{"id":"3cd3b6f4-ed8e-4397-bd08-fd201bfe3fba","arxiv_id":"2412.19433","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A hand-designed CNN block with multi-scale convolutions, residual connections and cross-branch feature reuse is reported to reach high accuracy on small image benchmarks.","lead":"This paper proposes two new convolutional blocks, ResFRI and Split-ResFRI, that combine inception-style multi-scale convolutions with residual connections and cross-branch feature passages. The authors report high accuracy on CIFAR-10 (97.94%), CIFAR-100 (85.91%) and Tiny ImageNet (70.54%) with small parameter counts, but the exact architecture and training settings are not fully specified.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim is undermined by test-set-selected pruning ratios and an under-specified architecture; reproduction with validation-based selection is needed.","rationale":"I read the paper as an empirical architecture paper whose central claim is SOTA image classification at matched model size. The reader's REJECT is supported: the comparison includes misattributed baselines (e.g., Table 1 lists PyramidNet under ref [12] which is Feature Pyramid Networks, and WRN-28-10 under the ASAM paper [27]), and the architecture is under-specified. My stress-test found the sharpest single vulnerability in the test-set-driven choice of pruning ratios: the paper explicitly selects different ratios per variant and per dataset based on test accuracy, and the winning margins over strong baselines are smaller than typical run-to-run variation. This makes the headline accuracies a selected maximum rather than a fair estimate. The proposed concrete test would settle it: if a validation-selected configuration reproduces the numbers with stable margins, the concern is resolved; otherwise the SOTA claim should be rejected. Since the reader's verdict already rejects, no change is needed. I do not see a soundness error in the architecture equations themselves; the problem is the empirical evidence chain.","tokens_in":19828,"tokens_out":5312,"duration_ms":46922,"concrete_test":"Obtain (or independently implement) the exact ResFRI/Split-ResFRI configurations, matching the published parameter counts. For each variant, select the pruning ratio on a held-out validation split rather than the test set, train with the stated SGD recipe, and report mean±std over at least three seeds. Check whether the validation-selected configuration still reaches 97.94% (CIFAR-10), 85.91% (CIFAR-100), and 70.54% (Tiny ImageNet), and still beats the cited baselines by more than one standard deviation. If the selected ratio changes or the margin disappears, the SOTA claim is an artifact of test-set tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ResFRI/Split-ResFRI achieves state-of-the-art accuracy at matched model size without extra data. The load-bearing requirement is that the reported numbers (97.94% CIFAR-10, 85.91% CIFAR-100, 70.54% Tiny ImageNet) are unbiased estimates of what a fully specified architecture achieves. This requirement is weakest at the pruning-ratio selection. In Section 3.9/Table 10, the ratios for the addition and concatenation versions are chosen on CIFAR-10 by comparing test error (0.7 for addition, 0 for concatenation), and Section 3.5/Table 5 reports both pruned and unpruned ResFRI-addition on Tiny ImageNet, with the better unpruned number (70.40) used in the comparison. Because the best of several configurations is picked using test labels, the quoted accuracy is a selection statistic, not a fair evaluation. The margins over the closest baselines are tiny (0.26% over DCDN on Tiny ImageNet; 0.04% over PyramidNet+ShakeDrop on CIFAR-10), and no error bars or validation-based selection are reported. This is compounded by Sections 2.1/3.1, which omit depth, channel widths, stem, downsampling, and per-stage block counts, so even the selected configuration is not reproducible from the paper. The SOTA claim therefore rests on an unverifiable and potentially overoptimistic number.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two CNN building blocks, ResFRI and Split-ResFRI, which combine four parallel convolution paths of different kernel sizes, 1x1 convolution information-interaction passages between adjacent paths, a residual connection, and a pooling layer; the Split variant divides input channels into 3/8, 3/8, 1/8, 1/8 groups to reduce parameters. The authors report top-1 errors of 2.06% on CIFAR-10, 14.09% on CIFAR-100, and 29.46% on Tiny ImageNet and claim state-of-the-art results among models of approximate size trained without extra data. Experiments also include MNIST, Fashion-MNIST, SVHN, and ablations on pruning ratio, pooling, residual connection, and interaction passages.","tokens_in":20113,"tokens_out":8207,"duration_ms":70819,"significance":"If the reported numbers are reproducible and the comparisons are fair, the architecture could be a useful efficient multi-scale CNN backbone, with the split version offering parameter savings at a small accuracy cost. The paper's strengths are the breadth of datasets, the component ablations, and the explicit attention to parameter and FLOP counts. However, the significance is currently limited because the central SOTA claim rests on unreproducible architecture details, test-set-selected hyperparameters, and comparison tables with mis-sourced or omitted baselines; until those are fixed, the contribution is not established.","major_comments":[{"comment":"The network is not fully specified. Equations (1)–(3) and Figures 2–4 define block-level operations but omit the overall network depth, channel widths per stage, stem, downsampling method, number of blocks per stage, and the exact placement of pooling and pruning operations. The implementation details in Section 3.1 give only the optimizer, batch size, and data augmentation, not the architecture. Without a complete specification or released code, the reported accuracies cannot be reproduced or verified, and it is impossible to tell whether the state-of-the-art claim depends on unstated implementation choices.","section":"§2.1, §3.1"},{"comment":"The pruning ratio and pooling type are selected using test-set performance. Section 2.3 and Eq. (4) fix the pruning ratio to 0.7 for the addition version and 0 for the concatenation version based on Table 10, which reports CIFAR-10 test errors; Section 3.5 and Table 5 similarly adopt the better of pruned and unpruned ResFRI-addition on Tiny ImageNet (70.40% versus 68.5%). Since the best of several configurations is chosen with test labels, the quoted accuracies are selection statistics. The margins over the closest baselines are small (0.04% over PyramidNet+ShakeDrop on CIFAR-10 and 0.26% over DCDN on Tiny ImageNet), and no error bars or validation-based selection are reported, so the central SOTA claim is currently not supported.","section":"§3.9, Eq. (4), Table 10"},{"comment":"The baseline comparisons are not reliable. Several numbers are cited to secondary sources (WRN-28-10 is cited to [27], the ASAM paper, rather than the original Wide-ResNet paper; Shake-Shake and PyramidNet+ShakeDrop are cited to [28], the SAM paper), and Table 2 lists WRN-28-10 (200 epochs) at 2.6% error but WRN-28-10 + SAM at 2.7%, which is inconsistent with the usual effect of SAM and indicates the numbers come from different training protocols. The tables also omit many modern efficient architectures at comparable FLOP counts, and Table 6 compares against test-time adaptation and ensemble methods (OCD variants) rather than standard training. The 'approximate model size' claim is further strained by DCDN having 77.79M parameters on Tiny ImageNet versus roughly 12–14M for the proposed models.","section":"Tables 1, 3, 5"},{"comment":"The ablation study confounds multiple factors. For example, 'ResFRI without AvgPooling layer (addition, pruning ratio 0.7)' is compared against the full addition version with pruning ratio 0.7, while 'ResFRI without AvgPooling layer (concatenation, pruning ratio 0)' is compared against the full concatenation version with pruning ratio 0; pooling and pruning ratio are thus not varied one at a time. Similarly, removing the residual connection in the addition version changes the parameter count from 12.1M to 8.9M, so the comparison mixes architectural change with capacity change. The ablation therefore does not cleanly identify which component drives the gains.","section":"§3.9, Table 10"}],"minor_comments":[{"comment":"The text mentions 'Split-ResPRI' where 'Split-ResFRI' is intended.","section":"§2.1"},{"comment":"The phrase 'start-of-the-art' should be 'state-of-the-art'.","section":"Contribution 4, §1"},{"comment":"The label 'MNSIT' should be 'MNIST'.","section":"Figure 5"},{"comment":"There are typos such as 'FashinMNIST', 'incpetion-like', 'practival', and 'dose not'; these should be corrected before resubmission.","section":"§3.8, §4, §3.10"},{"comment":"The table is titled 'Tiny Imagenet Dataset' but lists ViT/DeiT/OCD methods; the comparison is not a standard same-setting head-to-head and should be clearly labeled as such.","section":"Table 6"},{"comment":"The accuracy formula using TP/TN/FP/FN is written for binary classification; for multiclass datasets it should define overall accuracy or per-class average explicitly.","section":"Eq. (5)"},{"comment":"The original sources for WRN, Shake-Shake, and PyramidNet+ShakeDrop baseline numbers should be cited in addition to the papers from which the numbers were taken.","section":"References [27], [28]"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the architectural idea is not obviously invalid, but the current manuscript does not establish its central SOTA claim. The most serious issues are the under-specified architecture and the test-set-based selection of pruning ratio and pooling, compounded by unreliable baseline sourcing and missing comparisons to modern efficient models. I would not recommend acceptance until these are addressed with a complete model specification, code or full configuration, validation-based hyperparameter selection, error bars, and a fair, correctly cited comparison table. The header 'Preprint submitted to Journal of LATEX Templates' also suggests the template was not customized for a specific journal, which the authors should correct before formal submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a plausible architecture paper whose headline SOTA claim is not supported as written. The ResFRI/Split-ResFRI block is a sensible combination of inception parallel branches, Res2Net-style channel groups, and dense/residual passages; I believe the authors when they say they built it. What's new is the specific interaction passages (1x1 conv + pool + BN + ReLU) between adjacent groups. The ablations in Table 10 are the most useful part: they show pruning ratio, pooling choice, residual, and passages each matter, at least on CIFAR-10. That is honest legwork.\n\nThe soft spots are in the evaluation. First, the SOTA comparison is not credible: WRN-28-10 is cited to the ASAM paper and Shake-Shake to the SAM paper, which is wrong. Second, the architecture is under-specified: no depth, channel widths, stem, downsampling, or per-stage block counts, so no one can reproduce the result without guessing. Third, and most important, the reported configuration is selected using test labels. The pruning ratio 0.7 for addition and 0 for concatenation are chosen by comparing CIFAR-10 test errors (Table 10), and on Tiny ImageNet the paper reports both pruned (68.5) and unpruned (70.4) ResFRI-addition and then uses the unpruned number in the comparison. That turns the quoted accuracy into a selection statistic. The margins over DCDN and PyramidNet+ShakeDrop are 0.04-0.26%, so even a small optimistic bias matters. No error bars, no validation-selected hyperparameters.\n\nI'm not saying the block is useless; it may well be a decent efficient CNN building block. But the paper's load-bearing claim—state of the art at matched size—rests on numbers that are neither verifiable nor fairly selected. The right fix is to release code, fully specify the architecture, and choose hyperparameters on validation splits with repeated runs. The ablations are worth keeping.\n\nWho is this for? Someone who likes inception-style blocks and wants to see another combination attempt. It reads like a workshop paper or a preprint to mine for ideas, not a paper whose accuracy table should be trusted yet. My recommendation: don't reject on novelty grounds—the block arrangement is new enough—but treat the empirical claims as unsubstantiated until code and validation-based selection are provided. For peer review, I'd send it back for major revision rather than desk reject, because the method is coherent and the ablations are genuine. But I would not cite it for numbers.","headline":"A coherent block-design paper with a SOTA claim that is not supportable as written: pruning hyperparameters are test-selected, the architecture is under-specified, and some baselines are misattributed.","tokens_in":20631,"tokens_out":3231,"would_cite":false,"duration_ms":28895,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68T45"],"pacs":[],"model":"deepseek-v4-flash","headline":"Parallel convolution paths that reuse each other's features push a small CNN to 97.94% on CIFAR-10.","keywords":["ResFRI","Split-ResFRI","feature reutilization","inception network","multi-scale features","residual connection","image classification","CIFAR-100"],"falsifier":"Retrain ResFRI-concatenation on CIFAR-10 using the listed hyperparameters (SGD momentum 0.9, weight decay 0.0005, batch size 64, initial learning rate 0.01 halved on validation plateau, and the CIFAR augmentation schedule from [5]) with the full network configuration specified; if its top-1 error exceeds 2.06% on CIFAR-10 or 14.31% on CIFAR-100, or if a similarly sized listed baseline run under the same protocol surpasses 97.94% accuracy on CIFAR-10, the core state-of-the-art claim does not reproduce.","tokens_in":19637,"feed_emoji":"🖼️","tokens_out":5405,"duration_ms":45660,"temperature":0.7,"pith_summary":"This paper proposes a convolutional architecture built from residual feature-reutilization inception (ResFRI) blocks, plus a split variant called Split-ResFRI, and argues that it extracts multi-scale features more efficiently than existing CNN designs. The central claim is that, at comparable model size and without extra training data, these blocks reach state-of-the-art accuracy on standard image-classification benchmarks: 97.94% on CIFAR-10, 85.91% on CIFAR-100, and 70.54% on Tiny ImageNet. If true, the result shows that a purely convolutional, inception-style block with internal information-reuse passages can outperform much larger residual, dense, and wide models. The paper also reports ablations showing that removing the average pooling, residual connection, or inter-group passages lowers accuracy, and that the best pruning ratio depends on the fusion operation.","feed_headline":"Residual multi-scale CNN hits 97.94% on CIFAR-10","feed_subtitle":"Four parallel convolutions share features through residual passages, beating similarly sized models without extra data.","key_machinery":"The load-bearing object is the ResFRI block: four parallel convolutional combinations (1x1, 1x1-3x3, 1x1-5x5, and 1x1 plus 3x3 max-pooling paths, drawn from GoogLeNet's inception design), joined by information interaction passages that apply 1x1 convolution, pooling, BatchNorm, and ReLU to the previous branch's output and then add or concatenate it to the next branch's input, with an overall residual connection around the block. In Split-ResFRI, the input feature map is partitioned across branches in channel ratios 3/8, 3/8, 1/8, and 1/8 before processing. These passages are what reutilize features across scales, the residual connection eases optimization, and the channel split is what shrinks parameter count.","core_discovery":"The central discovery is that letting each of four parallel convolution branches receive not only the block input but also the processed output of the previous branch, through passages made of 1x1 convolution, pooling, BatchNorm, and ReLU, plus a global residual connection, produces richer multi-scale features than the standard inception design, which only concatenates branch outputs. Split-ResFRI divides the input channels into 3/8, 3/8, 1/8, and 1/8 shares before routing them to the four branches, cutting parameters and FLOPs while keeping accuracy close to the unsplit version. The authors report that ResFRI-concatenation reaches 2.06% top-1 error on CIFAR-10 and 14.31% on CIFAR-100, and that pruning the interaction passages at rate 0.7 helps the addition variant while the concatenation variant is best with no pruning.","pith_inferences":["The cross-branch reutilization pattern is not tied to GoogLeNet's four kernel choices; applying the same passage idea to other multi-branch blocks, such as Res2Net-style hierarchical splits or transformer attention heads, is a direct next test.","The 3/8, 3/8, 1/8, and 1/8 split ratio comes from GoogLeNet's channel budget; treating the split as a tunable hyperparameter per dataset could push accuracy further or shrink the model more, which the authors name as future work.","The ablation pattern suggests the passages and residual path carry independent gains, so stacking ResFRI blocks in a deeper or wider configuration may scale the improvement beyond the three datasets tested here.","Because Split-ResFRI keeps nearly all of ResFRI's accuracy at roughly two-thirds the parameters, the split idea is plausibly useful in deployment settings where memory or FLOPs are the binding constraint."],"forward_implications":["On CIFAR-10 and CIFAR-100, the four ResFRI and Split-ResFRI variants all beat ResNet-101 with far fewer parameters, while the GoogLeNet-sized Split-ResFRI models lead GoogLeNet by more than 2.8 percentage points on CIFAR-10.","On Tiny ImageNet, ResFRI-concatenation reaches 70.54% accuracy with 13.7M parameters, above the 70.28% of the 77.79M-parameter DCDN model and above FocusNet's 64.49%.","Ablations attribute part of the gain to each component: replacing max pooling with average pooling, keeping the residual connection, and keeping the inter-branch passages each improve accuracy on CIFAR-10.","The pruning result is fusion-dependent: the addition variant of ResFRI benefits from dropping 70% of interaction passages, while the concatenation variant is best with no pruning.","Because the blocks are self-contained inception-style units, the authors argue they can be dropped into existing inception-like networks and used as backbones for other vision tasks such as segmentation and detection."],"supporting_citations":[{"why":"Supplies the original inception structure, the parallel kernel-size paths, and the channel-number ratios used for the split ratios in Split-ResFRI.","marker":"[17]"},{"why":"Provides the residual-connection idea, the CIFAR training data augmentation strategy, and the ResNet baseline numbers the paper compares against.","marker":"[5]"},{"why":"Defines the Res2Net multi-scale residual design whose fixed 3x3 group convolutions the paper contrasts with its own variable kernel paths.","marker":"[11]"},{"why":"Motivates the dense-connectivity-like passages between branches and supplies DenseNet baselines for CIFAR, FashionMNIST, and SVHN.","marker":"[19]"},{"why":"Supplies the observation that dense connections can be redundant and motivates the pruning ratios used on the interaction passages.","marker":"[18]"},{"why":"Provides the WRN, PyramidNet, Shake-Shake, and PyramidNet+ShakeDrop baselines and the SAM technique used in the extra-technique comparisons.","marker":"[28]"},{"why":"Supplies FocusNet baselines on Tiny ImageNet and other datasets and the baseline numbers reused in the Tiny ImageNet comparison table.","marker":"[22]"},{"why":"Supplies DCDN baselines on CIFAR-10, CIFAR-100, and Tiny ImageNet, including the 70.28% Tiny ImageNet accuracy the proposed model is compared against.","marker":"[25]"},{"why":"Documents the impact of residual connections in inception networks and provides Inception-v4 and Inception-ResNet v2 baselines on CIFAR-100.","marker":"[10]"}],"fun_headline_variants":["Residual branches share features to boost image classification","Split-ResFRI trims parameters without sacrificing accuracy","Feature-sharing inceptions reach SOTA on CIFAR and Tiny ImageNet","97.94% CIFAR-10 accuracy from residual multi-scale CNN","Residual passages between convolutions enhance feature reuse"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy comparisons assume that the baseline numbers from earlier papers were measured the same way and that the paper's unstated network details (depth, channel widths, stem, and per-stage layer arrangement) do not change the reported results.","fun_headline_variants_meta":{"raw":{"variants":["Residual branches share features to boost image classification","Split-ResFRI trims parameters without sacrificing accuracy","Feature-sharing inceptions reach SOTA on CIFAR and Tiny ImageNet","97.94% CIFAR-10 accuracy from residual multi-scale CNN","Residual passages between convolutions enhance feature reuse"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1394,"prompt_tokens":935,"completion_tokens":459,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":375}},"tokens_in":551,"tokens_out":459,"duration_ms":4161,"temperature":1.0,"reasoning_tokens":375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:35:37.957772+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain ResFRI-concatenation on CIFAR-10 using the listed hyperparameters (SGD momentum 0.9, weight decay 0.0005, batch size 64, initial learning rate 0.01 halved on validation plateau, and the CIFAR augmentation schedule from [5]) with the full network configuration specified; if its top-1 error exceeds 2.06% on CIFAR-10 or 14.31% on CIFAR-100, or if a similarly sized listed baseline run under the same protocol surpasses 97.94% accuracy on CIFAR-10, the core state-of-the-art claim does not reproduce.","supporting_citations":[{"cited_title":"Szegedy, W","cited_arxiv_id":null,"evidence_quote":"Supplies the original inception structure, the parallel kernel-size paths, and the channel-number ratios used for the split ratios in Split-ResFRI."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the residual-connection idea, the CIFAR training data augmentation strategy, and the ResNet baseline numbers the paper compares against."},{"cited_title":"Huang, Z","cited_arxiv_id":null,"evidence_quote":"Motivates the dense-connectivity-like passages between branches and supplies DenseNet baselines for CIFAR, FashionMNIST, and SVHN."},{"cited_title":"Huang, S","cited_arxiv_id":null,"evidence_quote":"Supplies the observation that dense connections can be redundant and motivates the pruning ratios used on the interaction passages."},{"cited_title":"Foret, A","cited_arxiv_id":null,"evidence_quote":"Provides the WRN, PyramidNet, Shake-Shake, and PyramidNet+ShakeDrop baselines and the SAM technique used in the extra-technique comparisons."},{"cited_title":"Zhang, Z","cited_arxiv_id":null,"evidence_quote":"Supplies FocusNet baselines on Tiny ImageNet and other datasets and the baseline numbers reused in the Tiny ImageNet comparison table."},{"cited_title":"Patel, G","cited_arxiv_id":null,"evidence_quote":"Supplies DCDN baselines on CIFAR-10, CIFAR-100, and Tiny ImageNet, including the 70.28% Tiny ImageNet accuracy the proposed model is compared against."},{"cited_title":"Szegedy, S","cited_arxiv_id":null,"evidence_quote":"Documents the impact of residual connections in inception networks and provides Inception-v4 and Inception-ResNet v2 baselines on CIFAR-100."}],"review_version":1}