{"id":"d499e25b-f8e4-42e8-a102-095124c8603c","arxiv_id":"2608.04348","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"iStructTab reports that ordering tabular features via a graph-based descriptor score before transformer fusion improves multimodal image-table classification on most of six benchmarks, with uneven gains.","lead":"This paper proposes iStructTab, a model that decides an order for the table's columns, merges them with image features, and feeds that ordered sequence into a transformer with memory tokens. The model posts the best average rank over six image-plus-metadata benchmarks, but the gains are uneven, including a large accuracy loss versus a leading baseline on CheXpert.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GEDS is claimed to solve the CPP, but Algorithm 1's score is a per-feature norm with no demonstrated connection to Eq. (1); the discrete argsort prevents W from learning, so the ordering is an untrained batch-level heuristic whose contribution is never isolated against simple alternatives.","rationale":"The paper has real strengths: six held-out benchmarks, an informative component ablation (Table 3), and a transparent complexity model. My concern is not that the experiments are fabricated or that ordering is irrelevant; the 'w/o feature sequencing' row shows a large drop, which is genuine evidence that some structured ordering matters in OEMT. The load-bearing gap is that the paper's stated mechanism—GEDS as a CPP solver that reduces inter-feature dispersion—is not supported by any derivation, measurement, or comparison. The exact score s_j = ||σ(Â D W)_j||_2 with ascending argsort is an arbitrary heuristic in relation to the pairwise objective D(π) of Eq. (1). Because the permutation is discrete and the score vector is absent from L_total, the learnable matrix W cannot be trained; and because GEDS is applied per minibatch in Algorithm 2, the permutation and β targets are recomputed from batch statistics, so the 'optimal sequence' is not even stable at dataset level. These are correctness risks for the theoretical contribution and for attribution of the empirical gains, not for the existence of improved accuracy. The reader's CONDITIONAL verdict already requires closing this gap; my proposed check would settle whether the gap is purely presentational or genuinely invalidates the mechanism.","tokens_in":26526,"tokens_out":5239,"duration_ms":56053,"concrete_test":"On HAM10000 and DVM, hold the OEMT backbone, memory tokens, and Optuna budget fixed and compare four orderings: (a) Algorithm 1's GEDS order; (b) random permutation averaged over at least 5 seeds; (c) Fisher-score sorted ascending; (d) variance sorted ascending. For each ordering, report test accuracy, Avg. Rank/Regret, and the CPP cost D(π) computed with a concrete dissimilarity weight such as w_ij = |μ_i−μ_j| + |σ_i^2−σ_j^2|. Also run a single backward pass and check whether W in Eq. (9) receives a nonzero gradient; if not, the 'learnable' graph-convolution weight cannot be the mechanism. If GEDS does not produce the lowest D(π), or if any alternative ordering matches its accuracy within error bars, then the claim that GEDS implements the CPP objective and that its specific sequence explains the gains is falsifiable and currently unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that GEDS 'systematically determin[es] an effective feature sequencing' grounded in the Column Permutation Problem. But Algorithm 1 (steps 7–9) computes s_j = ||σ(Â D W)_j||_2 and sorts these per-feature norms in ascending order, whereas Eq. (1) defines D(π) as a sum of pairwise terms w_ij|π(i)−π(j)|. No derivation or measurement connects these objects: the paper never evaluates D(π), never defines w_ij operationally, and never compares πGEDS against random, Fisher, or variance-based orderings. The problem is compounded mechanically: the permutation is produced by an argsort, and the score vector s is not used in L_total, so the 'learnable' W in Eq. (9) receives no gradient through the discrete selection; the graph-convolution weight is effectively untrained. Moreover, Algorithm 2 applies GEDS per minibatch, so the mean/variance descriptors, the permutation, and the β target of Eq. (19) all change every minibatch, contradicting the dataset-level sequencing described in Eqs. (3)–(5). The Table 3 ablation shows that removing sequencing hurts, which is evidence that some ordering matters, but it does not establish that this specific ordering reduces dispersion or is better than cheap alternatives. The CPP grounding and the attribution of the reported gains to GEDS are therefore unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces iStructTab, a multimodal architecture for image-tabular classification. It proposes Graph-Enhanced Descriptor Sequencing (GEDS), which computes per-feature mean/variance descriptors, refines them via a graph convolution, and sorts the resulting L2 norms to produce a feature permutation πGEDS. This permutation is then used to reorder the fused feature vector before an Order-Aware Efficient Transformer with Memory Augmentation (OEMT) processes it. A sequencing loss aligns the model's predicted order scores with a linearly decreasing target β derived from πGEDS. The method is evaluated on six image-tabular benchmarks, with additional experiments on label noise, efficiency, calibration, and ablation. The paper reports the best average rank and lowest average regret among sixteen baselines, and a within-model ablation shows that removing feature sequencing degrades DVM accuracy substantially (96.85% to 80.80% untuned).","tokens_in":26698,"tokens_out":4417,"duration_ms":43126,"significance":"If the central claim is established, the paper would make a useful contribution by showing that explicit feature ordering, grounded in a combinatorial permutation objective, can improve multimodal fusion. The paper ships code and includes a broad empirical evaluation, which is commendable. The within-model ablation provides credible evidence that some form of feature sequencing matters for the proposed architecture. However, the specific grounding of GEDS in the Column Permutation Problem is not supported: the algorithm's score is never shown to relate to the dispersion objective D(π), and the reported gains are not isolated from simpler ordering heuristics. The significance of the contribution is therefore conditional on closing this gap.","major_comments":[{"comment":"The GEDS ordering score is never connected to the CPP objective D(π) in Eq. (1). The paper neither defines the weights w_ij operationally nor evaluates D(π) for πGEDS versus alternative orderings, so the abstract's claim that GEDS reduces inter-feature dispersion is unsupported. Please add a direct measurement of D(π) for πGEDS, random, original, variance-based, Fisher, and correlation-based orderings, and report both the dispersion values and the resulting classification accuracy.","section":"Section 3.1, Algorithm 1 steps 7–9, Eqs. (10)–(11)"},{"comment":"The graph-convolution weight W in Eq. (9) cannot receive a gradient through the discrete argsort in Eq. (11), and the score vector s is not used in the total loss L_total of Eq. (20). Consequently, the graph refinement is effectively an untrained fixed transformation, contradicting the claim that GEDS 'learns' a structured ordering. Please either remove the learnability claim, provide an analysis of W's role, or propose a differentiable relaxation that allows the ordering to be trained.","section":"Section 3.1, Eqs. (9)–(11) and Eq. (20)"},{"comment":"GEDS is applied independently to each minibatch in Algorithm 2, so the permutation π and the target β change at every step. This is inconsistent with the dataset-level optimal permutation described in Eqs. (3)–(5), and it means the sequencing loss in Eq. (20) aligns the model to a moving target. Please clarify whether πGEDS is intended as a fixed precomputed permutation or as a per-batch adaptive ordering, and study the effect of using a fixed training-set permutation instead.","section":"Algorithm 2, lines 5–7 vs. Section 3.1, Eqs. (3)–(5)"},{"comment":"The ablation row 'iStructTab w/o feature sequencing' drops accuracy to 80.80%, which shows that the ordering matters, but it does not establish that GEDS's specific ordering is responsible. The table lacks a comparison against the original column order, random permutation, or simple statistical orderings under the same OEMT backbone. Without these controls, the observed gains cannot be attributed to GEDS rather than to the mere act of imposing a deterministic order. Please add such comparisons.","section":"Section 4.E, Table 3"},{"comment":"On CheXpert, iStructTab (76.15%) underperforms STiL (88.60%) by 12.45 points, which is the largest gap between the top two methods on any dataset. The paper's headline 'best overall' claim rests on rank aggregation, but this substantial per-dataset deficit is not discussed. Please add a comment on why the method underperforms on this benchmark and whether the average-rank summary masks a systematic weakness in certain settings.","section":"Section 4.D1, Table 1, CheXpert row"}],"minor_comments":[{"comment":"Algorithm 1 returns only πGEDS, but Algorithm 2 line 5 unpacks '(π, s) ← GEDS(F)'. Please make the return values consistent.","section":"Algorithm 1 and Algorithm 2"},{"comment":"The heading 'G. Inference Level Ablation on Calibration and Robustness' is duplicated verbatim in the text; please remove the duplicate.","section":"Section 4.G heading"},{"comment":"The asterisk on iStructTab* in Figure 2 and the efficiency plots is never explained. Please define what the asterisk indicates.","section":"Section 4.D3 and Figure 2"},{"comment":"The row label 'iStructTab (ResNet-50 + GEDS)' omits OEMT, while other rows mention memory tokens and sequencing loss. Please clarify exactly which components are included in each ablation variant.","section":"Section 4.E, Table 3"},{"comment":"The Turing-style human-model evaluation reports no human entries; the section is a protocol description rather than a completed experiment. Please state this explicitly or move the material to future work.","section":"Supplementary Section I"},{"comment":"The code URL in the footnote (github.com/zadid6pretam/iStructTab) differs from the URL in Table B.5 (github.com/annonym414/multimodal-tabular-anon). Please unify these references.","section":"Code availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is already accepted at ICPR 2026 per the footnote, which is not itself a problem for this review, but it raises the bar for justifying major changes. The empirical core is strong enough that the identified gaps can likely be fixed with additional experiments, so major revision rather than rejection seems appropriate. One concern for the editor: the claimed CPP grounding is currently rhetorical, and the authors should be pressed to provide the dispersion measurements and ordering baselines before the scientific claims are accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the within-model ablation is the real contribution. If you ignore the CPP framing, the paper still shows that ordering features before a Linformer-based fusion helps on several benchmarks, and that a mismatched ordering loss is catastrophic. But GEDS as presented is not a learned solution to the Column Permutation Problem, and the paper's own CheXpert result contradicts the 'best overall' framing.\n\nWhat's new: GEDS (per-feature mean/variance descriptors, cosine-similarity graph convolution, L2-norm score, ascending sort) plus order-aware memory tokens and a sequencing loss is a concrete, reproducible recipe. The supplement is unusually complete: calibration, kNN probes, data efficiency, OOD confidence, and a noise-robustness table. The Table 3 ablation is the strongest evidence—removing sequencing drops DVM from 96.85 to 80.80 untuned, and applying a memory loss without an ordering signal collapses to 13.91. That is a genuine effect.\n\nSoft spots, in order of seriousness. First, the CPP grounding is not demonstrated. Equation (1) defines D(pi) as a pairwise weighted dispersion, but GEDS never evaluates D(pi), never defines w_ij operationally, and never compares pi_GEDS against random, Fisher, or variance-based orderings. So the paper's central conceptual claim is rhetorical. Second, the graph-convolution weight W in Eq. (9) receives no gradient—the permutation comes from an argsort, and s is not in the training loss. W is an untrained random projection; the 'graph-enhanced' description overstates the mechanism. Third, Algorithm 2 derives the permutation per minibatch, while Section 3 describes a dataset-level sequence. That inconsistency matters because the beta target changes every minibatch. Fourth, the comparisons are not fully controlled: iStructTab gets 20 Optuna trials per dataset, baselines use defaults, and Tables 1–2 have no error bars. The DVM win (99.29 vs 99.27) is noise-level. More importantly, CheXpert is a 12-point loss to STiL (76.15 vs 88.60), and the paper doesn't discuss it. Fifth, minor: two conflicting code URLs, no commit hash, and the 'Turing-style' evaluation has no human data yet.\n\nBottom line: this deserves a serious referee, but with major revision. Remove or re-frame the CPP claim, ablate against cheap orderings, make W actually learnable or drop it, report error bars, and address CheXpert. The empirical core is worth preserving; the current framing overclaims it.","headline":"Useful empirical evidence that feature ordering matters in image-tabular transformers, but the CPP framing is unsupported and the CheXpert gap undercuts the 'best overall' claim.","tokens_in":27407,"tokens_out":3878,"would_cite":false,"duration_ms":39423,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that ordering image-derived and tabular features before fusion is itself a learnable lever, and that a graph-based descriptor sequencing plus an order-aware transformer outperforms strong multimodal baselines across six…","keywords":["multimodal learning","feature sequencing","Column Permutation Problem","graph convolution","image-tabular fusion","transformer","feature ordering","label noise robustness"],"falsifier":"A reader could falsify the central claim by computing $D(\\pi_{\\mathrm{GEDS}})$ and $D(\\pi_{\\mathrm{random}})$, $D(\\pi_{\\mathrm{Fisher}})$, and $D(\\pi_{\\mathrm{corr}})$ on the same fused features of any benchmark; if ascending-norm ordering does not systematically yield lower $D$ than those alternatives, the CPP grounding collapses. Separately, an ablation that feeds the same pipeline the reversed or shuffled $\\pi_{\\mathrm{GEDS}}$ and finds no accuracy drop would falsify the claim that the sequencing itself carries the gain.","tokens_in":26149,"feed_emoji":"🧩","tokens_out":6707,"duration_ms":62356,"temperature":0.7,"pith_summary":"The paper claims that the order in which tabular and image-derived features are fed to a fusion transformer is not a nuisance but a learnable lever, and that choosing that order algorithmically improves prediction and robustness. It proposes iStructTab, which first computes per-feature statistical descriptors, refines them through a similarity graph, and sorts features by the norms of the refined descriptors to produce a permutation $\\pi_{\\mathrm{GEDS}}$. An order-aware transformer with memory tokens and a sequencing loss then consumes the features in that order. Across six image-plus-table benchmarks the resulting model reports the best average rank and the lowest average regret among sixteen methods, and it retains accuracy under noisy labels. If right, the paper establishes that structured feature sequencing can serve as a broadly effective component of multimodal fusion, not a niche preprocessing step.","feed_headline":"Learned feature order lifts image-plus-table accuracy","feed_subtitle":"Sorting fused features by graph-refined descriptors before a transformer sees them beats strong multimodal baselines on six benchmarks.","key_machinery":"The mechanism is GEDS plus OEMT. GEDS computes, for each fused feature column, a descriptor $\\mathbf{d}_j = [\\mu_j, \\sigma^2_j]^\\top$, row-normalizes the descriptor matrix $\\mathbf{D}$, builds a cosine-similarity graph with self-connections, symmetrically normalizes it to $\\hat{\\mathbf{A}}$, and applies one graph convolution $\\mathbf{D}' = \\sigma(\\hat{\\mathbf{A}}\\mathbf{D}\\mathbf{W})$. Each feature then receives a score $s_j = \\|\\mathbf{D}'_{j:}\\|_2$, and the permutation is $\\pi_{\\mathrm{GEDS}} = \\operatorname{argsort}(s_1, \\ldots, s_m)$ in ascending order. OEMT projects the reordered features into tokens, uses a learnable order-aware pooling matrix to compress $m$ tokens into $k \\ll m$ summary tokens, prepends $M$ learnable memory tokens, and processes the resulting sequence with a Linformer encoder that has linear attention. The classification head pools the memory tokens, and the total loss adds a sequencing term $\\lambda_{\\mathrm{FS}}\\|\\mathbf{s}' - \\boldsymbol{\\beta}\\|_2^2$ that penalizes deviation from a linear importance ranking $\\boldsymbol{\\beta}$ defined by $\\pi_{\\mathrm{GEDS}}$.","core_discovery":"The central claim is that multimodal fusion of images and tabular data can be formulated as a Column Permutation Problem (CPP), in which one seeks a permutation $\\pi$ of fused features that minimizes a dispersion cost $D(\\pi)=\\sum_{i<j} w_{ij}|\\pi(i)-\\pi(j)|$, and that a cheap graph-based heuristic produces such a permutation. The heuristic, GEDS, builds a two-dimensional descriptor per feature (mean and variance), normalizes the descriptors, forms a cosine-similarity graph, applies one graph convolution, and sorts features by the ascending L2 norms of the refined descriptors. The resulting order is enforced by the OEMT backbone through order-aware pooling, learnable memory tokens, and a sequencing loss that pushes predicted order scores toward a linear rank target. The paper's evidence is comparative: across six benchmarks, iStructTab attains the best average rank (1.50 ± 0.76) and lowest average regret (2.21 ± 4.59), beats prior methods under label noise, and its ablations show that removing feature sequencing collapses accuracy on DVM from 99.29% to roughly 81–83%.","pith_inferences":["Beyond the paper, the sequencing mechanism is not tied to image-plus-table input: any concatenated embedding whose columns are exchangeable could be run through GEDS, and a cheap test would be applying iStructTab-style ordering to text-plus-audio or multi-table fusion.","The paper does not establish that $\\pi_{\\mathrm{GEDS}}$ actually minimizes $D(\\pi)$; an editorial next experiment is replacing the heuristic with direct optimization of $D(\\pi)$, for instance by differentiable sorting, and checking whether the gains grow, shrink, or vanish.","The robustness under label noise may come partly from the sequencing regularization term rather than from the order itself; ablating $\\lambda_{\\mathrm{FS}}$ at high noise levels would separate those effects, but the paper does not report that ablation."],"forward_implications":["The six-dataset average rank of 1.50 ± 0.76 and average regret of 2.21 ± 4.59 imply that GEDS plus OEMT is consistently competitive across medical and consumer image-tabular benchmarks, not just on one favorable dataset.","Ablations on DVM show that removing feature sequencing drops accuracy to roughly 81–83% from 99.29%, which implies the ordering mechanism is a primary driver of the method's accuracy in that setting, not a small correction.","Under 60% label noise on HAM10000 the model reaches 76.42% versus 70.1% for the best prior method, implying that the sequencing-based representation degrades more gracefully under corrupted supervision.","Because GEDS costs $O(Nm + m^2)$ on fused features and OEMT uses linear attention, the sequence-enforcement pipeline adds modest overhead, implying it can be attached to existing multimodal encoders without a compute cliff.","Successful explicit feature ordering suggests that permutation-aware designs, whether by averaging over orders or learning an order, should be part of the comparison suite for future multimodal methods."],"supporting_citations":[{"why":"Formalizes the Column Permutation Problem with a Δ-evaluation function that penalizes pairwise variance or dissimilarity, grounding the paper's $D(\\pi)$ objective.","marker":"[49]"},{"why":"Shows explicit feature sequencing affects convergence and accuracy for tabular transformers, the prior result iStructTab extends to multimodal fusion.","marker":"[31]"},{"why":"Supplies the graph-convolution operation used to refine per-feature statistical descriptors in GEDS.","marker":"[47]"},{"why":"Provides the Linformer linear-attention encoder that makes OEMT efficient on long ordered feature sequences.","marker":"[65]"},{"why":"Sets the ResNet-plus-transformer encoder pipeline and serves as a leading multimodal baseline (TIP).","marker":"[16]"},{"why":"Defines the benchmarking protocol and is the strongest multimodal comparator (STiL); several ablations follow its recipe.","marker":"[17]"},{"why":"Provides the ResNet image encoder used for all image features in the fusion pipeline.","marker":"[34]"},{"why":"Provides the transformer attention machinery on which the order-aware encoder is built.","marker":"[62]"},{"why":"Establishes that column permutation problems are NP-hard, motivating the cheap heuristic GEDS rather than exact optimization.","marker":"[44]"}],"fun_headline_variants":["Graph-based feature sequencing boosts image-table fusion","Order-aware transformer beats baselines via graph-sorted features","Treating fusion as column permutation lifts multimodal accuracy","iStructTab: Graph-derived order for better image-table learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that sorting fused features by the ascending L2 norm of graph-convolved mean/variance descriptors genuinely approximates minimizing the dispersion objective $D(\\pi)$; the paper does not measure $D(\\pi)$ or compare this order with cheaper alternatives.","fun_headline_variants_meta":{"raw":{"variants":["Graph-based feature sequencing boosts image-table fusion","Order-aware transformer beats baselines via graph-sorted features","Treating fusion as column permutation lifts multimodal accuracy","iStructTab: Graph-derived order for better image-table learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1327,"prompt_tokens":901,"completion_tokens":426,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":363}},"tokens_in":517,"tokens_out":426,"duration_ms":4531,"temperature":1.0,"reasoning_tokens":363,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:19:07.875289+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could falsify the central claim by computing $D(\\pi_{\\mathrm{GEDS}})$ and $D(\\pi_{\\mathrm{random}})$, $D(\\pi_{\\mathrm{Fisher}})$, and $D(\\pi_{\\mathrm{corr}})$ on the same fused features of any benchmark; if ascending-norm ordering does not systematically yield lower $D$ than those alternatives, the CPP grounding collapses. Separately, an ablation that feeds the same pipeline the reversed or shuffled $\\pi_{\\mathrm{GEDS}}$ and finds no accuracy drop would falsify the claim that the sequencing itself carries the gain.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Linformer linear-attention encoder that makes OEMT efficient on long ordered feature sequences."},{"cited_title":"In: ECCV","cited_arxiv_id":null,"evidence_quote":"Sets the ResNet-plus-transformer encoder pipeline and serves as a leading multimodal baseline (TIP)."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"Defines the benchmarking protocol and is the strongest multimodal comparator (STiL); several ablations follow its recipe."},{"cited_title":"NeurIPS30(2017)","cited_arxiv_id":null,"evidence_quote":"Provides the transformer attention machinery on which the order-aware encoder is built."},{"cited_title":"WH Freeman (1979)","cited_arxiv_id":null,"evidence_quote":"Establishes that column permutation problems are NP-hard, motivating the cheap heuristic GEDS rather than exact optimization."}],"review_version":1}