedA2L nearly hyperparameter-free; the paper leaves this tuning to the user.","Extension: the reported gains shrink when the base algorithm already corrects drift (FedNTD, FedYogi), so a sensible deployment rule would engage FedA2L selectively, for base algorithms that lack such correction, rather than unconditionally.","Extension: an energy-accounting experiment that measures joules per round, rather than rounds to target, would directly test the edge-deployment motivation, since skipped rounds save both communication and battery."],"forward_implications":["FedA2L should reduce communication rounds to target accuracy by up to 59% relative to the best scheduler-based baseline, with the largest gains on deep networks and on sparse ring topologies where many baselines fail to converge.","Because FedA2L only modifies local step sizes, it can be dropped into existing DFL algorithms (FedAvg, FedProx, FedYogi, FedNTD, FedAWA, DFedSAM) without touching aggregation logic or adding communication overhead.","The bounded multiplier and sliding-window normalization guarantee that layer-wise learning rates stay within $(0, 2\\eta_0\\gamma_r)$ and converge back to the global decay schedule, so the method cannot amplify step sizes without bound.","The convergence analysis shows that FedA2L preserves the standard $\\mathcal{O}(1/\\sqrt{R})$ non-convex rate and the linear rate in the strongly convex case of the underlying DFL algorithm.","Under severe heterogeneity (Dirichlet $\\alpha=0.01$), 20–30 nodes, and sparse topologies, FedA2L is reported to maintain or widen its advantage, including in cases where baselines do not converge at all."],"supporting_citations":[{"why":"FedAvg is the vanilla aggregation baseline whose convergence FedA2L accelerates.","marker":"[3]"},{"why":"FedProx provides the client-drift regularized baseline; the paper shows FedA2L targets an orthogonal dimension.","marker":"[14]"},{"why":"FLAYER establishes that layer-wise treatment matters in personalized FL, motivating FedA2L's layer-granular approach.","marker":"[21]"},{"why":"FedYogi is a server-side adaptive optimizer that FedA2L must outperform while remaining server-free.","marker":"[22]"},{"why":"DFedHPO performs a static decentralized LR search, the direct DFL baseline that FedA2L's dynamic adaptation is measured against.","marker":"[24]"},{"why":"D-PSGD supplies the decentralized SGD convergence theory that FedA2L's theorems extend.","marker":"[28]"},{"why":"FedAWA uses adaptive aggregation weights; the paper shows layer-wise LR adaptation is complementary to it.","marker":"[36]"},{"why":"The weight-divergence concept FedA2L adapts into a per-layer metric comes from this non-IID FL analysis.","marker":"[40]"},{"why":"The consensus-mechanism work motivates FedA2L's aggregation-stability metric.","marker":"[41]"},{"why":"DFedSAM is a sharpness-aware baseline that FedA2L also accelerates on in the experiments.","marker":"[48]"}],"fun_headline_variants":["Layer-wise rates speed decentralized federated learning up to 4.94x","Per-layer rates trim federated learning rounds by 59%","FedA2L adapts per-layer rates, cutting communication rounds","Adaptive layer rates boost decentralized FL convergence 4.94x","Divergence-driven layer rates speed decentralized federated learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the default FedA2L hyperparameters, selected on the CIFAR-100 benchmark, transfer to unseen architectures and workloads — yet Section 5.6 shows convergence can degrade sharply when $R_{\\mathrm{warm}}$ or $\\xi$ deviates from those defaults.","fun_headline_variants_meta":{"raw":{"variants":["Layer-wise rates speed decentralized federated learning up to 4.94x","Per-layer rates trim federated learning rounds by 59%","FedA2L adapts per-layer rates, cutting communication rounds","Adaptive layer rates boost decentralized FL convergence 4.94x","Divergence-driven layer rates speed decentralized federated learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000982,"raw_usage":{"total_tokens":4203,"prompt_tokens":1018,"completion_tokens":3185,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":3095}},"tokens_in":634,"tokens_out":3185,"duration_ms":23146,"temperature":1.0,"reasoning_tokens":3095,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:23:34.838033+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FedA2L with the paper's default hyperparameters on a held-out architecture/dataset pair (for instance, a small transformer on CIFAR-100 with $\\alpha=0.1$) and compare rounds-to-target against the best scheduler baseline; if FedA2L fails to converge or loses the speedup, the transferability premise is refuted. A sharper version sweeps $R_{\\mathrm{warm}}$ over $\\{10,40\\}$ on that pair to see whether the bound reported in Table 6 (175 to 462 rounds on ResNet-18) reproduces on a second architecture.","supporting_citations":[{"cited_title":"Federated optimization in heterogeneous networks.Proceedings of Machine learning and systems, 2:429–450, 2020","cited_arxiv_id":null,"evidence_quote":"FedProx provides the client-drift regularized baseline; the paper shows FedA2L targets an orthogonal dimension."},{"cited_title":"Optimizing personalized federated learning through adaptive layer- wise learning","cited_arxiv_id":null,"evidence_quote":"FLAYER establishes that layer-wise treatment matters in personalized FL, motivating FedA2L's layer-granular approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"D-PSGD supplies the decentralized SGD convergence theory that FedA2L's theorems extend."},{"cited_title":"FedAWA: Adaptive optimization of aggregation weights in federated learning using client vectors","cited_arxiv_id":null,"evidence_quote":"FedAWA uses adaptive aggregation weights; the paper shows layer-wise LR adaptation is complementary to it."},{"cited_title":"Federatedlearningvia consensus mechanism on heterogeneous data: A new perspective on convergence","cited_arxiv_id":null,"evidence_quote":"The consensus-mechanism work motivates FedA2L's aggregation-stability metric."},{"cited_title":"Improving the model consistency of decentralized federated learning","cited_arxiv_id":null,"evidence_quote":"DFedSAM is a sharpness-aware baseline that FedA2L also accelerates on in the experiments."}],"review_version":1}