{"id":"bdf1144f-ea2c-456b-baf7-33bf1f951bc2","arxiv_id":"2501.19065","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"BEAT adaptively scales gradients of frequency-specific networks during training to balance learning speeds, with reported gains on some long-term forecasting benchmarks.","lead":"BEAT is a training-time framework that reweights per-frequency gradients for wavelet-based time-series forecasters. The paper reports benchmark gains, but the gains are not consistent across all datasets and the core mechanism is described inconsistently.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Table 3 contradicts the central claim: BEAT loses Traffic to iTransformer on every horizon, and the reported first-place counts (28 MSE, 30 MAE) exceed the 28 cells per metric available.","rationale":"The strongest claim is empirical, so the decisive issue is whether Table 3 actually shows consistent SOTA outperformance. It does not: Traffic favors iTransformer on all horizons, Weather 336 has a TimeMixer MSE win, and ECL 96 has an iTransformer MSE win. The '1st Count' row is internally impossible, listing 30 MAE firsts when only 28 MAE cells exist, and 28 MSE firsts despite visible counterexamples. This is not a question of outside consensus or missing code; the paper's own reported numbers contradict its headline. The reader's identified Eq. 7 sign issue is a reasonable secondary concern about the mechanism, but the empirical contradiction alone is sufficient to reject the central claim. Because the reader already recommended rejection and this concern reinforces that conclusion, the verdict remains unchanged.","tokens_in":13415,"tokens_out":4721,"duration_ms":41508,"concrete_test":"Recompute every cell of Table 3 and count first places per metric by scanning all 28 MSE and 28 MAE entries across the seven datasets and four horizons. If the audit reproduces non-BEAT best cells on Traffic, Weather 336, or ECL 96, or if any metric count exceeds 28, then the printed '1st Count' row and the 'consistently outperforms' claim are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that BEAT 'consistently outperforms state-of-the-art approaches' and 'achieves the highest number of first-place rankings' (Abstract, Section 4.1). This claim is contradicted by the paper's own Table 3. For Traffic, iTransformer is better than BEAT on MSE and MAE at every prediction length (e.g., horizon 96: 0.395/0.268 vs 0.459/0.278; horizon 720: 0.467/0.302 vs 0.510/0.312). On Weather 336, TimeMixer has lower MSE (0.251 vs 0.258), and on ECL 96, iTransformer has lower MSE (0.148 vs 0.149). The '1st Count' row claims 28 MSE first places and 30 MAE first places for BEAT, but there are only 7 datasets × 4 prediction lengths = 28 cells per metric, so 30 is impossible; moreover, the visible counterexamples show BEAT cannot have all 28 MSE firsts. Since the headline is an empirical superiority claim, these internal inconsistencies are decisive. The mechanism in Eq. 7 may also be mis-specified, but even a corrected mechanism would not rescue the printed results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes BEAT, a training-time gradient modulation scheme for frequency-decomposed time-series forecasting. The method decomposes the target series via multi-level discrete wavelet transform into approximation and detail coefficients, trains a separate WPMixer-style network per coefficient, monitors per-frequency prediction-error ratios, and scales each network's gradient by a coefficient c_v defined in Eq. (7). The authors evaluate on seven datasets (Weather, four ETT variants, Traffic, ECL) for horizons 96/192/336/720 and report MSE/MAE against eight baselines, plus ablations over wavelet type, decomposition level, discrepancy metric, and loss- versus gradient-modulation. The headline claim is that BEAT 'consistently outperforms state-of-the-art approaches' and achieves the highest number of first-place rankings across all prediction lengths and datasets.","tokens_in":13732,"tokens_out":4503,"duration_ms":38942,"significance":"If the mechanism and results were as stated, BEAT would be a simple, inference-free improvement applicable to any frequency-decomposed forecaster, and the study would be a useful contribution to training dynamics for multi-scale forecasting. The paper covers standard benchmarks and includes ablations of several design choices, which is a strength. However, two load-bearing issues undermine the manuscript: Eq. (7) does not implement the described suppression of fast learners, and Table 3's first-place counts are arithmetically impossible and contradicted by the table's own cell values. The central empirical claim is therefore not currently supported.","major_comments":[{"comment":"Equation (7) contradicts the stated intent. For r_v <= 1 the formula gives c_v = 1/r_v >= 1, and for r_v > 1 the sigmoid term is greater than 0.5 plus the offset 0.5, so c_v > 1. Thus the coefficient never falls below 1 and never 'moderates' a fast learner; it amplifies the gradients of all frequency networks, with the largest amplification assigned to the fastest learners (smallest r_v). The prose claims the opposite ('For a network with a faster learning process ... gradients will be decreased'). The formula needs to be corrected, and the prose condition 'where rv > 0' appears to be a typo for 'rv <= 1', but even that branch still amplifies.","section":"Methodology, Dynamical Gradient Balancer"},{"comment":"The '1st Count' row reports 28 MSE and 30 MAE first places for BEAT. Since Table 3 has only 7 datasets x 4 prediction lengths = 28 cells per metric, 30 is impossible; moreover, BEAT does not win all cells. For example, Traffic at horizons 96, 192, 336, and 720 is better for iTransformer (e.g., horizon 96: 0.395/0.268 versus BEAT 0.459/0.278; horizon 720: 0.467/0.302 versus 0.510/0.312); Weather at horizon 336 is better for TimeMixer (0.251 versus 0.258); and ECL at horizon 96 is better for iTransformer (0.148 versus 0.149). These internal inconsistencies invalidate the abstract and Section 4.1 claims of consistent superiority.","section":"Experiments, Table 3"},{"comment":"The comparison reports single-run point estimates without standard deviations, error bars, or significance tests, although many differences are at the third decimal (e.g., ECL horizon 96 MSE 0.149 versus 0.148). Given that the paper's central claim is a ranking claim over baselines, the results need multiple seeds and a significance check; otherwise 'first-place' outcomes may reflect noise rather than a genuine advantage.","section":"Section 4.1, Table 3"}],"minor_comments":[{"comment":"The table reference appears as 'Table??' and should be Table 2.","section":"Ablation Study, Effect of BEAT"},{"comment":"The text says 'nine state-of-the-art baseline methods' but lists eight methods (WPMixer, TimeMixer, iTransformer, TSMixer, PatchTST, Crossformer, TiDE, DLinear).","section":"Experimental Settings, Baselines"},{"comment":"The notation \\bar{g}L(\\theta_v^u) is overloaded; \\bar g_v should be defined as c_v g_v, and the prose 'Gradient Decent' should be 'Gradient Descent'.","section":"Methodology, Equation (8)"},{"comment":"The output order in Eq. (1) lists YD_f, YD_{f-1}, ..., YD_1 while the text later refers to i in {1, ..., f}; the ordering convention should be stated once to avoid ambiguity about which index corresponds to the highest frequency.","section":"Methodology, Equation (1)"}],"recommendation":"reject","confidential_remarks":"The manuscript's own Table 3 makes the central empirical claim untenable, and the gradient modulation equation reverses the described behavior. These are not presentation issues; they invalidate the paper's main conclusions. I would recommend rejection, although a corrected mechanism with corrected experiments could potentially be pursued as new work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is legitimate: apply gradient balancing from multimodal learning to the frequency sub-networks of a wavelet-based forecaster, so that fast-converging components don't overfit while slow ones underfit. That is a new application and a sensible direction. The paper does a reasonable job motivating the problem, and the ablations on wavelet type and decomposition level are the kind of systematic check you want to see.\n\nBut the paper as written does not hold together. The central claim is that BEAT \"consistently outperforms state-of-the-art approaches\" and has the most first-place rankings, and the paper's own Table 3 contradicts that. The 1st Count row says 28 MSE and 30 MAE firsts, but there are only 28 cells per metric (7 datasets × 4 horizons). 30 is impossible. And it isn't a small bookkeeping error: on Traffic, iTransformer beats BEAT on both metrics at every prediction length, and on ECL 96 and Weather 336 other baselines also win. So the headline empirical claim is false in the paper's own numbers.\n\nThe mechanism has a similar problem. Equation (7) is supposed to suppress gradients for fast learners and amplify them for slow learners, but as written it always outputs a coefficient ≥ 1. The r_v ≤ 1 branch gives 1/r_v ≥ 1, and the r_v > 1 branch gives a sigmoid shifted by +0.5, which is also > 1. Nothing is ever suppressed. This is not a minor typo: the explanation of the method in the text and the actual formula describe different behaviors. If the formula is what was run, the method is not what the paper says it is.\n\nWhat else weakens the case: no code is released, there are no error bars or significance tests, and the improvements over WPMixer are tiny (often 0.001–0.003 MSE), easily within run-to-run noise on these benchmarks. The ablation tables show the same small gaps. So even if the mechanism were fixed, the evidence that BEAT helps is thin.\n\nI'd sum it up as a promising idea attached to a manuscript that currently doesn't support its own conclusions. The authors could fix the equation, correct the count, and rerun with seeds, and then the paper might be a reasonable workshop contribution or a solid component of a larger study. But in its present form, it isn't there. I would not send it to a serious referee; I'd desk-reject and invite a corrected resubmission.","headline":"A plausible training-time idea undone by a self-contradictory results table and a gradient rule that does not match its stated intent.","tokens_in":14217,"tokens_out":2522,"would_cite":false,"duration_ms":25471,"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":"BEAT, a training-only gradient-balancing framework for wavelet frequency bands, reports the most first-place results across seven long-term forecasting benchmarks.","keywords":["long-term time-series forecasting","frequency-domain methods","wavelet decomposition","gradient modulation","balanced training","convergence monitoring","adaptive tuning","multivariate forecasting"],"falsifier":"Run BEAT with all modulation coefficients fixed to a constant $c_v = 1$ while keeping the Frequency-Specific Monitor active; if accuracy does not drop, the monitoring signal alone, rather than the gradient balancing, accounts for the gains. A direct check of Eq. 7 also shows that no positive discrepancy ratio $r_v$ produces $c_v < 1$, so the claimed 'moderating gradients for rapid learners' branch is never active, which would indicate that the stated mechanism does not match the implementation.","tokens_in":13243,"feed_emoji":"📈","tokens_out":10904,"duration_ms":83751,"temperature":0.7,"pith_summary":"BEAT (Balanced frEquency Adaptive Tuning) is a training-time framework for frequency-domain long-term time-series forecasting. It claims that training every wavelet frequency branch under one unified loss is inefficient because high-frequency branches converge too fast and low-frequency branches too slowly, so the model overfits one regime while underfitting the other. BEAT monitors each frequency's prediction error relative to the others, converts that discrepancy into a per-branch gradient modulation coefficient, and rescales back-propagated gradients to synchronize learning. If the claim is right, any wavelet-based forecaster can be upgraded at training time, with inference unchanged, and consistently beat state-of-the-art methods on long-horizon benchmarks.","feed_headline":"Gradient balancing between frequency bands tops 7 forecast benchmarks","feed_subtitle":"BEAT slows overfit high-frequency branches, boosts underfit low-frequency ones, and adds no inference cost.","key_machinery":"The load-bearing mechanism is the pair formed by the Frequency-Specific Monitor and the Dynamical Gradient Balancer. The monitor uses wavelet decomposition of the ground-truth target to compute a normalized per-frequency discrepancy ratio $r_v$ (Eq. 5), which is read as a relative learning-speed signal for each branch. The balancer converts $r_v$ into a gradient modulation coefficient $c_v$ via Eq. 7 and applies it in the parameter update as $\\bar{g}_v = c_v g_v$ (Eq. 8). The design intent is that branches with small errors relative to the mean (fast learners) receive damped gradients while branches with large errors (slow learners) receive boosted gradients, bringing all frequency branches into synchronized convergence.","core_discovery":"The central claim is that asynchronous learning across wavelet frequency bands is a measurable, correctable cause of suboptimal long-term forecasts, and that correcting it through training-time gradient modulation yields consistent gains. BEAT decomposes the ground-truth target $Y$ into an approximation coefficient $Y_A$ and detail coefficients $Y_{D_i}$ via discrete wavelet transform, computes per-branch MSE discrepancies $\\delta_A$ and $\\delta_{D_i}$, takes the mean $\\mu$ over the detail coefficients, and defines discrepancy ratios $r_v = \\delta_v / \\mu$ (Eq. 5). The Dynamical Gradient Balancer then multiplies each branch's back-propagated gradient by a coefficient $c_v$ from Eq. 7, with the stated intention of damping branches that have learned quickly and boosting branches that lag. In experiments on seven datasets with prediction horizons $\\{96, 192, 336, 720\\}$, BEAT reports the highest number of first-place MSE and MAE results among WPMixer, TimeMixer, iTransformer, TSMixer, PatchTST, Crossformer, TiDE, and DLinear.","pith_inferences":["Because BEAT is training-only and does not alter the base architecture, the same discrepancy-ratio monitor could in principle be attached to Fourier-based or other multiband forecasters, not just wavelet-based WPMixer.","The monitoring signal uses ground-truth wavelet coefficients only during training, which suggests a natural curriculum: a branch could be frozen or detached once its discrepancy ratio stabilizes, saving compute without changing the architecture.","The reported sensitivity of the best wavelet type and decomposition level to the dataset hints that the monitor itself could be used to select those hyperparameters automatically, something the paper does not explore.","Benchmark reporting that shows only aggregate MSE and MAE may hide frequency-level imbalance; reporting per-frequency losses would make it easier to see whether BEAT's gains come from the low-frequency branch, the high-frequency branch, or both."],"forward_implications":["Frequency-decomposition forecasters trained with BEAT should exhibit more synchronized convergence across wavelet bands, reducing overfitting of high-frequency branches and underfitting of low-frequency branches.","Because BEAT changes only the training loop, the inference-time cost of the base forecaster is unchanged.","The per-frequency discrepancy ratio gives a real-time, training-set-based signal for convergence, overfitting, and underfitting without requiring extra validation data.","On the seven reported benchmarks, BEAT attains more first-place MSE and MAE results than nine state-of-the-art baselines across prediction lengths $\\{96, 192, 336, 720\\}$."],"supporting_citations":[{"why":"Supplies WPMixer, the wavelet-based forecaster BEAT is instantiated on and the primary baseline it improves upon.","marker":"Murad, Aktukmak, and Yilmaz 2025"},{"why":"Provides the multiresolution wavelet decomposition theory used to split targets into approximation and detail coefficients.","marker":"Mallat 1989"},{"why":"Supplies the wavelet families (Daubechies, Coiflets, Biorthogonal, Symlets) evaluated as decomposition bases.","marker":"Daubechies 1992"},{"why":"RevIN instance normalization used before wavelet decomposition and after reconstruction.","marker":"Kim et al. 2021"},{"why":"The on-the-fly gradient modulation approach that FSM and DGB adapt to per-frequency balancing.","marker":"Peng et al. 2022"},{"why":"Provides the ETT datasets (ETTh1, ETTh2, ETTm1, ETTm2) used in the benchmarks.","marker":"Zhou et al. 2021"},{"why":"Provides the ECL electricity-consumption dataset used in the benchmarks.","marker":"Li et al. 2019"},{"why":"The TSLib benchmark codebase used to implement and evaluate BEAT against baselines.","marker":"Wang et al. 2024b"}],"fun_headline_variants":["BEAT tunes gradients per frequency to top seven forecast benchmarks","Per-frequency gradient balance beats long-term forecast rivals","Adaptive frequency gradient balancing wins on seven datasets","BEAT balances frequency learning speeds to top forecasting benchmarks","Gradient tuning per frequency band sets forecast records"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method depends on one assumption: that each frequency band's forecast error relative to the average reliably marks it as a fast or slow learner, and that the gradient-scaling rule then dampens fast learners and boosts slow learners exactly as intended.","fun_headline_variants_meta":{"raw":{"variants":["BEAT tunes gradients per frequency to top seven forecast benchmarks","Per-frequency gradient balance beats long-term forecast rivals","Adaptive frequency gradient balancing wins on seven datasets","BEAT balances frequency learning speeds to top forecasting benchmarks","Gradient tuning per frequency band sets forecast records"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000704,"raw_usage":{"total_tokens":3179,"prompt_tokens":952,"completion_tokens":2227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":2153}},"tokens_in":568,"tokens_out":2227,"duration_ms":14894,"temperature":1.0,"reasoning_tokens":2153,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:26:45.545260+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BEAT with all modulation coefficients fixed to a constant $c_v = 1$ while keeping the Frequency-Specific Monitor active; if accuracy does not drop, the monitoring signal alone, rather than the gradient balancing, accounts for the gains. A direct check of Eq. 7 also shows that no positive discrepancy ratio $r_v$ produces $c_v < 1$, so the claimed 'moderating gradients for rapid learners' branch is never active, which would indicate that the stated mechanism does not match the implementation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies WPMixer, the wavelet-based forecaster BEAT is instantiated on and the primary baseline it improves upon."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multiresolution wavelet decomposition theory used to split targets into approximation and detail coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet families (Daubechies, Coiflets, Biorthogonal, Symlets) evaluated as decomposition bases."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The on-the-fly gradient modulation approach that FSM and DGB adapt to per-frequency balancing."}],"review_version":1}