{"id":"ce3358fb-0d71-453d-8d6b-35b6691891db","arxiv_id":"2505.20890","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Training quantized models on low-frequency images plus frequency-aware batch normalization at test time improves both compression and domain-shift robustness.","lead":"CoDA is a method that trains quantized neural networks on low-frequency (blurred) versions of images and then adapts them at test time using frequency-split batch normalization. On corruption benchmarks such as CIFAR10-C and ImageNet-C, it reports accuracy gains of up to 7.96 and 5.37 percentage points over a full-precision test-time adaptation baseline, while using 4 to 16 times fewer bits.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training radius r is selected on the target corrupted benchmarks (CIFAR10-C/ImageNet-C), so the headline gains over full-precision TTA baselines may be inflated by test-set tuning.","rationale":"I read the strongest claim as the empirical one: CoDA, trained on LFC and adapted with FABN, outperforms full-precision TTA baselines under significant compression. The most load-bearing vulnerability is not the covariance approximation in Eq. 5, which is an unverified but plausibly benign technical detail; even if the covariance term is nonzero, the normalization could still work in practice. Rather, the training/filtering radius r is selected using the target corrupted benchmarks, which is a circular step. The paper's own Table 2 and Supplementary Section C show the radius was chosen after inspecting CIFAR10-C/ImageNet-C accuracy; the headline comparisons are against the weakest FP TTA baseline (NORM), and no sensitivity analysis of the full CoDA pipeline over r is provided. A missing control (standard TTA on the same LFC-trained model) is also important for isolating FABN's contribution, but it would weaken the novelty claim about FABN without necessarily overturning the headline compression+adaptation finding; the radius leakage directly threatens the validity of the headline numbers. If the suggested a-priori radius selection test preserves the gains, the central claim likely stands. The paper has considerable independent support: comprehensive tables across architectures/bitwidths, ablations, loss-landscape and t-SNE analyses, and per-corruption results, which is why a conditional verdict rather than rejection seems appropriate.","tokens_in":33138,"tokens_out":10577,"duration_ms":99670,"concrete_test":"Fix r without using CIFAR10-C/ImageNet-C: for example, choose the smallest r that retains at least 95% of clean source accuracy on a held-out validation split, or select r on a disjoint subset of corruptions (e.g., severities 1-2) and evaluate on severities 3-5. Re-run Tables 4, 5, and 6 with this a-priori r and report CoDA's accuracy vs. the best full-precision TTA baseline (not just NORM). If the gains over the best baseline drop below the claimed 7.96/5.37%p or become negative, the headline is inflated by test-set tuning. Also report a radius sensitivity sweep (e.g., r = 1/8, 1/4, 1/2 of input size) for the full CoDA pipeline to show the results are not peaked at the chosen value.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Supplementary Section C states that r=8 (CIFAR10) and r=56 (ImageNet) are set 'to ensure that LPF preserves a more robust and lower frequency range,' with Table 2 (which reports accuracy on CIFAR10-C/ImageNet-C for different radii) as the basis of the frequency analysis. This is test-set leakage: the radius is a hyperparameter of both LFC QAT and FABN, and it is chosen using the same corrupted test distributions on which CoDA is evaluated. On CIFAR10-C, r=8 is indeed the best LFC-QAT radius (LSQ 65.45, LQ 67.53 vs. r=4/12/16 in Table 2). On ImageNet-C the best radius in Table 2 is r=28 for both LSQ (21.90) and LQ (25.23), yet the paper selects the proportional r=56 (20.82/25.06), so the choice is not simply 'best on test' but it is still informed by test accuracy. No sensitivity analysis is reported for the full CoDA pipeline as a function of r, leaving open the possibility that the headline improvements (7.96%p on CIFAR10-C, 5.37%p on ImageNet-C, both against the weakest FP TTA baseline NORM) shrink or invert for other defensible radius choices. Because r controls both training and test-time filtering, this is a load-bearing circular step for the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoDA, a frequency-composition framework that jointly addresses model compression and test-time domain adaptation. During training, CoDA applies quantization-aware training (QAT) to low-frequency components (LFC) of images; at test time, it adapts batch-normalization statistics through Frequency-Aware Batch Normalization (FABN), which filters intermediate activations into LFC and HFC, updates LFC statistics with source-initialized EMA, uses current-batch HFC statistics, and combines them by summing means and variances. The method is evaluated on CIFAR10-C, ImageNet-C, ImageNet-R, and ImageNet-Sketch with LSQ and LQ quantizers at 2/4/8 bits across ResNet, MobileNet, and EfficientNet, and is combined with NORM, TENT, and SAR. The headline claims are accuracy improvements of 7.96pp on CIFAR10-C and 5.37pp on ImageNet-C over full-precision TTA baselines, with 4-16x model-size reduction.","tokens_in":33446,"tokens_out":3294,"duration_ms":35647,"significance":"If fully substantiated, CoDA would be an practically important demonstration that heavily quantized models can be more robust under domain shift than larger full-precision models with standard TTA. The paper's strengths include extensive experiments, per-corruption results with standard deviations for CIFAR10-C and ImageNet-C, integration with multiple TTA baselines, multiple architectures and bitwidths, model-size calculations, and ablations of the two main components. However, the central quantitative claims currently rest on a frequency radius selected using the same corrupted test distributions used for evaluation, and the specific benefit of FABN over standard TTA on the same LFC-trained model is not isolated. These issues must be resolved before the significance of the results can be assessed.","major_comments":[{"comment":"The frequency radius r is selected using target-domain accuracy on the corrupted test sets. Table 2 reports CIFAR10-C and ImageNet-C accuracy as a function of r, and Section C states that r=8 (CIFAR10) and r=56 (ImageNet) were chosen from this analysis. Since r controls both LFC QAT training (Sec. 3.1-3.2) and FABN test-time filtering (Sec. 3.3), the headline gains of 7.96pp and 5.37pp over full-precision TTA may be inflated by test-set leakage. Please select r on source-domain validation data (or a clean split), and report a sensitivity analysis of the full CoDA pipeline across a range of r values on both datasets.","section":"Supplementary Section C and Table 2"},{"comment":"The paper does not compare FABN against standard TTA applied to the same LFC-trained model. In Tables 4 and 5, the TTA baselines (NORM, TENT, SAR) are applied to FFC-trained models, while CoDA rows use LFC-trained models with FABN; the FABN-only row in Table 9 is based on an FFC-trained model with FABN. This does not establish that FABN is better than simply running NORM, TENT, or SAR on an LFC-QAT model. Please add experiments with each standard TTA method applied to the same LFC-trained model and compare against FABN.","section":"Tables 4, 5, and 9"},{"comment":"Equation (5) sets σ̂²_t = σ̂²_lfc,t + σ̂²_hfc,t, which assumes that the cross-covariance between the low-frequency and high-frequency feature activations is zero. No feature-space verification is provided. Please measure the correlation between f_lfc and f_hfc in representative BN layers, or use the exact variance of the sum, and report whether the results change.","section":"Equation (5), Section 3.3"},{"comment":"The EMA coefficient α in Equation (3) is a key hyperparameter of FABN, but its value is never reported anywhere in the main text or supplementary material, and no sensitivity analysis is given. Please report α and show how performance varies with it.","section":"Equation (3), Section 3.3"}],"minor_comments":[{"comment":"There are several typos that should be corrected, including 'informastion' in Section 3.2.1, 'adpated' in the Table 3 caption, 'should should' in Section 3, and 'Domian' in the heading of Supplementary Section G.","section":"Throughout"},{"comment":"The last column header 'A VG' appears to be a typo for 'Avg'; please correct it.","section":"Tables 13 and 14"},{"comment":"Figure 2 is dense and the connection between the frequency-domain filtering and the BN statistics update is hard to follow; labeling the equations and the train/test phases more explicitly would help.","section":"Figure 2"},{"comment":"The terms 'FFC' and 'full-frequency' are used interchangeably; please define the abbreviation at first use to avoid ambiguity.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"The test-set-based selection of the frequency radius is the main methodological concern. If the authors can provide a source-domain-based selection and demonstrate that the headline results are robust across reasonable radius values, I would be willing to reconsider. The claimed novelty as the first method to jointly address QAT and TTA is somewhat strong given prior work on quantized-model robustness, but the empirical integration is substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CoDA is a serious empirical paper with a first-of-its-kind combination: training quantized models on low-frequency components and then adapting BN statistics with frequency-decomposed target features. The experimental coverage is genuinely wide—multiple architectures, bitwidths, corruption benchmarks, continual shift, and an ablation showing both components contribute. The claim that quantized models can match or beat full-precision TTA baselines under 4-16x compression is the kind of result the edge-deployment community cares about.\n\nBut the headline number is softer than it looks. The radius r is chosen based on accuracy on CIFAR10-C and ImageNet-C (Table 2, App. C). That is effectively tuning a hyperparameter on the test distribution used for evaluation. No sensitivity analysis is reported for the full CoDA pipeline as a function of r, so the 5-8%p gains over FP baselines may be optimistic. This is not fatal—the method plausibly still helps for reasonable radii—but it is a real weakness for the abstract's claims.\n\nTwo smaller problems. First, FABN is compared against NORM/TENT/SAR on the FFC-trained model, not on the same LFC-trained model, so the individual contribution of FABN over a simpler TTA method on the same base is not isolated. Second, Eq. 5 adds LFC and HFC means and variances, dropping the cross-covariance term; there is no feature-level check that LFC and HFC activations are approximately uncorrelated. This may be acceptable in practice, but it is unstated.\n\nThe abstract's 'full-precision TTA baseline' is NORM, the weakest of the three compared. Against TENT or SAR the gains shrink to a few points or near zero. That is a presentation issue, not a scientific one, but it should be fixed.\n\nIf I were editor, I would send it to review: the core combination is new and the experiments are extensive enough to justify referee time. I would ask the authors to report r sensitivity, add a same-base TTA comparison for FABN, and clarify which baseline the abstract refers to. For a reader interested in quantized test-time adaptation, this is worth a look; for me, I would wait for the revised version before citing the headline numbers.","headline":"A strong empirical combo of LFC training and frequency-aware BN adaptation for quantized models, but the headline gains rely on test-set-tuned radius and the weakest baseline.","tokens_in":33983,"tokens_out":3860,"would_cite":true,"duration_ms":38157,"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":"CoDA, a frequency-composition training and test-time adaptation pipeline, lets heavily quantized networks match or beat full-precision networks under domain shift.","keywords":["quantization-aware training","test-time adaptation","frequency decomposition","domain shift","batch normalization","model compression","low-frequency components","corruption robustness"],"falsifier":"Compute, on a corrupted test set, the covariance between the low- and high-frequency bands of the BN input features at each layer. If that cross-covariance is large relative to the individual variances, the additive rule $\\hat{\\sigma}_t^2 = \\hat{\\sigma}_{\\mathrm{lfc},t}^2 + \\hat{\\sigma}_{\\mathrm{hfc},t}^2$ drops real signal and FABN should be miscalibrated. A second check: run ordinary prediction-time batch-statistics adaptation without the band split on the same LFC-trained model; if it matches FABN's accuracy, the frequency split is not the active ingredient.","tokens_in":32947,"feed_emoji":"🧠","tokens_out":6789,"duration_ms":65693,"temperature":0.7,"pith_summary":"The paper's claim is that compression and test-time adaptation should be solved as one problem, and that frequency content is the key that ties them together. It proposes CoDA, which restricts quantization-aware training to low-frequency image components and then adapts the compact model at test time with a frequency-aware batch-norm scheme that treats low- and high-frequency features differently. The payoff is empirical: heavily quantized models become more accurate under domain shift than larger full-precision models with standard adaptation. A sympathetic reader would take away the practical prospect of smaller on-device models that are not only cheaper but also more reliable when the deployment environment drifts.","feed_headline":"Low-frequency training lets 2-bit models beat full-precision ones","feed_subtitle":"A training-and-test recipe keeps quantized nets accurate under domain shift while cutting model size 4-16x.","key_machinery":"The mechanism is frequency decomposition: a 2D Fourier transform with radial low-pass and high-pass filters splits each image into $x_{\\mathrm{lfc}}$ and $x_{\\mathrm{hfc}}$ with $x = x_{\\mathrm{lfc}} + x_{\\mathrm{hfc}}$. CoDA trains with LFC-only images during quantization-aware training (LFC QAT), and at test time applies Frequency-Aware Batch Normalization (FABN), which band-pass filters each BN layer's input feature, evolves low-frequency statistics from the source running mean and variance via an exponential moving average, takes high-frequency statistics directly from the current batch, and adds the two. The additive recombination of means and variances is the step that carries the argument.","core_discovery":"The central discovery is that a quantized model trained only on low-frequency image components learns a flatter, more domain-invariant representation, and that at test time adapting a band-split batch normalization, with source-initialized EMA statistics for low frequencies and current-batch statistics for high frequencies, lets a 2-bit model surpass full-precision TTA baselines by up to 7.96 percentage points on CIFAR10-C and 5.37 percentage points on ImageNet-C while using 4 to 16 times fewer bits. The paper presents this as a unified pipeline that wraps existing QAT and TTA methods rather than replacing them.","pith_inferences":["The method's core split could be stress-tested layer by layer: if low- and high-frequency activations are strongly correlated at deeper layers, the additive statistics in Eq. (5) would need cross-covariance corrections, and the advantage should shrink there.","One testable extension is to apply the frequency split to the affine-parameter branch of TTA, giving low- and high-frequency channels independent adaptation rates; the paper's compatibility results suggest this but do not explore it.","LFC QAT's flatter loss landscape hints that low-frequency filtering might serve as a cheap regularizer for quantized training generally, potentially complementing sharpness-based penalties without extra gradient computation."],"forward_implications":["A 2-bit quantized model trained and adapted with CoDA can beat full-precision TTA baselines on CIFAR10-C and ImageNet-C, by up to 7.96 and 5.37 percentage points, while using 4-16x fewer bits.","CoDA is additive rather than competitive: it raises the accuracy of existing QAT methods such as LSQ and LQ, and of TTA methods such as NORM, TENT, and SAR, when combined with any of them.","The improvement transfers to lightweight architectures such as MobileNet-v3 and EfficientNet-b0, which also surpass their full-precision TTA counterparts.","Because the low-frequency statistics remain anchored to the source distribution, CoDA holds up at small test batch sizes where NORM, TENT, and SAR degrade sharply.","Even under continually changing corruptions, CoDA is competitive with CoTTA, a method explicitly designed for continual domain shift."],"supporting_citations":[{"why":"Provides the learned-step-size quantization-aware training method (LSQ) that CoDA wraps and compares against.","marker":"[10]"},{"why":"Provides the learned-quantization QAT method (LQ) used as the second base quantizer.","marker":"[45]"},{"why":"Supplies the CIFAR10-C and ImageNet-C corruption benchmarks on which the headline accuracy gains are measured.","marker":"[17]"},{"why":"Provides the prediction-time batch-normalization TTA baselines whose statistics-update idea FABN extends.","marker":"[28, 31]"},{"why":"Provides the entropy-minimization TTA baseline and the affine-parameter updates with which CoDA is shown to combine.","marker":"[34]"},{"why":"Provides the sharpness-aware TTA baseline (SAR) that CoDA also improves when combined.","marker":"[30]"},{"why":"Supplies the curriculum insight that low-frequency components are learned earlier and transfer better, motivating LFC QAT.","marker":"[39]"},{"why":"Provides evidence that CNNs process low- and high-frequency components differently, grounding the frequency split.","marker":"[36]"}],"fun_headline_variants":["2-bit nets beat full-precision via frequency-split adaptation","Quantized models adapt better: train low-freq, test high-freq","Frequency composition unifies compression and domain adaptation","CoDA: 4-16x smaller nets, up to 8% more accurate under shift","Low-freq training, high-freq test: robust tiny models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a low/high-frequency split that separates general from domain-specific information in input images does the same inside hidden-layer activations, and that adding the two bands' means and variances loses nothing important.","fun_headline_variants_meta":{"raw":{"variants":["2-bit nets beat full-precision via frequency-split adaptation","Quantized models adapt better: train low-freq, test high-freq","Frequency composition unifies compression and domain adaptation","CoDA: 4-16x smaller nets, up to 8% more accurate under shift","Low-freq training, high-freq test: robust tiny models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1642,"prompt_tokens":922,"completion_tokens":720,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":626}},"tokens_in":538,"tokens_out":720,"duration_ms":6170,"temperature":1.0,"reasoning_tokens":626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:45:08.217392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, on a corrupted test set, the covariance between the low- and high-frequency bands of the BN input features at each layer. If that cross-covariance is large relative to the individual variances, the additive rule $\\hat{\\sigma}_t^2 = \\hat{\\sigma}_{\\mathrm{lfc},t}^2 + \\hat{\\sigma}_{\\mathrm{hfc},t}^2$ drops real signal and FABN should be miscalibrated. A second check: run ordinary prediction-time batch-statistics adaptation without the band split on the same LFC-trained model; if it matches FABN's accuracy, the frequency split is not the active ingredient.","supporting_citations":[{"cited_title":"Lq-nets: Learned quantization for highly accurate and compact deep neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the learned-quantization QAT method (LQ) used as the second base quantizer."},{"cited_title":"Efficienttrain: Exploring generalized curriculum learning for training visual backbones","cited_arxiv_id":null,"evidence_quote":"Supplies the curriculum insight that low-frequency components are learned earlier and transfer better, motivating LFC QAT."},{"cited_title":"High-frequency component helps explain the generalization of convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Provides evidence that CNNs process low- and high-frequency components differently, grounding the frequency split."}],"review_version":1}