{"id":"f43c578f-c6b8-4789-8635-db75465a5ee2","arxiv_id":"2505.01057","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"GeloVec adds a Chebyshev-distance attention mechanism to U-Net and claims mIoU gains that its own tables do not consistently support.","lead":"This paper proposes GeloVec, an attention module for image segmentation that adds Chebyshev-distance-based geometric weighting to a U-Net. The authors report small gains in segmentation accuracy on three niche datasets, but the paper's own tables do not match the gains stated in the abstract.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (3.6) subtracts a per-pixel scalar inside softmax; shift invariance makes the geometric Chebyshev modulation a no-op, so the claimed attention smoothing cannot work as described.","rationale":"The reader's weakest assumption concerns the reproducibility and internal consistency of the reported numbers. That is a valid and serious problem, as the abstract's 2.1/2.7/2.4 gains do not match the +0.1 and +1.7 gains visible in Tables 1 and 3, and the LSDSC IoU is 84.3 in the text but 85.4 in Table 2. However, I found a more fundamental and more specific flaw in the proposed method itself: Equation (3.6) subtracts a per-pixel scalar D_norm from every key logit for a given query, and because softmax is invariant to uniform row shifts, this term cancels exactly. Thus the Chebyshev distance metric, which the paper presents as the core innovation for stabilizing attention, does not influence the attention weights at all. This is not a question of missing hyperparameters or unclear baselines; it is a mathematical property of the equation as written. The same conclusion would hold even if the experimental protocol were perfectly documented and the tables were corrected. I therefore disagree with the reader's identification of the weakest assumption, though I reach the same overall verdict. The central claim of geometric attention smoothing cannot stand as described. A simple numerical check on random tensors would settle the issue definitively, and no reimplementation or new dataset is needed.","tokens_in":8716,"tokens_out":7208,"duration_ms":72531,"concrete_test":"Implement Equations (3.3)-(3.6) as written on random feature maps with fixed Q, K, and D_norm; compute the softmax attention weights for lambda = 0 and lambda = 1. If the two attention matrices are identical to floating-point precision (e.g., maximum absolute difference below 1e-6), the geometric modulation has no effect and the central mechanism is a no-op.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing flaw is internal to the described mechanism, not merely a matter of reproducibility. Equation (3.4) defines D_norm as a per-pixel scalar: it takes the max of the Chebyshev distances over neighboring positions and applies a 1x1 convolution and sigmoid at that spatial location. Equation (3.6) then computes A_raw = Softmax(QK^T / sqrt(d_k) - lambda * D_norm). For a fixed query position, D_norm is a single scalar that is subtracted identically from every key logit in that row. Since softmax is shift-invariant, subtracting a constant from all logits in a row leaves the resulting attention distribution unchanged. Therefore, lambda * D_norm has no effect on the attention weights at all; the Chebyshev distance does not modulate attention as the abstract and Section 3.2 claim. The geometric 'attention smoothing' is mathematically inert in the attention pathway, and any observed benefit would have to come from the separate edge gate in Equation (3.5), which is not the paper's stated contribution. This is an internal inconsistency, independent of the missing training details and the contradictory IoU values in Tables 1-3 versus the abstract.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeloVec, a module inserted at multiple encoder stages of a U-Net with ResNet-34 backbone for semantic segmentation. The module combines an orthogonal basis transform, a Chebyshev-distance geometric adaptive sampling term, an edge gate, and a softmax attention mechanism modulated by a geometric distance. The authors report mIoU gains over SegFormer on three datasets (CUB-200-2011, LSDSC, FSSD) and claim that Riemannian geometry provides theoretical guarantees on segmentation stability. The central claim is that geometric smoothing of attention improves boundary preservation and intra-class coherence.","tokens_in":9031,"tokens_out":2461,"duration_ms":25590,"significance":"If the proposed mechanism worked as described, a geometry-based attention smoothing layer that is computationally cheap and pluggable into CNN segmentation backbones would be a useful contribution. The paper is also transparent in giving explicit equations for the proposed modules and in reporting numbers for four baseline architectures, which is commendable. However, the central mechanism is mathematically inert as written: the Chebyshev term added inside the softmax is cancelled by softmax shift invariance. In addition, the experimental numbers contradict each other between the abstract, the text, and the tables, and no training protocol or statistical evaluation is provided. These issues affect the central claims directly, not just the presentation.","major_comments":[{"comment":"The geometric modulation of the attention weights is a no-op. D_norm in Eq. (3.4) is, for each spatial query position, a single scalar: it is the maximum of the local Chebyshev distances over neighbors, passed through a 1x1 convolution and a sigmoid at that spatial location. In Eq. (3.6), this scalar is subtracted from every key logit in the row of the attention matrix for that query. Since softmax is invariant to adding a constant to all logits in a row, subtracting lambda times D_norm changes no attention probability. Therefore, the claimed Chebyshev-distance geometric smoothing of the attention pathway cannot work as described; any observed benefit would have to come from the separate edge gate in Eq. (3.5), which is distinct from the paper's stated core contribution. This is a load-bearing internal inconsistency.","section":"Sec. 3.2, Eq. (3.6)"},{"comment":"The reported numbers are internally contradictory. The abstract and Section 1 claim mIoU gains of 2.1%, 2.7%, and 2.4% over state-of-the-art methods on CUB-200-2011, LSDSC, and FSSD, respectively. Table 1 shows GeloVec at 83.6 vs. SegFormer at 83.5, a gain of only 0.1 points on CUB-200-2011. Table 3 shows 82.9 vs. 81.2, a gain of 1.7 points on FSSD. Additionally, Section 4 states an LSDSC IoU of 84.3% for GeloVec, while Table 2 lists 85.4%. These discrepancies mean the abstract's headline figures are not reproducible from the tables in the same manuscript, and the high-level claim of consistent 2+ point gains is unsupported.","section":"Abstract vs. Sec. 4, Tables 1-3"},{"comment":"The evaluation section provides no training hyperparameters, optimizer, learning rate schedule, batch size, number of epochs, loss function, data split details, or number of runs for any method. No error bars or statistical significance tests are reported. This is insufficient for a claim of consistent superiority over state-of-the-art baselines, especially when the reported differences are as small as 0.1 mIoU. The absence of ablations is also critical: the paper never isolates the contributions of the orthogonal basis transform, the geometric adaptive sampling, the edge gate, and the attention modulation, so the attribution of any observed gains to the proposed 'geometric smoothing' is not established even if the numbers were reliable.","section":"Sec. 4, Experimental protocol"},{"comment":"The paper repeatedly asserts that GeloVec has a 'mathematical foundation in Riemannian geometry' and provides 'theoretical guarantees on segmentation stability,' but no Riemannian metric, manifold structure, or stability theorem is defined or proved anywhere. The only geometric elements are the Chebyshev distance in Eq. (3.3) and the L2 normalization in Eq. (3.2). Moreover, Eq. (3.2) calls the reshaped normalized vectors an 'orthogonal basis,' but L2 normalization does not by itself make a set of vectors mutually orthogonal. These statements are thus assertion rather than derivation, and the claimed guarantees are not supported by the presented mathematics.","section":"Secs. 3 and 5, Theoretical claims"}],"minor_comments":[{"comment":"There are recurring typos such as 'spacial' for 'spatial' (e.g., Section 3.1 and Section 5), and the phrase 'L-norm' in Section 3.2 should be 'L-infinity norm' or 'Chebyshev norm' for clarity.","section":"Throughout"},{"comment":"The notation W_i for the learnable sampling weight is ambiguous: it is unclear whether the same weight vector is used across all feature channels d, and how the product W_i*(F_pc,d - F_pi,d) is computed when W_i is a scalar, vector, or tensor. The equation and surrounding text should specify the exact dimensions.","section":"Eq. (3.3)"},{"comment":"The architecture overview figure is referenced but not discussed in enough detail in the text; in particular, the placement of the four GeloVec variants (GeloVecLow, Mid, High, VeryHigh) is described verbally but not marked in the figure, which makes the multi-scale integration harder to follow.","section":"Fig. 1"},{"comment":"The visual results in Figures 2 and 3 are presented as qualitative evidence, but no explanation is given for what is being displayed in Figure 2 beyond 'attention values distribution,' and the caption of Figure 3 does not specify which dataset or which images are shown. This limits the usefulness of the qualitative comparison.","section":"Sec. 4"}],"recommendation":"reject","confidential_remarks":"The manuscript's central mechanism is mathematically inert as written (Eq. 3.6), and the reported results are internally inconsistent (abstract vs. tables vs. text). These are not presentation-level issues; they strike at the core claims. Even with a corrected attention formulation, the absence of any training protocol or ablations would require a substantially new experimental section. I cannot see a path to acceptance within the scope of the current manuscript. The paper also asserts theoretical guarantees without providing them, which is a scope mismatch for a serious venue. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this one has a load-bearing math error. The geometric distance term in Eq. (3.6) cannot influence attention, because for a fixed query position D_norm is a single scalar subtracted from every logit, and softmax is shift-invariant. The stress-test note is correct. That leaves the paper's stated mechanism—distance-modulated attention—without any mechanism. The separate edge gate in Eq. (3.5) is a plausible component, but the paper doesn't frame it as the main contribution.\n\nWhat is actually new: the specific combination of a 1x1-projected, L2-normalized basis transform, weighted Chebyshev edge gating, and a distance penalty in attention is not in the prior work I know. The edge-preservation idea is reasonable, and the literature review is broad. The writing is readable. So it's not a pure repackaging.\n\nThe soft spots are not minor. First, the math error above is internal, not a reproducibility issue. Second, the abstract promises 2.1/2.4 point gains over state of the art, but Table 1 shows +0.1 over SegFormer on CUB-200, Table 3 shows +1.7, and the LSDSC text says 84.3 while Table 2 lists 85.4. Without error bars or training details, these numbers can't be taken as reliable. Third, the Riemannian geometry guarantee is asserted, not derived—no metric is defined, no theorem is stated. Fourth, the baselines (SegFormer MiT-B1, HRNetV2-W18) are not current state of the art. There is also no code.\n\nWho this is for: maybe a reader curious about geometric edge gating in CNNs. But as a submission, the central claim fails on its own equations, and the empirical record is inconsistent. I would not spend referee time on it in current form; it needs a rewritten mechanism, a corrected derivation, and a proper evaluation before it deserves review.\n\nRecommendation: desk reject.","headline":"GeloVec's headline gains don't survive contact with its own equations: the geometric attention modulation is a softmax no-op, and the tables contradict the abstract.","tokens_in":9523,"tokens_out":3177,"would_cite":false,"duration_ms":30974,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","68T45","15A66","53A05","14R05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"GeloVec claims geometric smoothing via Chebyshev distances and an orthogonal basis transform stabilizes attention and lifts segmentation accuracy on three benchmarks.","keywords":["semantic segmentation","geometric smoothing","Chebyshev distance","attention mechanism","orthogonal basis transform","feature extraction","UNet","ResNet-34"],"falsifier":"Run GeloVec and the four baselines on the same three datasets with shared training settings and multiple seeds, then compare mean IoU; if GeloVec does not beat SegFormer by the reported margins, or if the numbers require unpublished configuration choices to line up, the central claim fails.","tokens_in":8511,"feed_emoji":"🧩","tokens_out":11777,"duration_ms":102893,"temperature":0.7,"pith_summary":"GeloVec is a proposal for treating attention in semantic segmentation as a geometric operation: instead of letting attention scores drift near object boundaries, it smooths features by computing Chebyshev distances in a higher-dimensional orthogonal feature space and uses those distances to modulate attention and preserve edges. The paper claims this geometric smoothing stabilizes attention, improves boundary precision, and keeps intra-class regions homogeneous, and reports that a UNet with a ResNet-34 backbone equipped with GeloVec modules beats U-Net, DeepLabV3+, HRNet, and SegFormer on Caltech Birds-200, LSDSC, and FSSD, with mean IoU gains of 2.1, 2.7, and 2.4 percentage points respectively. If the reported numbers hold, the practical significance is that a lightweight geometric block added between encoder stages can improve segmentation without switching to a heavier backbone or a transformer.","feed_headline":"Chebyshev-distance smoothing lifts segmentation scores on three benchmarks","feed_subtitle":"A lightweight UNet module beats SegFormer and other baselines on bird, fish, and flood images.","key_machinery":"The load-bearing object is the GeloVec module, whose steps are given in equations (3.1)-(3.6): an Orthogonal Basis Transform (OBT) projects features into a normalized higher-dimensional orthogonal basis; Geometric Adaptive Sampling (GAS) computes the Chebyshev distance $D_{\\mathrm{chebyshev}}(p_c,p_i)=\\max_{d}|W_i(F_{p_c,d}-F_{p_i,d})|$ between a center pixel and weighted dilated neighbors; a $1\\times1$ convolution plus sigmoid converts the max over neighbors into a normalized distance $D_{\\mathrm{norm}}$; and an edge gate $G_{\\mathrm{edge}}=\\sigma(\\mathrm{conv}(D_{\\mathrm{norm}}))$ blends original and edge features. The attention logits subtract $\\lambda D_{\\mathrm{norm}}$ before softmax. The Chebyshev max over channels is what gives the module its boundary sensitivity: a single channel jumping at an edge raises the distance and suppresses attention there, while flat regions keep small distances and stay homogeneous.","core_discovery":"The central claim is that replacing plain attention with a geometry-aware smoothing module yields more coherent segmentations. GeloVec inserts a module after each encoder stage of a UNet; the Orthogonal Basis Transform projects features into a normalized higher-dimensional orthogonal basis, and Geometric Adaptive Sampling compares each pixel with dilated neighbors using a modified Chebyshev distance, taking the maximum absolute per-channel difference. The normalized distance gates an edge-preserving blend and is subtracted from the scaled dot-product attention logits, so locations with large geometric change are treated as boundaries rather than being smoothed over. The paper reports mean IoU gains of 2.1, 2.7, and 2.4 percentage points over state-of-the-art methods on CUB-200-2011, LSDSC, and FSSD, with table IoU values of 83.6, 85.4, and 82.9 and the largest advantages in precision.","pith_inferences":["If the reported gains reproduce, the same geometric smoothing block could be inserted into transformer-based segmentation heads, where boundary instability is also reported; the paper does not test that configuration.","A component ablation, removing the edge gate, the Chebyshev modulation, or the orthogonal transform one at a time, would isolate how much of the improvement is boundary preservation versus intra-class smoothing.","The normalized distance map $D_{\\mathrm{norm}}$ could be exported as an interpretable edge prior for other dense-prediction tasks such as depth estimation or contour detection, an application the paper does not explore."],"forward_implications":["On the three datasets studied, GeloVec reports higher IoU, F1, precision, and recall than every baseline listed, so the geometric module is claimed to improve all four metrics at once.","The precision values (92.1, 90.7, and 89.6) are the largest gaps over the baselines, which the paper attributes to the geometric distance reducing false positives at boundaries.","GeloVec achieves these numbers with a ResNet-34 backbone, lighter than DeepLabV3+'s ResNet-50 and HRNet's HRNetV2-W18, implying the gain comes from the module rather than a larger encoder.","Because the module sits between encoder stages and its refined features travel through skip connections, the geometric smoothing affects both the semantic and the detail path of the UNet.","The abstract claims the transformations are parallelizable and lossless, so the approach is presented as adding accuracy without a meaningful compute penalty."],"supporting_citations":[{"why":"Supplies the U-Net encoder-decoder design that GeloVec augments with geometric modules between encoder stages.","marker":"[31]"},{"why":"Establishes the fully convolutional approach to dense prediction that motivates the end-to-end segmentation architecture.","marker":"[25]"},{"why":"Contributes the DeepLabV3+ baseline and the dilated-convolution context strategy that GeloVec is compared with.","marker":"[10]"},{"why":"Defines the scaled dot-product attention that GeloVec modifies by subtracting a geometric distance term.","marker":"[38]"},{"why":"Defines the Caltech-UCSD Birds-200-2011 dataset used for the first evaluation table.","marker":"[39]"},{"why":"Defines the large-scale fish segmentation dataset (LSDSC) used for the second evaluation table.","marker":"[37]"},{"why":"Defines the Flood Semantic Segmentation Dataset used for the third evaluation table.","marker":"[19]"},{"why":"Spatial pyramid pooling is cited as inspiration for the multispatial transformation in the GeloVec module.","marker":"[16]"},{"why":"Feature pyramid networks are cited as inspiration for the multi-scale feature handling behind the multispatial design.","marker":"[22]"}],"fun_headline_variants":["Geometry-aware smoothing lifts segmentation mIoU by ~2.4 points","Chebyshev-distance attention block sharpens boundaries in UNet","GeloVec: higher-dimensional geometric smoothing beats three benchmarks","Adaptive Chebyshev gates preserve edges while smoothing interiors","Lightweight geometric smoothing gains 2-3 mIoU on three datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central comparison assumes the reported IoU numbers were produced by the described architecture under a consistent, reproducible protocol, yet no optimizer, epochs, splits, or run count are given, and the abstract, text, and tables give conflicting values such as 84.3 versus 85.4 for LSDSC.","fun_headline_variants_meta":{"raw":{"variants":["Geometry-aware smoothing lifts segmentation mIoU by ~2.4 points","Chebyshev-distance attention block sharpens boundaries in UNet","GeloVec: higher-dimensional geometric smoothing beats three benchmarks","Adaptive Chebyshev gates preserve edges while smoothing interiors","Lightweight geometric smoothing gains 2-3 mIoU on three datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001191,"raw_usage":{"total_tokens":4922,"prompt_tokens":962,"completion_tokens":3960,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":3871}},"tokens_in":578,"tokens_out":3960,"duration_ms":22700,"temperature":1.0,"reasoning_tokens":3871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:27:27.543473+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GeloVec and the four baselines on the same three datasets with shared training settings and multiple seeds, then compare mean IoU; if GeloVec does not beat SegFormer by the reported margins, or if the numbers require unpublished configuration choices to line up, the central claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the fully convolutional approach to dense prediction that motivates the end-to-end segmentation architecture."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the DeepLabV3+ baseline and the dilated-convolution context strategy that GeloVec is compared with."},{"cited_title":"Gomez, Lukasz Kaiser, and Illia Polosukhin","cited_arxiv_id":null,"evidence_quote":"Defines the scaled dot-product attention that GeloVec modifies by subtracting a geometric distance term."},{"cited_title":"The Caltech-UCSD birds-200-2011 dataset","cited_arxiv_id":null,"evidence_quote":"Defines the Caltech-UCSD Birds-200-2011 dataset used for the first evaluation table."},{"cited_title":"A large-scale dataset for fish segmentation and classification","cited_arxiv_id":null,"evidence_quote":"Defines the large-scale fish segmentation dataset (LSDSC) used for the second evaluation table."},{"cited_title":"Flood semantic segmentation dataset, 2022","cited_arxiv_id":null,"evidence_quote":"Defines the Flood Semantic Segmentation Dataset used for the third evaluation table."},{"cited_title":"Spatial pyramid pooling in deep convolutional networks for visual recognition","cited_arxiv_id":null,"evidence_quote":"Spatial pyramid pooling is cited as inspiration for the multispatial transformation in the GeloVec module."},{"cited_title":"Feature pyramid networks for object detection","cited_arxiv_id":null,"evidence_quote":"Feature pyramid networks are cited as inspiration for the multi-scale feature handling behind the multispatial design."}],"review_version":1}