{"id":"58c34595-8d9f-42e3-8afa-20b4e1f42067","arxiv_id":"2505.05605","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A frequency-adaptive learning rate that slows updates to rare embedding IDs reduces multi-epoch overfitting for some sparse ad-conversion objectives, but not the sparsest, and its advantage disappears after days of continual training.","lead":"Ads models at Pinterest represent advertisers and products with large lookup tables; retraining over the same data makes rare IDs memorize noise and hurt next-day predictions. This paper tests scaling each table row's learning rate by log frequency, finds it helps some sparse objectives but not the sparsest, and shows daily fresh data makes the fix unnecessary.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run comparisons and no confidence intervals make the reported AUC/loss gains (0.07-0.20%) indistinguishable from run-to-run noise.","rationale":"The paper is a well-scoped industrial case study with explicit honesty about limitations: FAL's failure on the sparsest objective, the vanishing of gains under continual training, and the reliance on a private production dataset. The concern I raise is not about internal consistency or novelty but about the empirical support for the headline effect. The paper itself flags AUC noise in Section 4.4, yet the evaluation uses a single run per configuration. Because the effect sizes are fractions of a percent and the sparsest objective is the most noisy, the reported differences could easily be artifacts of initialization or data ordering. The paper does provide one form of independent support: the Sparse Optimizer has been launched to production, and both FAL and MEDA cross an internal +0.10% AUC threshold for online experiments. But the production launch is not a controlled replication, and the threshold does not quantify variance. A repeated-seed study would settle whether FAL's gains are real; if they are, the method is a credible, cheap alternative to MEDA. This matches the reader's verdict, so I recommend no change.","tokens_in":12875,"tokens_out":6584,"duration_ms":71010,"concrete_test":"Re-run the Aug-Dec 2024 2-epoch training for the baseline, FAL, and MEDA configurations with at least 5 different random seeds (varying model initialization and data-shuffle order) while keeping all other hyperparameters fixed. Compute per-objective end-of-epoch test loss and the Section 5.4 cumulative AUC gain for each run, then report means and 95% bootstrap confidence intervals for the paired differences (FAL - baseline, MEDA - baseline). If the intervals for the objectives where FAL is claimed to help (e.g., p(add-to-cart|click)) include zero, the central claim is not supported. A secondary check: also run a global learning-rate-decay control (e.g., multiply all embedding LRs by 1/log(1+max accumulated frequency)) to test whether any residual effect is specific to per-row frequency scaling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that FAL reduces multi-epoch overfitting and matches MEDA except on p(checkout|click)—rests on differences that are small and possibly within run-to-run noise. Section 4.4 states 'AUC values can be very noisy,' yet every result in Figures 5-9 comes from a single training run per configuration, with no confidence intervals, repeated seeds, or significance tests. The cumulative AUC metric (Section 4.4) is a ratio of sums over sequential days, which does not account for autocorrelation, and the claimed gains are 0.07% to 0.20%. Similarly, end-of-second-epoch test-loss differences between FAL and baseline (e.g., 0.23% to 0.83% lower loss on p(add-to-cart|click)) have no error bars. On the sparsest objective, where noise is highest, FAL is reported to be no better than baseline, which is itself an indication that the method's effect does not dominate noise on exactly the metric of interest. Without a noise model, the observed differences cannot be attributed to the method; the conclusion that a per-row log-frequency learning rate mitigates multi-epoch overfitting is therefore empirically unsupported as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports industrial lessons from Pinterest Ads conversion models on two challenges: slow convergence of embedding tables due to gradient sparsity, and multi-epoch overfitting ('one-epoch phenomenon'). It proposes a Sparse Optimizer that applies a higher layer-specific learning rate to embedding tables, and a Frequency-Adaptive Learning Rate (FAL) that scales the learning rate of each embedding row by its relative log frequency. The authors evaluate both methods on two large production datasets (May–Aug 2024 and Aug–Dec 2024), comparing FAL against an embedding-re-initialization baseline (MEDA) and against 1-epoch and 2-epoch baselines. They report faster convergence, reduced test-loss jumps at epoch boundaries for most objectives, cumulative AUC gains of 0.07–0.20% during continual training, and a conclusion that multi-epoch overfitting mitigation may be unnecessary when fresh data is available.","tokens_in":13062,"tokens_out":4879,"duration_ms":58574,"significance":"If the empirical claims hold, the paper makes a useful contribution: FAL is a simple, near-parameter-free per-row learning-rate schedule that could cheaply mitigate multi-epoch overfitting, and the demonstration that overfitting severity tracks label density in a multi-task model is valuable for practitioners. The paper's strengths include its industrial scale (298M parameters, 60 embedding tables, two disjoint datasets), a direct comparison to an external baseline (MEDA), an implementation-ready algorithm, and an unusually honest discussion of when multi-epoch overfitting mitigation is unnecessary. However, the central empirical support is currently weak because all comparisons are based on single training runs with no confidence intervals or significance tests, on a metric the paper itself calls noisy. The promising ideas deserve publication, but only after the statistical support is strengthened or the claims are appropriately softened.","major_comments":[{"comment":"The central empirical claims are supported only by single training runs per configuration. Section 4.4 states that 'AUC values can be very noisy,' yet Figures 5–9 report loss and cumulative AUC point estimates without confidence intervals, repeated seeds, or significance tests. The reported effects are small (loss increases reduced by 0.2–0.5 percentage points; cumulative AUC gains of 0.07–0.20%), and the sparsest objective, where noise is highest, is exactly where FAL fails to beat the baseline. Without a noise model—for example, bootstrap over days for the cumulative metric or repeated runs with different seeds—the observed differences cannot be attributed to the methods rather than to run-to-run variation. Please add repeated runs or at least a variance analysis over the daily evaluation points, and temper statements such as 'significantly higher' in Section 5.1.","section":"§4.4, §5.2–5.3, Figures 5–9"},{"comment":"FAL's log scaling is selected based on a single comparison reported in the first Table 1 ('Log Scaling (Control) 0.00%; Linear Scaling -0.13%'). This data-dependent choice is a design decision, not a fitted parameter, but it means the method's success relies on a hyperparameter selected on the same cumulative AUC metric with no repeated runs. Please include a small sensitivity analysis over the scaling form (e.g., log vs linear vs sqrt) and over the Sparse Optimizer multiplier (50x) used in Section 5.1, or explicitly present the log choice as a heuristic whose uncertainty is not quantified.","section":"§3.2, Table 1, Algorithm 1"},{"comment":"The Sparse Optimizer's advantage over the retrained baseline is reported as a cumulative AUC gain of 0.10% and over the production model as 0.017%, with no significance test. The production comparison is additionally confounded by the production model having four extra months of continual training, which the authors interpret as a disadvantage for the Sparse Optimizer; that is not a controlled comparison. Please state whether these differences are within the day-to-day variability visible in Figure 4 and provide paired statistics or a different evaluation design if the 'significantly higher cumulative AUC' claim is to be maintained.","section":"§5.1, Figure 4"},{"comment":"The cumulative AUC gain is defined as the ratio of sums of daily AUCs, not as an average of per-day gains. This formula does not account for autocorrelation across days and gives no direct estimate of the variance of the cumulative gain; two methods whose daily AUCs differ by a small correlated shift will show a small aggregate difference that could be reproducible, while methods with large daily swings could appear similar. Report per-day differences with standard errors (or a bootstrap over days) and specify the number of evaluation days for each figure. This will also help interpret the 'gains vanish after continual training' conclusion in Section 5.4.","section":"§4.4, Eq. (cumulative AUC gain)"}],"minor_comments":[{"comment":"The label 'Table 1' is used twice: once for the log-versus-linear scaling comparison in Section 3.2 and once for the label-density table in Section 4.1. Please renumber the tables.","section":"§3.2 and §4.1"},{"comment":"The text refers to 'FAD' once in the sentence 'and FAD, which punishes convergence speed'; this should be 'FAL'.","section":"§3.2"},{"comment":"The phrase 'the increase is insignificant' for the 0.13% total-loss jump is not supported by a statistical test and conflicts with the later use of 'significantly higher' in Section 5.1; make the use of 'significant' consistent with the evidence presented.","section":"§5.2"},{"comment":"The x-axis is labeled 'Epoch' but the range is 0.2 to 1.8; clarify where the epoch boundary occurs and whether evaluation points are sampled within epochs or only at epoch boundaries.","section":"Figures 5 and 7"},{"comment":"Algorithm 1 accumulates frequencies over the entire training run, so in the second epoch the learning-rate scale reflects the first epoch's frequencies plus ongoing accumulation. Please clarify whether F_T is reset between epochs; if it is not reset, the schedule is history-dependent in a way that is not discussed.","section":"Algorithm 1"},{"comment":"Reference [3] is cited as an arXiv preprint; if a peer-reviewed version exists, update the citation to the published venue.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a useful and honest industry case study, but the single-run comparisons are the main obstacle. The authors themselves note that AUC values are noisy, so the lack of confidence intervals or repeated runs makes the central ranking of methods (FAL, MEDA, baselines) difficult to verify. I would support acceptance after the authors either add a variance analysis (even a small-scale one) or explicitly reframe the paper as an experience report rather than a comparative validation, softening claims such as 'significantly higher' and 'comparable to MEDA' accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the paper's core idea—FAL, a frequency-adaptive learning rate that slows down infrequent embedding rows—is a real inversion of FA-SGD's logic and is cheap to implement. Also valuable is the per-objective breakdown showing multi-epoch overfitting severity tracks label sparsity. But the empirical support is thinner than the prose implies: every comparison rests on one training run per configuration, on a metric the authors themselves call noisy, with no confidence intervals or significance tests.\n\nWhat's genuinely good: FAL is simple, parameter-free once you choose log scaling, and the authors are honest about its failure on the sparsest objective and about gains evaporating after a few days of continual training. That candor is rare in industrial papers. The comparison to MEDA is fair, and the observation that overfitting varies by head in a multi-task model is a useful, transferable insight. The Sparse Optimizer is not novel, but the 50x embedding learning rate is a reasonable practical finding.\n\nSoft spots: the statistical grounding is the weak point. The cumulative AUC gain metric (a ratio of sums over days) does not account for autocorrelation, and differences of 0.07–0.20% are exactly the kind of numbers that disappear with a different random seed. The log-vs-linear scaling choice in Table 1 is a single comparison with no variance. The 50x multiplier is hand-tuned. No code or data release, so external verification is impossible. These do not make the central claims false—the effect is visually consistent across two datasets and multiple objectives—but they make the paper a preliminary case study, not a definitive controlled experiment.\n\nWho benefits: practitioners in ads and recommendation systems who want a cheap mitigation for multi-epoch overfitting, and researchers studying the one-epoch phenomenon. The paper deserves a serious referee; an editor should send it out rather than desk-reject. But the referee should demand confidence intervals, a noise model, or at least repeated runs on a subsample before the FAL effect is treated as established.\n\nMy recommendation: engage with it, but treat the numbers as indicative, not conclusive.","headline":"A useful industrial case study with a genuinely new trick in FAL, but the quantitative claims rest on single runs of a noisy metric and need stronger statistical support.","tokens_in":13668,"tokens_out":1999,"would_cite":false,"duration_ms":26484,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A frequency-adaptive learning rate reduces multi-epoch overfitting in ads conversion models as well as embedding re-initialization, except on the sparsest objective.","keywords":["ads recommendation systems","click-through rate prediction","conversion prediction","embedding tables","multi-epoch overfitting","frequency-adaptive learning rate","multi-task learning"],"falsifier":"Run each configuration (1-epoch baseline, 2-epoch baseline, FAL, and MEDA) with several independent seeds and shuffles on the same training window, and compute confidence intervals for the epoch-boundary loss jumps and cumulative AUC gains; if the FAL-versus-baseline and FAL-versus-MEDA gaps on objectives such as $p(\\text{checkout}|\\text{click})$ overlap zero, the central empirical claim collapses.","tokens_in":12648,"feed_emoji":"📉","tokens_out":10230,"duration_ms":94907,"temperature":0.7,"pith_summary":"The paper tries to establish two connected results about training large embedding tables in a production advertising conversion model. First, giving embedding tables a much higher layer-specific learning rate than the rest of the network (the Sparse Optimizer) speeds convergence and improves offline AUC. Second, the paper proposes Frequency-Adaptive Learning Rate (FAL), which scales each embedding row's learning rate by its relative log frequency so that rare rows learn slowly, and claims this mitigates multi-epoch overfitting about as well as embedding re-initialization on all objectives except the sparsest. It also argues that multi-epoch overfitting is not a universal crisis: its severity tracks label density, and the benefits of treating it mostly disappear after a few days of continual training on fresh data.","feed_headline":"Adaptive learning rates tame multi-epoch overfitting in ads models","feed_subtitle":"Lowering learning rates for rare IDs cuts epoch-boundary loss jumps with minimal overhead and no dense-layer changes","key_machinery":"The load-bearing object is the row-wise frequency-adaptive learning rate (FAL) schedule: each embedding row's learning rate is the table-level learning rate multiplied by $\\log(F_T[i]+1)/\\max_j \\log(F_T[j]+1)$. FAL accumulates per-row frequencies during training as a 32-bit integer tensor, adds about 3.125% memory overhead at embedding dimension 32, and works with gradient clipping and adaptive optimizers because it only scales the gradient. The complementary object is the Sparse Optimizer, a layer-specific multiplier (tuned to 50x) on the embedding-table learning rate that counters slow convergence from gradient sparsity. Together they split the frequency spectrum: the Sparse Optimizer accelerates high- and medium-frequency rows, while FAL slows down the low-frequency rows that drive overfitting.","core_discovery":"On its own terms, the paper discovers that multi-epoch overfitting in a multi-task conversion model is concentrated in low-frequency embedding rows, and that a row-wise learning-rate schedule can counteract it. FAL sets the learning rate of embedding row $i$ in table $T$ to $\\eta^*_T \\cdot \\log(F_T[i]+1) / \\max_j \\log(F_T[j]+1)$, where $F_T[i]$ is the cumulative frequency of that row; the multiplier is applied to the row's gradient before the optimizer update. On the May-Aug 2024 training window, FAL shrinks the epoch-boundary loss jump for $p(\\text{add-to-cart}|\\text{click})$ from +1.19% to +0.23% and for $p(\\text{checkout}|\\text{click})$ from +0.75% to +0.65%, and ends batch training with lower loss than the 2-epoch baseline on those objectives. On a second, 25% larger Aug-Dec window, FAL matches embedding re-initialization (MEDA) on every objective except the sparsest, $p(\\text{checkout}|\\text{click})$, where MEDA ends 0.27% lower in loss; over continual training, FAL's cumulative AUC gains on $p(\\text{checkout}|\\text{click})$ are 0.07% and 0.20% on the two windows. The paper reads these results as showing both that FAL is a viable anti-overfitting tool and that, once fresh daily data is available, the plain 2-epoch baseline catches up, so solving multi-epoch overfitting may be unnecessary when overfitting is mild.","pith_inferences":["The FAL formula depends only on cumulative row exposure, so the same schedule could be applied to any embedding-based model, not just ads conversion; the paper demonstrates it on a single production architecture, so transferability is untested.","Because the method treats log-frequency as a proxy for overfitting propensity, a natural extension is to replace it with a per-row uncertainty estimate or gradient-norm history; that might fix the sparsest objective, where FAL fails to beat the baseline.","The paper's continual-learning result implies that end-of-batch loss is the wrong yardstick for judging multi-epoch fixes: an evaluator should report loss and AUC after several days of fresh-data training, where the differences between methods largely collapse.","The interaction between the Sparse Optimizer's 50x embedding learning rate and FAL's down-scaling suggests a single frequency-dependent net learning-rate curve; sweeping the scaling exponent could find a schedule that beats MEDA even on $p(\\text{checkout}|\\text{click})$."],"forward_implications":["A per-row log-frequency learning-rate schedule is a cheap, drop-in mitigation for multi-epoch overfitting: one integer counter per embedding row, no change to the dense network or optimizer state.","Overfitting severity in multi-task models is tied to label density: sparser objectives such as $p(\\text{checkout}|\\text{click})$ need stronger treatment, while dense heads like $p(\\text{click})$ can be left on a plain schedule.","The plain 2-epoch baseline catches up with FAL and MEDA after several days of continual training on fresh data, so in a batch-then-continual production setup anti-overfitting machinery is only needed when overfitting is severe or data is scarce.","FAL keeps test loss approximately monotonically decreasing through the second epoch, which enables mid-epoch early stopping; MEDA's re-initialization creates a discontinuity that rules this out.","A frequency-aware extension of MEDA or regularization, one that re-initializes or penalizes mainly frequent rows, could combine the strengths of both methods and address the sparsest objective where FAL is weakest."],"supporting_citations":[{"why":"Supplies the MEDA embedding re-initialization method that FAL is compared against, and the observation that downstream layers adapt to biases from infrequent rows.","marker":"[3]"},{"why":"Provides the characterization of multi-epoch overfitting and the finding that a lowered learning rate reduces overfitting at the cost of peak performance, which FAL is designed to avoid for frequent rows.","marker":"[32]"},{"why":"Introduces frequency-aware per-row learning rates (FA-SGD); FAL inverts its direction by decreasing rather than increasing the learning rate for infrequent rows.","marker":"[10]"},{"why":"Supplies the adaptive per-parameter learning-rate family, including the observation that higher learning rates for rarely seen features improve convergence.","marker":"[2]"},{"why":"Defines the Adam optimizer that is the base training dynamic in all experiments, and whose embedding-table learning rate the Sparse Optimizer multiplies by 50.","marker":"[9]"}],"fun_headline_variants":["Frequency-aware learning rates tame embedding overfitting","Smaller steps for rare IDs shrink epoch-boundary loss spikes","Adaptive LR for embeddings cuts multi-epoch loss jumps","Overfitting fix: scale embedding LR by feature frequency","Frequency-adaptive LR offers a new anti-overfitting approach"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported single-run differences in test-loss jumps and 0.07% to 0.20% cumulative AUC gains are real method effects rather than run-to-run noise, since the evaluation reports no repeated seeds, confidence intervals, or significance tests.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-aware learning rates tame embedding overfitting","Smaller steps for rare IDs shrink epoch-boundary loss spikes","Adaptive LR for embeddings cuts multi-epoch loss jumps","Overfitting fix: scale embedding LR by feature frequency","Frequency-adaptive LR offers a new anti-overfitting approach"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000366,"raw_usage":{"total_tokens":2102,"prompt_tokens":1214,"completion_tokens":888,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":830,"completion_tokens_details":{"reasoning_tokens":810}},"tokens_in":830,"tokens_out":888,"duration_ms":9339,"temperature":1.0,"reasoning_tokens":810,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:02:22.330101+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run each configuration (1-epoch baseline, 2-epoch baseline, FAL, and MEDA) with several independent seeds and shuffles on the same training window, and compute confidence intervals for the epoch-boundary loss jumps and cumulative AUC gains; if the FAL-versus-baseline and FAL-versus-MEDA gaps on objectives such as $p(\\text{checkout}|\\text{click})$ overlap zero, the central empirical claim collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive per-parameter learning-rate family, including the observation that higher learning rates for rarely seen features improve convergence."}],"review_version":1}