{"id":"b9142c33-5daa-4887-a2ec-40e7f446eb6f","arxiv_id":"2511.18089","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"TTA reports improved C-index on five TCGA cohorts by aligning histology and genomics to shared prototypes with unbalanced optimal transport, then enforcing modality distinctiveness with anchor-based contrastive regularization.","lead":"A cancer-prognosis AI that combines tumor slide images with gene data first aligns the two modalities against shared prototypes, then uses contrastive training to preserve each modality's unique signal. On five TCGA cancer cohorts it reports higher concordance indices than earlier multimodal survival models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA gains may be inflated by hyperparameter selection on the same test folds; the +2.6% C-index margin over MMP needs validation on a held-out split.","rationale":"The central claim is empirical: TTA achieves 0.693 overall C-index and beats MMP by +2.6%. A claim like this lives or dies by the evaluation protocol. The paper's ablations and hyperparameter tables (Tables 6,7,9; Sec. C) sweep a large number of knobs—prototype count, KL weight, instance loss weight, contrastive weight, contrastive temperature, bag size, batch size, learning rate, loss function, epochs—and select the configuration that maximizes the average C-index on the same five cohorts used for the final comparison. No separate validation split is mentioned in Sec. 4. This is selection on the test data. With small cohorts (STAD n=318, CRC n=296) and large stds, the selected configuration can easily shift the overall C-index by several points, which is precisely the size of the claimed improvement over MMP (+2.6%). The self-referential UOT pseudo-label issue identified by the reader is a real secondary concern, but it is partly mitigated by the ablation structure and is common in self-training; the hyperparameter-selection issue directly undermines the headline comparison. If the authors add a validation split or nested CV and the margin persists, the claim is credible. Therefore we keep the reader's CONDITIONAL verdict.","tokens_in":24601,"tokens_out":3295,"duration_ms":33666,"concrete_test":"Request authors to perform a nested or split validation: for each of the 5 folds, hold out ~10% of the training data as a validation set; tune all hyperparameters (K, γ, ρ schedule, λ_inst, λ_contrast, τ_r, etc.) on this validation set; retrain the final model on train+val; then evaluate on the test fold. Report C-index for TTA and all baselines under identical protocol. If TTA's overall C-index no longer exceeds MMP by ≥2% or loses its rank on BRCA/STAD/KIRC, the headline claim should be tempered. As a secondary check, report per-fold paired DeLong or bootstrap C-index tests between TTA and MMP/LD-CV AE.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Across Tables 6, 7, 9 and Section C, the authors sweep prototype count K (16/32/50), KL weight γ (0.1–1.0), instance loss weight λ_inst (0.1–1.0), contrastive weight λ_contrast (0.02–1.0), contrastive temperature τ_r (0.05–1.0), bag size (2048/4096), batch size (32/64), learning rate, loss function (Cox/NLL), and epochs, reporting the overall average C-index over the same five TCGA cohorts/folds used for the final Table 1 comparison. The paper states only '5-fold site-stratified cross-validation' (Sec. 4) and does not mention holding out a validation split for model selection. Thus the reported 0.693 overall C-index and the +2.6% margin over MMP are the result of selecting hyperparameters on the test folds. This selection bias can inflate performance, especially with small cohorts (STAD n=318, CRC n=296) and high per-fold variance (std up to ±0.13–0.18). The central empirical claim—consistent improvement over recent multimodal survival methods—is therefore not established as stated. No paired significance tests are reported for Table 1, so even raw differences could be within fold-level noise. The min–max framing is a narrative issue, and the UOT self-loop is a secondary concern, but test-fold tuning directly determines whether the headline margin is real.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Together-Then-Apart (TTA), a multimodal survival-analysis framework that first aligns histology and genomic tokens to a shared prototype bank through unbalanced optimal transport with curriculum mass ('Together'), and then applies anchor-based contrastive regularization to preserve modality-specific distinctiveness ('Apart'). The method is evaluated on five TCGA cohorts with paired WSI and gene-expression data. The paper reports an overall C-index of 0.693 (Table 1), a +2.6% improvement over the second-best method MMP, along with extensive ablations, hyperparameter sensitivity analyses, and interpretability visualizations of the learned prototypes.","tokens_in":25052,"tokens_out":4444,"duration_ms":45995,"significance":"The core design idea—explicitly balancing cross-modal alignment against modality-specific distinctiveness—is coherent and relevant to a long-standing issue in multimodal survival analysis, where over-alignment can degrade performance. The paper provides considerable empirical scaffolding: five cohorts, ablations of both stages (Tables 2–4), detailed component ablations (Tables 6–7), backbone robustness (Table 8), training-parameter sensitivity (Table 9), and a formal equivalence proof for the proposed UOT reformulation (Appendix A.4). If the reported gains survive a properly separated model-selection protocol, the method would be a useful contribution. At present, however, the central empirical claim is not established because the evaluation protocol appears to select hyperparameters on the same folds used to report the headline comparisons, and no significance testing is provided. The self-supervised instance-level loss also raises a circularity concern that needs empirical diagnosis.","major_comments":[{"comment":"The paper does not provide any statistical significance testing for the C-index differences in Table 1. The standard deviations in the table are large (e.g., ±0.079 for TTA on STAD, ±0.131 on CRC), and the reported overall margin over MMP is only +2.6%. With per-fold variability of the magnitude shown, a paired test across the 5 folds (or a permutation test) is needed to establish that the observed ranking is not within noise. This is especially important because the paper claims TTA 'consistently improves' over recent methods, but Table 1 shows TTA is second on BLCA and CRC, and the per-dataset gains are not all large relative to the fold-level variability.","section":"§4, Table 1"}],"minor_comments":[{"comment":"The paper repeatedly describes the method as a 'min–max' optimization, but the total objective in Eq. (20) is a weighted sum of a survival loss, a contrastive loss, and an instance-level CE loss. There is no adversarial maximization or saddle-point optimization. Either the min–max language should be softened to 'balanced objectives,' or the authors should clarify in what formal sense the objective constitutes a min–max problem.","section":"Title/Abstract, Eq. (20)"},{"comment":"The abstract ends with 'jointly achieved in for robust, interpretable...' — 'in for' is a typo.","section":"Abstract"},{"comment":"Reference [19] is listed as 'Zhou et al. Cross-modal translation and alignment for survival analysis. In ICCV, 2023.' The author list and title are incomplete; this should be a full citation.","section":"References"},{"comment":"The paper states 'Our code will be available at here.' This is a placeholder. If code is to be released, provide a working repository or remove the statement; the absence of code is not itself a flaw, but the placeholder suggests the manuscript is not final.","section":"Abstract / Code availability"},{"comment":"The configuration row 'SK multi-head numbers' and 'Pseudo-label CE weights' are not defined in the main text until later in Appendix C.1. For clarity, define these in Sec. 3.2 when the multi-head consistency mechanism is first mentioned.","section":"Table 5"},{"comment":"The notation π_i^m for the i-th row of Q* is used before defining the row normalization; it would help to state explicitly that each row of Q* (after removing the sink) is treated as a probability vector, or to define the normalization used before the CE term.","section":"Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"The paper targets a timely problem and the two-stage framework is appealing, but the empirical claim as presented is not reliable because model selection appears to be conducted on the same folds used for the final comparison. The authors need to rerun the evaluation with a held-out validation split (or nested CV) and provide significance tests. The UOT self-loop is a genuine methodological concern that is not resolved by the current ablations. I would be willing to reconsider after these points are addressed; the theoretical inconsistency between Eq. (8) and Appendix A.3 also needs a clear fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know about this paper because it does something concrete and useful: it proposes TTA, a survival model that first aligns histology and genomics to shared prototypes via unbalanced optimal transport with a curriculum mass schedule, then diversifies with an anchor-based contrastive loss. The two-stage 'Together, Then Apart' idea is clearly motivated around the over-alignment collapse problem, and the components are combined in a way that isn't in the existing literature. The reported C-index on five TCGA cohorts (0.693 overall, +2.6% over MMP) is the headline.\n\nWhat's genuinely good: the ablations are thorough. They toggle the two stages, shared vs. separate prototypes, joint vs. separate OT, anchor refinement vs. contrastive loss, and sweep hyperparameters like K, KL weight, contrastive temperature, bag size, and loss function. The UOT derivation in the appendix, including the equivalence proof for the curriculum-mass sink, is sound as far as I can tell. The visualizations of prototype assignments and pathway-prototype interactions add interpretability, which matters in computational pathology.\n\nThe soft spots are real, and the biggest one is experimental validation. The hyperparameters appear to be selected on the same five cohorts/folds that are then used for the final comparison (Tables 6, 7, 9 and Appendix C all report average C-index over the same benchmark splits). There is no held-out validation split or nested cross-validation. Given per-fold standard deviations up to ±0.18, the +2.6% margin over MMP is not robustly established. No paired significance tests are reported either. Second, the code link is a placeholder, so none of this is reproducible as written. Third, the instance-level pseudo-labels are self-referential: the UOT plan is computed from the same token-prototype logits that it then supervises (Eqs. 3–11 vs. Eq. 14). That is a self-training loop. It may be fine in practice, but the paper doesn't analyze whether the alignment stage is capturing real structure or just reinforcing its own biases. Fourth, the min–max framing is narrative; the 'max' is just a negative contrastive loss. Minor.\n\nThe core method is not obviously wrong, and the empirical scaffold is extensive. This paper deserves a serious referee, but the headline claim is not fully established as stated. Send it to peer review with a clear request: release code, add a validation split or nested CV, report paired significance tests, and temper the min-max language. If those are addressed, this would be a solid contribution to multimodal survival analysis.","headline":"TTA is a genuinely useful multimodal survival framework with extensive ablations, but the headline C-index margin rests on test-fold hyperparameter tuning and no code, so the reported gains need external validation.","tokens_in":735,"tokens_out":710,"would_cite":true,"duration_ms":29999,"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":"Multimodal cancer survival prediction improves when histology and genomics are first aligned to shared prototypes and then pushed apart by contrastive anchors, reaching a 0.693 overall C-index on five TCGA cohorts.","keywords":["multimodal survival analysis","whole-slide images","genomic profiles","unbalanced optimal transport","prototype alignment","contrastive learning","representation distinctiveness","TCGA"],"falsifier":"Rerun the five TCGA benchmarks with the instance-level pseudo-label cross-entropy (Equation 14) removed. If the overall C-index stays near 0.693, the pseudo-label loop is not load-bearing; if it falls toward the 0.686 of UOT without curriculum mass, the alignment gain is mostly self-training. A second check is to freeze the transport plan using a separately pretrained encoder, so the assignments cannot adapt to the classifier being trained.","tokens_in":24472,"feed_emoji":"🧬","tokens_out":4684,"duration_ms":49414,"temperature":0.7,"pith_summary":"Survival models that fuse whole-slide images and gene expression often over-align the two modalities, blurring the very signals that matter for prognosis. This paper claims the right order is Together, Then Apart: first align both modalities onto shared prototypes using an unbalanced optimal transport plan, then enforce modality-specific distinctiveness with anchor-based contrastive learning. On five TCGA cancer cohorts the proposed TTA framework reaches an overall C-index of 0.693, outperforming the runner-up by 2.6%, and it avoids cases where prior multimodal methods fall below strong single-modality baselines. Ablations indicate that both stages contribute roughly equally and complement each other.","feed_headline":"Align first, then split: formula beats cancer-survival benchmarks","feed_subtitle":"Two-stage TTA model tops five cancer benchmarks, scoring 0.693 concordance vs 0.667 for the next-best method.","key_machinery":"The load-bearing machinery is a shared prototype bank combined with an unbalanced optimal transport (UOT) solver that allocates only a curriculum-controlled fraction of mass to real prototypes while routing the rest to a zero-cost sink, making assignments lenient early and stricter later. That transport plan is then reused as soft pseudo-labels. The Apart stage adds learnable modality anchors plus an InfoNCE-style contrastive regularizer, which pushes the two modalities apart after alignment.","core_discovery":"TTA's central claim is that survival prediction is best served by explicitly sequencing alignment and distinctiveness rather than blending them into one fusion objective. The Together stage maps WSI patch tokens and pathway tokens to K=32 shared learnable prototypes, solving a semi-relaxed unbalanced optimal transport problem over the concatenated tokens; the resulting plan is used both to aggregate prototype representations and as soft pseudo-labels that supervise token-to-prototype predictions. The Apart stage refines these prototype tokens with modality-specific anchors and a contrastive objective that attracts each modality to its own anchor and repels the other modality's anchor, counte","pith_inferences":["One testable extension: remove Equation (14), the instance-level soft cross-entropy on UOT pseudo-labels, and rerun the five-cohort benchmark. If the gain persists, the Together stage works through structure-aware aggregation; if it disappears, the benefit is largely self-training rather than alignment.","The transport pseudo-labels are generated from the same logits they supervise, so a cleaner proof of concept would use assignments from a frozen, independently trained encoder to rule out self-reinforcement bias.","The same Together-Then-Apart recipe could apply beyond survival, e.g., to vision-language models, where over-alignment is a known cause of modality confusion and where a curriculum-controlled transport step might serve as a regularizer.","The reported gains are numerically modest in absolute terms; whether the 0.693 concordance translates into clinically actionable risk stratification on independent cohorts is an open question not settled by the paper. "],"forward_implications":["If the principle holds, multimodal survival models can be designed as an explicit min–max trade-off between semantic coherence and modality distinctiveness rather than picking one or the other.","UOT with curriculum mass appears to be a reusable mechanism for handling noisy, heterogeneous correspondences in other multimodal biomedical tasks.","Learned shared prototypes provide a window into which morphological and pathway-level patterns jointly drive risk, which could support hypothesis generation in cancer biology.","The framework avoids the observed failure mode where several multimodal methods underperform WSI-only baselines on CRC, suggesting alignment pressure can be selectively relaxed.","The design can be transferred to any paired-modality survival setting, not only histology and transcriptomics, as long as both modalities can be tokenized and projected to a shared space. "],"fun_headline_variants":["TTA: align first, then diverge, for sharper cancer survival reads","Sequenced alignment: first common, then unique, lifts survival AI","Two-stage TTA beats blended fusion on cancer survival","Align to share, then split to keep: cancer survival boost","First find common ground, then preserve uniqueness for prognosis"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the unbalanced optimal transport plan—computed from the model's own token-to-prototype logits and then used as soft pseudo-labels to supervise those same logits—captures genuine shared survival structure rather than merely enforcing self-consistency.","fun_headline_variants_meta":{"raw":{"variants":["TTA: align first, then diverge, for sharper cancer survival reads","Sequenced alignment: first common, then unique, lifts survival AI","Two-stage TTA beats blended fusion on cancer survival","Align to share, then split to keep: cancer survival boost","First find common ground, then preserve uniqueness for prognosis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00054,"raw_usage":{"total_tokens":2413,"prompt_tokens":719,"completion_tokens":1694,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":1622}},"tokens_in":463,"tokens_out":1694,"duration_ms":11105,"temperature":1.0,"reasoning_tokens":1622,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:48:27.537994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the five TCGA benchmarks with the instance-level pseudo-label cross-entropy (Equation 14) removed. If the overall C-index stays near 0.693, the pseudo-label loop is not load-bearing; if it falls toward the 0.686 of UOT without curriculum mass, the alignment gain is mostly self-training. A second check is to freeze the transport plan using a separately pretrained encoder, so the assignments cannot adapt to the classifier being trained.","supporting_citations":[],"review_version":1}