{"id":"271e5a55-4c7a-498d-8bd0-75a5393fb0d6","arxiv_id":"2506.10084","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"DeepTraverse is a weight-tied residual network plus squeeze-and-excitation attention, framed as depth-first search, with claimed efficiency gains that rest on a questionable ImageNet subset comparison.","lead":"This paper proposes a new image-classification network, DeepTraverse, whose design is loosely inspired by depth-first search, with repeated shared layers that refine features and an attention step that reweights channels. If the reported numbers held, the network would beat several well-known efficient models at lower parameter counts, but the headline ImageNet comparison appears to compare a 100-class subset against full 1000-class baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 compares DeepTraverse on a 100-class ImageNet subset against baselines on full ImageNet-1k, so the 83.16% headline is not a valid state-of-the-art claim and the central performance argument is unsupported.","rationale":"The reader's weakest assumption—that the ImageNet-1k subset results are compared against full-ImageNet baselines—is exactly the load-bearing concern I identify. The strongest claim is an empirical comparison claim, so its validity depends on matched benchmarks. Table 3 violates this, and the paper's own text acknowledges the different label space. This alone invalidates the headline result, independent of other issues. The remaining weaknesses (no code, relabeled known components, unmeasured interpretability) reinforce rejection but are not necessary to the verdict. The equations in Section 3 are internally coherent, and the CIFAR results would be plausible if reproduced, but the paper as presented does not support the claimed ImageNet-1K state-of-the-art performance. I therefore concur with the reader's REJECT verdict.","tokens_in":12151,"tokens_out":3178,"duration_ms":38466,"concrete_test":"Train the same DeepTraverse architecture on the full ImageNet-1k training set (1000 classes) with a standard supervised recipe and evaluate on the official validation set; compare Top-1/Top-5 with the Table 3 baselines under an identical protocol. If full-ImageNet accuracy does not reproduce a comparable advantage (e.g., cannot reach the claimed 83.16% or clearly beat StarNet/FasterNet on the same 1000-class metric), the central ImageNet-1k claim is an artifact of the 100-class subset.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim—state-of-the-art accuracy per parameter on ImageNet-1K—rests entirely on Table 3, but Table 3 is not a valid comparison. DeepTraverse was evaluated on a 100-class subset of ImageNet-1k (Section 4.3, Table 3), while the baseline numbers are standard full-ImageNet-1k results (ResNet50 78.76, GhostNet 80.34, StarNet 74.16 are canonical 1000-class validation accuracies). The paper itself confirms the mismatch by explaining that StarNet is 'not as well-suited for this particular 100-class ImageNet subset.' A 100-class subset has a different label space, different per-class difficulty, and a different chance baseline, so the 83.16% Top-1 figure is not comparable to any full-ImageNet number. Every downstream claim of best accuracy or a new Pareto frontier on ImageNet-1K therefore fails at the comparison stage. No code or training configuration is supplied, so the result cannot be independently checked, but the label-space mismatch alone is sufficient to invalidate the headline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DeepTraverse, a convolutional backbone whose core building block (DFSBlock) combines a recursive 'exploration' module (DFS-EB) with a channel-recalibration 'backtracking' module (DFS-BB), inspired by depth-first search. The authors claim state-of-the-art accuracy-per-parameter on CIFAR-10/100, ImageNet-64, and ImageNet-1K, with a headline 83.16% top-1 accuracy on a 100-class ImageNet-1K subset using 5.04M parameters. The architecture is described in Eqs. (2)-(12), and experiments are reported in Tables 1-3.","tokens_in":12425,"tokens_out":5554,"duration_ms":57603,"significance":"If the performance claims were valid, the paper would present a parameter-efficient architecture with an appealing algorithmic interpretation and a clear potential for mobile and resource-constrained vision. The paper provides no code, no hyperparameter configuration, and no ablations, so the only evidence is the tables. The central performance comparison for ImageNet-1K is invalid because Table 3 mixes a 100-class subset of ImageNet with full 1000-class references, so the headline 'state-of-the-art' result cannot be taken at face value. The contribution of the DFS-inspired components is also not isolated, so the explanatory claim is unsupported as submitted.","major_comments":[{"comment":"The headline claim of 83.16% top-1 accuracy on ImageNet-1K is based on a 100-class subset of ImageNet, while every baseline number in Table 3 (ResNet50 78.76, GhostNet 80.34, etc.) is a canonical full 1000-class ImageNet validation accuracy. The text itself acknowledges this mismatch when it remarks that StarNet is 'not as well-suited for this particular 100-class ImageNet subset.' A 100-class subset has a different label space and chance baseline than the full 1000-class benchmark, so any accuracy comparison between the two is invalid. This invalidates the abstract's 'state-of-the-art' claim, the Section 4.3 analysis, and the conclusion's performance summary.","section":"Table 3, Section 4.3"},{"comment":"The architecture's key hyperparameters—recursive depth R, bottleneck reduction ratio r, number of stacked DFSBlocks, and channel widths—are never specified in the main text; the text defers to 'supplemental materials' that are not provided. No ablations are reported that remove or vary the recursive parameter sharing (Eqs. (4)-(5)) or the recalibration module (Eqs. (6)-(8)). Without these, the observed accuracy differences cannot be attributed to the DFS-inspired mechanism rather than to depthwise separable convolutions, network width, or the unspecified training recipe, so the central explanatory claim is unsupported.","section":"Section 4, Sections 3.2-3.4"},{"comment":"The training recipe is described only as 'provided by previous work on the same datasets' with no citation to that work and no details on augmentations, optimizer, weight decay, or learning-rate schedule beyond '100 epochs' and 'initial learning rate 0.1.' No code is released. Combined with the missing hyperparameter values, the reported results in Tables 1-3 are not reproducible, which is a load-bearing issue for an empirical architecture paper.","section":"Section 4, 'Datasets and Experiment setting'"},{"comment":"Several baseline entries appear inconsistent with standard published configurations (e.g., DenseNet with 0.60M parameters on CIFAR-100, EfficientNet with 0.12 GFLOPs on CIFAR-100 but 0.34 GFLOPs on CIFAR-10 for the same model family). The paper does not state which variant of each baseline was used or how parameter and FLOP counts were measured, making the claimed Pareto dominance in Sections 4.1 and 4.2 difficult to verify.","section":"Tables 1 and 2"}],"minor_comments":[{"comment":"The DFS-BB module is structurally identical to the Squeeze-and-Excitation block (Hu et al., CVPR 2018), including adaptive average pooling, a two-layer bottleneck with reduction ratio r, sigmoid, and channel-wise multiplication; the authors should cite that work and clarify what is new in the 'backtracking' formulation.","section":"Section 3.3, Eqs. (6)-(8)"},{"comment":"Reference [20] is cited for DenseNet, but the cited paper is CondenseNet; the bibliography entry should be corrected to the original DenseNet paper.","section":"References"},{"comment":"The term 'ImageNet64' is used without specifying whether it is Tiny ImageNet or a downsampled ImageNet-1k; the experimental setting mentions 'Tiny ImageNet with resolution of 64x64,' so the dataset should be defined consistently across the text and table.","section":"Section 4.2"},{"comment":"The figures contain small, hand-drawn-style annotations; labels such as 'Iter = 1', 'Iter = n', and 'Backtrack Focus' are hard to read and should be enlarged or typeset.","section":"Figures 1 and 2"},{"comment":"Reference [9] is a specialized paper on using DFS to minimize connected components; a standard algorithms textbook or survey would be a more appropriate citation for the definition of depth-first search.","section":"Section 2, reference [9]"}],"recommendation":"reject","confidential_remarks":"This submission is not ready for journal publication. The central performance claim rests on an invalid benchmark comparison (100-class ImageNet subset versus full 1000-class baselines), and the absence of code, hyperparameters, and ablations precludes verification of the reported results. The paper has the flavor of an extended abstract; if the authors rerun on full ImageNet-1k, provide ablations that isolate the DFS components, and release code, a resubmission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the headline result is a mirage. Table 3 compares DeepTraverse on a 100-class ImageNet subset against baselines on the full 1000-class ImageNet-1k, so the 83.16% Top-1 figure is not comparable. The stress-test note is right, and the paper itself confirms it in Section 4.3 by explaining StarNet's shortfall as due to the 100-class subset. Once you remove that table, the central claim of state-of-the-art accuracy per parameter collapses.\n\nWhat is actually here: the architecture is a recursive weight-tied residual refinement block (Eqs. 4-5) followed by a Squeeze-and-Excitation block (Eqs. 6-8). Both components are established. The paper relabels them with DFS vocabulary and adds nothing architectural beyond the combination. That combination is not crazy — a shared-depth residual block plus channel recalibration is a reasonable recipe, and the CIFAR-100 results are plausible for such a small model — but the paper does not cite SE or any recursive/weight-tied network, and it offers no ablations isolating R or r, no code, and no config values in the main text. The interpretability claim is entirely unmeasured, just a cartoon in Figure 1.\n\nThe CIFAR and ImageNet-64 numbers may be fine, but without the training recipe details, code, or ablations I cannot trust them. The statement that the training recipe is 'provided by previous work' without naming it is not reproducible by any standard.\n\nWho is this for? A reader tracking efficient CNN designs might skim it, but the invalid comparison and the overloaded novelty claims make it a poor use of referee time in its current form. If the authors fixed the comparison, added ablations, cited SE and the recursive literature, and shipped code, there might be a workshop or second-tier paper here. As is, I would desk-reject it. I would not cite it.","headline":"The ImageNet-1k claim is a mirage: Table 3 compares a 100-class subset against full 1000-class baselines, and the rest of the paper is a known architecture combo with no ablations.","tokens_in":12924,"tokens_out":2516,"would_cite":false,"duration_ms":28332,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DeepTraverse embeds depth-first search into a vision backbone and reports top accuracy per parameter.","keywords":["depth-first search","vision backbone","efficient neural networks","recursive exploration","channel recalibration","image classification","parameter-efficient architecture","algorithmic priors"],"falsifier":"Train DeepTraverse and its listed baselines on the same full ImageNet-1k training set with the same epochs and resolution, then compare top-1 accuracy on the official 1,000-class validation set; if DeepTraverse no longer outperforms StarNet or FasterNet at comparable parameter counts, the headline accuracy advantage is an artifact of the 100-class subset rather than a property of the architecture.","tokens_in":11951,"feed_emoji":"🔍","tokens_out":6950,"duration_ms":78336,"temperature":0.7,"pith_summary":"The paper proposes DeepTraverse, a vision backbone that organizes feature extraction the way depth-first search organizes a tree traversal: it repeatedly applies a shared transformation to deepen the representation along a path, then recalibrates channel importance using global context, and repeats this exploration/backtrack cycle in stacked blocks. The authors' central claim is that this algorithmic prior yields better accuracy per parameter than conventional and efficient backbones, including ResNet, GhostNet, MobileNetV3, ShuffleNetV2, EfficientNet, FasterNet, and StarNet across CIFAR-10, CIFAR-100, ImageNet-64, and a 100-class ImageNet-1k subset. Their headline numbers are 83.16% top-1 and 96.54% top-5 accuracy on that ImageNet subset with 5.04M parameters, 73.84% on CIFAR-100 with 0.26M parameters, and 71.50% on ImageNet-64 with 0.59M parameters. If the claim holds, it would show that structuring a network around a classical search metaphor is a viable efficiency principle rather than a mere analogy.","feed_headline":"A DFS-inspired vision net beats efficient backbones with fewer parameters","feed_subtitle":"Search-inspired structure, not added parameters, may be the efficiency lever: 83.16% top-1 on an ImageNet subset with 5.04M parameters.","key_machinery":"The load-bearing object is the DFSBlock, built from two named components. DFS-EB (Exploration Block) performs recursive residual refinement, $F_i = F_{i-1} + \\Phi_{\\text{recursive}}(F_{i-1})$, with $\\Phi_{\\text{recursive}}$ sharing parameters across $R$ iterations so extra depth costs almost no extra parameters. DFS-BB (Backtrack Block) is a channel recalibrator: global average pooling yields a channel descriptor $z$, a two-layer bottleneck maps it to a per-channel weight $s = \\sigma(W_2 \\delta(W_1 z))$, and $F' = F \\odot s$ rescales the features. Stacked with residual shortcuts, these blocks instantiate the paper's DFS analogy: iterative deepening during exploration and context-driven reweighting during backtracking.","core_discovery":"DeepTraverse's core claim is that visual representation learning can be structured as an explicit depth-first search over feature space. The DFS Exploration Block (DFS-EB) starts from a feature map produced by an extraction layer and applies the same recursive block $R$ times, each time adding a learned residual; because the recursive weights are shared, the effective depth grows without proportional parameter growth. The DFS Backtrack Block (DFS-BB) then pools global spatial information, computes a channel-wise attention vector through a bottleneck, and rescales the features before the next block, mimicking a backtracking step that prunes or reinforces what was found. The paper argues that the composition of these two mechanisms produces features that are progressively refined and contextually reweighted, and reports that the resulting network reaches the best accuracy-per-parameter among the models compared on the benchmarks tested.","pith_inferences":["A natural next test is full ImageNet-1k training; the paper's ImageNet numbers come from a 100-class subset, so the architecture's real-world advantage over baselines is still open.","If the recursive-depth effect is real, intermediate stages $F_1,\\ldots,F_R$ should show progressively finer or more class-specific features, which could be tested with probing classifiers or attention visualizations at each recursion depth.","The parameter-shared recursion makes DeepTraverse a natural candidate for adaptive computation, where the network selects how many recursions to run per input to trade accuracy for latency at inference time.","The design invites comparison with other weight-tied or recurrent architectures; if the DFS metaphor carries explanatory weight, breadth-first or best-first search could be operationalized similarly.","A concrete interpretability study would measure whether DFS-BB's channel weights align with human-annotated object parts; the paper does not report such a study."],"forward_implications":["Deepening a block by increasing $R$ costs shared-parameter compute rather than new weights, so accuracy can in principle be traded against speed without changing model size.","The DFS-BB's channel reweighting gives the architecture an explicit, inspectable locus where global context modifies local features, which supports the paper's interpretability claim.","Because the blocks are built from depthwise separable convolutions, batch norm, and pointwise mixing, the design can be dropped into existing training pipelines without new operators.","On the reported benchmarks, the per-parameter advantage implies that efficient backbones may have more headroom than their layer-count suggests.","Training DeepTraverse at wider scales on CIFAR-100 (82.20% with 14.26M parameters) suggests the DFS structure scales beyond lightweight regimes without a linear parameter penalty."],"supporting_citations":[{"why":"Supplies the depth-first search algorithm whose traversal and backtracking logic the architecture operationalizes.","marker":"[9]"},{"why":"Provides the residual learning mechanism and the ResNet baselines that DeepTraverse compares against.","marker":"[19]"},{"why":"GhostNet is a key efficient baseline used across all four benchmark tables.","marker":"[18]"},{"why":"StarNet is the main ImageNet-1k comparison in Table 3, where DeepTraverse reports its largest lead.","marker":"[23]"},{"why":"FasterNet serves as a modern efficient baseline on ImageNet-64 and ImageNet-1k.","marker":"[22]"},{"why":"EfficientNet/EfficientNetV2 supplies a standard efficiency baseline on CIFAR and ImageNet tables.","marker":"[21]"},{"why":"MobileNetV3 is a baseline for lightweight efficiency comparisons on CIFAR, ImageNet-64, and ImageNet-1k.","marker":"[13]"},{"why":"ShuffleNetV2 is included as a practical guideline-driven efficient CNN baseline.","marker":"[16]"},{"why":"WideResNet provides the wide-model baseline that DeepTraverse's wide version outperforms on CIFAR-100.","marker":"[60]"}],"fun_headline_variants":["DFS vision net beats heavier models with fewer parameters","Search-inspired backbone wins on accuracy-per-parameter","DeepTraverse: depth-first search as a vision backbone","Algorithmic vision net: efficient, structured, top-ranked","DFS-based features sharpen visual recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline ImageNet-1k results were measured on a 100-class subset of ImageNet, while the baselines' published numbers are for the full 1,000-class task, so the comparison assumes the subset is no easier than the full benchmark.","fun_headline_variants_meta":{"raw":{"variants":["DFS vision net beats heavier models with fewer parameters","Search-inspired backbone wins on accuracy-per-parameter","DeepTraverse: depth-first search as a vision backbone","Algorithmic vision net: efficient, structured, top-ranked","DFS-based features sharpen visual recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1202,"prompt_tokens":942,"completion_tokens":260,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":188}},"tokens_in":558,"tokens_out":260,"duration_ms":4111,"temperature":1.0,"reasoning_tokens":188,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:35:05.567717+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DeepTraverse and its listed baselines on the same full ImageNet-1k training set with the same epochs and resolution, then compare top-1 accuracy on the official 1,000-class validation set; if DeepTraverse no longer outperforms StarNet or FasterNet at comparable parameter counts, the headline accuracy advantage is an artifact of the 100-class subset rather than a property of the architecture.","supporting_citations":[{"cited_title":"A depth-first search-based algorithm for the minimization of the largest connected component in networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the depth-first search algorithm whose traversal and backtracking logic the architecture operationalizes."},{"cited_title":"Ghostnet: More features from cheap operations,","cited_arxiv_id":null,"evidence_quote":"GhostNet is a key efficient baseline used across all four benchmark tables."},{"cited_title":"Rewrite the stars,","cited_arxiv_id":null,"evidence_quote":"StarNet is the main ImageNet-1k comparison in Table 3, where DeepTraverse reports its largest lead."},{"cited_title":"Run, don’t walk: Chasing higher FLOPS for faster neural networks,","cited_arxiv_id":null,"evidence_quote":"FasterNet serves as a modern efficient baseline on ImageNet-64 and ImageNet-1k."},{"cited_title":"Searching for mobilenetv3,","cited_arxiv_id":null,"evidence_quote":"MobileNetV3 is a baseline for lightweight efficiency comparisons on CIFAR, ImageNet-64, and ImageNet-1k."}],"review_version":1}