{"id":"dfbf1326-703c-4aa7-937c-4090ea1fcc17","arxiv_id":"2607.16802","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A sigmoid concordance loss whose value approximates one minus the C-index stays coupled to ranking performance throughout training, while likelihood-based survival losses provably and empirically decouple from it.","lead":"The authors propose a sigmoid-based survival loss whose numeric value tracks the concordance index during training, unlike standard likelihood losses whose value can fall while the ranking does not improve. Across 18 datasets it matches likelihood baselines in discrimination while giving a reliable, nearly free signal for early stopping and checkpoint selection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The missing link: Corollary 2 asserts, without proof, that once ranking converges the likelihood gradient is predominantly rank-preserving — this is what turns existence of one rescaling direction into the claim that likelihood loss values become uninformative.","rationale":"Reader's weakest_assumption identifies Corollary 2, and I agree that is the single load-bearing gap. The paper's contribution has two connected parts: (1) SCL value is approximately 1−C (Theorem 1) — this is well supported by the proof sketch, the convex linear special case, and the empirical 0.96–0.99 correlations; (2) likelihood losses structurally decouple from C-index — the existence of rank-preserving directions is rigorously shown, but the trajectory-level conclusion is not. The central practical message ('the loss value itself is an unreliable signal for likelihood losses during expensive end-to-end training') depends on the unproved Corollary 2; without it, one could imagine a likelihood trajectory whose gradient mostly changes ranks early and preserves ranks late, or the opposite, and the theory would not discriminate. The paper's own A.6 limitation statement concedes exactly this missing arbitrary-direction strengthening. The Cox 'strictly proper' misclassification is a smaller correctness issue but does not damage Proposition 2. Empirical results are extensive, reproducible-in-principle, and would survive correction of the theory, so the conditional verdict is appropriate; no further adjustment needed.","tokens_in":27645,"tokens_out":6275,"duration_ms":64831,"concrete_test":"Instrument the released code on METABRIC with the Cox loss. At each epoch, extract the validation risk vector f and the training-loss gradient g with respect to risk outputs; decompose g into g∥ = (⟨g,f⟩/⟨f,f⟩)f (the rescaling/rank-preserving component) and g⊥ = g − g∥. Record r_t = ‖g⊥‖/‖g∥‖ and validation C-index across training. Corollary 2 predicts r_t → 0 after C-index plateaus; if r_t stays large or oscillates while the loss and C-index continue to move together/differently, the missing trajectory assertion is contradicted. A complementary analytical check: in a two-parameter linear Cox model, solve gradient flow and show dC(f_t)/dt is not identically zero whenever C is near its limit; this would refute the 'predominantly rank-preserving' claim as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The practical headline is that likelihood losses cannot be trusted as value proxies because their values decouple from the C-index during training. What is actually proved (Proposition 2 / Theorem 2, Appendix A.4) is weaker: for every score there exists some direction — e.g., rescaling f → αf — along which the Cox loss strictly decreases while C(αf)=C(f). Existence of a rank-preserving descent direction does not imply that gradient descent, SGD, or Adam follows it. The full gradient also has components that change pair orderings; C moves along those. Corollary 2 supplies the needed bridge: 'once the ranking has approximately converged, the remaining likelihood gradient lies predominantly in the rank-preserving subspace.' No proof or quantitative control is given for the rank-changing component, and no argument shows that approximate convergence of C bounds the rank-changing gradient. Appendix A.6 explicitly declines to make the arbitrary-direction statement, saying the rescaling case 'already establishes the core claim'; but the core training-time claim needs the trajectory statement, not merely an orbit direction. Separately, A.4 calls the Cox partial likelihood a strictly proper survival likelihood; the Cox partial likelihood is not a strictly proper scoring rule for the conditional event distribution, though Proposition 2's proof does not depend on properness. The empirical evidence (SCL rank correlations 0.96–0.99 vs likelihood ≤0.53) supports the phenomenon on the tested trajectories, so the concern is about theoretical overreach, not the experimental core.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sigmoid concordance loss (SCL) for deep survival prediction, defined as a soft count of misordered comparable pairs, and argues that its value tracks the C-index during training. The authors prove that the population SCL equals one minus the C-index up to an O(τ) temperature term (Theorem 1), and they prove that likelihood losses such as the Cox partial likelihood admit rescaling directions along which the loss decreases while the C-index is exactly unchanged (Proposition 2/Theorem 2). They then use this to claim that likelihood loss values decouple from the concordance index during training, particularly late in training. Empirically, across eighteen datasets from four modalities under a fixed five-fold, three-seed protocol, SCL achieves discrimination comparable to likelihood baselines, with loss-metric rank correlations of 0.96–0.99 versus at most 0.53 for likelihood losses, and selection regret below 0.011 on all datasets. Code and preprocessing pipelines are released.","tokens_in":27972,"tokens_out":6038,"duration_ms":59969,"significance":"The paper addresses a real and underappreciated problem: when validation C-index evaluation is too expensive, practitioners rely on loss values for early stopping and model selection. The proposed SCL is simple, architecture-agnostic, and backed by a strong empirical protocol with a fixed backbone per modality, multiple datasets, multiple seeds, several metrics, and released code. If the value-tracking property is taken as the design criterion, the paper makes an original contribution. The main caveat is that the theoretical attack on likelihood losses is only partially established: the existence of a rank-preserving descent direction does not by itself prove that actual gradient-based training follows such a direction, and the paper's Corollary 2, which is the bridge to the training-time conclusion, is asserted without proof. The empirical results remain informative, but the theoretical claims need to be either strengthened or explicitly reframed.","major_comments":[{"comment":"The statement of Theorem 2 contains an invalid inference: 'there is a descent direction of L along which the concordance index is first-order stationary, so gradient flow on L moves within the class of order-equivalent scores while the concordance index does not change.' Gradient flow follows the negative full gradient, not an arbitrary tangent descent direction, and nothing in the proof constrains the full gradient to the tangent space of the orbit. This overstates the theoretical result. The proved content — existence of a rank-preserving descent direction — is correct but much weaker than the training-trajectory conclusion drawn from it.","section":"A.4 / Theorem 2"},{"comment":"The paper repeatedly classifies the Cox partial likelihood as a 'strictly proper survival likelihood.' The Cox partial likelihood is not a strictly proper scoring rule for the conditional event distribution; it is a partial likelihood that uses risk sets and does not score the full survival distribution. This does not invalidate Proposition 2, whose rescaling computation is independent of properness, but the terminology is inaccurate and appears in the abstract, Section 3.5, and Appendix A.4. The authors should say that the Cox partial likelihood and the listed discrete-time likelihoods are likelihood-based objectives, and reserve 'strictly proper' for the full-data negative log-likelihoods where it is justified.","section":"A.4 / A.1"}],"minor_comments":[{"comment":"Assumption A1 says the score gap has a density bounded by B 'whenever f is non-degenerate and X has a density.' This is not generally true without additional regularity on f; a continuous non-degenerate f can still push a continuous X distribution onto a distribution with no density. Add smoothness/injectivity conditions or state A1 directly as an assumption.","section":"A.1"},{"comment":"The phrase 'value equals one minus the C-index up to a temperature term' should specify that the equality is asymptotic as τ → 0 and that at τ = 0.1 the residual term is not quantified. The abstract and conclusion state this slightly more strongly than the theorem proves.","section":"3.3 / Theorem 1"},{"comment":"The checkpoint selection protocol (best validation Harrell C-index) is favorable to all methods, but this is stated only in the supplementary material. Consider stating it explicitly in the main text, since it affects interpretation of the discrimination results.","section":"4.4"},{"comment":"Some per-dataset standard deviations in the supplementary tables are missing or shown without ± (e.g., Table S1a 'Ours-Sigmoid' 0.725±0.000, and some cells such as '0.934' or '0.995' with no deviation). Please ensure all entries have consistent formatting.","section":"Table S1 / Table 2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this for the empirical story, not the theorem statements. The core finding — that a temperature-0.1 sigmoid concordance loss has a validation loss that tracks the C-index with Spearman correlations around 0.96–0.99 across 18 datasets, while Cox, MTLR, DeepHit, and TripleSurv losses do not — is credible and practically important for anyone training expensive end-to-end survival models where per-epoch C-index evaluation is costly. The protocol is controlled: same backbone per modality, five-fold CV, three seeds, multiple metrics, and they honestly report that SCL matches rather than beats likelihood losses on discrimination.\n\nThe new things are the value-monotonicity framing, the explicit proof that rescaling a risk score strictly decreases the Cox likelihood while leaving the C-index exactly unchanged, and the demonstration that a sigmoid (unlike a squared hinge) has a value that approximates 1−C up to a temperature term. Theorem 1 and Proposition 2 are basically correct elementary statements. The ablation with a likelihood anchor, where coupling collapses as soon as any positive anchor is added, gives the theory some predictive content.\n\nSoft spots: the headline narrative goes beyond what is proved. Theorem 2 shows there exists a rank-preserving descent direction; it does not show gradient descent follows it. Corollary 2's claim that 'once the ranking has approximately converged, the remaining likelihood gradient lies predominantly in the rank-preserving subspace' is asserted, not proven, and it is exactly what turns the existence result into the claim that likelihood loss values become uninformative. The O(τ) bound in Theorem 1 is not controlled tightly enough at τ=0.1 to guarantee the observed coupling, and the finite-sample bound is admitted to be sketched. Also, the Cox partial likelihood is not a strictly proper scoring rule for the conditional event distribution, but the proof of Proposition 2 does not depend on that — a sloppy label, not a load-bearing flaw. None of this sinks the empirical conclusion; the loss-metric correlations are measured on actual training trajectories, so the phenomenon is real on the tested setups. What needs fixing is the theory-to-trajectory bridge.\n\nThis paper deserves a serious referee. The useful audience is survival practitioners in medical imaging and anyone who uses validation loss for early stopping or checkpoint selection. I would send it to review, expecting major revision on the theory section: correct the overstatements, either prove a trajectory statement or clearly label it as a conjecture supported empirically, and benchmark against a couple of existing smooth concordance surrogates. The code is promised; a release with a commit hash would help reproduction.","headline":"The empirical loss-metric coupling is real and worth knowing; the theory overreaches exactly where Corollary 2 bridges an existence result to a training-trajectory claim.","tokens_in":28441,"tokens_out":2655,"would_cite":true,"duration_ms":25858,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62N01","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper's central claim is that the value of a survival loss can be trusted as a proxy for the C-index only if it is value-monotone, and that standard likelihood losses are provably not; the proposed sigmoid concordance loss satisfies thi","keywords":["survival analysis","concordance index","value-monotonicity","sigmoid concordance loss","likelihood decoupling","censored data","early stopping","model selection"],"falsifier":"Run a likelihood-trained deep survival model on any public dataset, logging the validation C-index and the validation loss every epoch. If, after the validation C-index has visibly plateaued, one can find a run in which the validation loss and validation C-index remain strongly rank-correlated (say above 0.8) for many epochs, the paper's claim that likelihood losses are uninformative late in training would be false in that regime; if instead the loss moves while the C-index is flat, the decoupling mechanism is confirmed.","tokens_in":27509,"feed_emoji":"📈","tokens_out":10672,"duration_ms":90114,"temperature":0.7,"pith_summary":"Deep survival models are judged by the concordance index but trained almost exclusively on likelihood objectives. This paper argues that the mismatch is structural, not incidental: every strictly proper survival likelihood admits a rescaling direction along which the loss decreases while the C-index stays exactly constant, so after the ranking has roughly converged the loss value stops telling you anything about the metric. The remedy is a sigmoid concordance loss (SCL) whose value equals one minus the C-index up to a temperature term, which makes lower loss imply higher concordance throughout training. In eighteen datasets spanning tabular, CT/PET, MRI, and pathology data, the SCL matches the discrimination and calibration of likelihood baselines while its validation value tracks the C-index with rank correlation 0.96–0.99 (versus -0.03 to 0.53 for likelihood losses), and checkpoint selection by the loss alone loses at most 0.011 of C-index. If true, this gives practitioners in expensive end-to-end imaging a loss they can monitor and select on without recomputing the metric every epoch.","feed_headline":"One survival loss tracks the C-index all through training","feed_subtitle":"Likelihood losses decouple from ranking once it plateaus; a sigmoid loss stays correlated at 0.96-0.99.","key_machinery":"The load-bearing object is the sigmoid concordance loss (SCL), a temperature-smoothed mean over comparable pairs of the indicator of a wrong ordering. It is constructed so its value approximates one minus the C-index rather than merely sharing its optimum; the key identity is Lτ = 1 − C + O(τ), which gives value-monotonicity. The companion mechanism is rank-invariance of the C-index: C(φ∘f)=C(f) for any strictly increasing φ, so the C-index's level sets are order-equivalence classes. Because likelihood losses are not invariant to such reparameterizations, their gradients carry a component along these rank-preserving directions, and that is the mechanism that decouples the loss value from the","core_discovery":"The central discovery is two-sided. On the negative side, the paper proves that every strictly proper survival likelihood admits an explicit, always-available direction—rescaling the risk score—along which the loss strictly decreases while the C-index, a rank statistic, is exactly unchanged; a general theorem extends this to all monotone reparameterizations. On the positive side, the proposed sigmoid concordance loss Lτ (mean over comparable pairs of σ(−(f_i−f_j)/τ)) satisfies Lτ = 1 − C + O(τ), so for small temperature the loss value itself estimates the misordering rate. Empirically, the SCL's validation trajectory follows the validation C-index with near-perfect rank correlation, whereas","pith_inferences":["The theoretical argument isolates score rescaling, so models whose final-layer scale, batch-normalization gain, or temperature drifts during training are the most exposed to decoupling; a diagnostic experiment that freezes or carefully controls these scale parameters could show that likelihood losses couple better than reported.","Because the SCL optimizes only the ordering, it discards information about event-time magnitude; a natural extension is to use SCL as a monitor and selector while training a likelihood, or to add a calibration head, to get both trustworthy value-tracking and calibrated survival curves.","A full proof of the finite-sample bound the paper sketches would turn value-monotonicity from an empirical observation into a rigorous guarantee with explicit sample-size requirements, which would strengthen the practical case for routine use of the loss.","The SCL's insensitivity to score magnitude could make it brittle when risk scores are badly calibrated across folds; testing the loss under distribution shift or with deliberately mis-scaled scores would reveal whether value-monotonicity survives outside the training distribution."],"forward_implications":["If SCL is value-monotone, early stopping and checkpoint selection can be driven by the loss value alone, eliminating the per-epoch validation C-index computation that dominates cost in end-to-end imaging.","A practitioner who chooses the minimum-validation-loss checkpoint under SCL obtains essentially the same test C-index as one who selects by the metric itself, with selection regret at most 0.011 on all reported datasets.","Because every strictly proper survival likelihood is non-invariant to monotone rescaling, no likelihood-based training loss can be made value-monotone; adding any positive likelihood weight to SCL re-introduces decoupling, as the hybrid ablation shows.","The same rank-invariance argument applies to any rank-based evaluation metric optimized through a proper scoring rule, so the decoupling phenomenon extends beyond survival analysis to AUC- or NDCG-style training."],"fun_headline_variants":["Loss value mirrors C-index during training","Sigmoid loss: 0.96–0.99 rank correlation with C-index","Proof: likelihood losses decouple—sigmoid loss doesn't","Train with a loss that tracks ranking, not just fit","A survival loss that never decouples from the C-index"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that likelihood losses become uninformative late in training rests on an unproved step: once the ranking has approximately converged, the remaining likelihood gradient is asserted, not shown, to lie predominantly in directions that preserve the ranking, so that the loss keeps moving while the C-index does not.","fun_headline_variants_meta":{"raw":{"variants":["Loss value mirrors C-index during training","Sigmoid loss: 0.96–0.99 rank correlation with C-index","Proof: likelihood losses decouple—sigmoid loss doesn't","Train with a loss that tracks ranking, not just fit","A survival loss that never decouples from the C-index"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001068,"raw_usage":{"total_tokens":4372,"prompt_tokens":867,"completion_tokens":3505,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":3418}},"tokens_in":611,"tokens_out":3505,"duration_ms":26008,"temperature":1.0,"reasoning_tokens":3418,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T19:53:09.305680+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a likelihood-trained deep survival model on any public dataset, logging the validation C-index and the validation loss every epoch. If, after the validation C-index has visibly plateaued, one can find a run in which the validation loss and validation C-index remain strongly rank-correlated (say above 0.8) for many epochs, the paper's claim that likelihood losses are uninformative late in training would be false in that regime; if instead the loss moves while the C-index is flat, the decoupling mechanism is confirmed.","supporting_citations":[],"review_version":1}