{"id":"060d36ce-5ff6-40ea-9b67-6f67d8dcfbe1","arxiv_id":"2607.20374","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ARROW is a new streaming algorithm that reduces minibatch variance for MMD and CORAL by reweighting each incoming batch to match an exponential moving average of alignment statistics.","lead":"This paper introduces ARROW, an online variance-reduction method for the MMD and CORAL domain-adaptation losses, which reweights streaming minibatches to match moving-average reference statistics. A smart generalist would read it because it promises stable domain adaptation in streaming or incremental settings where offline variance-reduction methods cannot run.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ARROW's variance-reduction claim rests on an unproven substitution: aligning minibatch statistics to a stale EWMA reference does not necessarily reduce true gradient error under model or data drift, and unconstrained QP weights can amplify gradients.","rationale":"The reader identified the same load-bearing assumption: the EWMA reference is treated as a valid stand-in for the true time-varying D_m, and the unconstrained weights are not analyzed. I agree, and I sharpen the issue: the reference is stale not only under distribution shift but also under ordinary model parameter evolution, because D_m depends on the current feature extractor. The paper provides no control-variate correction or bias bound, so the core variance-reduction guarantee is not established. The static experiments, including the Monte Carlo simulation with fixed data and no training, cannot detect this failure mode. The unconstrained QP is a second concrete hazard because the optimal weights can be large without regularization. These are correctness risks, not mere missing ablations; however, they do not warrant changing the reader's CONDITIONAL verdict, since the experiments do show competitive practical performance and the theoretical gap is addressable.","tokens_in":13305,"tokens_out":6696,"duration_ms":62749,"concrete_test":"Run ARROW in a real training loop on one DomainBed task (e.g., Spawrious or a synthetic two-Gaussian stream) exactly as in Section 3.2. Every 50 iterations, freeze the model and compute the true D_m on a large held-out source/target sample; also record \\tilde D_m, the raw \\hat D_m, the solved weights u,v, and the resulting gradient error ||\\hat\\nabla_w - \\nabla L_disc||. Plot ||\\tilde D_m - D_m|| and ||\\hat\\nabla_w - \\nabla L_disc|| over time. If ||\\tilde D_m - D_m|| is comparable to or larger than ||\\hat D_m - D_m||, or if max ||u||_2, ||v||_2 grow sharply relative to the unweighted case, the EWMA surrogate and unconstrained reweighting are not supporting the claimed variance reduction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is the substitution in Section 2.1: the paper argues that since \\hat\\nabla L_disc - \\nabla L_disc is proportional to \\hat D_m - D_m, one can reduce gradient error by driving a weighted minibatch statistic \\hat D_w toward an EWMA reference \\tilde D_m, as in Eq. (4). This requires \\tilde D_m to be an accurate estimate of the true current D_m. No theorem or bound is given. In fact, \\tilde D_m is a moving average of past noisy minibatches computed under past model parameters. During training, D_m changes not only with input distribution shifts but also as the feature extractor parameters evolve, so \\tilde D_m is a lagged/stale statistic. With alpha=0.1, the effective memory is only about 10 minibatches, and there is no SVRG-style correction term to restore unbiasedness. Minimizing distance to a stale reference can reduce variance around the wrong target, increasing the true gradient error. Second, the weight optimization in Section 2.2 is an unconstrained QP (Eq. 15): the paper explicitly says no constraints are imposed on u and v. When \\tilde D_m is poorly represented in the span of the current minibatch features, optimal weights can be arbitrarily large, inflating the weighted loss gradient and destabilizing training. The experiments use static DomainBed benchmarks and a stationary Monte Carlo simulation, so neither the parameter-drift lag nor the weight-explosion failure mode is exercised.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ARROW, an online stochastic variance reduction method for the MMD and CORAL domain-adaptation losses. ARROW maintains an exponentially weighted moving average (EWMA) of the discrepancy statistic and, at each minibatch, solves a quadratic program that reweights the current minibatch so that its weighted discrepancy statistic matches the EWMA reference. For CORAL, a relaxation using linear combinations of outer products is proposed. The paper reports Monte Carlo variance comparisons, DomainBed accuracy experiments on Spawrious, Office-Home, and Humpbacks, and wall-clock runtimes. The authors claim that ARROW is the first online SVR method for MMD/CORAL and that it is competitive with offline SVR methods in variance reduction, accuracy, and speed.","tokens_in":13771,"tokens_out":6486,"duration_ms":57953,"significance":"If the central mechanism were proven, ARROW would fill a genuine gap: existing SVR methods for MMD/CORAL require full-dataset access and are unsuitable for streaming, distributed, or incremental UDA. The algorithmic construction is concrete and the quadratic formulation in Eq. (15) is mathematically well-defined. The experiments cover three benchmarks and compare a broad set of sampling-based SVR baselines. However, the paper's load-bearing premise — that aligning a weighted minibatch statistic to an EWMA reference reduces the true gradient error — is asserted rather than proved, and the empirical evaluation does not exercise the streaming/drift regime that motivates the method. The unconstrained weight optimization also raises stability concerns that are not addressed. The idea is promising and the experimental setup is useful, but the current evidence is insufficient to support the full claim.","major_comments":[{"comment":"The central variance-reduction claim is not established. The paper argues that ||∇̂L_disc − ∇L_disc|| ∝ ||D̂_m − D_m|| and then substitutes the EWMA reference D̃_m for the true D_m. This is a three-step leap: (i) the chain-rule proportionality is only schematic, since the gradient also involves the Jacobian of the feature embedding, which changes with model parameters; (ii) the cited smoothness conditions in Zhao & Zhang (2014) and Liu et al. (2020) concern stratified sampling for finite-sum SGD and do not cover a time-varying reference; (iii) D̃_m is a biased, lagged average of past minibatches computed under past model parameters, and with α=0.1 and a 0.01 coefficient cutoff it retains only about 22 minibatches. Under model or data drift, minimizing ||D̂_w − D̃_m|| can increase ||∇̂L − ∇L|| by moving the estimate toward a stale target. There is no SVRG-style correction term and no boun","section":"§2.1, Eq. (4)"},{"comment":"The unconstrained quadratic program is ill-posed as stated. The Gram matrices K_ss and K_tt can be singular, and since no constraints or regularization are imposed on u and v, the objective can be unbounded below when the linear term has a component in the null space, or the optimal weights can be arbitrarily large. The paper explicitly declines KMM-style simplex constraints, but gives no bound on ||u|| or ||v||. Large weights can inflate the weighted discrepancy gradient and destabilize training, which is precisely the failure mode SVR is meant to prevent. The claim that the weighted statistic can 'perfectly align' with the reference if the reference lies in the linear span does not address the case where it does not, nor the effect of large weights on the downstream model update.","section":"§2.2, Eq. (15)"},{"comment":"Two issues affect the CORAL extension. First, the relaxation that replaces covariance estimators with fixed-center linear combinations of outer products has no approximation guarantee: the optimized Σ̂_w is not a proper covariance estimate when u and v are arbitrary, and the paper provides no bound relating the relaxed objective to the true covariance-difference error. The statement that the problem is 'solvable to 0 if the linear span contains D̃_CORAL' is about algebraic span, not about estimating the true D_CORAL. Second, the algorithm is under-specified: after solving Eq. (23), the paper never states how the weights enter the actual model gradient — whether the update uses the gradient of ||D̂_w||², the gradient of ||D̂_w − D̃_m||², or something else. This is essential for the claimed variance-reduction behavior.","section":"§2.3, Eqs. (20)–(23)"},{"comment":"The experimental evidence is only partially consistent with the paper's accuracy claims, and it does not test the streaming premise. On Spawrious MMD, ARROW (65.2) is far below ORDERED (76.1) and PSDA (73.7); on Office-Home CORAL, ARROW (49.3) is below PSDA (52.7); only Humpbacks shows a clear advantage for ARROW. The claim that ARROW is 'competitive' is therefore selective. More importantly, all benchmarks are static DomainBed datasets and the Monte Carlo simulation in Figure 1 is stationary Gaussian data; there is no experiment with time-varying source/target distributions or feature drift during training. Since the paper's central motivation is online/streaming SVR, the unique selling point of ARROW is not actually exercised. A streaming benchmark with controlled shifts, or at least a synthetic non-stationary experiment, is needed to support the online claim.","section":"§3, Tables 1–3"}],"minor_comments":[{"comment":"Office-Home results report only the average standard error across splits, not per-cell standard errors, unlike Tables 1 and 3. This makes row-level comparisons difficult and should be fixed.","section":"Table 2"},{"comment":"The variance curves are shown without error bars or confidence bands, even though the simulation uses 1,000 samples. Statements like 'comparable to VaRDASS' would be more convincing with uncertainty estimates.","section":"Figure 1"},{"comment":"The statement that ARROW's runtime is 'completely independent of dataset size' is overstated: the per-minibatch cost is independent of dataset size, but total runtime scales with the number of streamed minibatches. The paper reports wall-clock times only for fixed-size offline datasets, so the scaling claim is untested.","section":"§3.3"},{"comment":"Several key prior works are self-citations to manuscripts dated 2026 (Napoli & White 2026a,b; Napoli 2026; Baumann et al. 2026) that may not yet be publicly available. Please mark them as 'in press' or provide arXiv identifiers/availability information, since the novelty claim 'first online SVR' depends on these comparisons.","section":"§1, References"},{"comment":"Minor typographical issue in the column header 'A verage'; also, the random search over minibatch size uses k ∼ 2^Uniform(3,7), so k ranges from 8 to 128; the QP cost for the largest k is not discussed.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is built on a chain of same-author manuscripts (2026a,b, Napoli 2026) that appear to be unpublished or very recent. The editor may wish to verify their availability and whether the 'first online SVR' claim survives once those works are examined. No code release is mentioned; given that the central mechanism is currently unproven, release of code and a non-stationary experiment would materially help the revision. I do not see a fatal error in the quadratic formulations themselves, but the missing theoretical justification and the untested streaming setting make acceptance at the current stage premature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes ARROW, an online SVR method for MMD and CORAL. What is genuinely new is the combination of EWMA reference statistics with minibatch instance reweighting, plus a linear relaxation for CORAL that makes the weight optimization tractable. The MMD formulation is a correct convex QP, and the experiments show ARROW is at least competitive with offline methods, sometimes better (Humpbacks 93.0 vs PSDA 91.0 for CORAL). The writing is clear and the empirical work is fairly extensive across three benchmarks and many baselines.\n\nThe soft spots are real. The central claim in Section 2.1 is that aligning the weighted minibatch statistic to the EWMA reduces true gradient error, but the proportionality argument skips over the feature extractor's Jacobian, which is not constant. The EWMA is a lagged statistic under model drift, and no bias or error bound is given. The unconstrained weights could in principle inflate the gradient magnitude; there is no analysis of weight behavior or conditioning. Alpha and the buffer cutoff are fixed without sensitivity checks, and the buffer truncation error is not analyzed. Most importantly, the 'streaming' setting is not actually exercised: all three benchmarks are static DomainBed sets, and the Monte Carlo simulation is stationary Gaussian data. So the method's behavior under distribution shift or model evolution is untested.\n\nThe citation pattern is heavily self-referential, but the prior work is directly on target and ARROW is not defined in terms of those results, so I would not call it circular. Still, a reviewer should ask whether the 'first online SVR' claim holds against a broader online variance-reduction literature, which is not surveyed.\n\nWho is this for? Anyone working on UDA, especially on making MMD/CORAL usable in streaming or incremental settings. It deserves a serious referee, but the revision should either prove the variance-reduction property under reasonable assumptions, or honestly frame the method as a heuristic. I would also ask for a non-stationary experiment, such as a shifting target distribution or a corrupting feature extractor, before accepting the streaming claim. This is a useful paper, not a fatal one.","headline":"A plausible online variance-reduction method for MMD/CORAL with solid experiments, but the core mechanism is asserted rather than proven and the streaming setting is not actually tested.","tokens_in":14173,"tokens_out":4591,"would_cite":true,"duration_ms":44099,"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":"A new online method, ARROW, reweights each minibatch to a moving reference to stabilize MMD and CORAL domain adaptation on streaming data.","keywords":["online variance reduction","domain adaptation","maximum mean discrepancy","CORAL","instance reweighting","exponential moving average","streaming data","stochastic optimization"],"falsifier":"Run ARROW on a stream with an abrupt mid-stream distribution shift, and measure the actual gradient error $\\|\\hat{\\nabla}L - \\nabla L\\|$ at each step. If the error spikes sharply after the shift and does not recover, the EWMA reference lags too much. A second check is to replace the EWMA with an oracle that has access to the true population statistic: if ARROW's variance reduction is much worse than the oracle's, the reference, not the reweighting, is the bottleneck.","tokens_in":13218,"feed_emoji":"🏹","tokens_out":6025,"duration_ms":44624,"temperature":0.7,"texified_at":"2026-08-05T21:35:51.702246+00:00","pith_summary":"This paper claims that the noisy gradients of the MMD and CORAL domain-adaptation losses can be variance-reduced in a purely online, streaming setting, something previous variance-reduction methods could not do because they required access to the whole dataset. The proposed ARROW method keeps an exponentially weighted moving average of the alignment statistics and, at each step, chooses per-sample weights that pull the current minibatch's statistic toward that reference. Because the weights are unconstrained, the weight-solving step becomes a convex quadratic program for MMD and a linear least-squares problem for CORAL, both solvable with standard techniques. The author argues this makes variance-reduced domain adaptation feasible for online, distributed, or incremental learning, and reports that ARROW matches offline methods in variance reduction, speed, and target accuracy.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":3301,"prompt_tokens":710,"completion_tokens":2591,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":1976}},"feed_headline":"ARROW cuts gradient noise in streaming domain adaptation","feed_subtitle":"Reweighting each minibatch to a moving reference makes MMD/CORAL training stable without full dataset access.","key_machinery":"The central mechanism is instance reweighting of the minibatch alignment statistic: each incoming sample receives an arbitrary (not necessarily convex) weight so that the weighted kernel-mean difference (for MMD) or weighted covariance difference (for CORAL) matches a running EWMA reference statistic. The reference tracks the true population statistic over time, and the per-step weight optimisation is a convex quadratic program (MMD) or a linear least-squares problem (CORAL) that can be solved with standard linear-algebra routines.","core_discovery":"On its own terms, the central claim is that stochastic variance reduction for the MMD and CORAL losses does not require finite-sum structure or offline access to the data. It can be achieved by replacing the unknown true alignment difference $D$ with an exponentially weighted moving-average estimate $\\tilde{D}$ and, at each timestep, solving a reweighting problem that aligns the weighted minibatch statistic to that estimate. For MMD this is a convex quadratic in the sample weights; for CORAL, after a relaxation that fixes the covariance centres, it becomes a linear least-squares problem. The paper reports that this online scheme delivers variance reduction and target-domain accuracy competitive","pith_inferences":["A natural extension would be to swap the EWMA for other online estimators with decreasing step sizes, which might give provable convergence guarantees for the reweighting scheme.","Since the weights are unconstrained, they could overfit a single minibatch to the reference and inflate gradient norms; adding weight regularisation or clipping is a concrete, testable variant.","ARROW could be combined with within-minibatch variance-reduction tricks, such as antithetic pairing, to further lower gradient noise without needing full-dataset access.","A direct measurement of true gradient error, rather than the proxy such as the distance between the weighted statistic and the EWMA, would reveal exactly when the EWMA proxy fails, such as under abrupt domain shifts."],"forward_implications":["If ARROW is correct, domain-adaptation training can be stabilised on streaming or unbounded data without storing or revisiting past samples.","The variance-reduction step's runtime depends only on the minibatch size and a small buffer of historical minibatches, not on dataset size, making it viable for extremely large or infinite streams.","The same reweighting-to-a-moving-reference mechanism may extend to other non-additive losses beyond MMD and CORAL.","Because the reference is updated continuously, the method can track time-varying distributions, a capability offline SVR methods lack.","In the reported experiments, integrating ARROW into MMD/CORAL training yields target accuracy competitive with substantially more complex domain-adaptation methods."],"fun_headline_variants":["ARROW: first online variance reduction for streaming domain adaptation","Adaptive reweighting cuts gradient noise without full dataset access","Moving-average references make MMD/CORAL training stable online","Streaming domain adaptation achieves offline-level variance reduction","Relaxed reweighting makes online variance reduction tractable"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that aligning the weighted minibatch statistic to the EWMA reference faithfully proxies for reducing the true discrepancy-gradient error; if the EWMA lags behind a distribution shift, or if unconstrained weights distort the gradient, the claimed variance reduction is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["ARROW: first online variance reduction for streaming domain adaptation","Adaptive reweighting cuts gradient noise without full dataset access","Moving-average references make MMD/CORAL training stable online","Streaming domain adaptation achieves offline-level variance reduction","Relaxed reweighting makes online variance reduction tractable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1285,"prompt_tokens":666,"completion_tokens":619,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":410,"completion_tokens_details":{"reasoning_tokens":537}},"tokens_in":410,"tokens_out":619,"duration_ms":5657,"temperature":1.0,"reasoning_tokens":537,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:59:43.262321+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ARROW on a stream with an abrupt mid-stream distribution shift, and measure the actual gradient error $\\|\\hat{\\nabla}L - \\nabla L\\|$ at each step. If the error spikes sharply after the shift and does not recover, the EWMA reference lags too much. A second check is to replace the EWMA with an oracle that has access to the true population statistic: if ARROW's variance reduction is much worse than the oracle's, the reference, not the reweighting, is the bottleneck.","supporting_citations":[],"review_version":1}