{"id":"108c910b-1da7-4ae8-8566-04391f3bd8f0","arxiv_id":"2412.19989","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Caesar lowers federated learning communication costs by tailoring model and gradient compression ratios to device staleness and data importance, with only 0.68% accuracy loss.","lead":"Caesar is a federated learning system that adjusts how much to compress the global model and each device's gradient based on device staleness and local data importance, and also tunes per-device batch sizes to cut idle waiting. In tests on 120 physical devices, it cut communication traffic by roughly 25 to 38 percent while losing only 0.68 percent in final accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5)'s uniform-Φ0 assumption is violated in the paper's own non-IID and real-world OPPO-TS experiments, so the importance ranking may over-compress rare-class gradients and the traffic savings may not be robust.","rationale":"The reader's verdict is CONDITIONAL, and I agree with the identified weakest assumption. The device-importance formula is the hinge of the gradient-compression component, and the paper's own datasets violate its uniform-distribution premise. A wrong importance ranking directly undermines the mechanism by which Caesar claims to save traffic: important gradients are over-compressed, accuracy falls, and more rounds are needed, inflating traffic. The proposed check can falsify or validate this in one clean re-run. The paper has genuine strengths — physical testbeds, a clear ablation study, and consistent results across four datasets — so rejection is not warranted. However, sensitivity to the uniform assumption should be a stated condition for acceptance, which the current CONDITIONAL verdict already reflects.","tokens_in":28887,"tokens_out":5213,"duration_ms":55008,"concrete_test":"Re-run the OPPO-TS experiment (Table 3, row 4) with Eq. (4) computed against the empirical global label distribution Φ0 = Σ_i n_i Φ_i / Σ_i n_i instead of the uniform vector, keeping Eqs. (5)-(6) unchanged. Compare the traffic cost at the same AUC target of 0.65 and record the Spearman rank correlation between the two importance orderings. If the traffic cost changes by more than ~5% or the rank correlation is below 0.8, the uniform-Φ0 assumption is load-bearing for the headline traffic-savings claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The device-importance model in Eqs. (4)-(6) is the load-bearing component of Caesar's gradient-compression mechanism. Eq. (4) computes KL divergence to a fixed uniform distribution Φ0 = 1/K, and Eq. (5) turns this into a scalar C_i that directly controls the gradient compression ratio via Eq. (6). The paper asserts 'the global data distribution is typically uniform' without justification. In the evaluations, this assumption is violated: the Dirichlet splits (Section 6.1) produce realized global distributions that are not exactly uniform, and the OPPO-TS dataset is real user data with no reason to be uniform. Under a non-uniform global distribution, a device whose local data is concentrated on a rare-but-critical class will have large KL to uniform, hence small e^{-KL}, hence low importance; Eq. (6) then assigns it a high compression ratio. The rare-class gradients are aggressively compressed, which can suppress exactly the information needed to maintain test accuracy on tail classes. Because the traffic claim is measured at a fixed target accuracy (Table 3), any accuracy loss caused by this mis-ranking forces additional communication rounds, eroding the claimed 25.54%–37.88% traffic reduction. The paper reports no error bars, no seed variation, and no sensitivity analysis over λ or the choice of Φ0, so the empirical results do not currently rule out this failure mode.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Caesar is a federated learning framework that adaptively compresses both downlink global-model updates and uplink local gradients. The server sets each device's model compression ratio from its participation staleness (Eq. 3), compresses via a hybrid Top-K sparsification plus 1-bit quantization scheme with local-model-based recovery, and computes a device importance score from sample volume and KL divergence to a uniform reference distribution (Eqs. 4-5), converting importance ranks into gradient compression ratios (Eq. 6). A greedy batch-size rule (Eq. 9) attempts to equalize per-round completion times. Experiments on 80 NVIDIA Jetson devices and 40 OPPO smartphones across four datasets report 23.94%-65.06% traffic reductions relative to FedAvg, FlexCom, ProWD, and PyramidFL at a common target accuracy, with only a 0.68% final-accuracy loss relative to full-precision communication.","tokens_in":29206,"tokens_out":6407,"duration_ms":66477,"significance":"The core idea of Caesar—tying compression fidelity to staleness and data quality rather than only to device capability—is timely and likely useful for cross-device FL systems. The paper's strengths include physical implementation on two hardware platforms, a wrapper design that is method-agnostic to the underlying gradient compression operator, and the absence of fitted constants for the headline traffic claims: λ=0.5 is fixed by default and the compression bounds [0.1, 0.6] are adopted from prior work. If the reported traffic reductions reproduce under multi-seed evaluation and under a data-dependent global-distribution estimate, this would be a solid systems contribution. The current evidence is limited by the evaluation methodology and by an unsupported uniform-global-distribution assumption in the importance model.","major_comments":[{"comment":"The assertion that \"the global data distribution is typically uniform\" (p_h^0 = 1/K) is load-bearing for Caesar's gradient compression, and it is contradicted by the paper's own evaluation settings. In the Dirichlet-based splits (§6.1), the aggregate realized data distribution is only approximately uniform for large N, and in OPPO-TS the data are assigned by real user IDs with no reason to be uniform. Under a non-uniform global distribution, a device with a large volume of rare-class data can receive a low importance score because KL(Φ_i || uniform) is large, and Eq. (6) then assigns it a high compression ratio. Since Table 3 measures traffic at a fixed target accuracy, any accuracy loss from over-compressing such gradients can force extra rounds and erode the claimed 25.54%-37.88% traffic reduction. I request a sensitivity analysis over λ and over the reference distribution Φ0 (e.g., a server-side estimate from reported label frequencies), and at least one experiment with a deliberately skewed global class distribution.","section":"§4.2, Eqs. (4)-(5)"},{"comment":"The evaluation reports single runs with no random seeds, confidence intervals, or error bars. The target accuracy/AUC in Table 3 is described as \"the highest achievable by all schemes,\" which appears to be chosen after observing the runs; comparing at a post hoc ceiling can favor the scheme that degrades least at that point and can misstate the traffic-to-accuracy ratio. Please prespecify target accuracies before running, report results across at least three seeds with variance, and show traffic versus a range of target accuracies rather than a single point. Without this, the precise percentages in the abstract and Table 3 are not yet convincing.","section":"Table 3 and §6.1-§6.2"},{"comment":"The ablation study does not cleanly support the stated contributions. Caesar takes 8,615s and 115.57GB to reach 80% on CIFAR-10; Caesar-BR (no deviation-aware compression) takes 12,043s and 228.31GB; Caesar-DC (no batch-size regulation) takes 17,818s and 147.59GB. The text says adaptive batch-size regulation gives a 2.09× speedup, but 17,818/8,615 ≈ 2.07, and says deviation-aware compression gives a 1.39× speedup, but 12,043/8,615 ≈ 1.40; the traffic-reduction contributions (21.69% vs 49.38%) are computed against different baselines. The conclusion that deviation-aware compression \"plays a more critical role\" is therefore not cleanly established. Please provide a full 2×2 ablation and report all ratios against the same full-Caesar baseline.","section":"§6.4"}],"minor_comments":[{"comment":"Typo: \"gradiens\" should be \"gradients\" in the overview sentence.","section":"§3"},{"comment":"Typo: \"dose not exceeds\" should be \"does not exceed\" in the explanation following Eq. (9).","section":"§4.3"},{"comment":"The number of staleness clusters E is a tunable complexity-accuracy knob but is never specified in the experiments; please report the value used or a sensitivity study.","section":"§4.1"},{"comment":"The axis label and legend in Figure 1(b) are partly garbled in the PDF rendering; please regenerate the figure.","section":"Figure 1"},{"comment":"Minor capitalization issue: \"Meanwhile, We impose\" should be \"Meanwhile, we impose.\"","section":"§7"},{"comment":"The manuscript does not mention code release; given the 3k-line implementation, providing an artifact would substantially strengthen reproducibility.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"This is a competent systems paper with a plausible mechanism and a genuine hardware prototype; the requested revisions are within scope. I would not reject on the uniform-distribution concern alone, but the quantitative headline claims need multi-seed evidence and a robustness test under a non-uniform reference distribution. The post hoc target selection in Table 3 is the most serious methodological issue and should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real systems paper with a coherent story and serious hardware evaluation, but the evidence is not yet tight enough to pin down the headline traffic numbers. The genuinely new part is the combination of staleness-aware model download compression, data-property-aware gradient upload compression, and batch-size tuning. Individually the pieces have antecedents—PyramidFL does rank-based compression, MergeSFL does batch-size regulation—but the integration is clean and the paper makes a reasonable case for why the pieces reinforce each other.\n\nWhat it does well: the preliminary experiments are instructive, the implementation on 40 smartphones and 80 Jetson devices is substantial, and the evaluation covers four datasets, including an industrial OPPO-TS, plus ablations and a simulated scale-up. The model-recovery trick—using the local model to reconstruct 1-bit parameters—is neat and clearly explained. The paper cites the relevant prior work and doesn't oversell its novelty too hard.\n\nSoft spots, roughly in order of importance. First, no error bars and no multiple seeds anywhere. On physical testbeds that is expensive, but the headline 25–38% traffic saving is a single-run number, and the accuracy differences between schemes are small enough that noise could matter. A few seeds on CIFAR-10 would materially strengthen the work. Second, the target accuracy in Table 3 is chosen post hoc as the highest all schemes can achieve. The learning curves in Figures 5 and 6 partly mitigate this, but a pre-registered target or a sweep over targets would be cleaner. Third, the importance model in Eqs. (4)–(5) assumes the global data distribution is uniform. That is exactly true for CIFAR-10, roughly true for HAR and Speech, but likely false for OPPO-TS. The worry that rare-class devices are over-compressed is plausible, and the paper gives no sensitivity analysis over λ or the choice of Φ0. This does not sink the paper, because the empirical results may still hold, but it is a genuine analytical gap. Fourth, no code or data release, which makes the system hard to reproduce.\n\nOverall, the central argument—deviation-aware compression gives most of the accuracy at a fraction of the traffic—holds up in spirit. The numbers in the figures and Table 3 are internally consistent, and the ablation shows both components matter. I would send this to a serious referee. The fixes are straightforward: add seed variation, be transparent about target selection, and test the importance model's sensitivity.","headline":"A solid FL systems paper with a clean combination of compression heuristics and real-hardware evaluation; the traffic-savings claim is plausible but needs error bars and sensitivity analysis on the importance model.","tokens_in":29717,"tokens_out":4266,"would_cite":true,"duration_ms":42290,"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":"Caesar cuts federated-learning traffic by up to 38% at same accuracy","keywords":["federated learning","gradient compression","model compression","data heterogeneity","model staleness","communication efficiency","batch size optimization","device importance"],"falsifier":"Run Caesar on a dataset with a deliberately skewed global label distribution, so the true $\\Phi_0$ is far from uniform, and compare against the same system using the true distribution in the KL term; if the uniform-assumption version over-compresses high-contribution devices, its final accuracy or traffic savings should visibly degrade.","tokens_in":28703,"feed_emoji":"📉","tokens_out":6941,"duration_ms":62448,"temperature":0.7,"pith_summary":"Caesar is a federated learning framework built on a simple claim: the damage done by compression comes from model and gradient deviation, so compression ratios should be device-specific rather than fixed or capability-based. For each device, it sets the download compression ratio from the staleness of the device's local model and the upload ratio from the importance of the device's local data, then tunes batch sizes so faster devices do not wait idly. The paper reports that this reduces traffic by roughly 25.54% to 37.88% compared with compression-based baselines at the same target accuracy, while losing only 0.68% in final accuracy against full-precision communication. That matters because communication, not computation, is the usual bottleneck in mobile federated learning.","feed_headline":"Caesar cuts federated-learning traffic by up to 38% at same accuracy","feed_subtitle":"Staleness-aware model downloads and importance-aware gradient uploads keep accuracy loss to 0.68%.","key_machinery":"The carrying mechanism is a pair of compression-ratio policies. For the downlink, the model compression ratio for device $i$ in round $t$ is $\\rho_{M,i}^t=(1-\\delta_i^t/t)\\rho_M^{\\max}$, where $\\delta_i^t$ is the number of rounds since that device last participated; the server then applies Top-K sparsification plus 1-bit quantization, and the device recovers the 1-bit entries using its local model, replacing sign-mismatched or out-of-range values with the average absolute value. For the uplink, device importance is $C_i=\\lambda n_i/n_{\\max}+(1-\\lambda)e^{-\\mathrm{KL}(\\Phi_i\\|\\Phi_0)}$ with $\\lambda=0.5$ and a uniform global distribution $\\Phi_0$, and a rank-based rule assigns smaller gradient compression ratios to higher-importance devices. A third mechanism, a greedy batch-size assignment, keeps each device's round time near the fastest device so the synchronous barrier does not waste time.","core_discovery":"The paper's central claim is that a low-deviation compression schedule can make federated learning nearly as accurate as uncompressed training while spending much less traffic. It argues that two sources of error dominate: devices with outdated local models need a faithful copy of the global model, and devices whose local gradients are most aligned with the global objective need faithful uploads. Caesar therefore gives stale devices smaller model compression ratios, gives data-rich and distributionally aligned devices smaller gradient compression ratios, and uses a greedy batch-size rule to equalize round durations. The measured consequence is that Caesar reaches a given target accuracy with 23.94% to 65.06% less traffic than four baselines across the tested datasets, keeping the final accuracy gap to full-precision communication at 0.68%.","pith_inferences":["If the global label distribution is not uniform, the fixed uniform $\\Phi_0$ in the KL term could mis-rank devices; estimating it from the server's aggregated view or from recent participants would be a natural, testable correction.","The importance values are computed once before training; in federated settings where device populations or data distributions drift, periodic recomputation of importance may be needed to keep the rankings valid.","Because the savings come from ratios applied to whole models, the method should benefit larger models even more; testing on a modern transformer-scale model would show whether the trend holds.","The importance ranking could also be reused by a participant-selection strategy, choosing the most important devices more often would likely reduce rounds further, though this is not what the paper itself claims."],"forward_implications":["At the target accuracy, federated deployments can spend roughly a quarter to a third less traffic than compression-based baselines, and up to about 65% less than individual baselines in the reported settings.","Frequently participating devices can accept high model compression ratios without hurting training, while rarely seen devices receive more faithful models, reducing bad updates from model obsolescence.","Devices with balanced label distributions and large sample volumes retain more gradient information, so the global model keeps learning the knowledge those devices carry.","Batch-size regulation shortens the synchronized round, so faster devices spend less time idle and the wall-clock time to target accuracy drops.","Because the compression schedule is method-agnostic, the same deviation-aware ratios can be applied on top of sparsification, quantization, or tensor-decomposition compressors."],"supporting_citations":[{"why":"supplies the federated averaging protocol and the uncompressed baseline against which Caesar measures accuracy loss.","marker":"[5]"},{"why":"supplies the rank-based gradient compression-ratio strategy that Caesar adapts, and acts as a comparison baseline.","marker":"[36]"},{"why":"supplies a capability-aware compression baseline that assigns ratios by network condition, the main alternative Caesar improves on.","marker":"[25]"},{"why":"supplies a quantization-based baseline whose bit width is set by bandwidth, another capability-aware comparison.","marker":"[51]"},{"why":"supplies Top-K sparsification, the concrete compressor Caesar uses for models and gradients, and a prior adaptive compression scheme.","marker":"[26]"},{"why":"supplies the convergence foundation for Top-K sparsification that underlies the compression component.","marker":"[38]"},{"why":"supplies the 1-bit quantization idea used in the model download compression and recovery step.","marker":"[24]"},{"why":"supplies the batch-size regulation insight and the heterogeneous data and latency setup used to evaluate the system.","marker":"[39]"},{"why":"supports the premise that devices with balanced labels and larger sample volume carry gradients crucial for convergence.","marker":"[30]"}],"fun_headline_variants":["Caesar cuts FL traffic up to 38%, accuracy loss just 0.68%","Low-deviation compression: Caesar trims FL traffic by 38%","Staleness-aware model downloads and importance-based uploads cut FL traffic 38%","Caesar achieves 38% traffic cut with only 0.68% accuracy loss in FL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Eq. (5) correctly ranks devices by their contribution to global convergence using only sample volume and KL divergence from an assumed uniform global data distribution; if that ranking is wrong, important gradients get compressed too hard and the reported accuracy-traffic tradeoff weakens.","fun_headline_variants_meta":{"raw":{"variants":["Caesar cuts FL traffic up to 38%, accuracy loss just 0.68%","Low-deviation compression: Caesar trims FL traffic by 38%","Staleness-aware model downloads and importance-based uploads cut FL traffic 38%","Caesar achieves 38% traffic cut with only 0.68% accuracy loss in FL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000651,"raw_usage":{"total_tokens":2991,"prompt_tokens":953,"completion_tokens":2038,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":1947}},"tokens_in":569,"tokens_out":2038,"duration_ms":15139,"temperature":1.0,"reasoning_tokens":1947,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:41:00.583016+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Caesar on a dataset with a deliberately skewed global label distribution, so the true $\\Phi_0$ is far from uniform, and compare against the same system using the true distribution in the KL term; if the uniform-assumption version over-compresses high-contribution devices, its final accuracy or traffic savings should visibly degrade.","supporting_citations":[{"cited_title":"signsgd: Compressed optimisation for non- convex problems","cited_arxiv_id":null,"evidence_quote":"supplies the 1-bit quantization idea used in the model download compression and recovery step."},{"cited_title":"Communication-efficient learning of deep net- works from decentralized data","cited_arxiv_id":null,"evidence_quote":"supplies the federated averaging protocol and the uncompressed baseline against which Caesar measures accuracy loss."},{"cited_title":"Pyramidfl: A fine-grained client selection framework for efficient federated learn- ing","cited_arxiv_id":null,"evidence_quote":"supplies the rank-based gradient compression-ratio strategy that Caesar adapts, and acts as a comparison baseline."},{"cited_title":"To talk or to work: Flexible communication compression for en- ergy efficient federated learning over heterogeneous mobile edge devices","cited_arxiv_id":null,"evidence_quote":"supplies a capability-aware compression baseline that assigns ratios by network condition, the main alternative Caesar improves on."},{"cited_title":"Bitwidth heterogeneous federated learning with progressive weight dequantization","cited_arxiv_id":null,"evidence_quote":"supplies a quantization-based baseline whose bit width is set by bandwidth, another capability-aware comparison."},{"cited_title":"Adaptive control of local updating and model compres- sion for efficient federated learning","cited_arxiv_id":null,"evidence_quote":"supplies Top-K sparsification, the concrete compressor Caesar uses for models and gradients, and a prior adaptive compression scheme."},{"cited_title":"The convergence of sparsi- fied gradient methods","cited_arxiv_id":null,"evidence_quote":"supplies the convergence foundation for Top-K sparsification that underlies the compression component."},{"cited_title":"Mergesfl: Split federated learning with feature merging and batch size regulation","cited_arxiv_id":null,"evidence_quote":"supplies the batch-size regulation insight and the heterogeneous data and latency setup used to evaluate the system."},{"cited_title":"FedDisco: Federated Learning with Discrepancy-Aware Collaboration","cited_arxiv_id":"2305.19229","evidence_quote":"supports the premise that devices with balanced labels and larger sample volume carry gradients crucial for convergence."}],"review_version":1}