{"id":"600d6a7a-6b04-40e9-9e69-6b0f034be8f5","arxiv_id":"2507.04613","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HiLa aligns multiple survival-related language prompts with patch- and region-level slide features using optimal transport, contrastive learning, and cross-level gating, achieving higher C-index on three TCGA cohorts.","lead":"A new computer vision method, HiLa, combines language prompts generated by GPT-4o with hierarchical slide image analysis to predict cancer survival better than previous models on three public datasets. It gains a few concordance-index points over prior state of the art, though a central training equation appears misprinted.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed C-index gains over VLSA are within reported fold-to-fold variance; without per-fold paired significance tests, the headline SOTA superiority is not established.","rationale":"We weighed several candidate concerns. The loss in Eq. (7) is ambiguous: if c_n=1 denotes censored (as suggested by the phrase 'right censor status'), the equation is correct; if c_n=1 denotes an event, it is reversed. The paper never defines the convention, so this is a genuine reproducibility issue but not a demonstrated error. The OPL cosine-similarity assumption flagged by the reader is plausible, and the ablations show the OPL component changes the C-index, so the mechanism is doing something. The most load-bearing concern is that the central claim of superiority is not statistically grounded. The reported gains are all less than one standard error of the difference, and the only reported significance tests are within-method, not between-method. Because the abstract and Section 3.2 make direct superiority claims, this is the weakest link in the argument. A single concrete check—per-fold paired testing—can settle it. We recommend keeping the reader's conditional verdict: the manuscript should either provide significance evidence or soften the SOTA claims; the method structure is coherent, so this is not a rejection.","tokens_in":10558,"tokens_out":8318,"duration_ms":94342,"concrete_test":"Request from the authors the per-fold C-index values for every method on all three datasets (or the patient-level risk scores enabling bootstrapping). On each dataset, compute a paired one-sided test between HiLa and VLSA across the five folds (e.g., Wilcoxon signed-rank test, or a paired t-test, reporting the bootstrap 95% CI of the mean difference). If the 95% CI includes zero in any dataset, the claimed advantage of 2.1–3.7 points is not statistically supported there; if all three CIs exclude zero, the concern is resolved. Also report the same test for the overall average across datasets using the three per-dataset differences.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is empirical: HiLa is state-of-the-art, beating VLSA by 2.1–3.7 C-index points per dataset (Table 1). The only support is a single 5-fold cross-validation with mean ± std. From the reported numbers, the differences are small relative to the noise: HiLa−VLSA = 0.037/0.021/0.034 on BRCA/LUAD/UCEC, while approximate standard errors of these differences (√(σ_HiLa²+σ_VLSA²)) are ≈0.049/0.074/0.064. No per-fold results, p-values, or confidence intervals are given. The Kaplan-Meier log-rank tests in Fig. 2 compare high- and low-risk groups within a single method, not HiLa versus VLSA, so they do not substantiate the cross-model advantage. If the per-fold differences are not statistically significant, the phrases 'SOTA', 'surpasses', and 'superior' overstate what is demonstrated. This is independent of whether Eq. (2)–(3) is a valid alignment mechanism: even if the mechanism is sound, the headline quantitative claim may be noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HiLa, a hierarchical vision-language collaboration framework for whole-slide-image survival prediction. It generates multiple cancer- and level-specific language prompts with an LLM, aligns them to visual tokens at both patch and region levels through an optimal-transport-based Optimal Prompt Learning (OPL) module, and then combines Cross-Level Propagation (CLP) and Mutual Contrastive Learning (MCL) to enforce hierarchical cooperation. The final risk score is obtained from selected visual tokens via a discrete-time survival negative log-likelihood. Experiments are conducted on three TCGA datasets (BRCA, LUAD, UCEC) with 5-fold cross-validation; the paper reports an average C-index of 0.671 and claims state-of-the-art performance over vision-only and vision-language baselines, with ablations attributing gains to each proposed component.","tokens_in":10782,"tokens_out":7920,"duration_ms":87451,"significance":"If the empirical claims hold, HiLa is a potentially useful step in vision-language survival prediction: it addresses two under-explored issues (multi-prompt alignment and hierarchical WSI modeling), uses a reasonable experimental setup on public TCGA data, and provides an ablation ladder from a V-only baseline to the full model. The method is for the most part clearly described and the components are properly motivated. At the same time, the significance is currently tempered by two load-bearing concerns: the printed survival NLL in Eq. (7) is inconsistent with the standard censored/uncensored likelihood, and the headline C-index gain over VLSA is reported only as per-dataset mean±std without paired significance testing. The absence of direct evidence that the OPL-selected tokens correspond to survival-relevant morphology additionally limits the interpretability claims. With the requested corrections and supporting analyses, the contribution would be much stronger.","major_comments":[{"comment":"The survival negative log-likelihood is printed with the censored and uncensored terms reversed relative to the standard discrete-time likelihood. The correct NLL is Σ_n [ -c_n log S_n(t_n−1|H_F^n) − c_n log h_n(t_n|H_F^n) − (1−c_n) log S_n(t_n|H_F^n) ]; Eq. (7) instead puts -log S_n(t_n) on uncensored patients and -log S_n(t_n−1) - log h_n(t_n) on censored patients. If the implementation follows the printed equation, training optimizes the wrong target; if the implementation uses the standard form, the text must be corrected. This is a load-bearing part of the method and must be fixed.","section":"2.3, Eq. (7)"},{"comment":"The headline claim that HiLa surpasses VLSA by 2.1–3.7 C-index points is not supported by the reported statistics. With 5-fold cross-validation, the per-dataset differences are comparable to or smaller than the apparent fold-level noise; for example, LUAD shows a difference of 0.021 against a standard-error approximation of about 0.074 from the reported fold standard deviations. No per-fold paired differences, confidence intervals, or p-values are provided. The Kaplan-Meier log-rank tests in Fig. 2 compare high- vs low-risk groups within a single method, not HiLa versus VLSA, so they do not substantiate the cross-model superiority. Please add paired significance testing or per-fold difference distributions and temper the abstract/conclusion claims accordingly.","section":"Table 1, Sec. 3.2"},{"comment":"The claim that OPL selects survival-relevant visual tokens rests on the assumption that cosine similarity between PLIP text prompts and HIPT visual embeddings is a valid measure of survival relevance. Because CLP and MCL operate on the OPL-selected tokens, a systematic mismatch between the prompt space and slide-image space would propagate through the whole pipeline. The paper provides no direct evidence for this assumption. Please add qualitative token-selection visualizations, a retrieval or alignment accuracy measure, or a control experiment with shuffled or incongruent prompts to support the claim that the selected tokens carry survival information. The marginal distributions u and v in Eq. (2) should also be specified.","section":"2.2, Eqs. (2)–(3), Fig. 1"}],"minor_comments":[{"comment":"The expression for the selected token set writes the dimension as (M × a%) × d; this should be (M × r%) × d to match the top-r% notation used in the text.","section":"2.2, after Eq. (3)"},{"comment":"The abbreviation CPM is used in the ablation descriptions and Table 2 (Models F and G), while the method is named CLP in Section 2.2; please use one consistent abbreviation throughout.","section":"3.3, Table 2"},{"comment":"The notation L_P = {L_P^i}_{i=1}^{N_P × d} is dimensionally confusing; it should be L_P ∈ R^{N_P × d} (and likewise L_R) rather than a set indexed by a product of two numbers.","section":"2.1, Language Prompts Generation"},{"comment":"The hardware description says \"a GeForce GTX 3090 GPUs\"; this should be singular. Please also state the exact LLM query template, the number of generated prompts per level (N_P, N_R), and whether the generated prompt sentences are shared across all patients or vary per slide.","section":"3.1, Implementation Details"},{"comment":"The adaptation of CoOp for the vision-language baseline is only described as cosine similarity plus top-r% selection; it is unclear whether the compared CoOp uses learned prompt context vectors or a fixed zero-shot prompt, and whether the same OPL hyperparameters are used. Please clarify to make the comparison reproducible.","section":"3.2, CoOp baseline"}],"recommendation":"major_revision","confidential_remarks":"The main barriers to acceptance are the incorrect printed survival NLL and the lack of paired statistical evidence for the headline SOTA claim; both are fixable within the scope of a revision. The paper would also benefit from sharing code or per-fold predictions to enable verification. The topic is well within the scope of a computational pathology or medical imaging venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine engineering step forward for vision-language survival prediction. It takes HIPT hierarchical features, adds LLM-generated prompts at patch and region levels, uses optimal transport for prompt-token alignment, and ties the two levels together with a gated propagation and a contrastive consistency loss. The ablation table is the best part: each added component moves the C-index up consistently (0.612 to 0.671 overall), which makes the design choices credible. The authors also use three public TCGA cohorts and standard 5-fold CV.\n\nThat said, there are two real soft spots. First, Eq. (7) as printed is wrong: for uncensored patients it keeps log S(t) instead of log S(t-1)+log h(t), and for censored patients it uses log S(t-1)+log h(t) instead of log S(t). I assume this is a typesetting error, because the cited loss (Zadeh and Schmid) is standard, but it cannot be left as is. Second, the headline \"SOTA\" claim is not backed statistically. The reported gaps over VLSA are 2.1-3.7 C-index points; with 5-fold CV and no per-fold values, the reader can't tell whether those gaps are real. The stress-test's computation overstates the noise (it uses the std of a single fold as if it were the SE of the mean), but even using reasonable SEs (sigma/sqrt(5)) the differences are not significant at 0.05. I'd want paired per-fold results, a Wilcoxon test, or confidence intervals before accepting \"surpasses.\"\n\nTwo smaller issues: no code or generated prompts are released, which hurts reproducibility, and there's no qualitative evidence, such as token-level heatmaps or retrieval examples, that the OPL-selected tokens actually correspond to survival-relevant morphology. The final C-index is the only support for the alignment mechanism. The comparison set is also thin: only one recent VL method (VLSA, an arXiv preprint) and several older MIL baselines.\n\nOverall, the method is coherent, the ablations are convincing, and the problem is worth solving. If Eq. (7) is corrected and the significance analysis is added, I'd be comfortable with it at a decent venue. It deserves a serious referee; I'd send it out rather than desk-reject.","headline":"Solid incremental VL-survival framework with a clean ablation story, but the SOTA claim lacks significance tests and Eq. (7) appears to swap censored and uncensored terms.","tokens_in":11302,"tokens_out":4126,"would_cite":false,"duration_ms":42380,"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":"Matching whole-slide images to many LLM-generated text prompts at two tissue scales improves cancer survival prediction, reaching an average C-index of 0.671 across three TCGA datasets.","keywords":["survival prediction","whole-slide images","vision-language models","optimal transport","hierarchical features","computational pathology","contrastive learning"],"falsifier":"Run HiLa on a TCGA dataset with the OPL selection replaced by random selection of the same number of tokens, keeping all other components identical; if the C-index does not fall substantially, the claim that optimal prompt alignment is responsible for the gain is falsified. A complementary check is to visually inspect the top-ranked tokens per prompt and verify they correspond to the attribute the prompt describes.","tokens_in":10355,"feed_emoji":"🔬","tokens_out":4720,"duration_ms":49247,"temperature":0.7,"pith_summary":"This paper proposes HiLa, a framework for predicting cancer patient survival from gigapixel whole-slide images by aligning image content with multiple text prompts that describe survival-related visual attributes. The authors argue that prior vision-language survival models align slides to a single oversimplified prompt with plain cosine similarity, missing fine-grained associations, and that they ignore the hierarchical structure of WSIs. HiLa instead builds patch-level and region-level visual features, generates many attribute-specific prompts with a large language model, and matches prompts to visual tokens with optimal transport. On three TCGA datasets, the method reports an average concordance index of 0.671, outperforming the best prior vision-language survival model by 2.1 to 3.7 points per dataset. If the result holds, it suggests that richer language supervision, applied at multiple image scales, is a practical way to extract survival-relevant morphology from unlabeled slide regions.","feed_headline":"Two-scale text-image matching boosts cancer survival prediction","feed_subtitle":"LLM prompts and optimal transport pick survival-relevant tissue, beating prior vision-language models by 2.1-3.7 C-index points.","key_machinery":"The central mechanism is Optimal Prompt Learning (OPL), which treats the alignment of language prompts to visual tokens as an optimal-transport problem: it computes a cost matrix from cosine similarities, solves for an optimal matching plan, and scores each visual token by its summed matching probability across all prompts; the top r% tokens are kept. Around this are two hierarchical modules: Cross-Level Propagation (CLP), which uses a gating mechanism to let selected patch-level tokens recalibrate region-level tokens before region-level OPL; and Mutual Contrastive Learning (MCL), which enforces consistency between patch-level and region-level prototypes for the same patient through a memory-queue contrastive loss. The architecture uses pretrained patch-level and region-level feature extractors for hierarchical visual tokens and a pathology vision-language text encoder for the prompts.","core_discovery":"HiLa's central claim is that hierarchical vision-language collaboration—matching many LLM-generated survival-attribute prompts to visual tokens at both patch and region levels, then propagating and contrasting information between the two levels—produces discriminative slide representations that improve survival prediction beyond both vision-only MIL methods and existing vision-language survival models. The paper reports an average C-index of 0.671 across BRCA, LUAD, and UCEC, with per-dataset gains of 3.7, 2.1, and 3.4 points over the second-best VLSA baseline, together with statistically significant Kaplan-Meier stratification. The ablation shows each component contributes: multiple prompts over a single prompt, optimal-prompt learning over cosine similarity, region-level over patch-only, and cross-level propagation and mutual contrastive learning add further gains. In short, the discovery is that survival-relevant signal in WSIs is spread across morphological scales and multiple linguistic attributes, and optimal-transport alignment can harvest both.","pith_inferences":["The paper's framing suggests that LLM-generated prompts encode a prior over what pathologists look for; a natural extension, untested here, is to see whether prompts derived from actual pathology reports or clinical guidelines change performance.","Because OPL selects tokens by alignment to text, it implicitly provides an interpretability map per attribute; one could validate whether the top tokens per prompt correspond to the described morphology, something the paper does not report.","The mutual contrastive loss assumes patch-level and region-level prototypes for the same patient should be consistent; an alternative assumption, that they should be complementary, might lead to different design choices that could be explored.","The reported gains come from TCGA diagnostic slides at a single magnification; whether the hierarchy transfers to other scanners, stains, or resolutions is untested and would be a useful stress test."],"forward_implications":["If HiLa generalizes, vision-language survival models should move from single-prompt cosine alignment to multi-prompt optimal-transport alignment, since the ablation shows this is a major source of gain.","Hierarchical modeling at patch and region scales should become a standard component in WSI survival pipelines, as the region level provides complementary global context and cross-level interaction adds further improvement.","The OPL token-selection procedure may serve as a weakly supervised way to localize survival-relevant morphology in gigapixel slides, because selected tokens are precisely those matched to attribute-specific prompts.","Cross-level propagation and mutual contrastive learning may transfer to other hierarchical gigapixel pathology tasks, such as subtyping or grading, where local cellular morphology and global tissue organization both matter.","The per-dataset gains over the prior best vision-language survival model are consistent across all three cancer types, suggesting the approach is not tuned to a single tissue or tumor type."],"supporting_citations":[{"why":"Supplies the pretrained patch- and region-level feature extractors that produce the hierarchical visual tokens central to HiLa.","marker":"[24]"},{"why":"Supplies the text encoder that converts generated prompt sentences into language embeddings used for vision-language alignment.","marker":"[3]"},{"why":"The prior vision-language survival model that HiLa compares against and claims to beat by 2.1-3.7 C-index points.","marker":"[21]"},{"why":"Prior optimal-transport and prototyping methods that inspire the Optimal Prompt Learning module.","marker":"[25, 26]"},{"why":"Supplies the negative log-likelihood survival loss used to supervise the final risk prediction.","marker":"[27]"},{"why":"Adapted as a vision-language baseline using cosine similarity, providing the comparison that shows OPL's advantage.","marker":"[30]"}],"fun_headline_variants":["Hierarchical vision-language collaboration lifts survival prediction","Cross-scale text-image alignment improves cancer survival prediction","Vision-language hierarchy outperforms prior survival methods","Optimal prompt learning sharpens survival prediction from slides","Survival prediction gains from hierarchical vision-language alignment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the similarity between LLM-generated survival-attribute sentences and pretrained visual embeddings is a reliable guide to which tissue patches and regions actually matter for survival; the paper's only direct evidence for this is the final C-index improvement.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical vision-language collaboration lifts survival prediction","Cross-scale text-image alignment improves cancer survival prediction","Vision-language hierarchy outperforms prior survival methods","Optimal prompt learning sharpens survival prediction from slides","Survival prediction gains from hierarchical vision-language alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000888,"raw_usage":{"total_tokens":3886,"prompt_tokens":1053,"completion_tokens":2833,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":2763}},"tokens_in":669,"tokens_out":2833,"duration_ms":23110,"temperature":1.0,"reasoning_tokens":2763,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:43:51.851465+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HiLa on a TCGA dataset with the OPL selection replaced by random selection of the same number of tokens, keeping all other components identical; if the C-index does not fall substantially, the claim that optimal prompt alignment is responsible for the gain is falsified. A complementary check is to visually inspect the top-ranked tokens per prompt and verify they correspond to the attribute the prompt describes.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained patch- and region-level feature extractors that produce the hierarchical visual tokens central to HiLa."},{"cited_title":"and Zou, J.: A visual–language foundation model for pathology image analysis using medical twitter","cited_arxiv_id":null,"evidence_quote":"Supplies the text encoder that converts generated prompt sentences into language embeddings used for vision-language alignment."},{"cited_title":"Interpretable Vision-Language Survival Analysis with Ordinal Inductive Bias for Computational Pathology","cited_arxiv_id":"2409.09369","evidence_quote":"The prior vision-language survival model that HiLa compares against and claims to beat by 2.1-3.7 C-index points."},{"cited_title":"and Schmid, M., Bias in cross-entropy-based training of deep survival networks","cited_arxiv_id":null,"evidence_quote":"Supplies the negative log-likelihood survival loss used to supervise the final risk prediction."},{"cited_title":"and Liu, Z.: Conditional prompt learning for vision-language models","cited_arxiv_id":null,"evidence_quote":"Adapted as a vision-language baseline using cosine similarity, providing the comparison that shows OPL's advantage."}],"review_version":1}