{"id":"67ae3eba-7f30-4f29-9aff-1611c07e7381","arxiv_id":"2411.08014","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Applying tanh, softsign, or scaling smoothing to ResNet activations yields stylization quality comparable to softmax-based SWAG, though the evidence is only qualitative.","lead":"This paper re-implements three classic neural style transfer algorithms and tests new activation-smoothing tricks to make ResNet-based style transfer work better. It is an incremental contribution that confirms the existing SWAG method with alternative smoothing functions, but only offers qualitative results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that activation smoothing 'greatly improves' stylization and that tanh/softsign/scaling match softmax rests entirely on visual inspection of a few images (Section 6.4, Figures 7–9); no quantitative or statistical support is provided.","rationale":"The reader's weakest assumption is exactly the one I identify: the paper relies on visual inspection of a small set of stylized images to conclude that smoothing transformations 'greatly improve' quality and that tanh, softsign, and scaling are equivalent to softmax. I agree that this is the load-bearing weakness. My stress-test adds a concrete, quantitative protocol that would either support or refute the claim: a controlled multi-pair comparison with paired statistical tests and perceptual metrics. I see no additional internal inconsistency that would force a separate rejection reason; the primary reason to reject is the insufficiency of the evidence for the central empirical claim. Therefore the reader's verdict of REJECT remains appropriate, and no adjustment is needed.","tokens_in":5678,"tokens_out":2002,"duration_ms":20266,"concrete_test":"Conduct a controlled, pre-registered evaluation over at least 20 content-style pairs. For each pair, run the same optimization procedure with identical hyperparameters (iterations, learning rate, loss weights) for five conditions: un-smoothed ResNet, ResNet+softmax, ResNet+tanh, ResNet+softsign, and ResNet+scale(0.001). Compute quantitative metrics for each output: style loss (Gram distance to the style image), content preservation (e.g., LPIPS or VGG feature distance to the content image), and a perceptual quality score. Report means with standard errors and perform a paired non-parametric test (e.g., Wilcoxon signed-rank) comparing each smoother against un-smoothed ResNet and softmax against the other three smoothers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is the empirical assertion that multiplying activations by 0.001, tanh, or softsign 'can reach the similar effect as the softmax function' and that smoothing transformations 'greatly improve' stylization quality. This assertion is load-bearing because the entire extension over SWAG is the claim that these alternative, partly hyperparameter-free smoothers work comparably. The evidence offered in Section 6.4 is qualitative inspection of Figure 7 (and references to Figures 8–9, which are not present in the text). No quantitative metrics such as style loss, content preservation, LPIPS, or user ratings are reported; no error bars or significance tests appear; and the number of content-style pairs shown is very small, leaving selection bias uncontrolled. In addition, the comparison is not shown to be controlled: the un-smoothed ResNet baseline may use different effective loss weighting or optimization budgets, since the text only states the style loss weight (10^12) for the smoothed cases. Thus the central claim that smoothing transformations improve quality, and the specific equivalence among softmax, tanh, softsign, and scaling, is unsupported by the presented evidence. This is a correctness risk rather than a mere presentation issue, because the claim is empirical and the manuscript's own experimental record does not substantiate it. The template artifact in Section 2 ('The next subsections provide instructions...') and the mismatched references further indicate the manuscript is unfinished, but the evidential gap is the primary obstacle to accepting the claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper describes reimplementations of three established neural style transfer (NST) methods—image-based optimization (Gatys et al.), fast feed-forward style transfer (Johnson et al.), and AdaIN-based arbitrary style transfer (Huang and Belongie)—and then proposes an extension of the SWAG framework in which the softmax activation smoothing is replaced by multiplying activations by 0.001, tanh, and softsign. The paper's central claim, stated in the abstract and Section 5, is that these smoothing transformations greatly improve stylization quality and that tanh, softsign, and scaling achieve similar effects to softmax. The evidence for this claim is limited to qualitative visual comparison of a small number of images in Section 6.4, with no quantitative metrics, no code release, and no controlled baselines.","tokens_in":5967,"tokens_out":4046,"duration_ms":40035,"significance":"If substantiated, the observation that simple activation smoothers such as tanh and softsign perform comparably to softmax in SWAG-based style transfer would be a useful practical finding, since these alternatives are hyperparameter-free and computationally cheap. However, the manuscript does not substantiate this claim: there are no quantitative metrics, no error bars, no statistical tests, no controlled baselines, and no released code. The remaining content is a tutorial-level reimplementation of existing methods with no new theoretical contribution, so the empirical claim is the paper's entire novel contribution. The presence of template text and mismatched references further indicates the manuscript is not in a publishable state.","major_comments":[{"comment":"The paper's central claim—that smoothing transformations 'greatly improve' stylization quality and that tanh, softsign, and scaling 'can reach the similar effect as the softmax function'—is supported only by visual inspection of four images. No quantitative metrics (e.g., style loss, content preservation, LPIPS, user ratings), no error bars or significance tests, and no code are provided. The text also refers to 'Figure 9' while discussing results meant to be in Figure 7, and Figures 8 and 9 are never presented, so the actual comparison cannot be examined. Because the paper's sole novel contribution is this empirical claim, the absence of measurable or reproducible evidence is a load-bearing defect.","section":"Section 6.4, Figure 7"},{"comment":"The claim that multiplying activations by 0.001, tanh, and softsign produce similar effects to softmax is stated without any hypothesis or theoretical rationale, and Section 6.4 provides only qualitative images as support. Moreover, the preference for tanh and softsign over scaling is justified by their being 'hyperparameter-free,' but the paper does not test sensitivity to the scaling coefficient (0.001), and the scaling method is itself a free parameter in the comparison. Without an ablation over the scaling constant or a quantitative similarity measure, the equivalence claim is unsupported.","section":"Section 5, Eqs. (13)-(14)"},{"comment":"The manuscript contains leftover ACM template text ('The next subsections provide instructions on how to insert figures, tables, and equations in your document') in Section 2, and several citations are clearly mismatched: references [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], and [19] point to unrelated works (steganography, user identity linkage, path planning, etc.) rather than to the claims they are attached to. These errors prevent verification of the described methods and indicate the manuscript is an incomplete draft.","section":"Section 2 and Reference list"}],"minor_comments":[{"comment":"The layer index is inconsistently denoted: the text says 'in layer 𝐼' where the layer index is ℓ; please use a consistent notation.","section":"Section 2.2, Eq. (2)"},{"comment":"The loss network is described as VGG-16 in Section 3.2 but VGG-19 elsewhere; please clarify which architecture was used for the fast style transfer experiments.","section":"Section 3.2"},{"comment":"Hyperparameters for the image-based NST experiments are reported, but the number of runs or random seeds is not, so the reader cannot assess variability of the results.","section":"Section 6.1"},{"comment":"The sentence 'As we can learn from Figure 9' appears to refer to Figure 7; figure numbering is inconsistent throughout the manuscript.","section":"Section 6.4"},{"comment":"References to '5b,14' and '(15)' for content images and video results are confusing; these should be replaced by proper figure and video references.","section":"Section 6.2"},{"comment":"The abstract claims improved 'robustness,' but no robustness experiments (e.g., perturbations, noise, different style-content pairs) are reported.","section":"Abstract and Section 5"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to have been assembled from template material and an unrelated reference list; the core experiment is a small set of qualitative images. This is more of a course project report than a research paper. The lack of quantitative evaluation, controlled baselines, and code would require a major rewrite and new experiments to make the central claim supportable. I also note a potential scholarly integrity concern in the citation list, which appears reused from other works."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an unfinished manuscript that re-implements Gatys, Johnson, and AdaIN NST and tests scaling, tanh, and softsign as substitutes for softmax in SWAG. The re-implementations are standard, and the smoothing comparison is a legitimate small experiment. The problem is that the paper's only real claim—that these alternative smoothers match softmax and that smoothing greatly improves quality—rests on a handful of qualitative images with no metrics, no error bars, and an uncontrolled baseline. The template text in Section 2, the citations to steganography papers for style loss and AdaIN, and the reference to a non-existent Figure 9 confirm the manuscript is not in a submittable state.\n\nWhat's new: the specific empirical comparison of tanh, softsign, and scaling against softmax in SWAG is not reported in the cited literature, so the finding is technically new. If the equivalence holds, it would save practitioners from implementing softmax and would remove a hyperparameter. That is a small practical convenience, no more. The authors do not oversell it as a new theory or framework; they call it an extension.\n\nWhere it falls down: the evidence. No quantitative style/content loss numbers, no LPIPS or user study, no released code. The text gives the content weight (1) and style weight (10^12) for the smoothed cases, but never states the weights for the un-smoothed ResNet baseline. Without that, the 'greatly improves' claim could be an artifact of loss weighting. The comparison is by eye on four images, with no control for selection bias. The references are a mess—[8]–[18] mostly point to steganography, RL navigation, and unrelated papers, so the related-work placement is unreliable. The text refers to Figure 9 when only Figure 7 is present, and Figure 8 is never shown, so the paper is internally inconsistent.\n\nThe central argument—that alternative smoothers work comparably—might well be true, but the paper does not demonstrate it. This is not a subtle flaw in an otherwise sound paper; the load-bearing evidence is missing entirely.\n\nWho it's for: someone curious whether SWAG's smoothing can be simplified, but they would be better served by reading the original SWAG paper and writing a few lines of code. A serious referee should not spend time on this version. The authors need to fix the manuscript, add quantitative comparisons over many image pairs, and release code before resubmission.\n\nRecommendation: desk reject, with an invitation to resubmit if the empirical claims are actually backed by data.","headline":"The paper re-implements three known NST methods and tests tanh, softsign, and scaling as replacements for softmax in SWAG, but the only new empirical claim rests on a few uncontrolled qualitative images, and the manuscript is unfinished.","tokens_in":6518,"tokens_out":2540,"would_cite":false,"duration_ms":42169,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-12T21:59:59.188489+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}