{"id":"2f4ff06d-1491-44c0-966e-9281bac39cba","arxiv_id":"2509.01431","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Mamba-CNN reports PC 0.9187, MAE 0.2022, and RMSE 0.2610 on SCUT-FBP5500, but its 'Mamba' block is a depthwise conv plus sigmoid gate, not an SSM.","lead":"This paper introduces Mamba-CNN, a hybrid network for facial beauty scoring that combines convolutional blocks with a sigmoid gating path, and reports a small state-of-the-art improvement on SCUT-FBP5500. The name suggests a state space model, but the gating path is a depthwise convolution followed by sigmoid, not a real Mamba or SSM.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim depends on unspecified split/evaluation protocol; paper provides no code or per-fold stats to make the comparison checkable.","rationale":"The reader's weakest_assumption correctly identifies the most load-bearing condition: protocol comparability of the SOTA comparison. The manuscript gives no fold identifiers, no per-fold statistics, no code, no evaluation script, and no statement of how the Table 1 baselines were generated. Without these, the claimed improvement over R3CNN cannot be checked, and the empirical claim is the central one. The architectural 'Mamba' naming concern is real but secondary: §4.1 explicitly says the gating mechanism is a simplification without SSM state-space parameterization or selective scan, so the novelty claim is overstated, but that alone would not overturn a correctly measured empirical result. The reader's CONDITIONAL verdict is appropriate: the paper should be accepted only if the protocol is documented and reproducible artifacts are provided. No new concern changes the verdict, so I mark it UNCHANGED.","tokens_in":9384,"tokens_out":4415,"duration_ms":51670,"concrete_test":"Ask the authors for the exact split definition, per-fold predictions, and the evaluation script, then independently rerun the strongest baseline (R3CNN, ResNeXt-50) under Mamba-CNN's exact protocol — same folds, same train/validation partition, same augmentation, same score normalization/denormalization. If R3CNN reproduces 0.9142/0.2120/0.2800 under that protocol and Mamba-CNN still yields 0.9187/0.2022/0.2610, the SOTA claim survives. If R3CNN's numbers change under Mamba-CNN's protocol, the headline comparison is invalid and the paper should report metrics on the official single split used by the baselines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that Mamba-CNN sets a new state-of-the-art on SCUT-FBP5500 (Abstract; §3.2, Table 1) — rests entirely on the numbers PC 0.9187 / MAE 0.2022 / RMSE 0.2610. The necessary condition for this claim is that these numbers were produced under the same evaluation protocol as the baselines, especially R3CNN (PC 0.9142 / MAE 0.2120 / RMSE 0.2800). That condition is not established. Section 2.1 states the model 'adheres to the dataset’s prescribed cross-validation split,' but the paper never says which fold(s) are used, whether metrics are averaged over folds, how the validation set for early stopping is carved from the training data, or how the Table 1 baselines were evaluated. If Mamba-CNN used a different split or a different number of folds than the cited baselines, the reported margin could be an artifact of protocol mismatch. No code, per-fold results, or evaluation script are provided, so the comparison cannot be independently verified. The paper's own §4.1 also concedes that the MambaBlock is 'a simplified, SSM-inspired gating mechanism' without the 'selective scan mechanism... state-space parameterization and discretization' of the original Mamba, so the architectural novelty is overstated; however, the empirical SOTA claim is the load-bearing one and it is unverifiable as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Mamba-CNN, a hybrid architecture for facial beauty prediction on the SCUT-FBP5500 benchmark. The network combines a hierarchical CNN backbone of inverted-residual-style MambaBlocks—where each block uses a depthwise convolution followed by sigmoid as an 'SSM-inspired' gate—with a multi-scale feature pyramid and a fully connected regression head. Training uses MSE loss, AdamW, augmentation, and early stopping. The paper claims state-of-the-art results (PC 0.9187, MAE 0.2022, RMSE 0.2610) and presents an ablation showing that both the gating mechanism and the feature pyramid contribute to performance.","tokens_in":9764,"tokens_out":5847,"duration_ms":67377,"significance":"If the reported results are reproducible under the same evaluation protocol as the cited baselines, Mamba-CNN would be a competitive hybrid model for facial beauty prediction, with an internally consistent ablation (§3.4) and a clearly described training procedure (Algorithm 1). The external SCUT-FBP5500 benchmark and external baselines mean the central performance claim is not circular. However, the SOTA claim is only as strong as the protocol transparency behind Table 1. The paper provides no code, per-fold results, or statistical uncertainty estimates, and the margin over R3CNN is small (PC +0.0045). Moreover, the paper openly states in §4.1 that the MambaBlock is a simplified gating mechanism without the selective scan, state-space parameterization, or discretization of Mamba, so the 'global context' and 'new architectural paradigm' claims are not supported by the described mechanism.","major_comments":[{"comment":"The SOTA claim in the Abstract and §3.2 depends on the numbers in Table 1 being produced under the same evaluation protocol as the baselines. Section 2.1 only says the model 'adheres to the dataset’s prescribed cross-validation split,' and §2.5 says metrics are computed after denormalization, but the paper never states which split or fold is used, whether metrics are averaged over folds, how the validation set for early stopping (Algorithm 1) is carved from training data, or how the Table 1 baselines, especially R3CNN, were evaluated. Since the margin over R3CNN is small, a different split or normalization convention could change the ranking. Please specify the exact split, report per-fold results, and provide an evaluation script or enough protocol detail to make the comparison checkable.","section":"§2.1, §2.5, Table 1"},{"comment":"No uncertainty estimates or repeated-run statistics are reported. The comparison is based on single numbers for each model. For example, the PC improvement over R3CNN is 0.9187 versus 0.9142, a difference of 0.0045; without error bars or per-fold results, it is not possible to assess whether this difference is significant. Report mean and standard deviation over multiple runs, or at least per-fold numbers, and state whether early stopping is applied independently per run.","section":"§3.2, Table 1"},{"comment":"The MambaBlock as described is an inverted residual block whose 'SSM-inspired gating' branch is a 3x3 depthwise convolution followed by a sigmoid. The paper itself concedes in §4.1 that it 'does not incorporate the full complexity of the selective scan mechanism found in the original Mamba architecture, such as its state-space parameterization and discretization.' Therefore the Abstract's claims that the model 'mod[s] global context' and captures 'long-range spatial relationships' are not established by the described mechanism, which has only a local 3x3 receptive field. The empirical results may still be valid, but the architectural framing and the 'new architectural paradigm' claim in §5 need to be scaled back unless a global-context analysis is provided.","section":"§2.3, §4.1"},{"comment":"The title and the motivation emphasize efficiency, but the paper reports no parameter counts, FLOPs/MACs, or inference-time comparison against the baseline CNNs or ViTs. Since the efficiency-versus-accuracy trade-off is the stated motivation (Abstract, §1), the reader cannot verify that Mamba-CNN is more efficient than a standard CNN or less costly than a ViT. Add a complexity table (parameters, MACs, throughput) or remove the efficiency claim.","section":"Title, §1, §3.1"}],"minor_comments":[{"comment":"The normalization description says the original minimum and maximum scores were stored for denormalization, but Algorithm 1 and the evaluation protocol do not mention this step explicitly. Please clarify where and how denormalization is performed.","section":"§2.1"},{"comment":"Add a column describing the evaluation protocol used for each baseline (split, input size, and any normalization), so that the reader can see that all methods are compared on equal terms. Also include the source or footnote for each baseline number.","section":"Table 1"},{"comment":"The Pearson correlation formula is mis-rendered; the square root sign should cover the product of the two sums. Please fix the notation.","section":"§2.5"},{"comment":"Reference [30] (R3CNN) is incomplete: it lacks volume/article number and page range. Some other references also have inconsistent formatting; please check them against the journal style.","section":"References"},{"comment":"The label 'SSM Gating (DW-Conv + Sigmoid)' is misleading because the component is not an SSM. Consider renaming it to 'gating branch' or 'Mamba-inspired gating' consistently with §4.1.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The key risk is the verifiability of the SOTA claim, not the architecture itself. If the authors provide exact split details, per-fold results, and code or an evaluation script, the paper could become acceptable. I would not reject solely because the MambaBlock is not a true SSM, since the paper is transparent about that limitation; however, the abstract and conclusion should be aligned with that transparency. I also noticed a heavy reliance on the author's own prior work, but the core comparison is external, so this is not a circularity concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a modest architecture paper with an overstated name and an uncheckable headline number. The MambaBlock is not a state space model; it's a MobileNetV2-style inverted residual block with a parallel depthwise-conv-plus-sigmoid gate. The authors admit as much in §4.1. So the 'new architectural paradigm' language in the abstract should be ignored.\n\nWhat's actually good: the writing is clear, the ablation is internally consistent, and the empirical direction (gating + multi-scale pooling helps FBP) is plausible. If the numbers hold, it's a small SOTA improvement over R3CNN. The limitations section is unusually candid, including the admission about the simplified gating and the limited dataset.\n\nThe soft spot that matters: the SOTA claim rests entirely on PC 0.9187 / MAE 0.2022 / RMSE 0.2610, but no code, no per-fold results, and no explicit statement of which split or how the Table 1 baselines were evaluated. The paper says it adheres to the dataset's prescribed cross-validation split, but that isn't enough to reproduce the comparison. If the baselines used different folds or a different test split, the margin could vanish. This is not a manufactured concern; it's the load-bearing assertion and it is currently unverifiable. The ablation table has the same issue—no variance or fold breakdown, so the claimed ordering of ablations could be noise.\n\nThe citation pattern leans heavily on the author's own prior FBP papers, but that is not inherently a flaw. The Mamba designation is a branding issue more than a correctness issue, since the paper explicitly frames the block as SSM-inspired.\n\nFor whom: someone working on facial beauty prediction might get a useful data point; anyone else won't. The paper deserves a serious referee because the empirical claim is concrete and contestable, not because the architecture is conceptually deep.\n\nRecommendation: send to peer review, but insist on code or detailed per-fold results and a clearer definition of the evaluation protocol before the SOTA claim can be accepted. Renaming or reframing 'Mamba' would also help.","headline":"The SOTA claim is unverifiable as reported, and the 'Mamba' block is really a depthwise conv plus sigmoid gate, but the paper is honest, the ablation is coherent, and the result is worth refereeing if the author releases code and per-fold numbers.","tokens_in":10190,"tokens_out":2097,"would_cite":false,"duration_ms":23725,"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 CNN block with Mamba-inspired gating raises the top score on facial beauty prediction.","keywords":["facial beauty prediction","Mamba state space models","hybrid CNN architecture","selective gating","SCUT-FBP5500","multi-scale feature pyramid","facial attractiveness regression"],"falsifier":"Retrain Mamba-CNN and R3CNN on the exact same split of SCUT-FBP5500 with identical preprocessing and denormalization; if the Pearson correlation gap shrinks to under about 0.002 or the reported numbers do not reproduce, the claimed state-of-the-art result would not hold.","tokens_in":9302,"feed_emoji":"📈","tokens_out":6546,"duration_ms":67089,"temperature":0.7,"pith_summary":"The paper proposes Mamba-CNN, a hybrid architecture for facial beauty prediction that grafts a Mamba-inspired selective gating mechanism onto a convolutional backbone. The gate multiplies the main feature path by the output of a depthwise-conv-plus-sigmoid branch, letting the network amplify cues it deems relevant to attractiveness while suppressing background clutter, without the quadratic cost of full attention. On the SCUT-FBP5500 benchmark, Mamba-CNN reports a Pearson correlation of 0.9187, a mean absolute error of 0.2022, and a root mean square error of 0.2610, which the paper argues sets a new state of the art across all three metrics. An ablation attributes the largest gain to the gating branch and a smaller additive gain to a multi-scale feature pyramid, supporting the paper's thesis that content-aware selective modulation is a promising replacement for global attention in nuanced visual regression tasks.","feed_headline":"Mamba-CNN tops facial beauty benchmark with 0.9187 correlation","feed_subtitle":"Selective gating plus multi-scale pooling beats previous models on all three metrics.","key_machinery":"The MambaBlock is the core building block: an inverted residual structure that first expands channels with a 1x1 convolution, applies a 3x3 depthwise convolution, and then modulates the result with a parallel gating branch made of another 3x3 depthwise convolution plus a sigmoid activation. The element-wise multiplication of gate and main path gives the network input-dependent, spatially selective re-weighting. A multi-scale feature pyramid—adaptive average pooling to 1x1, 2x2, and 4x4, followed by concatenation—feeds multi-resolution context into a fully connected regression head. The residual skip connection (when dimensions match) and the inverted residual design keep training stable, whi","core_discovery":"The central discovery is that a simplified, content-aware gating path—a 3x3 depthwise convolution followed by a sigmoid, multiplied element-wise onto the main feature stream—captures much of the benefit of global attention for facial beauty regression. Inserted into an inverted residual block (expand, depthwise, gated multiply, project, residual skip), this 'MambaBlock' lets the network apply a spatially varying gain to its own features. Combined with a multi-scale feature pyramid that pools at 1x1, 2x2, and 4x4 and concatenates the results before a fully connected head, the architecture yields the reported metrics. The paper's ablation shows the gate alone raises Pearson correlation from 0.","pith_inferences":["Because the 'Mamba' in MambaBlock is only a sigmoid gate and does not actually run a selective scan or state-space parameterization, the paper's success would suggest that for FBP the essential ingredient is content-dependent feature gain, not long-range mixing; a true 2D SSM might add little on this benchmark.","The paper reports no parameter count, FLOPs, or training-time comparison; if the model is as compact as the MobileNetV2-style design suggests, it would strengthen the claim that it is 'efficient' relative to ViTs.","A natural testable extension: replace the fixed adaptive-average pooling scales with a learned multi-scale attention fusion, which might further improve PC beyond 0.9187.","The dependence on a single dataset means the gating benefit could be an artifact of SCUT-FBP5500's particular label distribution; cross-dataset validation on more diverse face sets would clarify that."],"forward_implications":["If the reported numbers reproduce, Mamba-CNN becomes the top-performing model on SCUT-FBP5500 on all three standard metrics, surpassing R3CNN (PC 0.9187 vs 0.9142, MAE 0.2022 vs 0.2120, RMSE 0.2610 vs 0.2800).","The large ablation gain from the gating branch (PC +0.0105) indicates that selective spatial re-weighting, not just deeper or wider convolutions, drives the improvement.","The architecture's linear-complexity gate makes it a practical candidate for high-resolution facial images and real-time beauty-scoring applications where quadratic attention is too costly.","The hybrid CNN+SSM design principle, if it generalizes, offers a template for other fine-grained regression tasks that need both local texture and global configuration."],"supporting_citations":[{"why":"Supplies the SCUT-FBP5500 benchmark and its prescribed cross-validation split, which defines the data and evaluation setting for all results.","marker":"[21]"},{"why":"R3CNN is the strongest baseline in Table 1; the state-of-the-art claim depends on beating its PC, MAE, and RMSE numbers.","marker":"[30]"},{"why":"DyAttenConv is a dynamic attention baseline that the gating mechanism is directly contrasted against in the comparison table.","marker":"[29]"},{"why":"Mamba is the source of the selective state space model inspiration that motivates the paper's gating design.","marker":"[19]"},{"why":"Provides the ResNet/ResNeXt baseline and residual learning principles that the MambaBlock builds on and compares against.","marker":"[13]"}],"fun_headline_variants":["Mamba-CNN: new SOTA for facial beauty prediction","Mamba-CNN hits 0.9187 PC on facial beauty","Selective gating powers Mamba-CNN to beauty SOTA","Mamba-CNN: gated SSM achieves 0.9187 on beauty"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The state-of-the-art claim assumes that every baseline in Table 1 was evaluated with the same train/test split, preprocessing, and score denormalization as Mamba-CNN, but the paper never specifies those protocol details.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-CNN: new SOTA for facial beauty prediction","Mamba-CNN hits 0.9187 PC on facial beauty","Selective gating powers Mamba-CNN to beauty SOTA","Mamba-CNN: gated SSM achieves 0.9187 on beauty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000985,"raw_usage":{"total_tokens":4015,"prompt_tokens":746,"completion_tokens":3269,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":3202}},"tokens_in":490,"tokens_out":3269,"duration_ms":24837,"temperature":1.0,"reasoning_tokens":3202,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:32:30.456380+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain Mamba-CNN and R3CNN on the exact same split of SCUT-FBP5500 with identical preprocessing and denormalization; if the Pearson correlation gap shrinks to under about 0.002 or the reported numbers do not reproduce, the claimed state-of-the-art result would not hold.","supporting_citations":[{"cited_title":"Regression Guided by Relative Ranking Using Convolutional Neural Network (R3CNN) for Facial Beauty Prediction","cited_arxiv_id":null,"evidence_quote":"R3CNN is the strongest baseline in Table 1; the state-of-the-art claim depends on beating its PC, MAE, and RMSE numbers."}],"review_version":1}