{"id":"40d8ee94-c0dd-4247-91ac-b3ddbd72391f","arxiv_id":"2607.16260","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An adaptive Mamba architecture with dynamic cross-modal weighting and semantic reordering of WSI patches reports ~2% C-Index gains over baselines on five TCGA datasets.","lead":"AdaSurvMamba is a neural network that combines tumor slide images and genomic data to predict cancer survival, using adaptive weighting of the two data types and reordering of image patches into semantically coherent sequences before a Mamba state-space scan. It reports consistent but small C-Index gains over baseline methods on five TCGA cancer cohorts, with no significance tests and no external validation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing Mamba baselines (SurvMamba, ME-Mamba) make the claimed average C-Index improvement over SOTA unsubstantiated; the comparison set is incomplete.","rationale":"The reader's verdict is CONDITIONAL, and our analysis confirms that the empirical claim is not yet fully supported. We identified the omission of the closest Mamba-based baselines as the single most load-bearing concern because it directly undermines the 'best competing method' against which the average improvement is measured. The reader's weakest_assumption lists this as one of three fragile premises; we agree with that but single it out as the primary threat. A concrete reimplementation test can settle whether the reported gains are genuine. Other issues (hyperparameter selection, lack of significance tests) are secondary and can be addressed alongside. The conditional verdict remains appropriate, hence no change to the reader's verdict.","tokens_in":8985,"tokens_out":4244,"duration_ms":36881,"concrete_test":"Reimplement (or obtain official code for) SurvMamba [5] and ME-Mamba [23], and evaluate them on the same five TCGA cohorts under the same unified protocol: identical data splits, feature extractors (ResNet-50 and UNI), and 5-fold cross-validation with the same discrete-time NLL objective. Compute C-Index for each fold and perform paired significance tests (e.g., bootstrap or permutation across folds). If either baseline matches or exceeds AdaSurvMamba on any cohort, or if the average advantage falls below one standard deviation of the paired fold differences, the central claim of consistent SOTA improvement fails as currently stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the reported empirical SOTA improvement: an average C-Index gain of 2.32% over the best competing method (Section 3.2, Table 1). However, the two most relevant Mamba-based multimodal survival methods, SurvMamba [5] and ME-Mamba [23], are explicitly excluded because their official implementations were not publicly available. This is a load-bearing gap. The paper's introduction motivates AdaSurvMamba as addressing limitations of these very methods, so they are the natural comparators. Without them, the 'best competing method' in Table 1 may not be the actual best, and the claimed margin could shrink or even reverse. The justification about unavailability is not decisive; the authors could reimplement from the papers or use the original reported results if the evaluation protocol is compatible. As it stands, the empirical evidence does not support the assertion of 'consistent improvements over existing methods' against the closest state of the art.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"AdaSurvMamba proposes an adaptive framework for multimodal survival analysis combining WSI patches and genomic profiles. It introduces DSIR, a dual-scale gating mechanism that reconstructs features using sequence- and token-level importance weights, and SAS, which reorganizes tokens into semantics-based sequences via a shared prototype pool and modulates the Mamba state-transition step size. The method is evaluated on five TCGA cohorts against unimodal and multimodal baselines, reporting mean C-Index from 5-fold cross-validation and claiming consistent improvements (e.g., 2.32% average gain under UNI features). Code is provided.","tokens_in":9292,"tokens_out":8967,"duration_ms":76988,"significance":"The conceptual contribution is timely: moving from static fusion and fixed scanning paths to input-dependent gating and semantically reordered scanning is a plausible way to adapt Mamba to gigapixel pathology and heterogeneous genomic tokens. The paper ships a public implementation, includes ablations for the main modules, and evaluates on multiple public cohorts, which are strengths. However, the claimed empirical advantage rests on comparisons that omit the two closest Mamba-based baselines, on C-Index deltas that are partly within one standard deviation, and on hyperparameters selected on the same cohorts used for the final numbers. These issues must be resolved before the significance of the result can be assessed.","major_comments":[{"comment":"The two most relevant comparators, SurvMamba [5] and ME-Mamba [23], are explicitly excluded because their official implementations were not available. Since the introduction motivates AdaSurvMamba by the limitations of these methods, they are the direct state of the art. The claim of 'consistent improvements over existing methods' is therefore not supported against the closest alternatives. Please add reimplementations, original reported numbers on compatible protocols, or at least a sensitivity comparison, and soften the claim if this is not possible.","section":"§3.2, Table 1"},{"comment":"No statistical significance testing is reported. Several improvements over the best baseline are within one standard deviation of the baseline or of the proposed method (e.g., BRCA UNI: 0.701±0.039 vs 0.690±0.019; GBMLGG UNI: 0.866±0.017 vs 0.852±0.032). Since only five folds are used, report fold-level paired comparisons or confidence intervals to show the gains are not noise.","section":"§3.2, Table 1"},{"comment":"The hyperparameters T_g=4, T_p=16, r=32, d_p=64, λ=0.5 are selected through ablations performed on BLCA and BRCA with ResNet50 features, then used to produce the final Table 1 results on all five cohorts, including BLCA and BRCA, and also under UNI features. This is selection on the evaluation cohorts and can inflate the reported gains. Use a nested/held-out protocol, report selection and test cohorts separately, or validate externally before claiming consistent gains.","section":"§3.1 / §3.3"},{"comment":"The headline average improvement of 2.32% (and 2.43% for ResNet50) is not reproducible from Table 1. Using the natural per-cohort relative improvement (ours−best)/best yields 2.51% (UNI) and 2.49% (ResNet50); using absolute C-index points yields 1.78 and 1.74 percentage points. Please state the exact formula and correct the numbers if needed.","section":"§3.2"}],"minor_comments":[{"comment":"L_m is used in R_m ∈ R^{L_m×T_m} but never defined; please define it as the token count for modality m.","section":"§2.3"},{"comment":"Gumbel-Softmax outputs continuous approximations, not discrete weights. Clarify whether hard straight-through assignments are used for the argsort and how gradients flow through the reordering.","section":"§2.3"},{"comment":"The shapes of G_raw and G_ctx after the per-modality softmax are not stated; make the tensor dimensions explicit.","section":"§2.2, Eqs. (3)–(4)"},{"comment":"The ablation only varies sizes up to 8 for T_p and does not explore larger T_g; if T_g=4,T_p=16 is claimed optimal, include neighboring configurations (e.g., (8,16)) or justify the grid.","section":"Table 5"},{"comment":"The figure is very dense; a notation table for X, C, w, G, P, R, Δ would improve readability.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The core ideas are reasonable and the code release is a plus, but the empirical case needs substantial strengthening. I would ask for a revised version that either adds the missing Mamba baselines and proper significance testing or explicitly narrows the claims to 'improvement over the included baselines'. The hyperparameter-selection issue is fixable but requires re-running experiments under a cleaner protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"AdaSurvMamba is a useful, well-engineered next step for Mamba-based multimodal survival analysis. The DSIR dual-scale gating and SAS semantic reordering are genuinely new architectural components, and the paper backs them with careful ablations and released code. The reported C-Index gains are consistent across five TCGA cohorts, but the evaluation is incomplete in one important way: the two closest prior methods — SurvMamba and ME-Mamba — are absent because their official code wasn't public. That makes the \"best competing method\" claim weaker than the abstract suggests.\n\nWhat's good: the design is motivated, the modules are clearly explained, and the ablations demonstrate that each piece contributes. The ablation comparing semantic scan to 1-way, 2-way, and random scans is particularly convincing. The experimental setup is standard — five TCGA cohorts, two feature extractors, multiple baselines — and the code is available.\n\nWhere it's soft: first, the missing Mamba baselines. The authors note the reason, but reimplementing from the papers or citing original reported numbers would substantially strengthen the claim. Second, several of the headline deltas (e.g., BRCA UNI: 0.701 vs 0.690; GBMLGG: 0.866 vs 0.852) are within one standard deviation of the best baseline, and no significance tests are reported. Third, the optimal prototype counts and loss weight were tuned on BLCA/BRCA and then applied to all five cohorts, so the advantages on those two cohorts are partly selection artifacts. Neither issue is disqualifying — the architecture is plausible and the ablations support the design — but they mean the precise magnitude of the improvement over the true state of the art is uncertain.\n\nI'd send this to peer review. The right call is to accept with major revision, asking for the missing Mamba comparisons or a clear explanation, plus significance testing or confidence intervals. The paper deserves a serious referee.","headline":"Worth engaging with: plausible architectural novelty and solid ablations, but the SOTA claim is undercut by missing Mamba baselines and thin margins.","tokens_in":9792,"tokens_out":2740,"would_cite":true,"duration_ms":25836,"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":"Adding patient-level importance gating and semantic token reordering to a state-space model improves cancer survival prediction across five cohorts.","keywords":["survival analysis","whole slide images","genomics","multimodal fusion","state space models","Mamba","cancer prognosis","selective scanning"],"falsifier":"Perform repeated cross-validation with paired bootstrap significance tests, on all five cohorts, comparing AdaSurvMamba against the best competing baseline and the two closest Mamba-based models; if the improvement is not consistently outside the baseline's standard deviation, the central claim fails.","tokens_in":8897,"feed_emoji":"🧬","tokens_out":9807,"duration_ms":73002,"temperature":0.7,"pith_summary":"This paper attempts to establish that cancer survival prediction from whole-slide images and genomic profiles improves when the model dynamically adjusts how strongly the two modalities interact and scans tokens in a semantic rather than a fixed physical order. The authors argue that standard fusion assumes a static interaction strength and standard state-space scanning breaks the continuity of scattered medical features. They introduce two mechanisms: DSIR, a dual-scale gating that reweights modalities per patient and per token, and SAS, a semantic reordering with an explicit step-size modulation. Across five public cancer cohorts, they report an average C-Index improvement of 2.32% over the best competing method with a pathology foundation model, which matters because better risk stratification directly affects treatment decisions.","feed_headline":"2.32% average C-Index gain from adaptive fusion and semantic scanning","feed_subtitle":"Weights each modality per patient and reorders tokens semantically to improve cancer survival risk scores.","key_machinery":"The load-bearing machinery is the Dual-Scale Importance-Aware Reconstruction (DSIR) combined with Semantic Aggregation Scanning (SAS). DSIR computes a sequence-level importance weight from pooled modality representations and a token-level gate logit from co-attended context, then multiplies the two to reconstruct each token before the state-space scan. SAS learns a shared prototype pool, uses Gumbel-Softmax routing to assign every token to a prototype, sorts tokens so same-prototype tokens are adjacent in the input sequence, and adds a prototype-weighted bias to the continuous step-size parameter of the Mamba state-space model. Together they make the sequence order and the information absorp","core_discovery":"AdaSurvMamba claims that replacing static multimodal fusion and rigid scanning with adaptive patient- and token-level gates plus semantically reordered state-space scanning yields consistent survival-prediction gains on five cancer cohorts. The paper's central evidence is a table of concordance indices showing the proposed model ahead of all compared unimodal and multimodal baselines on every cohort, with an average improvement of roughly 2.3% under both tested feature extractors. The authors attribute this to the DSIR module, which reconstructs each modality's input by combining global importance weights with local softmax gates, and the SAS module, which clusters tokens via a shared protot","pith_inferences":["In our view, the DSIR gating mechanism could transfer to other multimodal medical tasks, such as radiology-image-plus-report or clinical-notes-plus-lab-values, where the dominant modality varies from patient to patient.","The SAS semantic scanning may also improve Mamba-based models on other unordered or sparse input domains, such as point clouds or graph-structured data, where physical ordering is arbitrary.","A testable extension would be to evaluate AdaSurvMamba on external cohorts with different staining protocols and scanners; if the semantic prototypes need retraining, that would indicate overfitting to the training distribution.","One could also decouple the two modules and measure their individual contribution on new datasets, which the paper only does for two cohorts, to check whether the reported gains are dataset-specific."],"forward_implications":["If the reported gains are reproducible, adaptive per-patient fusion becomes a practical alternative to fixed late fusion for histology-genomics survival models.","The semantic reordering before scanning suggests that token order itself, not just token content, is a tunable resource in state-space models.","The 2.32% average C-Index improvement could translate into better risk stratification for the five studied cancer types if the result generalizes.","The framework runs on a single GPU, so it is accessible to laboratories without large-scale computing resources.","The shared prototype pool provides a lightweight way to align pathology and genomic tokens without heavy cross-attention."],"fun_headline_variants":["Adaptive fusion and semantic scanning lift survival prediction by 2.3%","Patient-specific gates and token reordering beat static fusion in cancer prognosis","Semantic scanning and adaptive fusion boost survival prediction by 2.3% on five cohorts","Adaptive weights per patient and semantic token order improve survival prediction","New model dynamically weighs modalities and scans semantically to predict survival"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim rests on assuming that the reported C-Index improvements are true algorithmic gains rather than artifacts of hyperparameter tuning, incomplete baseline comparison, or random variation in cross-validation.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive fusion and semantic scanning lift survival prediction by 2.3%","Patient-specific gates and token reordering beat static fusion in cancer prognosis","Semantic scanning and adaptive fusion boost survival prediction by 2.3% on five cohorts","Adaptive weights per patient and semantic token order improve survival prediction","New model dynamically weighs modalities and scans semantically to predict survival"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00099,"raw_usage":{"total_tokens":4044,"prompt_tokens":763,"completion_tokens":3281,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":3186}},"tokens_in":507,"tokens_out":3281,"duration_ms":18774,"temperature":1.0,"reasoning_tokens":3186,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T09:39:46.445041+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Perform repeated cross-validation with paired bootstrap significance tests, on all five cohorts, comparing AdaSurvMamba against the best competing baseline and the two closest Mamba-based models; if the improvement is not consistently outside the baseline's standard deviation, the central claim fails.","supporting_citations":[],"review_version":1}