{"id":"d25851cd-0fc4-4865-b97a-d70d9cb090b2","arxiv_id":"2507.10999","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A lightweight CNN with multi-scale spatial mixing and wave-based channel aggregation achieves competitive ImageNet and COCO accuracy at low parameter counts.","lead":"This paper introduces SpaRTAN, a lightweight convolutional network for image recognition that reaches 77.7% accuracy on ImageNet with only 3.8 million parameters. If the numbers hold, it offers a parameter-efficient alternative to heavier vision models for phones and edge devices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline accuracy/efficiency claim rests on unmatched training recipes and input resolutions; the 256x256 + 300-epoch SpaRTAN numbers are compared to 224x224 published baselines, so the architecture-level advantage is not yet established.","rationale":"The reader's weakest assumption was that training schedule, resolution, and augmentation differences could explain the claimed accuracy/parameter advantages, and my stress-test supports that. I read the paper in good faith: the architecture is described in enough detail to reimplement, the ablations in Tables IV-VIII are internally consistent, and the parameter counts are plausible. The localized concern is external validity: Table II mixes the authors' 256x256 results with published 224x224 baselines, and the paper's 300-epoch recipe with strong augmentation is not the recipe used for most cited baselines. The same issue propagates to COCO because the SpaRTAN backbones carry the stronger ImageNet pretraining while the ResNet RT-DETR baselines use standard pretraining. These are not internal contradictions, but they are direct threats to the strength of the headline claim, whose margins (0.5 top-1, 1.2 AP) are comparable to the likely recipe/resolution effects. I do not see grounds to reject: the paper contains controlled internal ablations, and the released code can settle the comparison. I therefore keep the reader's CONDITIONAL verdict; the condition should be a matched-recipe re-benchmark of the baselines at matched resolutions, plus reproduction from the code release.","tokens_in":13313,"tokens_out":8292,"duration_ms":96124,"concrete_test":"Using the released code, reproduce SpaRTAN-T at 224 and 256, then train MogaNet-XT and ConvNeXt-XT under the exact SpaRTAN recipe (300 epochs, AdamW, batch 2048, RandAugment mag 7, Mixup 0.2, CutMix, Random Erasing, Stochastic Depth) at both resolutions and evaluate all models at both resolutions. If either baseline reaches 77.7% at 256 or 77.1% at 224, the claimed architecture-level advantage is not established. Separately, pretrain ResNet-18 and ResNet-34 with the same ImageNet recipe and plug them into RT-DETR; if ResNet-34 reaches 50.0 AP, the COCO gain is explained by pretraining rather than by SpaRTAN.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract and Table II is that SpaRTAN-T reaches 77.7% at 3.8M/1.0 G with notable parameter efficiency. The load-bearing assumption is that the comparison against published baselines is controlled. It is not: SpaRTAN-T is quoted at 256x256 input (1.08 GFLOPs), while most baselines in Table II are at 224x224, and SpaRTAN uses a 300-epoch recipe with RandAugment, Mixup, CutMix, Random Erasing, and Stochastic Depth that most cited baselines did not use. At 224x224 SpaRTAN-T is 77.1%; the extra 0.6 points quoted in the abstract come from the higher-resolution and higher-cost setting. The COCO comparison in Table III has the same structure: SpaRTAN backbones are pretrained with this strong 300-epoch recipe, whereas the published ResNet RT-DETR numbers use standard, much shorter ImageNet pretraining. The margins at issue (0.5 on ImageNet vs MogaNet-XT at 256, 1.2 AP vs ResNet-34 on COCO) are the same order of magnitude as the unstated recipe/resolution effects, so the architecture-level claim is not yet isolated. The internal ablations (Table IV) give some support but only against the authors' own baseline and cannot validate the absolute comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SpaRTAN, a lightweight CNN architecture composed of two modules: a spatial SMixer that combines small and large effective receptive fields via stacked convolutions with different dilations, and a wave-based CMixer that treats channels as real/imaginary wave components and aggregates them through a superposition mechanism guided by a maximally activated channel. The authors report 77.7% ImageNet-1k top-1 accuracy with 3.8M parameters and about 1.0 GFLOPs, and 50.0% COCO AP when used as an RT-DETR backbone with 21.5M parameters, claiming superior parameter efficiency over existing lightweight models. The paper includes architectural ablations, Grad-CAM visualizations, and a public code release.","tokens_in":13620,"tokens_out":3670,"duration_ms":44142,"significance":"If the accuracy and efficiency claims survive controlled comparison, SpaRTAN would be a genuinely parameter-efficient convolutional design, especially because it avoids large kernels and high MLP expansion ratios. The paper's strengths include a clear engineering motivation, internal ablation studies that support the module-level choices (Table IV through Table VIII), a public code release, and evaluation on two standard benchmarks. However, the central quantitative claims currently rest on comparisons against baselines trained at different input resolutions and with different training recipes; this fairness gap is the main barrier to accepting the architecture-level conclusions.","major_comments":[{"comment":"The headline result of 77.7% top-1 accuracy with 3.8M parameters and \"approximately 1.0 GFLOPs\" corresponds to SpaRTAN-T evaluated at 256x256 input (1.08 GFLOPs), whereas most baselines in Table II are reported at 224x224 input. At 224x224 the same model achieves 77.1% with 0.83 GFLOPs. The abstract's presentation is therefore misleading without specifying the input resolution. More importantly, the comparison is not controlled for training recipe: SpaRTAN is trained for 300 epochs with RandAugment, Mixup, CutMix, Random Erasing, and Stochastic Depth, while many cited baselines (e.g., ResNet-18, PVT-T, ConvNeXt-XT) use shorter or different schedules. Since the margins over MogaNet-XT and ConvNeXt-XT are 0.5 and 0.2 points respectively, these unstated differences could account for the advantage. Please either retrain key baselines under the same recipe, or substantially qualify the comparison claims.","section":"Section IV-A.2, Table II, Abstract"},{"comment":"The COCO comparison compares SpaRTAN backbones pretrained on ImageNet for 300 epochs with a strong augmentation recipe against RT-DETR ResNet-18/34 baselines that use standard (much shorter) ImageNet pretraining as published in [38]. The detector training is matched, but the backbone pretraining is not. The reported 1.2 AP improvement over ResNet-34 and 2.1 AP over ResNet-18 are therefore confounded by pretraining schedule and augmentation. Please retrain the ResNet backbones with the same 300-epoch recipe, or present a controlled comparison where only the architecture differs. Without this, the claim that the backbone architecture is responsible for the detection gains is not established.","section":"Section IV-B, Table III"},{"comment":"The wave-based CMixer is not specified at the level needed for reproduction. The text explains the conceptual wave interpretation, but the actual forward computation of W(·) is left underspecified: the definition of Fmax as \"the maximum value of C\" is ambiguous (presumably a channel-wise max across the channel dimension?), the split of channels into sine and cosine halves is not formalized, the role of the complex weight Wc in the actual tensor operations is not written out, and the promised \"linear approximation of sinusoidal waves using a point convolution with a non-linear activation function\" is not described concretely. The paper should provide a precise layer-by-layer definition of the wave-based aggregation module, or a pseudocode block, to make the contribution self-contained despite the public code.","section":"Section III-C, Equations (7)-(9)"}],"minor_comments":[{"comment":"The text contains a typographical spacing error in \"77. 7%\" which should be \"77.7%\".","section":"Abstract"},{"comment":"The column headers \"APval APval 50 APval 75 APval S APval M APval L\" are unclear due to missing subscripts and spacing; please reformat them as AP, AP50, AP75, APS, APM, APL.","section":"Table III"},{"comment":"The augmentation \"Random Resized Crop\" should be typeset consistently with the standard name \"RandomResizedCrop\" to avoid ambiguity.","section":"Section IV-A.1"},{"comment":"The table reports kernel sizes 3 and 5, but the text refers to replacing a 5x5 dilation-2 convolution with stacked 3x3 dilation-2 convolutions; please clarify that the \"5\" entry denotes the single 5x5 kernel and the \"3\" entry denotes the stacked 3x3 replacement.","section":"Table V"},{"comment":"The feature maps FH and FL are introduced in the text as outputs of the high- and low-frequency branches, but they are not formally defined before appearing in Equation (6); please add explicit definitions.","section":"Section III-B, Equation (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-written and the internal ablations are consistent, but the headline comparisons are not yet controlled for training recipe and input resolution. If the authors can provide same-recipe baseline comparisons or clearly delimit the claims to their specific training setup, the contribution would be publishable. The novelty of the wave-based CMixer is moderate but sufficient for a venue accepting incremental architectural improvements; the main risk is that the reported margins may shrink or disappear under controlled comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: SpaRTAN is a sensible lightweight CNN backbone that combines a two-branch spatial mixer with a wave-motivated channel mixer. The engineering is careful and the ablations are informative. But the headline numbers compare a 256x256, 300-epoch model against published 224x224 baselines, so the claimed efficiency advantage is not yet established.\n\nWhat's genuinely good: The architecture is described in enough detail to reimplement. The SMixer's stacked 3x3 dilated convolutions are a reasonable stand-in for a 5x5 kernel, and Table V shows the trade-off. The wave-based CMixer is a fresh application of complex-weight modulation to channels, and the ablation in Table IV shows it adds 0.9% over a vanilla MLP and 1.0% over MLP+SE. The paper also reports throughput in Table VIII, which too many architecture papers skip. Code is promised on GitHub. That counts.\n\nSoft spots, in order of importance. First, the comparison is uncontrolled. The abstract's 77.7% is at 256x256, and the paper says 'unless stated otherwise' 224, but the abstract omits the exception. At 224, SpaRTAN-T gets 77.1, which still beats ConvNeXt-XT and matches EfficientNet-B0, but the FLOPs advantage shrinks (0.83 vs 0.39 for EfficientNet). The COCO comparison uses SpaRTAN backbones pretrained with the same 300-epoch recipe, while the ResNet baselines are the published RT-DETR numbers, which use standard pretraining. That 1.2 AP gap could partly be recipe. Second, all results are single runs, so the 0.5-point margins over MogaNet-XT are noise-level. Third, the novelty claim that this is the first to treat channels as waves is hard to defend when Wave-MLP already treats image patches as waves and uses complex weights; the channel-specific application is new but the framing oversells it.\n\nI don't think these problems sink the paper. The ablations are against the authors' own baseline and they show consistent gains. But the absolute numbers should be presented with the resolution caveat, and the authors should either train baselines under their own recipe or report results at matching resolutions. The paper deserves a serious referee, and a good reviewer will push on the experimental protocol rather than the architecture.\n\nRecommendation: send to peer review with a request for controlled comparisons and multi-seed results. I'd bring it to reading group as a case study in benchmarking hygiene.","headline":"A capable lightweight backbone whose reported gains are real but not yet isolated from resolution and training-recipe differences.","tokens_in":14174,"tokens_out":2525,"would_cite":false,"duration_ms":28564,"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 lightweight CNN with 3.8M parameters reaches 77.7% top-1 accuracy on ImageNet-1k and, as a detection backbone, lifts RT-DETR to 50.0% AP on COCO.","keywords":["lightweight CNN","multi-order spatial interactions","channel aggregation","wave-based mixer","image classification","object detection","parameter efficiency"],"falsifier":"Train SpaRTAN-T and the leading baselines (e.g., ConvNeXt-XT, MogaNet-XT, Swin-1G) under identical conditions with 224x224 input and the same 300-epoch recipe, then check whether SpaRTAN-T still matches or beats their published accuracy at equal or lower parameter counts. A failure to maintain the advantage would indicate that the reported gains stem from training details rather than the proposed modules.","tokens_in":13108,"feed_emoji":"📊","tokens_out":2076,"duration_ms":24159,"temperature":0.7,"pith_summary":"SpaRTAN is a convolutional architecture that aims to close the accuracy gap between small, efficient networks and much larger models by explicitly capturing mid-order spatial interactions. Its two core modules are a spatial SMixer that combines different kernel sizes and dilations to extract both fine-grained and global features, and a wave-based CMixer that treats channel features as superposed waves, reinforcing channels aligned with the most informative ones. The paper reports that SpaRTAN-T achieves 77.7% top-1 accuracy on ImageNet-1k with only 3.8M parameters and roughly 1.0 GFLOPs, and that it outperforms the ResNet-34-based RT-DETR by 1.2% AP on COCO while using fewer parameters. The significance is that a purely convolutional, small-kernel design can rival attention-based and large-kernel models without heavy parameter or compute cost.","feed_headline":"A 3.8M-parameter CNN hits 77.7% on ImageNet","feed_subtitle":"SpaRTAN uses stacked convolutions and a wave-based channel mixer to beat much larger models with fewer parameters.","key_machinery":"The central mechanism is a two-module block. The spatial SMixer decomposes the input into fine-grained and global feature streams, then applies convolutions with different receptive fields (a 3x3 convolution and a dilated 3x3 convolution with dilation 2, equivalent to a 5x5 effective field) to capture multi-order spatial information. The wave-based CMixer reformulates each channel as a complex-valued wave, expressed via Euler's formula, and applies a superposition mechanism that amplifies channels aligned with the maximally activated channel while suppressing trivial ones; a complex weight modulates the waves, enabling dynamic context-aware channel aggregation. Together these modules replace the standard MLP-based mixer in modern CNN blocks, allowing a lower expansion ratio and better parameter efficiency.","core_discovery":"The paper claims that a carefully designed convolutional network can extract the middle-order spatial interactions that both CNNs and transformers tend to undervalue, and that doing so yields a superior accuracy-efficiency trade-off. The discovery is demonstrated by SpaRTAN, which uses stacked 3x3 convolutions with dilation to cover multiple receptive fields and a channel-aggregation module modeled on wave superposition. On ImageNet-1k, SpaRTAN-T reaches 77.7% at 256x256 input with 3.8M parameters, and SpaRTAN-XT reaches 74.4% with only 2.2M parameters. As the backbone for RT-DETR, SpaRTAN-T achieves 50.0% AP on COCO, exceeding the ResNet-34 variant by 1.2% while using 10M fewer parameters. The authors attribute these gains to the combination of the spatial SMixer and wave-based CMixer, which together improve parameter utilization and reduce channel-wise redundancy.","pith_inferences":["The paper's implicit claim that multi-order interactions are the cause of the accuracy gains is not directly evidenced; the gains could also stem from the specific training recipe, data augmentation, or the hybrid convolution strategy.","A direct comparison with baselines trained under identical settings (same resolution, same epochs, same augmentation) would be a stronger test of whether the architecture itself, rather than the training setup, is responsible for the improvements.","The wave-based CMixer's reliance on the maximally activated channel as a reference point could be sensitive to outlier channels; testing on adversarial or distorted inputs might reveal failure modes.","If the middle-order interaction hypothesis is correct, SpaRTAN's design could be combined with other lightweight modules or formally analyzed to quantify interaction order, but the paper does not provide such a mechanism."],"forward_implications":["If the reported results are correct, a small-kernel convolutional architecture can match or beat attention-based and large-kernel transformers in both classification and detection at a fraction of the parameters and FLOPs.","The wave-based channel aggregation suggests that modeling channels as waves with learnable complex weights is a viable alternative to high-expansion-ratio MLPs, reducing redundancy without sacrificing accuracy.","The architecture's efficiency at low parameter counts could make it applicable to mobile and embedded settings where both accuracy and latency matter.","The claimed mid-order interaction capture implies that such networks may be more robust to occlusion and background clutter, as suggested by the Grad-CAM visualizations."],"supporting_citations":[{"why":"ResNet-18 is the baseline that SpaRTAN compares against in ImageNet classification and serves as a parameter/FLOP benchmark.","marker":"[3]"},{"why":"EfficientNet-B0 is a competing efficient architecture that SpaRTAN-T outperforms while using fewer parameters.","marker":"[4]"},{"why":"ConvNeXt motivates the modern CNN design and provides a strong baseline (ConvNeXt-XT) on ImageNet.","marker":"[13]"},{"why":"HorNet's recursive gated convolutions are the starting point for the multi-order spatial interaction concept that SpaRTAN modifies.","marker":"[17]"},{"why":"MogaNet provides the feature decomposition module that SpaRTAN adapts and the multi-order gated aggregation baseline.","marker":"[18]"},{"why":"Wave-MLP is the source of the wave/phase formulation that SpaRTAN's channel mixer builds upon.","marker":"[21]"},{"why":"Squeeze-and-Excitation blocks are used in the SMixer and CMixer for channel recalibration.","marker":"[31]"},{"why":"RT-DETR is the object detection framework used to evaluate SpaRTAN as a backbone, with ResNet-18 and ResNet-34 baselines.","marker":"[38]"}],"fun_headline_variants":["SpaRTAN: 3.8M params, 77.7% ImageNet, 50% COCO AP","Wave-based channel mixing makes a tiny CNN competitive","SpaRTAN: A 3.8M-parameter CNN that beats ResNet34 on COCO","Multi-order spatial features: how a small CNN gets big accuracy","77.7% ImageNet with 3.8M params? Meet SpaRTAN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's headline results are only meaningful if the comparison against published baselines is fair; the main risk is that SpaRTAN's 256x256 input, 300-epoch training schedule, and strong augmentation recipe, rather than the architecture itself, are what drive the accuracy and parameter-efficiency advantages.","fun_headline_variants_meta":{"raw":{"variants":["SpaRTAN: 3.8M params, 77.7% ImageNet, 50% COCO AP","Wave-based channel mixing makes a tiny CNN competitive","SpaRTAN: A 3.8M-parameter CNN that beats ResNet34 on COCO","Multi-order spatial features: how a small CNN gets big accuracy","77.7% ImageNet with 3.8M params? Meet SpaRTAN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000982,"raw_usage":{"total_tokens":4216,"prompt_tokens":1044,"completion_tokens":3172,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":3059}},"tokens_in":660,"tokens_out":3172,"duration_ms":25757,"temperature":1.0,"reasoning_tokens":3059,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:19:58.777401+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SpaRTAN-T and the leading baselines (e.g., ConvNeXt-XT, MogaNet-XT, Swin-1G) under identical conditions with 224x224 input and the same 300-epoch recipe, then check whether SpaRTAN-T still matches or beats their published accuracy at equal or lower parameter counts. A failure to maintain the advantage would indicate that the reported gains stem from training details rather than the proposed modules.","supporting_citations":[{"cited_title":"A convnet for the 2020s,","cited_arxiv_id":null,"evidence_quote":"ConvNeXt motivates the modern CNN design and provides a strong baseline (ConvNeXt-XT) on ImageNet."},{"cited_title":"Hornet: Efficient high-order spatial interactions with recursive gated convolu- tions,","cited_arxiv_id":null,"evidence_quote":"HorNet's recursive gated convolutions are the starting point for the multi-order spatial interaction concept that SpaRTAN modifies."},{"cited_title":"Moganet: Multi-order gated aggregation network,","cited_arxiv_id":null,"evidence_quote":"MogaNet provides the feature decomposition module that SpaRTAN adapts and the multi-order gated aggregation baseline."},{"cited_title":"Detrs beat yolos on real-time object detection,","cited_arxiv_id":null,"evidence_quote":"RT-DETR is the object detection framework used to evaluate SpaRTAN as a backbone, with ResNet-18 and ResNet-34 baselines."}],"review_version":1}