{"id":"ce63e288-743f-4b5f-8dcf-74b1c5e4aaef","arxiv_id":"2411.19647","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"CAdam masks Adam updates when momentum and gradient disagree in sign, improving robustness to distribution shift and label noise while preserving Adam's convergence rate.","lead":"Researchers at Tencent and Tsinghua introduce CAdam, a modified Adam optimizer that skips parameter updates when the momentum and the latest gradient point in opposite directions, aiming to ignore noise and react faster to distribution shifts. The method is a drop-in Adam replacement and is reported to improve recommendation metrics in large-scale A/B tests.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The convergence proof is for a fixed deterministic objective, so it does not cover the stochastic, non-stationary online setting CAdam targets; the sign gate is a data-dependent nonlinearity that no existing lemma controls under noise.","rationale":"I found no internal inconsistency in Lemma B.2/B.3 for deterministic settings: the mask only reduces the step norm, and with exact gradients the masked error pi_{t,Xi_t} is indeed bounded by the unmasked pi_t because when the mask is off, |m_i - g_i| = |m_i| + |g_i| >= |g_i|. The load-bearing weakness is scope: the theorem's assumptions are static and deterministic, while the paper's headline claims are for online, non-stationary, noisy environments. The Limitation section honestly states the deterministic setting, but the conclusion then overstates the theory's reach. The stochastic extension is non-trivial because the sign gate depends on the noisy gradient itself and acts as a truncation, not an unbiased projection. The reader's CONDITIONAL verdict already captures this gap; I only broaden it to include the static-objective mismatch. A concrete analytic or numerical check on a one-dimensional Gaussian-gradient example would determine whether the gated error term can be controlled; absent that, the paper should either add the analysis or explicitly scope Theorem 4.3 to offline optimization. The production A/B evidence is suggestive but, without per-scenario statistical detail, does not close the theoretical gap.","tokens_in":13599,"tokens_out":13841,"duration_ms":134503,"concrete_test":"Re-derive Lemma B.3 under the standard stochastic model g_t = grad f(x_t) + xi_t with E[xi_t]=0 and bounded variance, and compute on a single coordinate the ratio E[||pi_{t,Xi_t}||^2] / E[||pi_t||^2], where pi_t = m_t - g_t. Concretely, set true gradient mu = 0.01, Gaussian noise sigma = 1, beta1 = 0.9, and run CAdam and Adam on one-dimensional quadratic loss with noisy gradients for 10^5 steps. Record average squared update magnitude and final gradient norm. If CAdam's gated update saturates at a positive floor or the ratio exceeds 1, the deterministic proof does not extend to stochastic gradients, and the paper's online-learning claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Limitation A admits the analysis is deterministic. More importantly, Theorem 4.3 is about a single fixed function f: it bounds the average squared gradient of a static nonconvex objective. CAdam's motivation and deployment are streams with distribution shifts and noise, where there is no fixed f to converge to and no regret or tracking bound is proved. A static deterministic convergence rate is compatible with zero or negative benefit under drift; it does not support the conclusion that CAdam 'enables seamless replacement in real-world systems.' The stochastic extension is not a formality. The mask I(m_t*g_t > 0) is a sign-dependent gate on the same samples used in the update. With noisy gradients, the event {m_t,i * g_t,i > 0} is correlated with g_t,i, so the gated update is a truncated, biased function of the gradient. Lemma B.3's core control is ||pi_{t,Xi_t}|| <= ||pi_t||; under stochastic gradients the masked noise term when the gate is off is -g_t,i, and its variance can exceed the Adam error ||pi_t||^2 in low-signal coordinates. The proof supplies no stochastic descent inequality, and the theorem's delta-dependent probability notation only makes the mismatch more visible. Thus theoretical support for CAdam in its target setting is absent; the paper should either provide a stochastic/dynamic analysis or explicitly scope Theorem 4.3 to offline optimization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CAdam, a confidence-based variant of Adam that masks coordinate-wise updates whenever the running momentum and the current gradient disagree in sign (Algorithm 1, line 14). The authors claim that CAdam preserves Adam's convergence rate under relaxed nonconvex smoothness assumptions (Theorem 4.3), adapts faster to distribution shifts, is more robust to label noise, and outperforms Adam and other optimizers in offline benchmarks and in a seven-scenario production A/B test. The convergence proof is developed in Appendix B by adapting the framework of Li et al. [24] to the masked update, and the authors explicitly acknowledge in the Limitation section that the analysis is currently restricted to a deterministic setting.","tokens_in":13838,"tokens_out":5219,"duration_ms":48096,"significance":"If the empirical and theoretical claims were fully established, CAdam would be a practically attractive drop-in replacement for Adam: it introduces no new hyperparameters, the update rule is simple, and the reported nine-month production deployment across sixteen scenarios is a strong real-world endorsement. The paper also deserves credit for being explicit about its limitation that the convergence analysis is deterministic and that no formal superiority guarantee over Adam is proved. However, the central theoretical result does not cover the stochastic, non-stationary online setting that motivates the paper, and the reported offline gains on Criteo are small relative to the reported standard deviations. The production A/B results are valuable but are presented without statistical detail, and one table contains an arithmetic inconsistency. The paper would need a substantive revision to make the theoretical and empirical support match the strength of the claims.","major_comments":[{"comment":"The paper's target setting is online learning with distribution shifts and noisy labels, as stated in the abstract and Section 1, but Theorem 4.3 is a convergence bound for a single fixed nonconvex objective f under the deterministic smoothness condition of Assumption 4.2. The Limitation section explicitly says that the analysis is established under a deterministic setting. A bound on the average squared gradient of a static f does not imply any regret or tracking guarantee for a shifting sequence of objectives f_t, nor does it control the effect of stochastic gradient noise. As written, the theoretical support for CAdam in its intended deployment scenario is absent. The authors should either provide a stochastic or dynamic analysis, or explicitly scope the theoretical claims to offline deterministic optimization and soften the online-learning conclusions accordingly.","section":"§4, Theorem 4.3 and Appendix A"},{"comment":"The proof that the mask does not break convergence relies on the deterministic inequality ||π_{t,Ξ_t}|| ≤ ||π_t|| in Lemma B.3. Under stochastic gradients this inequality is not sufficient: the event {m_{t,i} g_{t,i} > 0} is correlated with g_{t,i}, so the masked update is a truncated, biased function of the gradient, and the variance of the masked noise term can exceed the Adam error ||π_t||^2 in low-signal coordinates. The proof supplies no stochastic descent inequality, and the probability notation in Theorem 4.3 is not connected to any source of randomness in the deterministic proof. A concrete fix would be to prove a descent lemma under unbiased stochastic gradients or to state the theorem as a deterministic result without the δ-probability language.","section":"Appendix B, Eq. (4), Eq. (5), Lemma B.3"},{"comment":"Algorithm 1 line 15 uses the AMSGrad-style denominator sqrt(v_hat_max,t) + epsilon, while the proof in Eq. (4) analyzes the update with sqrt(v_hat_t) + epsilon and never mentions v_hat_max. Consequently, the proof does not cover the AMSGrad variant CAmsGrad that is reported in Table 1 and Table 4, even though the paper presents CAmsGrad as a main empirical contribution. The proof should either explicitly handle the v_hat_max recursion or the theorem should be restricted to the non-AMSGrad version of CAdam.","section":"Algorithm 1 vs. Appendix B, Eq. (4)"},{"comment":"The headline empirical claim that CAdam and CAmsGrad 'surpass' Adam is not supported by the reported statistics. In Table 1, Adam has average AUC 80.92 and CAdam 80.94, but for several models the differences are smaller than the reported standard deviations (e.g., DeepFM: 80.87±.011 vs 80.88±.008; DNN: 80.89±.003 vs 80.90±.002). No significance tests, paired comparisons, or confidence intervals are reported, and the authors state only that results are averaged over three seeds. Given that the paper claims a 0.05% GAUC increase is nontrivial, the statistical evidence for the Criteo result needs to be made explicit rather than asserted.","section":"§3.3, Table 1"},{"comment":"The production A/B table has an arithmetic inconsistency that needs correction or explanation. For the seven listed scenarios, the arithmetic mean of the Adam GAUC values is 85.91%, not the reported 85.34%, and the mean of the CAdam values is 86.17%, not the reported 85.64%; the reported average improvement of 0.30% is also larger than the 0.26% implied by the per-scenario differences. If the 'Average' column is a traffic-weighted or otherwise non-arithmetic mean, the weighting should be described; otherwise the production results are not reproducible from the table as printed.","section":"§3.4, Table 3"}],"minor_comments":[{"comment":"The notation in the algorithm mixes masked and unmasked quantities: line 14 overwrites hat_m_t with the masked vector, but the text in Section 2 describes the update as using m_t/sqrt(v_t), which is not exactly what the algorithm computes. Please unify the notation between the algorithm, the descriptive text, and the proof.","section":"§2, Algorithm 1"},{"comment":"The numerical noise experiment is described differently in Section 3.1 (randomly modifying the function value) and in Appendix D.1 (randomly masking each gradient dimension with probability 0.5). Please clarify which procedure was actually used, since the two descriptions lead to different noise models.","section":"§3.1 and Appendix D.1"},{"comment":"The text mentions that CAdam exhibits 'lower regret' under distribution shifts, but no regret measure is defined or plotted in Figure 2. Adding a quantitative regret or tracking-error curve would strengthen the numerical illustration.","section":"§3.1 and Figure 2"},{"comment":"Reference [12] has a malformed author list ('Jean-Baptiste Tien, joycenv') and should be corrected to the proper Criteo Display Advertising Challenge citation.","section":"References [12]"},{"comment":"The hyperparameter selection protocol is described at different levels of detail: Section D.3 states that Adam's optimal learning rate was used for both Adam and CAdam, while Section D.4 describes a per-optimizer grid search. Please state explicitly for each experiment whether the learning rate was tuned separately for CAdam, since this affects the fairness interpretation of the comparisons.","section":"§3.3, D.4"},{"comment":"The paper reports that CAdam has been deployed in 16 scenarios but provides A/B results only for 7 scenarios; please indicate whether the remaining 9 deployments were evaluated similarly and whether the reported nine-month stability claim refers to all 16 scenarios.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's core algorithmic idea is simple and potentially useful, and the production deployment evidence is a genuine strength. However, the main theoretical contribution is scoped to a deterministic setting that does not match the paper's stated online-learning motivation, and the empirical tables contain both statistical and arithmetic weaknesses. I believe these issues can be fixed within the scope of a revision, so I recommend major revision rather than rejection. The editor may also want to ask the authors to make the production A/B methodology (traffic splits, metric definitions, significance tests) more explicit, as the current Table 3 is not self-consistent as printed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: CAdam is a one-line Adam variant that masks out coordinates where momentum and current gradient disagree, and the paper makes a decent empirical case that this helps under noise and distribution shift. The theoretical case is narrower than the advertising: the convergence proof is for a fixed deterministic objective and says nothing about non-stationary online settings. The authors say so in the Limitation section, so the gap is disclosed, but the abstract and conclusion still lean on the theory.\n\nWhat's actually new: the mask rule itself is the same as the concurrently-posted Cautious Optimizers, which the paper cites. CAdam's contribution is the online-learning framing plus a convergence analysis in the relaxed-smooth framework of Li et al. The proof is mostly delegation—two lemmas supplied—but it is honest and the lemmas are straightforward. No fitted parameters, no self-citations, and they include compute and ablation details. The production A/B results across 16 scenarios and nine months are real evidence, even if no significance tests are reported.\n\nWhere it wobbles: First, the Criteo table shows gains that are mostly within one standard deviation of Adam; the average is 0.94 vs 0.92, and CAmsGrad does a bit better. \"Consistently outperforms\" is too strong. Second, the deterministic convergence theorem is about a static objective and does not cover the stochastic, drifting streams CAdam is designed for. The stress-test worry is legitimate: the mask is a data-dependent gate correlated with the noisy gradient, and no lemma in the appendix controls that interaction. But the authors explicitly flag the deterministic-only proof in Limitation A, so this is a known limitation rather than a hidden one. Third, the production GAUC gains have no error bars, which is common in industry papers but limits what a reviewer can conclude.\n\nBottom line: this is a useful empirical paper about a plausible, cheap optimizer tweak, with modest novelty and disclosed theoretical limits. It deserves a serious referee, and the right outcome is probably acceptance with tightened claims and either a stochastic analysis or a clearer scope statement.\n\nRecommendation: send to peer review, but the referee should push on the stochastic gap and the strength of the empirical claims.","headline":"CAdam is a clean empirical paper about a one-line Adam variant whose deterministic-only theory is narrower than the abstract suggests, but the authors disclose the gap and the production evidence is real.","tokens_in":14434,"tokens_out":2069,"would_cite":true,"duration_ms":19462,"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":"CAdam replaces Adam's update with a per-coordinate confidence gate that pauses updates when momentum and gradient disagree, and the paper argues this makes online learning adapt faster to distribution shifts and more resilient to noise.","keywords":["online learning","Adam optimizer","distribution shift","label noise","confidence mask","adaptive moment estimation","convergence analysis","recommendation systems"],"falsifier":"Run CAdam versus Adam on a stationary smooth nonconvex problem with stochastic gradients having zero mean (for example, a quadratic with symmetric label noise), and track the average squared gradient norm $\\frac{1}{T}\\sum_{t=1}^T \\|\\nabla f(x_t)\\|^2$. If CAdam's average stays above Adam's at large $T$, or if its updates stall on coordinates where momentum and gradient differ only because of noise, then the claimed same convergence rate does not transfer to the stochastic setting.","tokens_in":13380,"feed_emoji":"⚙️","tokens_out":7649,"duration_ms":59566,"temperature":0.7,"pith_summary":"The paper introduces CAdam, an optimizer that changes Adam's update rule with one per-coordinate condition: a parameter is updated only when the momentum $m_t$ and the current gradient $g_t$ point in the same direction ($m_t \\odot g_t > 0$). The paper's aim is to show that this sign-agreement gate lets an online learner tell real distribution shifts apart from transient noise, because a genuine shift makes the gradient keep disagreeing with stale momentum, while a noisy gradient disagrees for only a step or two. The authors claim the modification does not cost the convergence rate: Theorem 4.3 proves CAdam matches Adam's rate on deterministic nonconvex objectives, and experiments on rotated CIFAR-10, noisy labels, the Criteo CTR dataset, and production recommendation A/B tests show CAdam matching or beating Adam and other optimizers. The practical payoff is a drop-in replacement for Adam that needs no new hyperparameters and has been run in production across sixteen scenarios for over nine months.","feed_headline":"CAdam beats Adam on shifting data by pausing misaligned updates","feed_subtitle":"A per-parameter gate checks momentum-gradient agreement, preserving Adam's convergence rate while filtering noise and drift.","key_machinery":"The load-bearing object is the confidence mask $M_t = I(m_t \\odot g_t > 0)$, applied element-wise to the bias-corrected momentum before the update: $$\\theta_t = \\theta_{t-1} - \\$\\alpha$ \\frac{\\hat m_t \\odot M_t}{\\sqrt{\\hat v_t} + \\epsilon}.$$ The mask is hyperparameter-free: it either keeps Adam's original update or blocks a coordinate for one step. It does the argument's work in two places. Practically, it creates the pause-and-monitor behavior that the paper identifies with distinguishing distribution shift from noise. Theoretically, it only shrinks the update vector, so the displacement bound $\\|x_{t+1}-x_t\\| \\le \\alpha D$ still holds, which is why the proof can reuse Adam's convergence framework and only needs the two masked-update lemmas.","core_discovery":"CAdam's central claim is that the alignment between momentum and current gradient is a reliable confidence signal for online optimization. On coordinates where $m_t$ and $g_t$ agree, CAdam performs the exact Adam update; on coordinates where they disagree, it withholds the update for that step. The paper argues that this selective updating prevents stale momentum from carrying the model in an outdated direction after a distribution shift, and prevents a single corrupted gradient from moving the parameters, while repeated disagreements let the momentum decay and re-align with the new gradient trend. Under Assumptions 4.1 and 4.2, Theorem 4.3 gives an average squared-gradient bound of order $\\gamma^2$ after $T=\\max\\{(1-\\beta_1)^{-2}, O(\\alpha \\gamma^{-2})\\}$ iterations, matching the convergence rate of Adam in the deterministic setting; the proof supplies two lemmas bounding the masked update and substitutes them into the relaxed-smoothness framework. The empirical sections report consistent gains, including a $0.30\\%$ average GAUC improvement over Adam across seven production A/B tests.","pith_inferences":["If the deterministic proof does not transfer to stochastic gradients, the mask may still confer stability, but one would expect CAdam to be conservative on stationary stochastic problems: with symmetric zero-mean noise, roughly half the coordinates will have $m_t \\cdot g_t \\le 0$ at any step, effectively halving the update magnitude on those coordinates.","The sign-agreement gate is a generic component, so it likely transfers to other momentum-based optimizers such as Lion, SGDM, and RAdam; the paper's own ablations with CSGDM and CAmsGrad are initial evidence.","A testable extension is to tune the gate to a momentum-yield threshold rather than a hard sign: for high-noise regimes, requiring agreement over several consecutive steps could filter persistent noise, while for fast drift, a single-step mask may be too conservative.","In recommendation production, part of the observed GMV and GAUC gain may come from the optimizer dampening updates after abrupt user-behavior changes; a deployment experiment could measure whether CAdam's advantage grows with measured concept-drift rate or label-corruption rate."],"forward_implications":["CAdam can be swapped into a live system in place of Adam or AdamW with no hyperparameter changes, because the mask is the only added mechanism and it defaults to Adam on aligned coordinates.","Under sudden and continuous distribution shifts, the alignment ratio drops at the moment of the shift and recovers afterward; CAdam's accuracy drops more slowly and recovers faster than Adam's.","Under label noise, CAdam updates fewer parameters as noise increases, and on the noisy Criteo benchmark its average AUC drop is $-0.12$ compared with Adam's $-0.39$.","The convergence guarantee is limited to deterministic objectives: Theorem 4.3 establishes the same rate as Adam but does not claim a strictly better rate, and the paper leaves stochastic convergence analysis to future work.","In production recommendation scenarios, the average GAUC gain over Adam was $0.30\\%$, and CAdam has been running in 16 online scenarios for over nine months."],"supporting_citations":[{"why":"Defines Adam, the optimizer CAdam modifies and the primary baseline in every experiment.","marker":"[2]"},{"why":"Supplies the AMSGrad variant used in Algorithm 1 and the convergence baseline that CAdam builds on.","marker":"[8]"},{"why":"Provides the Criteo display-ad click dataset used for the public CTR benchmark.","marker":"[12]"},{"why":"Provides the FuxiCTR benchmarking setup and default hyperparameters used in the CTR experiments.","marker":"[13]"},{"why":"Provides the relaxed-smoothness convergence framework and the lemmas that CAdam's proof reuses.","marker":"[24]"}],"fun_headline_variants":["CAdam pauses updates when momentum and gradient diverge","CAdam: pause on disagreement, adapt to shifts faster","A confidence check makes Adam robust to online noise","CAdam gates updates to survive distribution shifts","Pause misaligned Adam steps for better online learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that CAdam keeps Adam's convergence rate assumes deterministic gradients, while the paper's target use is stochastic online learning; if the guarantee does not extend to stochastic gradients, the theoretical support for CAdam in the setting where it is claimed to help is absent.","fun_headline_variants_meta":{"raw":{"variants":["CAdam pauses updates when momentum and gradient diverge","CAdam: pause on disagreement, adapt to shifts faster","A confidence check makes Adam robust to online noise","CAdam gates updates to survive distribution shifts","Pause misaligned Adam steps for better online learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000801,"raw_usage":{"total_tokens":3561,"prompt_tokens":1021,"completion_tokens":2540,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":2465}},"tokens_in":637,"tokens_out":2540,"duration_ms":16532,"temperature":1.0,"reasoning_tokens":2465,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:58:49.185930+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CAdam versus Adam on a stationary smooth nonconvex problem with stochastic gradients having zero mean (for example, a quadratic with symmetric label noise), and track the average squared gradient norm $\\frac{1}{T}\\sum_{t=1}^T \\|\\nabla f(x_t)\\|^2$. If CAdam's average stays above Adam's at large $T$, or if its updates stall on coordinates where momentum and gradient differ only because of noise, then the claimed same convergence rate does not transfer to the stochastic setting.","supporting_citations":[{"cited_title":"Open benchmarking for click-through rate prediction","cited_arxiv_id":null,"evidence_quote":"Provides the FuxiCTR benchmarking setup and default hyperparameters used in the CTR experiments."}],"review_version":1}