{"id":"9640bd2b-73b3-421a-85c9-2a948e749101","arxiv_id":"2412.16487","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TMCN uses a Mamba selective fusion module (TMFN) and an average-similarity contrastive loss (AsCL) to achieve reported state-of-the-art multi-view clustering on four benchmarks.","lead":"This paper combines a Mamba selective state-space model with a cluster-aware contrastive loss for multi-view clustering, reporting large accuracy gains on four benchmark datasets. A reader might care because it suggests that selective sequence models can replace simple weighted fusion in multi-view learning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (16) as written is not a valid contrastive loss: the denominator subtracts e^{1/tau}, which can be far larger than the sum, making the log argument negative; the AsCL module and the SOTA claim depend on this formula.","rationale":"Good-faith reading: the paper proposes TMFN and AsCL to improve multi-view clustering, and its central claim is SOTA performance. For that claim to hold, both modules must be correctly specified. TMFN is a plausible use of Mamba's selective state-space mechanism, and the code release is a positive step. However, AsCL is the algorithmic core of the second contribution, and its printed definition in Eq. (16) is mathematically inconsistent: subtracting e^{1/tau} instead of removing the j=i unit term makes the denominator potentially negative. This is a correctness risk distinct from the reader's concern about S_ij being a reliable cluster-similarity proxy: even a perfect proxy would not rescue Eq. (16) as written. The proxy-reliability concern is valid and secondary, but the formula error is more immediate and can be settled directly by inspecting and running the released code. If the code uses a corrected loss, the paper still contains a formal error that must be fixed and the ablation redone; if the code uses Eq. (16) literally, training should fail or behave pathologically on small datasets with small tau. Either way, the SOTA claim is conditional on resolving this issue. I therefore keep the reader's CONDITIONAL verdict rather than accepting or rejecting outright: the architecture has independent plausibility, no formal verification or error bars are provided, and the code release makes the necessary check straightforward.","tokens_in":8135,"tokens_out":6536,"duration_ms":58607,"concrete_test":"Use the released repository to compute L_Asc exactly as printed in Eq. (16) on random initialized embeddings with tau=0.07 and N=551 (Prokaryotic size), without training; if the log argument is negative or NaN for any sample, the printed loss is not usable. Then compare this result against the code's actual AsCL implementation: if they differ, Eq. (16) is not the method actually run, and Table III's AsCL ablation needs to be re-run with the printed loss to determine which loss produces the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The AsCL loss in Eq. (16) is not internally consistent. Since S_ii=1 for every view, the j=i term in the denominator is e^{(1-S_ii)C/tau}=1, so removing the positive pair would require subtracting 1, not e^{1/tau}. For typical tau<1, e^{1/tau} > e ~ 2.718, and at tau=0.07 it is about 1.2e6, which can exceed reasonable denominator sums early in training; the log argument can then be negative or undefined. The sentence after Eq. (16) also misdescribes the equation: it says C increases with decreased S_ij, whereas Eq. (16) only changes the exponent weight (1-S_ij) multiplying C, not C itself. AsCL is credited with large ablation gains (Table III: removing AsCL drops ACC by 9.92 on Hdigit and by 2.71 on Cifar100), so the empirical SOTA claim rests on this loss being well-posed. If the released code implements a different, valid loss, the paper must state that Eq. (16) is a typo and re-verify the ablation attribution; if the code implements Eq. (16) literally, AsCL as defined cannot be the source of the reported gains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes TMCN, a deep multi-view clustering method combining autoencoder reconstruction, a 'Trusted Mamba Fusion Network' (TMFN) based on selective state-space models, and an 'Average-similarity Contrastive Learning' (AsCL) module that re-weights contrastive pairs using average cross-view cosine similarity. The authors claim state-of-the-art ACC/NMI/PUR on Hdigit, Cifar100, Prokaryotic, and Wiki, and report ablations showing both TMFN and AsCL provide large gains. The paper also includes convergence, parameter sensitivity, and t-SNE visualizations.","tokens_in":8413,"tokens_out":6165,"duration_ms":50161,"significance":"If validated, the use of Mamba-style selective gating for multi-view fusion and the shift from instance-level to cluster-level contrastive alignment would be a useful contribution to deep multi-view clustering. The paper promises code and provides a clear framework diagram. However, the AsCL loss appears mathematically ill-posed, the empirical state-of-the-art claim rests on single-run numbers without error bars, and at least one ablation configuration is under-specified. These issues are load-bearing, so the contribution cannot be accepted in its current form.","major_comments":[{"comment":"The AsCL loss is not well defined as written. Since S_ii=1 for every view, the j=i term in the denominator is e^{(1-S_ii)C(...)/τ}=e^0=1; removing that positive pair would require subtracting 1, not e^{1/τ}. For τ<1, e^{1/τ} is large and can exceed the denominator sum, making the argument of the log negative and the loss undefined. The sentence after Eq. (16) additionally misstates the dependence: C is not increased by decreased S_ij; only the exponent weight (1-S_ij) changes. Because Table III credits AsCL with substantial ablation gains (e.g., 9.92 ACC on Hdigit and 2.71 on Cifar100), the SOTA claim depends on this formula. Please correct the equation or clarify if the implementation uses a different loss, and re-verify the ablation results.","section":"Section II-C, Eq. (16)"},{"comment":"The comparative results are reported as single values with no standard deviations, repeated runs, or seeds. On Cifar100 the reported TMCN ACC of 0.9853 versus CoMVC's 0.6569 is a 32-point margin, and the MFLVC result of 0.1342 ACC with NMI 0.0070 is far below the behavior reported in the original MFLVC paper. These numbers need to be reconciled with the evaluation protocol. Please provide mean±std over multiple runs, a description of how the deep features are obtained for each dataset, and the exact training protocol.","section":"Section III-B, Table II"},{"comment":"The 'No-TMFN' ablation is under-specified. The text says the fused representation Z is the concatenation of all view-specific representations, but Section II-B Eq. (5) already concatenates sequence vectors before the Mamba network. It is unclear whether the baseline operates on the final one-dimensional embeddings z_i^m or on the sequence-level concatenation, and whether all other components are kept identical. Without this clarification, the ablation cannot isolate the contribution of the selective mechanism.","section":"Section III-C, Table III and Section II-B"}],"minor_comments":[{"comment":"The operations rea1, rea2, rea3, and rea4 are not defined in the text or figure; please define the segmentation/reshape operation explicitly.","section":"Section II-B, Eqs. (4)-(10)"},{"comment":"The clustering objective is written as min_{U,V} ||H - UV||_2^2 with constraints U1=1, U≥0 and called k-means, but this is not the standard k-means objective; please clarify the relationship or present the k-means update used.","section":"Section II-D, Eq. (18)"},{"comment":"The caption says 'convergence analysis and visualization analysis,' but only one set of curves is clearly described; please label which panel is the t-SNE visualization and which is the convergence curve.","section":"Figure 2"},{"comment":"The sentence 'Five state-of-the-art clustering methods' is followed by a list of five methods; the capitalization of 'Five' and the colon usage should be fixed.","section":"Section III-B"},{"comment":"The page range for [25] is typeset as 'pp. 162:21 090–21 110,' which is malformed; please correct it.","section":"Reference [25]"}],"recommendation":"major_revision","confidential_remarks":"The abstract includes a GitHub link; during review I could not verify that the code is accessible. Given the invalid equation in Section II-C and the unreported variance in Table II, I recommend that the editors require the authors to release the code and rerun the experiments with the corrected loss before reconsidering the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: TMCN is a reasonable attempt to bring Mamba's selective state-space gating into multi-view clustering, with a cluster-similarity-weighted contrastive loss (AsCL) that is genuinely new relative to CoMVC/MFLVC. The code release helps. But the central SOTA claim currently rests on an equation that looks invalid as written, and on single-run benchmark numbers that are hard to believe without error bars.\n\nThe new part is the TMFN fusion network: segment each view embedding into a sequence, concatenate, run through a Mamba-style selective SSM, and gate with a second branch. That's a sensible use of the selective mechanism to down-weight noisy view content. AsCL re-weights the contrastive denominator by average cross-view pairwise similarity, pushing same-cluster representations together – a nice idea that goes beyond instance-level contrastive alignment. The ablations show both modules contribute, and the paper is clearly organized.\n\nThe soft spots are real. Eq. (16) subtracts e^{1/tau} from the denominator. Since the j=i positive-pair term is e^{(1-S_ii)C/tau}=1, the correct removal term is 1, not e^{1/tau}. For typical tau<1, e^{1/tau} is large enough to make the log argument negative early in training. The sentence after the equation also misdescribes it: it says C increases with decreased S_ij, but the formula only scales the exponent by (1-S_ij). Because AsCL is credited with large ablation gains (9.92 ACC on Hdigit), this isn't cosmetic. The authors need to either correct the formula or confirm the code implements a valid variant.\n\nSecond, Table II has no standard deviations. The Cifar100 margin (ACC 0.9853 vs 0.6569 for CoMVC) is enormous, and MFLVC's reported 0.1342 ACC is far below that method's published results. That suggests baseline numbers were not properly verified. Third, hyperparameters like lambda, tau, and training details are not reported.\n\nOverall: the idea is coherent and worth pursuing, but the evidence as presented is not trustworthy yet. The code being public means the authors can fix this quickly. This deserves peer review with major revision; I would not desk-reject it.\n\nFor you: maybe take it to reading group as a case study in how a promising architecture can be undermined by a bad equation and sketchy baseline verification. I wouldn't cite it until the loss is corrected and numbers are reproduced.","headline":"A promising Mamba-based multi-view clustering architecture, but the central SOTA claim currently rests on a likely invalid loss equation and single-run numbers that need verification.","tokens_in":8977,"tokens_out":2504,"would_cite":false,"duration_ms":19759,"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":"Multi-view clustering becomes trustworthy when fusion is selective and contrastive learning follows clusters rather than instances; the paper reports 98.53% ACC on Cifar100, about 32 points above the best baseline.","keywords":["multi-view clustering","multi-view fusion","contrastive learning","selective state space model","Mamba","trusted fusion","deep clustering","unsupervised representation learning"],"falsifier":"Run the released code on Cifar100 and check whether ACC reproduces 98.53%; then, in the same setting, replace $S_{ij}$ with random weights in Eq. (16). If accuracy stays near 98%, the AsCL weighting is not driving the result; if accuracy collapses, the cross-view similarity proxy is load-bearing.","tokens_in":7915,"feed_emoji":"🧩","tokens_out":11211,"duration_ms":88819,"temperature":0.7,"pith_summary":"Deep multi-view clustering methods typically fuse views by averaging or concatenating embeddings, and align views by instance-level contrastive learning, so noise in one view or a sample whose views disagree can pull the consensus representation away from the true clusters. The paper tries to fix both failure modes at once: it introduces a Trusted Mamba Fusion Network (TMFN) whose selective state-space gate filters noisy and redundant view information, and an Average-similarity Contrastive Learning (AsCL) loss that replaces instance-level positive pairs with cluster-level similarity weighting. On four benchmarks the method reports state-of-the-art accuracy, including 98.53% ACC on Cifar100, about 32 percentage points above the best compared method, and ablations attribute the gains to both modules. The central claim is that selective fusion plus cluster-aware contrastive alignment is enough to make multi-view fusion trustworthy.","feed_headline":"Mamba fusion hits 98.5 percent multi-view clustering accuracy","feed_subtitle":"A selective state-space filter plus cluster-aware contrastive loss beats strong baselines by up to 32 points in accuracy.","key_machinery":"The load-bearing machinery is the pair (TMFN, AsCL). TMFN (Trusted Mamba Fusion Network) turns each view embedding into a sequence of tokens, concatenates the view sequences into one global sequence, and feeds it to a selective state-space model in which the matrices $B$ and $C$ depend on the input (Eq. 8), so the hidden state acts as a learned gate that filters which view information passes through; a second branch applies a gated nonlinearity, and the two are multiplied elementwise (Eq. 9) before conversion back to a fused vector. AsCL (Average-similarity Contrastive Learning) computes $S_{ij}$ as the average cosine similarity between the $i$-th and $j$-th samples over all view embeddings (Eq. 12), then uses $(1-S_{ij})$ as a weight on every negative term of the contrastive denominator (Eq. 16), so samples that look similar across views are not pushed apart. This replaces instance-level positives with cluster-level similarity as the unit of alignment.","core_discovery":"The central claim is that the 'untrusted fusion' problem in deep multi-view clustering has two identifiable causes — noisy or redundant view information and contrastive alignment that treats only the same sample as positive — and that both are fixable in one network. TMFN segments each view-specific embedding into a token sequence, concatenates the views, and runs them through a Mamba-style selective state-space model that gates information based on the input, suppressing irrelevant view content before converting back to a fused vector. AsCL then aligns the fused representation with each view-specific representation under a contrastive loss whose denominator reweights each other sample by $(1-S_{ij})$, where $S_{ij}$ is the mean cosine similarity of view embeddings across views; samples judged to be in the same cluster are down-weighted as negatives. The result, the paper reports, is a fused representation on which k-means achieves 0.9756 ACC on Hdigit, 0.9853 on Cifar100, 0.6715 on Prokaryotic, and 0.5691 on Wiki, beating all compared deep multi-view clustering methods.","pith_inferences":["A testable extension the paper leaves implicit is to make $S_{ij}$ iterative: re-estimate the similarity matrix from the fused representation as training proceeds. If AsCL's benefit really comes from cluster-aware weighting, an updated $S_{ij}$ should maintain or improve accuracy; if the fixed proxy is what matters, iterative refinement should make little difference.","The roughly 32-point gain on Cifar100 is far outside typical incremental improvements in multi-view clustering. If it replicates independently, it would suggest that instance-level contrastive objectives were doing active harm in deep multi-view clustering, and that cluster-weighted contrastive losses could transfer to other self-supervised representation learning problems.","TMFN's input-dependent gate is a natural candidate for handling missing or corrupted views, since the selective state-space model can in principle learn to suppress uninformative view tokens; the paper only evaluates complete views, so this remains an untested property of the architecture."],"forward_implications":["If the reported Cifar100 result is reproducible — 98.53% ACC against CoMVC's 65.69% — then previous contrastive multi-view clustering methods were not merely slightly suboptimal but were actively misaligning representations, and cluster-aware weighting should become a standard component.","Ablations across the four datasets show that removing TMFN costs 3.24–9.44 ACC points and removing AsCL costs 2.13–9.92 ACC points, so the two modules contribute complementary gains rather than one doing all the work.","Because AsCL uses only the cross-view average cosine similarity and the contrastive loss, it requires no labels and no cluster-count supervision during training, so the method applies to any multi-view dataset with arbitrary per-view feature dimensions.","The reported margins imply that the reconstruction loss alone is not enough to produce clustering-friendly fused representations; the selective gate and cluster-aware contrastive term carry most of the benefit."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space model whose input-dependent B and C matrices implement the trusted fusion gate in TMFN.","marker":"[27]"},{"why":"CoMVC/SiMVC are the strongest baselines on Cifar100 and the representation-alignment approach that AsCL extends and contrasts; its table row anchors the reported 32-point ACC gain.","marker":"[22]"},{"why":"DEMVC is the strongest baseline on Prokaryotic and the collaborative-training deep multi-view clustering method that TMCN outperforms by 11 ACC points.","marker":"[21]"},{"why":"MFLVC is the contrastive multi-view clustering baseline whose instance-level positive/negative design AsCL is meant to fix.","marker":"[23]"},{"why":"GCFAggMVC supplies the global cross-view feature aggregation baseline and the matrix-factorization clustering objective used on the fused representation.","marker":"[30]"},{"why":"The autoencoder reconstruction setup for view-specific embeddings comes from this dimensionality-reduction method.","marker":"[29]"},{"why":"Provides the k-means-as-matrix-factorization formulation used in the clustering module.","marker":"[32]"}],"fun_headline_variants":["Trusted Mamba fusion solves multi-view clustering's untrusted fusion","State-space fusion and cluster-aware contrast beat multi-view baselines","Mamba gating plus cluster-aware contrast hits 98.5% on Cifar100","Selective state-space model fuses views without noise and misalignment","Fusing views with Mamba and cluster-aware contrast sets state of the art"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the average cosine similarity $S_{ij}$ of view-specific embeddings correctly indicates whether samples $i$ and $j$ belong to the same cluster; if this proxy is wrong, AsCL will amplify the wrong pairings and the reported accuracy gains could disappear.","fun_headline_variants_meta":{"raw":{"variants":["Trusted Mamba fusion solves multi-view clustering's untrusted fusion","State-space fusion and cluster-aware contrast beat multi-view baselines","Mamba gating plus cluster-aware contrast hits 98.5% on Cifar100","Selective state-space model fuses views without noise and misalignment","Fusing views with Mamba and cluster-aware contrast sets state of the art"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000801,"raw_usage":{"total_tokens":3544,"prompt_tokens":988,"completion_tokens":2556,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":2458}},"tokens_in":604,"tokens_out":2556,"duration_ms":15836,"temperature":1.0,"reasoning_tokens":2458,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:31:00.126736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code on Cifar100 and check whether ACC reproduces 98.53%; then, in the same setting, replace $S_{ij}$ with random weights in Eq. (16). If accuracy stays near 98%, the AsCL weighting is not driving the result; if accuracy collapses, the cross-view similarity proxy is load-bearing.","supporting_citations":[{"cited_title":"Reconsider- ing representation alignment for multi-view clustering,","cited_arxiv_id":null,"evidence_quote":"CoMVC/SiMVC are the strongest baselines on Cifar100 and the representation-alignment approach that AsCL extends and contrasts; its table row anchors the reported 32-point ACC gain."},{"cited_title":"Deep embedded multi- view clustering with collaborative training,","cited_arxiv_id":null,"evidence_quote":"DEMVC is the strongest baseline on Prokaryotic and the collaborative-training deep multi-view clustering method that TMCN outperforms by 11 ACC points."},{"cited_title":"Multi-level feature learning for contrastive multi-view clustering,","cited_arxiv_id":null,"evidence_quote":"MFLVC is the contrastive multi-view clustering baseline whose instance-level positive/negative design AsCL is meant to fix."},{"cited_title":"Gcfagg: Global and cross-view feature aggregation for multi-view clustering,","cited_arxiv_id":null,"evidence_quote":"GCFAggMVC supplies the global cross-view feature aggregation baseline and the matrix-factorization clustering objective used on the fused representation."}],"review_version":1}