{"id":"93032abd-bcd8-4cb8-8b8a-3c0d90d02773","arxiv_id":"2608.01370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A lightweight gating network that fuses frozen pathology foundation models improves several downstream benchmarks, but one headline metric is contradicted by the paper's own table.","lead":"This paper introduces AdaFusion, a method that combines several pretrained AI models for pathology images and lets a learned gate decide how much to trust each model for each tissue sample. It reports higher accuracy on three medical benchmarks and uses gate weights to visualize where each model contributes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper's own Table 1 contradicts the central 'consistently outperforms' claim: H-optimus-0 alone reaches ATEC23 AUC 0.913 vs AdaFu-F's 0.910, and AdaFu-C trails Virchow v2 on HEST (0.368 vs 0.387). No significance tests are reported.","rationale":"Read in good faith: the AdaFusion idea—a lightweight sample-conditioned gate over frozen PFM features—is reasonable, and the interpretability maps are a useful addition. The code link and three benchmark evaluations are positive signs. However, the central claim is stated without qualification in the Abstract and Conclusion, and the paper's own results contradict it. This is not a matter of differing from consensus; it is a direct mismatch between claim and evidence. The reader identified pooling as the weakest assumption; I agree it is a real risk, but the more immediate issue is the reported numbers. The best fix is to soften the claim to 'often improves' or 'outperforms on most benchmarks,' add uncertainty quantification, and report fold-level results. Because these are substantive but addressable revisions, the CONDITIONAL verdict is appropriate; I would not change the reader's verdict.","tokens_in":8283,"tokens_out":7170,"duration_ms":62171,"concrete_test":"Run the released code to reproduce Table 1 for ATEC23, using the identical 5-fold splits and settings, and compute per-fold AUC for H-optimus-0 (original dims) and AdaFu-F (d=512) with a paired DeLong test. If the reproduced values match the table and H-optimus-0's mean AUC exceeds AdaFu-F's, the 'consistently outperforms' claim is false as stated. This can be done within the provided repository; if code is incomplete, request the authors to release fold indices and seeds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Most load-bearing concern: the abstract and §4.3 claim 'AdaFusion consistently outperforms individual PFMs and other fusion baselines,' but Table 1 shows the best AdaFusion variant (AdaFu-F, d=512) has ATEC23 AUC 0.910, below H-optimus-0's original-dim AUC 0.913; at d=64 it is 0.900. On HEST-Benchmark, AdaFu-C average PCC 0.368 is below Virchow v2's 0.387 (Table 2), and AdaFu-F's 0.399 is only a modest gain. No fold-level values, standard deviations, or paired significance tests are reported, so even the apparent PANDA gains are not shown to be statistically reliable. This is an internal inconsistency: the paper's own data fail to support the poster claim. The reader's pooling-loss concern (Eq. 1) is plausible but secondary; even if pooling were lossless, the reported numbers would not establish consistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AdaFusion, a lightweight adaptive fusion framework for combining features from six frozen pathology foundation models (PFMs). Each PFM's tile features are mean-pooled to a common low dimension (default d=64), concatenated, optionally masked during training, and reweighted by a sample-conditioned gating module in either model-wise (AdaFu-C) or channel-wise (AdaFu-F) mode. The fused representation is passed to a task head (ABMIL for slide-level classification, linear head for tile-level regression). The authors claim consistent improvements over individual PFMs and fusion baselines on three public benchmarks (ATEC23, PANDA, HEST-Benchmark), and additionally propose contribution-driven interpretability maps that attribute each region to the PFM with the highest gate-derived score. Code is released.","tokens_in":8630,"tokens_out":3748,"duration_ms":35393,"significance":"If the claims are sustained, AdaFusion would offer a practical and efficient way to combine heterogeneous PFM embeddings, and its contribution maps could help characterize model-specific biases. The use of three public benchmarks and the release of code are strengths. However, the central 'consistently outperforms' claim is not supported by the paper's own tables: several individual PFMs beat the best AdaFusion variant on ATEC23 AUC, and AdaFu-C underperforms Virchow v2 on HEST average PCC. No confidence intervals or significance tests are reported, so even the apparent PANDA gains are not established. The interpretability analysis is anecdotal. These issues affect the paper's main contribution claims.","major_comments":[{"comment":"The text states 'AdaFusion outperforms all PFMs and fusion baselines' on ATEC23, but Table 1 shows the best AdaFusion variant (AdaFu-F, d=512) achieves ATEC23 AUC 0.910, below H-optimus-0 at original dimension (0.913); AdaFu-C at d=64 achieves 0.901. Similarly, Table 2 shows AdaFu-C's average HEST PCC (0.368) is below Virchow v2 (0.387). The abstract's 'consistently outperforms individual PFMs and other fusion baselines' is not supported by the reported numbers. Please either restrict the claim to specific settings where it holds, or report additional evidence that establishes an overall advantage (e.g., wins/losses across tasks, significance tests, or effect-size summaries).","section":"§4.3, Table 1 and Table 2"},{"comment":"Only mean 5-fold scores are reported, without standard deviations, confidence intervals, or paired significance tests. The term 'significantly outperforming' in §4.3 is therefore unsupported. Please report fold-level results and run paired statistical tests (e.g., Wilcoxon signed-rank or bootstrap over slides/folds) for each benchmark, or explicitly avoid statistical language. This is load-bearing because the paper's headline claim is about consistent superiority.","section":"§4.2, §4.3, Tables 1–3"},{"comment":"The method's core compression step is a simple mean pooling to a fixed low dimension d=64. The paper does not validate that this pooling preserves the complementary information that motivates fusion, nor does it compare with alternative pooling or learned compression. This matters especially for PFMs like Prov-GigaPath that use dilated whole-slide features, where global/spatial context may be lost by per-tile mean pooling. The ATEC23 deficit versus H-optimus-0 may result from this compression. Please add ablation or analysis comparing pooling strategies and dimensions, or soften the claims about robust information preservation.","section":"§3.2, Eq. (1)"},{"comment":"The interpretability claim rests on contribution scores S_i defined by averaging gate weights (Eq. 5). No validation is provided that these scores correspond to actual model preferences or tissue phenotypes. The spatial maps and exemplar tiles in Fig. 2 are illustrative only. If contribution-driven interpretability is a headline contribution, quantitative evaluation is needed, e.g., consistency with expert-annotated tissue types, correlation with model-specific saliency maps, or human-preference studies.","section":"§3.4, §4.4, Fig. 2"}],"minor_comments":[{"comment":"The table formatting is difficult to read: the header is split across many columns without clear grouping, and the 'A TEC23' label should be 'ATEC23'. Please reformat for clarity.","section":"Table 1"},{"comment":"The dataset descriptions are brief. For ATEC23, the number of slides and the exact prediction task (effective vs. ineffective treatment) should be stated with a citation to the challenge. For PANDA, clarify whether the development set is used for both training and validation in each fold.","section":"§4.1"},{"comment":"The task abbreviations (ID, PR, P, A, SK, CO, RE, RC, LU, LY, Av g) are not all defined in the text. Please expand the table legend.","section":"Table 2"},{"comment":"The ablation study compares AdaFusion-Fine only against concatenation baselines. The effect of random masking is reported only for the ensemble, not for the gated variant. It would be informative to also ablate the coarse gate and masking in the same table.","section":"§4.3, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The empirical core is potentially useful, but the abstract and Section 4.3 overclaim relative to the reported numbers, and the lack of significance testing is a serious omission. With corrected claims, fold-level statistics, and a more careful treatment of the pooling and interpretability assumptions, the paper could become acceptable. The current version does not meet the bar because the central claim is contradicted by the paper's own tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is simple and worth taking seriously: compress frozen features from several pathology foundation models to a low dimension, then learn a sample-conditioned gate that reweights their contributions before the task head. That is a reasonable extension of attention/MoE-style fusion, and it ships with an interpretability hook — contribution maps that assign tissue regions to preferred PFMs. The authors test on three public benchmarks, include ablations, and promise code. That is real progress, not a paradigm shift.\n\nThe soft spots are significant, though. The abstract and Section 4.3 claim AdaFusion 'consistently outperforms' individual PFMs, but Table 1 shows the best AdaFu-F variant reaching ATEC23 AUC 0.910, below H-optimus-0's 0.913 at original dimension. On HEST, AdaFu-C's average PCC is 0.368, under Virchow v2's 0.387. Those are not consistent wins. Only mean 5-fold scores are reported, with no fold-level values, standard deviations, or paired significance tests, so even the apparent PANDA gains cannot be evaluated statistically. This is a load-bearing weakness: the paper's own data contradict its poster claim.\n\nThe pooling concern — that mean-pooling tile features to a 64-dimensional vector may erase spatial or global context, which could explain H-optimus-0's edge at original dimension — is plausible but secondary. Even if pooling were lossless, the missing uncertainty quantification would remain. The interpretability maps are post hoc readings of gate weights; they look qualitatively reasonable but are not validated against any ground truth, so they should be framed as illustrative, not evidence.\n\nNone of this sinks the method. The central argument that learned adaptive gating can help fuse complementary PFMs is credible and the ablations support the gate's role. The paper deserves a serious referee, but it needs revision: report fold-level results, add significance tests, include a stronger fusion baseline such as simple concatenation with a fine-tuned head, and soften the 'consistently outperforms' claim to match the data. I would not cite it as-is, but with those fixes it could be a useful practical contribution.","headline":"A sensible, lightweight fusion idea with decent experiments, but the headline 'consistently outperforms' is contradicted by the paper's own Table 1 and the statistics are too thin to support it.","tokens_in":9031,"tokens_out":1328,"would_cite":false,"duration_ms":14453,"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":"AdaFusion shows that a lightweight, sample-conditioned gate over compressed features from six frozen pathology foundation models consistently outperforms any single model and yields interpretable contribution maps.","keywords":["pathology foundation models","feature fusion","adaptive gating","multiple instance learning","computational pathology","interpretability","whole slide images"],"falsifier":"Run AdaFusion on a slide-level benchmark where the decisive signal is global tissue architecture, such as whole-slide staging or survival prediction, comparing its mean-pooled fusion against the best single PFM's native whole-slide features; if the individual model consistently wins, then the 64-dim pooling premise fails.","tokens_in":8240,"feed_emoji":"🧬","tokens_out":3178,"duration_ms":32606,"temperature":0.7,"pith_summary":"The paper tries to establish that complementary strengths of pathology foundation models can be coordinated by a cheap adaptive fusion gate, rather than by picking one model. It compresses each frozen model's tile embeddings to a fixed small dimension, then reweights models per sample before a downstream head. Across treatment-response classification, prostate grading, and gene-expression regression, this fused representation beats every individual model and earlier fusion baselines. The same gate also yields per-tile contribution maps that align with tissue phenotypes, offering a way to see what each model is attending to.","feed_headline":"Adaptive gate beats single pathology models on three benchmarks","feed_subtitle":"Sample-conditioned reweighting of compressed features lifts accuracy, AUC, and gene-expression correlation across three tasks.","key_machinery":"The adaptive fusion gate: per-model tile embeddings are mean-pooled to a unified dimension $d$ (default 64), concatenated, randomly masked during training, then passed through a lightweight gating module that emits either model-wise ($\\mathbb{R}^{N\\times 1}$) or channel-wise ($\\mathbb{R}^{N\\times d}$) weights; the reweighted features feed a linear head or an attention-based MIL aggregator. This gate is what makes the fusion sample-conditioned and is also the source of the interpretable model-wise contribution scores.","core_discovery":"AdaFusion is a lightweight framework that takes six frozen pathology foundation models with heterogeneous embedding dimensions, mean-pools tile features into 64-dimensional vectors, and learns a sample-conditioned gating module that reweights each model's contribution (coarse model-wise or fine channel-wise). The fused representation consistently improves accuracy and AUC on the ATEC23 bevacizumab-response and PANDA prostate-grading benchmarks and raises Pearson correlation on HEST gene-expression prediction, while contribution scores assigned to tiles produce spatial maps that agree with known morphological preferences. The paper argues this shows that PFM-specific blind spots can be counte","pith_inferences":["The authors do not isolate whether the performance gain comes from 'synergy' or simply from robust ensembling; a straightforward test would be to compare AdaFusion against a fixed, equally weighted average of the same compressed features on datasets where the individual models disagree most.","Because pooling collapses each tile to a mean vector, AdaFusion may underuse models whose value lies in global or spatial context, such as dilated whole-slide models; a token-level or attention-pooled variant is a natural extension the paper leaves untested.","The gate's contribution scores could be repurposed as a dataset-level diagnostic: if the same PFM always dominates across all tiles, that signals redundancy rather than synergy, and would justify shrinking the ensemble.","If the method were applied to out-of-distribution slides, the sample-conditioned weights might reveal which models carry reliable signal under distribution shift, an extension beyond the three in-domain benchmarks reported."],"forward_implications":["If the central claim holds, clinicians and computational pathology pipelines can combine several frozen PFMs with a small trainable gate and get better predictive performance than any single state-of-the-art PFM.","The compressed 6×64 representation matches or beats higher-dimensional feature combinations, implying fusion can be done with a large reduction in downstream compute and memory.","The contribution maps offer a mechanism for auditing which tissue structures each PFM responds to, which could inform model selection and quality control without knowing the PFMs' private training data.","The consistently positive results across classification, multi-class grading, and regression suggest the adaptive-gating principle transfers beyond any single task type.","The ablation shows that learned sample-conditioned gating, not just concatenation or random masking, is the source of the gain."],"supporting_citations":[{"why":"Supplies the UNI v2 pathology foundation model and its frozen tile embeddings, one of the six fused feature sources.","marker":"[2]"},{"why":"Supplies the Phikon-v2 feature extractor used as another PFM input to the fusion gate.","marker":"[6]"},{"why":"Supplies CONCH, the visual-language pathology foundation model whose high-dimensional embeddings are compressed by AdaFusion.","marker":"[14]"},{"why":"Supplies H-optimus-0, a large PFM that is both a fusion input and the strongest individual baseline on ATEC23.","marker":"[17]"},{"why":"Supplies Prov-GigaPath, the whole-slide dilated transformer PFM whose features test whether global-context embeddings survive pooling.","marker":"[23]"},{"why":"Supplies Virchow2, the mixed-magnification PFM used as another fusion input and individual baseline.","marker":"[25]"},{"why":"Provides the attention-based multiple-instance-learning aggregator used in the slide-level task heads.","marker":"[10]"},{"why":"Provides the PANDA prostate cancer grading benchmark whose 10,616 WSIs evaluate the multi-class classification claims.","marker":"[1]"},{"why":"Provides the HEST benchmark and its spatial gene-expression regression tasks used to evaluate the regression claims.","marker":"[11]"},{"why":"Provides the ATEC23 ovarian cancer bevacizumab-response challenge dataset used for binary classification evaluation.","marker":"[20]"},{"why":"Provides the Self-Attn features-fusion baseline that AdaFusion must outperform.","marker":"[7]"}],"fun_headline_variants":["Adaptive fusion gate beats single pathology models","Gating six PFMs outperforms any one on three benchmarks","Sample-conditioned fusion tops pathology benchmarks","AdaFusion: reweight frozen models, win on three tasks","Pathology model gating lifts accuracy and AUC"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that mean-pooling every tile embedding to a 64-dimensional vector preserves the complementary signal that makes each pathology model useful; if a model's value lives in spatial or global context, that pooling erases it before the gate can reweight it.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive fusion gate beats single pathology models","Gating six PFMs outperforms any one on three benchmarks","Sample-conditioned fusion tops pathology benchmarks","AdaFusion: reweight frozen models, win on three tasks","Pathology model gating lifts accuracy and AUC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1174,"prompt_tokens":690,"completion_tokens":484,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":409}},"tokens_in":434,"tokens_out":484,"duration_ms":5212,"temperature":1.0,"reasoning_tokens":409,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:14:48.853699+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run AdaFusion on a slide-level benchmark where the decisive signal is global tissue architecture, such as whole-slide staging or survival prediction, comparing its mean-pooled fusion against the best single PFM's native whole-slide features; if the individual model consistently wins, then the 64-dim pooling premise fails.","supporting_citations":[{"cited_title":"Nature Medicine30(3), 863–874 (2024) Synergistic Interactions among Pathology FMs via AdaFusion 11","cited_arxiv_id":null,"evidence_quote":"Supplies CONCH, the visual-language pathology foundation model whose high-dimensional embeddings are compressed by AdaFusion."},{"cited_title":"Nature Medicine28(1), 154–163 (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the PANDA prostate cancer grading benchmark whose 10,616 WSIs evaluate the multi-class classification claims."},{"cited_title":"Advances in Neural Information Process- ing Systems37, 53798–53833 (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the HEST benchmark and its spatial gene-expression regression tasks used to evaluate the regression claims."},{"cited_title":"Medical Image Analysis99, 103342 (2025)","cited_arxiv_id":null,"evidence_quote":"Provides the ATEC23 ovarian cancer bevacizumab-response challenge dataset used for binary classification evaluation."},{"cited_title":"bioRxiv pp","cited_arxiv_id":null,"evidence_quote":"Provides the Self-Attn features-fusion baseline that AdaFusion must outperform."}],"review_version":1}