{"id":"1c106836-60fd-4610-a4bd-4d8ecff786f5","arxiv_id":"2411.18995","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MVFormer combines a weighted blend of three normalizations with a three-branch multiscale convolutional token mixer, achieving modest top-1 accuracy gains on ImageNet-1K and downstream vision tasks.","lead":"This paper proposes MVFormer, a vision transformer that mixes batch, layer, and instance normalization and uses a three-scale convolutional token mixer to improve efficiency and accuracy. The authors report small gains over previous convolution-based transformers across image classification and dense prediction tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on 0.1-0.4% top-1 margins with no code, no seeds, and an unablated ResScale addition; a controlled multi-seed rerun is needed.","rationale":"The reader's weakest assumption is exactly the load-bearing point: the empirical SOTA claim is built on differences of 0.1-0.4 top-1 points, and the paper provides no repeated-seed variance, no code release, and no ablation for the extra ResScale component. I agree with the reader. I also checked for other potential objections. The MVN module is closely related to Switchable Normalization, which is not cited in Section 2.1; this weakens the novelty claim but does not by itself falsify the empirical performance claim. The pseudo-code in Algorithms 1-2 is not directly runnable as written (e.g., nn.BatchNorm with a tuple argument, nn.Linear applied to (B,C,H,W) tensors), which is a reproducibility concern rather than a correctness concern. The abstract's blanket phrase 'same or lower parameters and MACs' is not literally true for MVFormer-B (12.7G) versus VAN-B4 (12.2G) in Table 2, so the claim should be scoped to comparable or ConvFormer-style baselines. These secondary issues reinforce a CONDITIONAL verdict rather than a rejection. The decisive check is a controlled multi-seed rerun with and without ResScale; until then the central claim is plausible but unverified.","tokens_in":17271,"tokens_out":9343,"duration_ms":86351,"concrete_test":"Run MVFormer-T/S/B and ConvFormer-S18/S36/M36 in the same codebase with the same data pipeline, optimizer, 300-epoch schedule, augmentation, and with ResScale either included for all models or removed for all models, for at least 3 seeds each; report mean +/- std of top-1 per size. Treat the SOTA claim as established only if the MVFormer mean exceeds the ConvFormer mean by more than the pooled standard error at all three sizes, or if code and pretrained weights are released so this comparison can be performed independently.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MVFormer-T/S/B beat ConvFormer-S18/S36/M36 by 0.4/0.2/0.1 top-1 points at equal or lower parameters/MACs (Table 2). The paper reports one run per configuration, gives no confidence intervals, and releases no code. At this performance level, ImageNet-1K top-1 differences are known to shift by 0.1-0.3% with batch size, stochastic depth, augmentation, and small training-recipe changes. The paper also adds ResScale for the last two stages (Section 4.1) without ablating it, and it does not state whether the ConvFormer baseline numbers were retrained under the exact MVFormer recipe or taken from the original papers. If the baselines were not run under identical conditions, the reported margins may reflect recipe differences rather than the proposed modules. Since the entire SOTA conclusion depends on these small margins, the claim is load-bearing and currently unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MVFormer, a convolution-based vision transformer built on the MetaFormer scheme, with two new components: Multi-View Normalization (MVN), which computes a learnable weighted sum of batch-, layer-, and instance-normalized features, and Multi-View Token Mixer (MVTM), a multiscale depthwise convolution with local, intermediate, and global filters whose channel split ratios and global kernel sizes vary by stage. The model is evaluated on ImageNet-1K classification (83.4/84.3/84.6% top-1 for T/S/B variants), COCO object detection and instance segmentation, and ADE20K semantic segmentation, reporting small gains over ConvFormer baselines at matched or lower parameters and MACs. Ablations on a smaller xT model support the individual contributions of MVN and MVTM, and the learned normalization weights are analyzed.","tokens_in":17466,"tokens_out":7974,"duration_ms":62900,"significance":"If reproducible, the paper offers a modest architectural insight: combining multiple normalization views with stage-specific multiscale convolution can improve the efficiency-accuracy trade-off of convolution-based ViTs. The MVN module's consistent +0.2% improvement when plugged into several existing architectures (Table 7) is a useful empirical result, and the ablations in Tables 5-8 are fairly detailed. However, the headline SOTA claim rests on single-run differences of 0.1-0.4% top-1 in Table 2, with no variance estimates or released code; this currently limits the significance to a suggestive result rather than a demonstrated improvement.","major_comments":[{"comment":"The central claim that MVFormer-T/S/B outperforms ConvFormer-S18/S36/M36 relies on top-1 margins of 0.4%, 0.2%, and 0.1%, respectively. The paper reports a single run per configuration and provides no confidence intervals, multiple seeds, or released code. At this performance level, ImageNet-1K top-1 accuracy is known to vary by roughly 0.2-0.3% run-to-run under the DeiT-style recipe used here, so some of the reported margins are within the noise level. The authors should report mean +/- standard deviation over at least three independent runs for the main comparisons, or release code and checkpoints so that the results can be independently reproduced; without this, the 'outperforms SOTA' claim is not yet supported.","section":"Section 4.1, Table 2"},{"comment":"It is not stated whether the ConvFormer and other baseline numbers in Table 2 were retrained under the exact MVFormer recipe (300 epochs, batch size 4096, learning rate 4e-3, stochastic depth rates, RandAugment, Mixup/CutMix, and ResScale for the last two stages). If the baselines are taken from their original papers, differences in training recipes can easily shift top-1 accuracy by more than the reported margins; batch size and stochastic depth changes alone are known to move ImageNet top-1 by 0.1-0.3%. The authors must either retrain the relevant baselines under the same codebase and hyperparameters, or clearly state that published numbers are quoted and justify that the comparison is controlled.","section":"Section 4.1, Table 2"},{"comment":"The recipe for the main MVFormer models includes ResScale in the last two stages, but no ablation isolates its contribution. The ablation study in Table 5 uses the MVFormer-xT model with the training configuration of Table 10, which does not list ResScale; therefore the gains of the full models in Table 2 relative to ConvFormer could be partly due to ResScale rather than to the proposed MVN and MVTM modules. The authors should either add an ablation of ResScale on the same models, or remove it from the main comparison.","section":"Section 4.1 and Section 4.4"},{"comment":"The PyTorch-style pseudocode contains API errors: `nn.BatchNorm(affine_shape, affine=False)` and `nn.InstanceNorm(affine_shape, affine=False)` pass a shape where the number of features is expected, and `nn.LayerNorm(affine_shape, elementwise_affine=False)` may also be incorrect if `affine_shape` is a tuple. Since no code is released, this pseudocode is the only implementation reference and should be corrected, or the actual code should be provided.","section":"Appendix F, Algorithms 1 and 2"}],"minor_comments":[{"comment":"There is a typo, 'convonlution-based', which should be 'convolution-based'.","section":"Section 3.2.2"},{"comment":"The 'Global kernel size' row is garbled; it should be formatted to clearly show the per-stage values (e.g., 55x1/1x55, 27x1/1x27, 13x1/1x13, 7x7).","section":"Table 1"},{"comment":"The terms 'state-of-the-art' and 'SOTA' are used despite the absence of error bars and the small margins; consider softening to 'competitive' unless statistical significance is demonstrated.","section":"Abstract and Section 1"},{"comment":"The sentence 'MVN displays impressive generalization, significantly improving the original performance' uses 'significantly' without any statistical test; 'consistently improves by 0.2%' would be more accurate.","section":"Section 4.4.3"},{"comment":"The claim that MVN is 'the first study of a normalization integration paradigm in ViTs' is strong; please either temper it or support it with a more comprehensive review of related normalization-combination work.","section":"Section 2.1"},{"comment":"The main text says EMA is used when fine-tuning to 384x384 resolution, but Table 10 lists 'EMA decay rate None' in the training configuration; please clarify that EMA is applied only during fine-tuning.","section":"Section 4.1 and Table 10"}],"recommendation":"major_revision","confidential_remarks":"The paper is competently written and the ablations are fairly thorough, but the SOTA claim rests on very small margins without variance estimates or code. If the authors can supply multi-seed results and an ablation of ResScale, or retrain baselines under equal conditions, the paper may be acceptable. The current version, however, does not convincingly support the central 'outperforms' claim. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the headline is more incremental than the abstract suggests: MVN is mathematically Switchable Normalization (Luo et al. 2018) — a learnable weighted sum of BN, LN, and IN — which is not cited, and MVTM is a three-branch version of existing multiscale depthwise mixers like InceptionNeXt and SegNeXt. The stage-specific kernel and channel-split schedule is new but is essentially a hand-set hyperparameter. Second, the experiments are competent and well-executed, but the central SOTA claim over ConvFormer rests on 0.1–0.4% top-1 margins from single runs, with no code and no variance estimates. At that scale, ImageNet noise and training-recipe differences can flip the ordering.\n\nWhat the paper does well: the empirical coverage is solid — classification, detection, and segmentation, plus ablations and a generalization check of MVN on Swin, ConvNeXt, PoolFormer, and ResNet. The gains are consistently positive, small but real. The MVTM ablation (Table 8) gives a sensible decomposition of stage-specific split and global kernel adjustments. The learned-weight analysis (Fig. 3) is a nice sanity check that LN dominates, which fits prior experience.\n\nSoft spots, in order of importance. (1) The missing SN citation is a genuine novelty problem. The authors claim \"first normalization integration paradigm in ViTs,\" which is not accurate. They need to cite SN and show a head-to-head comparison. (2) The SOTA claim is load-bearing and unverified: no multi-seed runs, no error bars, no code. The paper doesn't say whether ConvFormer baselines were retrained under the exact MVFormer recipe (DeiT-style, ResScale for last two stages, 300 epochs, etc.) or taken from the original papers. ResScale is itself an unablated addition. If the baselines are published numbers, the margins may be recipe differences, not module effects. (3) Stage-specific configurations were presumably tuned on ImageNet validation; that's mild selection, but worth acknowledging. None of these are fatal to the engineering contribution; the method is plausible and the ablations internally support it. They just don't support the strength of the stated claim.\n\nWho this is for: people working on efficient convolutional ViT designs who want an incremental accuracy bump and a clear template for combining normalization and token mixing. It deserves a serious referee, but a responsible one should demand code, multi-seed numbers, identical-condition baselines, and a Switchable Normalization comparison before accepting the SOTA sentence.","headline":"A well-run incremental engineering paper whose headline SOTA claim rests on single-run margins; MVN is Switchable Normalization restated, and the missing citation needs fixing.","tokens_in":18045,"tokens_out":2655,"would_cite":false,"duration_ms":28069,"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":"Three normalization views push vision transformers past conv-based rivals at equal cost.","keywords":["vision transformer","multi-view normalization","token mixing","convolution-based ViT","MetaFormer","image classification","object detection","semantic segmentation"],"falsifier":"Train MVFormer-T and ConvFormer-S18 from scratch on ImageNet-1K with the same code, the same DeiT-style recipe, and at least five random seeds, then compare the mean top-1 accuracy with confidence intervals. If the 95% confidence intervals overlap or the ordering flips, the central superiority claim is falsified.","tokens_in":1274,"feed_emoji":"🖼️","tokens_out":3304,"duration_ms":51334,"temperature":0.7,"pith_summary":"This paper argues that efficient vision transformers leave a resource on the table in the normalization step. It proposes MVFormer, which replaces the single normalization inside each MetaFormer block with a learnable weighted blend of batch, layer, and instance normalization, and replaces the token mixer with a three-scale depthwise convolution whose receptive fields shrink and shift across stages. On ImageNet-1K, MVFormer-T/S/B reach 83.4%, 84.3%, and 84.6% top-1 accuracy, edging out the previous convolution-based state of the art by 0.1 to 0.4 percentage points with equal or fewer parameters and MACs, and it also reports higher mAP on COCO object detection and instance segmentation and higher mIoU on ADE20K semantic segmentation under similar budgets. The paper's central claim is that diversifying both the normalization view and the mixing scale is what buys these gains.","feed_headline":"Three-way normalization lifts ViTs past conv nets at same cost","feed_subtitle":"Blending batch, layer, and instance norm plus a three-scale mixer reaches 84.6% top-1 on ImageNet.","key_machinery":"The central object is the MVFormer block, a MetaFormer block in which the token-mixer and MLP subblocks both use multi-view normalization (MVN) as their normalization layer. MVN computes three differently normalized copies of the input (batch, layer, and instance) and combines them with learnable per-channel weights before a single affine transform, giving the token mixer access to batch-level, channel-level, and sample-level statistical views at negligible parameter cost. The token mixer, MVTM, is an inverted separable convolution with the depthwise layer split into three channel groups of different kernel sizes; stage specificity sets the channel split ratio (from 50:50:0 in stage 1 to 0:50:50 in stage 4) and shrinks the global kernel size (55x55 down to 7x7) across stages, steering early stages toward local mixing and later stages toward global mixing.","core_discovery":"MVFormer outperforms state-of-the-art convolution-based vision transformers on image classification, object detection, instance segmentation, and semantic segmentation while staying within the same or lower parameter and MAC budgets. The design combines two ideas: multi-view normalization (MVN), a learnable weighted sum of batch-, layer-, and instance-normalized features within every MetaFormer subblock, and a multi-view token mixer (MVTM), a channel-wise depthwise convolution split into local (3x3), intermediate (7x7), and global filters whose channel ratios and global kernel sizes are tuned per stage.","pith_inferences":["Because MVN also improves CNNs like ResNet50, the principle of mixing normalizations may be a general training stabilizer that does not depend on the token-mixing mechanism, so it could be tested in pure attention-based ViTs as well.","The derived benefit may stem mostly from the LN component, which dominates the learned weights in most blocks (per Figure 3); a controlled study that freezes LN and only adds a small BN or IN branch could isolate whether the synergy is genuinely mutual.","The stage-specific global kernel in stage 1 (55x55 on a 56x56 feature map) is effectively whole-image mixing, yet later stages use 7x7 on a 7x7 feature map, meaning the static convolution still cannot perform content-dependent global interactions; replacing the global filter with a lightweight attention mechanism in the last stage might further push efficiency and accuracy."],"forward_implications":["If the central claim holds, MVN is a cheap drop-in module: it improved top-1 accuracy by 0.2 percentage points when substituted for layer norm in Swin-T, ConvFormer-S18, ConvNeXt-T, and PoolFormer-S36, and by 0.2 points when substituted for batch norm in ResNet50.","MVTM's three-scale split is responsible for reliable gains beyond the baseline, and both stage-specific channel ratios and stage-specific global kernel sizes contribute independently; removing either or both degrades ImageNet top-1 accuracy by 0.08 to 0.15 points.","The strong downstream results imply the normalization and mixing diversity transfers well to dense prediction, not just classification: MVFormer-T under Mask R-CNN 1x reaches 46.2 box AP and 42.1 mask AP with fewer parameters and MACs than prior backbones.","At 384x384 fine-tuning, all three MVFormer variants keep or extend their advantage over ConvFormer, suggesting the benefits are not specific to 224x224 training.","The paper's ablation shows instance normalization alone hurts (78.83% top-1) but improves when combined with batch or layer norm, so the synergy among normalizations is what MVN is designed to exploit."],"supporting_citations":[{"why":"Supplies the ConvFormer baseline and the MetaFormer block architecture that MVFormer modifies","marker":"[59]"},{"why":"Provides the DeiT-style training recipe (300 epochs, augmentations, optimizer settings) used for ImageNet-1K experiments","marker":"[43]"},{"why":"Defines the inverted separable convolution module that MVTM builds on as its token-mixing backbone","marker":"[35]"},{"why":"Establishes the MetaFormer abstraction and the observation that local token mixing is preferred in early stages and global mixing in later stages","marker":"[58]"},{"why":"Defines batch normalization, one of the three normalization views in MVN","marker":"[18]"},{"why":"Defines layer normalization, the other primary normalization view in MVN","marker":"[1]"},{"why":"Defines instance normalization, the third view in MVN that contributes most in later stages","marker":"[45]"},{"why":"Introduces multiscale convolutional token mixing with parallel mixing paths, which MVTM extends to three scales","marker":"[11]"}],"fun_headline_variants":["MVFormer: Three norms, three scales beat conv ViTs","Three-way norm + multiscale mixer beats conv ViTs","Multi-view norms and token mixing outdo conv ViTs","MVN+MVTM: same cost, higher accuracy for ViTs","ViT with three-view norms and three-size mixer tops conv rivals"],"cache_read_input_tokens":20224,"weakest_assumption_plain":"The claim that MVFormer outperforms prior convolution-based vision transformers depends on the assumption that the reported accuracy gaps of 0.1 to 0.4 percentage points over the closest baselines are larger than run-to-run training noise from a single training run.","fun_headline_variants_meta":{"raw":{"variants":["MVFormer: Three norms, three scales beat conv ViTs","Three-way norm + multiscale mixer beats conv ViTs","Multi-view norms and token mixing outdo conv ViTs","MVN+MVTM: same cost, higher accuracy for ViTs","ViT with three-view norms and three-size mixer tops conv rivals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001131,"raw_usage":{"total_tokens":4689,"prompt_tokens":925,"completion_tokens":3764,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":3676}},"tokens_in":541,"tokens_out":3764,"duration_ms":26389,"temperature":1.0,"reasoning_tokens":3676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:38:46.466170+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MVFormer-T and ConvFormer-S18 from scratch on ImageNet-1K with the same code, the same DeiT-style recipe, and at least five random seeds, then compare the mean top-1 accuracy with confidence intervals. If the 95% confidence intervals overlap or the ordering flips, the central superiority claim is falsified.","supporting_citations":[{"cited_title":"Metaformer baselines for vision, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the ConvFormer baseline and the MetaFormer block architecture that MVFormer modifies"},{"cited_title":"Training data-efficient image transformers & distillation through at- tention","cited_arxiv_id":null,"evidence_quote":"Provides the DeiT-style training recipe (300 epochs, augmentations, optimizer settings) used for ImageNet-1K experiments"},{"cited_title":"Mobilenetv2: Inverted residuals and linear bottlenecks, 2019","cited_arxiv_id":null,"evidence_quote":"Defines the inverted separable convolution module that MVTM builds on as its token-mixing backbone"},{"cited_title":"Metaformer is actually what you need for vision","cited_arxiv_id":null,"evidence_quote":"Establishes the MetaFormer abstraction and the observation that local token mixing is preferred in early stages and global mixing in later stages"},{"cited_title":"In- stance normalization: The missing ingredient for fast styliza- tion, 2017","cited_arxiv_id":null,"evidence_quote":"Defines instance normalization, the third view in MVN that contributes most in later stages"},{"cited_title":"Segnext: Rethinking convolutional attention design for semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Introduces multiscale convolutional token mixing with parallel mixing paths, which MVTM extends to three scales"}],"review_version":1}