{"id":"885eaa8f-65a1-484c-9183-eb8ae38c01d8","arxiv_id":"2509.11034","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"csMIL adds K-means cluster sparsity to attention-based MIL, reporting CAMELYON16 AUC 0.951 and TCGA-NSCLC AUC 0.933, but with test-set-tuned hyperparameters and a borrowed Lasso bound.","lead":"A new whole-slide image classifier clusters tissue patches, attends within each cluster, then applies sparse weights to keep only diagnostic clusters and reports top accuracy on two pathology benchmarks. The paper also claims a compressed-sensing-style guarantee that few slides are needed to find the relevant clusters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA results are selected on the test set: K and γ are chosen by test performance with no validation set, so Tables 1/2 overstate csMIL's advantage.","rationale":"The reader's verdict is REJECT, and I agree with that verdict, but my load-bearing concern differs from the reader's stated weakest assumption. The weakest assumption in the reader's summary concerns cluster separability and RE/incoherence for the Lasso bound. That is indeed a conditional, unverified premise, and the Appendix itself acknowledges it: 'Non-linear Dependencies: The linear model assumption may not hold strictly; the bound serves as a theoretical idealization' and 'Cluster Quality: Poor clustering (violating RE/incoherence) invalidates the bound.' So the theoretical contribution is already explicitly conditional; it is an overstatement to call it a demonstration, but it is not the main reason to doubt the paper.\n\nThe more damaging issue is the experimental protocol. Section 5.1 describes only a 4:1 train/test split. Section 5.4 and 5.6 select K and γ by observing performance on what appears to be the same test data (Tables 3–4 and Figure 5 report accuracy values that match the final Table 1–2 results). No validation set is mentioned. This selection on the test set invalidates the reported SOTA comparisons. Baselines are reported at fixed defaults, so the comparison is not apples-to-apples. This is a concrete, protocol-level flaw rather than a modeling assumption, and it directly undermines the central empirical claim. Hence the verdict remains REJECT, and the concern is the selection-bias one, not primarily the cluster-separability assumption.","tokens_in":17413,"tokens_out":4172,"duration_ms":44966,"concrete_test":"Run a nested cross-validation or a fixed train/validation/test split: select K and γ using only validation folds, then evaluate the chosen configuration once on held-out test folds, reporting mean±std over folds. If CAMELYON16 Acc/F1/AUC fall to or below DTFD-MIL's 0.898/0.872/0.946, or the TCGA-NSCLC gains over DSMIL/DTFD-MIL shrink to within one standard deviation, the reported superiority is an artifact of test-set selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the state-of-the-art accuracy/F1/AUC in Tables 1 and 2. The paper's protocol (Section 5.1) splits each dataset only into training and test sets (4:1), with no validation set. Section 5.4 then reports accuracy/F1/AUC for K=2,3,5,10 (Tables 3-4) and Section 5.6 reports accuracy vs. γ over 15 values (Figure 5); the final Tables 1-2 use K=3 and the best γ found in these sweeps. Thus the reported metrics are the maximum over K and γ evaluated on the same test folds used for comparison. This is test-set-driven hyperparameter selection; it biases the headline numbers upward and makes the comparison to baseline methods (which are not shown to be tuned under the same protocol) unfair. The absence of error bars across folds further obscures the variance. The theoretical claim is less load-bearing because the paper itself flags in the Appendix ('Non-linear Dependencies...', 'Cluster Quality... invalidates the bound') that RE/incoherence and linearity are assumed, not established; but the empirical SOTA claim is presented unconditionally. If the test-set selection is the source of the gains, the central claim fails.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes csMIL, a multi-instance learning framework for whole-slide images that combines global K-means clustering over all instances, local assignment of each bag's instances to the K global centers, within-cluster attention pooling into K prototypes, and a bag-level aggregation z_i = sum_k beta_k z_i^k with an l1 penalty on beta. The authors argue that cluster-level sparsity discards irrelevant instance cohorts, improves interpretability, and reduces computational cost. They state a sparse-recovery guarantee (Theorem 1: O(s log K) bags suffice to recover s relevant clusters) and report state-of-the-art accuracy/F1/AUC on CAMELYON16 (0.905/0.876/0.951) and TCGA-NSCLC (0.854/0.855/0.933), together with qualitative pathology-informed visualizations of cluster weights.","tokens_in":17726,"tokens_out":6108,"duration_ms":72318,"significance":"If the empirical claim were supported, csMIL would be a simple and attractive addition to attention-based MIL: it repurposes standard K-means and l1 regularization into an interpretable cluster-selection mechanism, and the qualitative maps in Section 5.5 are a genuinely useful clinical-interpretability direction. The paper, however, does not currently provide a sound empirical foundation for the SOTA claim. The lack of a validation split and the explicit test-set-driven sweeps over K and gamma (Tables 3–4, Figure 5) bias the headline numbers upward. The theoretical section is also not a derivation for the actual model: the proof assumes squared-error linear regression and assumes, rather than establishes, restricted eigenvalue and incoherence conditions on learned cluster features; the appendix itself concedes that nonlinear dependencies and poor clustering invalidate the bound. The central insight is potentially publishable, but the current evidence is not sufficient.","major_comments":[{"comment":"The reported SOTA is obtained by test-set-driven hyperparameter selection. The protocol in §5.1 has only training and test folds (4:1), with no validation split. §5.4 sweeps K in {2,3,5,10} and reports test-fold Acc/F1/AUC (Tables 3–4); §5.6 sweeps 15 values of gamma and reports test accuracy (Figure 5). The final Tables 1–2 then use K=3 and the best gamma found in these sweeps. Selecting hyperparameters on the same test folds used for comparison makes the headline numbers optimistically biased and makes comparisons to baselines unfair. The paper must either use a proper validation/nested-CV procedure or clearly report the selection mechanism and per-fold variance; without this, the SOTA claim is unsupported.","section":"§5.1, §5.4, §5.6; Tables 1–4; Fig. 5"},{"comment":"Theorem 1 as stated does not follow from the actual objective. The appendix proof analyzes the Lasso problem min_beta (1/(2M)) sum_i (Y_i - beta^T z_i)^2 + gamma ||beta||_1, while the method's loss in Eq. (4) is sum_i CE(g(z_i), Y_i) + gamma ||beta||_1, with g mapping z_i to a probability simplex. The theorem is therefore a generic Lasso sample-complexity statement, not a recovery guarantee for csMIL as implemented. Moreover, the restricted eigenvalue and incoherence conditions are assumed on learned cluster-aggregated features rather than verified; the appendix itself states that nonlinear dependencies and poor clustering invalidate the bound. The authors should either prove the bound for the actual cross-entropy objective under explicit, checkable assumptions or substantially reframe the theoretical contribution.","section":"§4.3 and Appendix (Proof of Theorem 1)"},{"comment":"The method and the theory both rest on the assumption that diagnostically relevant and irrelevant instances reside in different clusters. The only support offered is the qualitative bar plot in Figure 1, with no cluster-purity, silhouette, or other cluster-quality metric, and no ablation in which this assumption is artificially violated. Since the pruning logic and the O(s log K) guarantee collapse if clusters mix relevant and irrelevant patterns or if learned features are collinear, the paper should quantify cluster quality on the actual learned features and test robustness to imperfect clustering.","section":"§3.2 and §4.3"}],"minor_comments":[{"comment":"The notation in Eq. (3) is hard to parse (h_n^T inside tanh), and the summation index in z_i^k = sum_{i=1}^{C_i^k} alpha_i^k h_i reuses i for both the bag index and the instance index. Please reindex instances (e.g., j) and clarify the dimensions.","section":"Eq. (3) and §3.3"},{"comment":"The opening sentence says the K-sweep is 'shown in Table 1 and 2 respectively,' but the actual K-sweep results are in Tables 3 and 4. Please correct the cross-reference.","section":"§5.4"},{"comment":"The caption is missing the gamma symbol ('Model Accuracy vs. [gamma]'), and the text says 'accuracy increases from 0.750 to 0.900' when it presumably means accuracy rises over the gamma range. Please clarify the axis description and the direction of the claim.","section":"Figure 5 and §5.6"},{"comment":"The TCGA-NSCLC dataset description (§5.1) does not match reference [54], which is a breast-cancer TCGA paper; a lung-cancer TCGA reference should be cited. Also, the abstract contains 'CAMEL YON16' (spacing typo).","section":"References and dataset description"},{"comment":"The appendix contains a placeholder sentence 'Appendix two text goes here' followed by the acknowledgment line. This should be removed before submission.","section":"Appendix"}],"recommendation":"major_revision","confidential_remarks":"The test-set-driven hyperparameter selection is the most serious issue; if the authors cannot demonstrate the same gains under a clean validation protocol, the paper should not be accepted. The theoretical section is currently a generic Lasso restatement rather than a proof for the proposed method. I am not recommending outright rejection because the core idea is modular and potentially useful, but the current submission's central claims are not adequately supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The method is straightforward: global K-means over all patches, per-slide assignment to those centers, within-cluster attention to get K prototype vectors, then a learned K-dimensional beta with L1 regularization to zero out unhelpful clusters. That is a reasonable assembly of existing parts, and the interpretability story is real—you can see which clusters get pruned, and the patch-level heatmaps align with tumor morphology. The complexity reduction is plausible too: since K<<N, instance-level aggregation is replaced by cluster-level aggregation, which matters for gigapixel slides. The modest accuracy gains over DTFD-MIL on CAMELYON16 (0.905 vs 0.898) and TCGA-NSCLC (0.854 vs 0.845) would be a small but genuine benefit, if the numbers were trustworthy.\n\nThe problem: they are not, and the stress-test note has it right. Section 5.1 splits each dataset 4:1 into train/test with no validation set. Section 5.4 then reports accuracy for K=2,3,5,10, and Section 5.6 sweeps gamma across 15 values—on the same test folds. The final tables use K=3 and the best gamma from those sweeps. That is test-set-driven model selection, and the baselines are not given the same advantage. Without error bars or per-fold results, the reported 0.7–1 point gains could easily be selection noise.\n\nThe theory is also thinner than the abstract suggests. Theorem 1 is the standard Lasso sample-complexity bound O(s log K), with RE and incoherence assumed for cluster features that are themselves learned. The appendix explicitly concedes that poor clustering invalidates the bound and that the linear model is an idealization. The paper does not hide this, but the contribution section calls it \"aligned with compressed sensing,\" which oversells it.\n\nAlso missing: the feature extractor is never specified (patch size is given, backbone is not), there is no code release, and \"validated by pathologists\" appears in the contributions without any supporting evidence in the text.\n\nOverall, the cluster-sparsity idea is worth a serious referee, but the empirical claim needs to be redone with a validation split or nested CV, error bars, and baseline tuning under the same protocol. If the numbers survive, this is a useful paper for the computational pathology community. As is, I would not cite it for the reported SOTA. Send it to review with the expectation of major revision.","headline":"Cluster-level sparsity is a sensible idea, but the headline SOTA is picked on the test set and the theory is standard Lasso; salvageable, not publishable as is.","tokens_in":18190,"tokens_out":2457,"would_cite":false,"duration_ms":30838,"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":"Whole-slide cancer classification improves when entire clusters of non-diagnostic patches are pruned, not just individual patches.","keywords":["multi-instance learning","whole-slide images","cluster-level sparsity","L1 regularization","attention mechanism","computational pathology","compressed sensing","weakly supervised classification"],"falsifier":"Run csMIL on a synthetic MIL dataset where positive instances are deliberately spread across every cluster; if recovering the signal requires more than O(s log K) bags or forces nonzero weights on all clusters, the separation assumption is falsified. On real slides, a simpler check: take a trained model, remove all patches from zero-weight clusters on held-out slides, and see whether predictions stay the same—if they shift, the sparsity mask is not truly selecting irrelevant tissue.","tokens_in":17295,"feed_emoji":"🔬","tokens_out":3535,"duration_ms":40931,"temperature":0.7,"pith_summary":"The paper tries to show that multi-instance learning on whole-slide images works better when pruning happens at the level of clusters, not individual patches. It groups every patch in a slide into K clusters defined by global centers, runs attention inside each cluster to form a prototype, then combines prototypes with learnable weights regularized by L1. A zero weight effectively discards an entire cluster, filtering out noisy tissue regions. The authors argue this needs O(s log K) slides to recover s relevant clusters, and report top accuracy, F1, and AUC on CAMELYON16 and TCGA-NSCLC.","feed_headline":"Cluster pruning lifts WSI cancer detection to 0.905","feed_subtitle":"L1-sparse cluster weights discard non-diagnostic tissue, beating attention MIL on CAMELYON16 and TCGA-NSCLC.","key_machinery":"The load-bearing object is the cluster-level sparse aggregation: K global cluster centers, within-cluster attention (ABMIL-style) that produces one prototype per cluster, and a shared sparse weight vector β with ℓ1 regularization. The zero or near-zero entries of β are the selection mechanism: they mark clusters of patches the model discards. The theory maps this to Lasso and compressed sensing, with the bag-level cluster prototypes as the design matrix and a restricted eigenvalue or incoherence condition as the bridge.","core_discovery":"The central claim is that sparse selection of cluster-level prototypes—not instance-level attention alone—is what makes weakly supervised whole-slide classification robust. csMIL creates K global cluster centers from all training patches, assigns each bag's patches to those centers, computes attention-weighted prototypes per cluster, and sums the prototypes with a K-dimensional weight vector β regularized by ℓ1. In the learned model, clusters whose β components shrink to zero are discarded, so each slide is classified from a small set of diagnostically relevant tissue patterns. The paper reports that this beats attention-based and instance-selection baselines on CAMELYON16 (0.905 accuracy, 0","pith_inferences":["Editorial inference: If the separation premise holds, the same cluster-pruning idea could transfer to other weakly labeled bag problems where relevant and irrelevant patterns are spatially coherent—a testable extension the paper does not run.","Editorial inference: The bound's dependence on learned features is the key gap; since the prototypes are data-dependent, the restricted eigenvalue and incoherence constants are not guaranteed, so a natural extension is to verify the eigengap on real feature embeddings.","Editorial inference: The paper's K-sensitivity results suggest a practical heuristic: choose K by the stability of β's support rather than final accuracy, since over-segmentation diluted features at K = 5 and K = 10."],"forward_implications":["If the central claim holds, whole-slide classifiers can drop entire non-diagnostic tissue classes without retraining, because β learns to zero them out.","The O(s log K) recovery statement means the number of slides needed scales only logarithmically with the number of clusters, so finer-grained clustering does not force proportionally more data.","Cluster-level sparsity maps give pathologists a direct visual: retained clusters are diagnostic regions, zero-weight clusters are noise, which is the paper's interpretability story.","Because K is much smaller than the number of instances, the per-step cost avoids the quadratic instance-pair cost of transformer MIL and the instance-level scoring cost of selection-based MIL."],"fun_headline_variants":["Cluster-level sparsity sharpens WSI cancer detection","Sparse cluster weights cut noise, boost slide diagnosis","csMIL: prune irrelevant clusters, improve WSI classification","L1-sparse cluster MIL beats attention on pathology slides"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes that diagnostically relevant and irrelevant tissue patches sort into distinct clusters, and that the cluster-aggregated features satisfy the restricted eigenvalue and incoherence conditions; if either fails, the sparsity-based pruning and the O(s log K) guarantee collapse.","fun_headline_variants_meta":{"raw":{"variants":["Cluster-level sparsity sharpens WSI cancer detection","Sparse cluster weights cut noise, boost slide diagnosis","csMIL: prune irrelevant clusters, improve WSI classification","L1-sparse cluster MIL beats attention on pathology slides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000287,"raw_usage":{"total_tokens":1531,"prompt_tokens":762,"completion_tokens":769,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":703}},"tokens_in":506,"tokens_out":769,"duration_ms":9847,"temperature":1.0,"reasoning_tokens":703,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T17:11:11.798095+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run csMIL on a synthetic MIL dataset where positive instances are deliberately spread across every cluster; if recovering the signal requires more than O(s log K) bags or forces nonzero weights on all clusters, the separation assumption is falsified. On real slides, a simpler check: take a trained model, remove all patches from zero-weight clusters on held-out slides, and see whether predictions stay the same—if they shift, the sparsity mask is not truly selecting irrelevant tissue.","supporting_citations":[],"review_version":1}