{"id":"4ef7ab84-ed36-4c2d-86b5-3c466f9cd78b","arxiv_id":"2507.12908","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Fremer forecasts cloud workloads by aligning frequency spectra via linear padding, filtering noise, and attending over frequency combinations.","lead":"Fremer is a frequency-domain Transformer for cloud workload forecasting that combines learnable linear padding, frequency filters, and complex-valued attention. It claims accuracy gains over existing models at lower compute and releases four new ByteDance workload datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline SOTA gains (5.5/4.7/8.6 in abstract) cannot be reproduced from the controlled workload comparison (§5.2: 2.9/2.5/3.5), and appear to rely on TFB baseline numbers copied from another environment; the SOTA claim hinges on this mismatch.","rationale":"The reader correctly identified baseline asymmetry and the use of externally referenced TFB numbers as a fairness risk. My read agrees but sharpens the point: the abstract's headline improvements are not even internally consistent with the paper's own controlled workload table, so the central claim currently rests on a measurement that is not reproducible from the reported experiments. This is a load-bearing reporting issue, not a design flaw: the proposed architecture is plausible, the model often wins on the controlled datasets, and the released code and datasets make a fair re-benchmark feasible. The concern could be settled by recomputing the headline numbers under one protocol, so it does not by itself warrant rejection; it does require a correction or explicit re-derivation before the SOTA claim is accepted. Hence the verdict remains conditional, with no change from the reader's assessment.","tokens_in":22277,"tokens_out":9068,"duration_ms":98516,"concrete_test":"Run one controlled re-benchmark: on the seven workload datasets plus Traffic, Electricity, and PEMS04, fix a single input length, normalization, and horizon protocol; run Fremer and all baselines (at minimum PatchTST, iTransformer, PDF, FITS, and FEDformer) under identical code paths with matched hyper-parameter search budgets; and recompute average relative improvements over the best baseline per dataset. As a first step, also recompute the average gains from Table 3 alone and from Table 7 alone to identify which table actually produces the abstract's 5.5/4.7/8.6 values. If the controlled numbers match Section 5.2's 2.9/2.5/3.5 rather than the abstract values, the SOTA claim must be revised to the controlled comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Fremer \"consistently outperforms baseline models, achieving average improvements of 5.5% in MSE, 4.7% in MAE, and 8.6% in SMAPE over SOTA models\" (abstract). The only controlled workload comparison, Table 3, does not support those numbers: Section 5.2 reports average gains of 2.9% MSE, 2.5% MAE, and 3.5% SMAPE against the best baseline per dataset, and recomputing from Table 3 gives roughly 2.9/2.7/3.3, not 5.5/4.7/8.6. The larger abstract figures must therefore come from Table 7, whose baseline columns are \"directly referenced from the TFB paper\" (Section 5.6) rather than measured under Fremer's protocol. That comparison is uncontrolled in a way that directly affects the claimed superiority: TFB numbers were produced with TFB's own input lengths, normalization, and default hyperparameters, while Section 5.1 states Fremer underwent its own hyper-parameter search, and the input length used for Fremer in Table 7 is not specified. Because the abstract's headline improvements are not traceable to a single controlled protocol, the \"surpassing all SOTA models\" claim is not currently established, even though Fremer may well be competitive and efficient.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Fremer, an encoder-only Transformer that forecasts in the frequency domain. It pads the input with a learnable linear layer (LLP) to align the DFT frequency grids of input and complete series, applies low- and high-pass frequency filters, and uses a complex-valued spectrum attention (CSA) mechanism over frequency combinations. The model is evaluated on four new ByteDance workload datasets, three Materna workload traces, and eleven TFB general-forecasting datasets, with additional efficiency, ablation, transfer, and Kubernetes HPA experiments. The authors claim average improvements of 5.5% MSE, 4.7% MAE, and 8.6% SMAPE over SOTA models, while using 0.57M parameters and lower computational cost.","tokens_in":22592,"tokens_out":7303,"duration_ms":78216,"significance":"If the claims are established, Fremer would be a valuable accuracy-efficiency tradeoff for cloud workload forecasting, and the open-sourced ByteDance datasets would be a useful community resource. The paper also gives a concrete treatment of DFT frequency-resolution misalignment, which is a real and underappreciated issue in frequency-domain forecasting. However, the headline superiority claim is currently not supported by a single controlled protocol: the workload-forecasting results in Table 3 show smaller gains and two datasets where Fremer is not best, while the larger gains in the abstract appear to rely on Table 7 whose baselines are copied from the TFB paper. The core CSA definition also has notational problems that obscure whether the mechanism actually performs attention over frequency combinations.","major_comments":[{"comment":"The abstract claims average improvements of 5.5% in MSE, 4.7% in MAE, and 8.6% in SMAPE over SOTA models, but the controlled workload comparison in §5.2 reports only 2.9%, 2.5%, and 3.5%. Recomputing from Table 3 gives roughly 2.8% MSE, 2.5% MAE, and 3.8% SMAPE against the best baseline per dataset, and Fremer is not the best on MT1 MSE (15.408 vs. iTransformer 15.205) or MT2 MSE (6.543 vs. iTransformer/Crossformer 6.467). The larger abstract numbers must come from Table 7, whose baselines are 'directly referenced from the TFB paper' rather than measured under Fremer's protocol. The authors should either report the headline numbers from the controlled workload experiment or clearly label the abstract claim as referring to the uncontrolled general-forecasting setting.","section":"Abstract and §5.2"},{"comment":"Table 7 cannot support the claim that Fremer 'outperforms all baseline models' on general forecasting. Baseline results are copied from TFB, so they were produced with different normalization, input lengths, hardware, and hyperparameter settings, while §5.1 states that Fremer underwent its own hyperparameter search. This is not a controlled comparison. Moreover, Table 7 itself shows counterexamples to the text's claim of superiority on strongly periodic datasets: on PEMS08, Fremer's MSE is 0.288 versus 0.213 for PatchTST and 0.176 for Crossformer. The authors should rerun the baselines under the identical protocol used for Fremer, or restrict the general-forecasting claims to descriptive observations.","section":"§5.6 and Table 7"},{"comment":"The definition of Complex-valued Spectrum Attention is dimensionally inconsistent as written. F_c is defined as an element of C^{L'}, and Q_h = F_c^T W_Q^h with W_Q^h ∈ C^{L'×l} gives Q_h ∈ C^{1×l}; the same holds for K_h and V_h. Then Q_h K_h^T is a 1×1 matrix, so the softmax is trivial and the 'attention' output is just V_h scaled by 1, with no mixing across frequency combinations. The later statement in §4.4 that Q,K,V ∈ C^{1×L'} has the same problem. The authors should provide the correct tensor dimensions (e.g., whether F_c is a matrix of combinations, or whether Q,K,V are L'×d matrices) and show how the claimed O((L')^2/H) complexity arises. As written, the central architectural mechanism is not well defined.","section":"§4.3, Eqs. (4)-(5)"},{"comment":"There is an internal contradiction about the experimental protocol. §5.1 says 'We implement Fremer and all baseline models utilizing the PyTorch framework' and that 'a uniform configuration is applied across all models,' while §5.6 says all baseline results for general forecasting are 'directly referenced from the TFB paper.' These two statements cannot both describe the same evaluation. The authors should state explicitly which baseline numbers were measured and which were imported, and ensure that the headline claims are based only on the measured comparisons.","section":"§5.1 and §5.6"}],"minor_comments":[{"comment":"There are several typos and formatting errors, including 'Univerisity' in the author affiliation, 'Alighment' in the §4.2 heading, and the malformed table entry '11.6431' in Table 3. A careful proofreading pass is needed.","section":"Global"},{"comment":"The sentence 'The dot-product Attention mechanism has a complexity of O(L)' contradicts the standard O(L^2) complexity and the paper's own later formula; this should be corrected to O(L^2).","section":"§4.4"},{"comment":"The ablation study is reported only as bar charts without numerical values or error bars, making it hard for readers to assess the magnitude of each component's contribution. Reporting the exact numbers or a small table would improve reproducibility.","section":"§5.3.1 / Figure 8"},{"comment":"The Kubernetes HPA experiment uses a single function instance from the FaaS dataset, so the 18.78% latency improvement and 2.35% resource reduction have no statistical characterization. The authors should state whether these results are representative and, ideally, report variability across multiple instances.","section":"§5.7 / Table 8"},{"comment":"The claim of being 'the first to address' frequency misalignment should be softened or supported by a discussion of related spectral interpolation and zero-padding techniques in the signal-processing and forecasting literature.","section":"§4.1"},{"comment":"The efficiency comparison includes only Transformer-based baselines, not all SOTA models such as FITS or DLinear. Since the abstract claims improved efficiency relative to SOTA overall, the table caption or text should clarify that the comparison is restricted to Transformer baselines.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising model and useful datasets, but the headline empirical claim is currently split across two incompatible experimental protocols. I would ask the editor to require the authors to either rerun the TFB baselines under the same protocol or remove the general-forecasting SOTA claim from the abstract. The CSA dimension issue in §4.3 should also be fixed before the paper can be evaluated fairly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Fremer is worth reading, and the model is worth taking seriously. The genuinely new piece is Learnable Linear Padding: instead of FFT-ing the input window and predicting the spectrum of the longer input-plus-forecast window on mismatched frequency grids, Fremer pads the input to L+T with a learnable linear layer before the FFT. That cleanly fixes a resolution-alignment problem that most frequency-domain forecasting papers implicitly ignore. The complex-valued attention over linear combinations of frequency points, with inverted multi-head projection, is also a reasonable design and it delivers real efficiency: about 0.57M parameters, with training and inference times several times faster than PatchTST on the IaaS/RDS tests. The four ByteDance workload datasets are a genuine community resource, and the code and data are released.\n\nThe main experimental claims are only partially supported. On the controlled workload comparison (Table 3), Fremer wins 5 of 7 MSE comparisons, and the average gains against the best baseline per dataset are 2.9% MSE, 2.5% MAE, and 3.5% SMAPE—reasonable but not headline material. The abstract's 5.5/4.7/8.6 figures must come from Table 7, where the baselines are 'directly referenced from the TFB paper' rather than re-run under Fremer's protocol. That is an uncontrolled comparison: different normalization, input lengths, and hyperparameter choices, while Fremer itself got a hyperparameter search focused on the filter threshold. The 'surpassing all SOTA models' claim in the abstract is therefore not established by the evidence presented. I'd also like to see error bars; there's no indication of variance across random seeds.\n\nThese are fixable presentation and protocol problems, not fatal flaws. The model is sensible, the ablation study supports each component (especially the LPF/HPF filters), and the efficiency numbers are reproducible from the released code. I'd send this to a serious venue, but the authors should be asked to either re-run the Table 7 baselines in the same environment or downgrade the abstract claims to match the controlled comparison.","headline":"Worth reading: the learnable padding trick is a genuine new idea and the datasets are useful, but the headline SOTA gains don't survive the controlled comparison and the abstract overclaims.","tokens_in":23114,"tokens_out":3066,"would_cite":true,"duration_ms":33466,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that a frequency-domain Transformer that aligns input and target spectra with learnable padding, attends over frequency combinations, and filters noise outperforms all tested forecasting models on cloud workload…","keywords":["workload forecasting","frequency domain","transformer","cloud computing","time series forecasting","frequency resolution alignment","complex-valued attention","auto-scaling"],"falsifier":"Re-run the full benchmark with every baseline hyper-parameter-tuned on the same validation split, using the same normalization, hardware, and training configuration, and re-run the general-forecasting comparisons in the same environment; if Fremer's average MSE advantage over the best baseline falls to near zero or reverses, the paper's central claim of consistent superiority fails.","tokens_in":22067,"feed_emoji":"📈","tokens_out":4819,"duration_ms":54610,"temperature":0.7,"pith_summary":"Fremer is a new forecasting model designed for cloud workloads with strong multi-period patterns. The paper's central claim is that working in the frequency domain—predicting the spectrum of the complete input-plus-horizon series—makes forecasting both more accurate and computationally cheaper than time-domain Transformers. The authors report average gains of 5.5% in MSE, 4.7% in MAE, and 8.6% in SMAPE over the best prior models across their datasets and public benchmarks, with large reductions in parameters and compute. They also release four workload datasets from their industrial cloud services and show that in a Kubernetes proactive auto-scaling test the model cuts average latency by 18.78%.","feed_headline":"Frequency transformer trims cloud workload forecast error by 8.6%","feed_subtitle":"Lightweight model also cuts Kubernetes autoscaling latency 18.78% with far fewer parameters.","key_machinery":"The core mechanism is Learnable Linear Padding, which pads an input series of length L to length L+T with a learned affine map so that the frequency bins of the input spectrum align with those of the complete input-plus-horizon spectrum. Around this, Fremer builds three other components: Complex-valued Spectrum Attention, which treats single frequency points as semantically weak and computes multi-head attention over linear combinations of frequencies, reducing complexity to O((L')^2/H); Frequency Filters, a low-pass filter that retains trend information while preventing overfitting by keeping it outside the backbone's input, and a high-pass filter that removes the highest-frequency noise; and Frequency Reversible Instance Norm, an adaptation of reversible instance normalization to the frequency domain that makes spectra with different global features comparable.","core_discovery":"Fremer's central claim is that forecasting in the frequency domain, applied to the spectrum of the input padded to the full forecast horizon, is both more accurate and more efficient than existing time-domain and frequency-domain approaches for cloud workload series. The paper identifies a previously unaddressed problem called frequency resolution misalignment: because the discrete Fourier transform samples frequencies at intervals of 1/L for input length L, the true periodic frequencies of the complete series are often not present in the input spectrum, so matching input and target spectra is inherently difficult. Fremer solves this with a learnable linear layer that pads the input to length L+T, aligns the frequency grids, then applies complex-valued attention over combinations of frequencies, plus low-pass and high-pass filters that retain trend and discard noise. On seven workload datasets and several public benchmarks, the authors report consistent improvements over state-of-the-art models, with average reductions of 5.5% in MSE, 4.7% in MAE, and 8.6% in SMAPE, while cutting parameter counts by roughly an order of magnitude and improving inference throughput.","pith_inferences":["Editorial extension: the frequency-resolution misalignment problem likely affects other frequency-domain forecasters that compare input and output spectra at different resolutions; the learnable linear padding idea could be lifted and inserted into those models as a general module.","Editorial extension: the finding that attention over frequency combinations works better than attention over individual frequency points suggests that harmonic structure, not individual peaks, is the right inductive bias for periodic workloads; this could be tested on non-cloud series with chirp-like or drifting periodicities.","Editorial extension: the Kubernetes test uses a single function instance under replay, so the claimed 18.78% latency gain may not directly scale to mixed multi-tenant clusters; replicating the test across the full FaaS dataset and multiple instance types would tell how robust the benefit is.","Editorial extension: because the low-pass filter threshold strongly influences overfitting, the threshold could be adapted per series or per dataset using validation loss, turning Fremer into a self-regularizing forecaster without extra training cost."],"forward_implications":["If the central claim holds, Fremer offers a better accuracy-efficiency tradeoff than existing Transformer forecasters on cloud workload series, making it feasible for large-scale deployments where hundreds of thousands of forecasts run per hour.","Channel-independent frequency attention transfers to unseen instances and new datasets, so the model can be used as a zero-shot forecasting backbone without retraining per instance.","In a Kubernetes proactive auto-scaling setup, using Fremer's forecasts reduces average latency by 18.78% and average pod consumption by 2.35% compared with a PatchTST-based autoscaler.","On general benchmarks with strong periodicity (Traffic, Electricity, PEMS04), Fremer outperforms baselines, suggesting the frequency-domain design generalizes beyond cloud workloads.","The frequency filters provide a simple, tunable mechanism to control overfitting: the low-pass threshold directly trades training fit against validation and test error."],"supporting_citations":[{"why":"Supplies the evaluation platform, the default baseline settings for all compared models, and the general-forecasting results that Section 5.6 references directly.","marker":"[39]"},{"why":"Provides the main frequency-domain Transformer baseline (FEDformer) and the Frequency Enhanced Attention mechanism that Complex-valued Spectrum Attention is contrasted against.","marker":"[61]"},{"why":"Establishes the channel-independent design principle that Fremer adopts and is the strongest Transformer baseline in the efficiency comparison.","marker":"[35]"},{"why":"Introduces reversible instance normalization, which Fremer adapts into Frequency Reversible Instance Norm for normalizing spectra.","marker":"[18]"},{"why":"Documents the low-frequency bias and overfitting tendency of frequency-domain Transformers, motivating Fremer's separate handling of low-frequency trend and high-frequency noise.","marker":"[38]"},{"why":"Provides a compact frequency-domain baseline with few parameters, used in the accuracy and efficiency comparisons.","marker":"[53]"}],"fun_headline_variants":["New frequency transformer beats all SOTA on cloud workload forecasting","Fremer: frequency-domain transformer cuts cloud forecast error by 8.6%","Lightweight frequency transformer: 10x fewer params, 8.6% better SMAPE","Frequency transformer boosts cloud autoscaling: 18.78% lower latency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparison against baselines is fair: Fremer gets a limited hyper-parameter search focused on the frequency filter threshold, whereas baselines run with default settings from the benchmark platform, and the general-forecasting results are taken directly from the platform's paper rather than re-run under identical conditions. If that premise fails, the claim of consistent superiority over all state-of-the-art models is weakened, though Fremer may still be competitive and efficient.","fun_headline_variants_meta":{"raw":{"variants":["New frequency transformer beats all SOTA on cloud workload forecasting","Fremer: frequency-domain transformer cuts cloud forecast error by 8.6%","Lightweight frequency transformer: 10x fewer params, 8.6% better SMAPE","Frequency transformer boosts cloud autoscaling: 18.78% lower latency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000449,"raw_usage":{"total_tokens":2300,"prompt_tokens":1019,"completion_tokens":1281,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":1197}},"tokens_in":635,"tokens_out":1281,"duration_ms":14411,"temperature":1.0,"reasoning_tokens":1197,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:35:16.398791+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full benchmark with every baseline hyper-parameter-tuned on the same validation split, using the same normalization, hardware, and training configuration, and re-run the general-forecasting comparisons in the same environment; if Fremer's average MSE advantage over the best baseline falls to near zero or reverses, the paper's central claim of consistent superiority fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the channel-independent design principle that Fremer adopts and is the strongest Transformer baseline in the efficiency comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the low-frequency bias and overfitting tendency of frequency-domain Transformers, motivating Fremer's separate handling of low-frequency trend and high-frequency noise."}],"review_version":1}