{"id":"095b4d70-b3d3-45b1-9ced-a37815ac2349","arxiv_id":"2606.22002","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Graph coverage selection on frozen embeddings outperforms baselines in balanced accuracy for one-shot data selection on MedMNIST datasets.","lead":"This paper introduces a one-shot graph-based method to select informative medical images for labeling using only embeddings from a frozen pretrained model. By building a k-NN graph and using a coverage kernel, it selects subsets that maximize manifold coverage without iterative training, potentially reducing annotation costs in medical AI.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Central claim depends on generic pretrained embeddings preserving medical-image manifold structure for k-NN graph and coverage kernel to yield useful selections.","rationale":"The reader's weakest_assumption directly names the same embedding-manifold condition that the method's correctness hinges on. Because the paper supplies no independent check of that condition, the concern remains the primary point of uncertainty; the reader's CONDITIONAL verdict therefore stands.","tokens_in":1769,"tokens_out":323,"duration_ms":23658,"concrete_test":"Re-run the full selection pipeline on the five MedMNIST datasets using a medical-domain encoder (e.g., MedCLIP or a fine-tuned ResNet on the same data) in place of the generic foundation model; if the balanced-accuracy advantage over baselines shrinks below statistical significance on the imbalanced ratios, the embedding assumption is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The empirical superiority (highest balanced accuracy on 9/10 conditions) is obtained by running greedy facility location on a two-term kernel derived from a k-NN graph built on frozen foundation-model embeddings. For this to translate into downstream classifier gains, especially on imbalanced MedMNIST subsets, the embedding space must retain the local geometry that the heat-diffusion approximation is meant to cover. No validation of embedding quality (e.g., neighborhood preservation metrics or comparison against domain-specific encoders) is described in the supplied abstract; if the generic embeddings distort or collapse the relevant manifold, the graph construction and resulting selections become arbitrary and the reported gains cannot be attributed to the proposed kernel.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes a one-shot data selection method for medical image classification that constructs a k-NN graph on frozen pretrained foundation-model embeddings, derives a two-term coverage kernel from the heat diffusion kernel to capture direct and two-hop neighborhoods, and applies greedy facility location to select class-balanced subsets maximizing manifold coverage. Evaluated on five MedMNIST datasets (histopathology, radiology, microscopy), the method reports the highest balanced accuracy on nine of ten dataset-ratio conditions versus training-dynamics and geometry-based baselines, with largest gains on imbalanced data, and releases code at the provided GitHub link.","tokens_in":1918,"tokens_out":562,"duration_ms":17311,"significance":"If the central empirical claim holds after addressing the embedding assumption, the work offers a training-free, one-shot alternative to active learning that efficiently handles annotation budgets in medical imaging, particularly for imbalanced classes via global graph structure. Credit is due for the code release (reducing circularity risk), the reduction to sparse operations with only k and one kernel hyperparameter, and the focus on cross-class manifold coverage that per-class methods miss.","major_comments":[{"comment":"Evaluation section: no validation is provided of whether the generic pretrained embeddings preserve local manifold structure on the MedMNIST datasets (e.g., neighborhood preservation metrics such as trustworthiness or continuity, or ablation against domain-specific encoders). This assumption is load-bearing for the central claim, because the k-NN graph and two-term kernel selections are only useful downstream if the embedding geometry aligns with the data manifold; distortion would render the reported balanced-accuracy gains un-attributable to the coverage kernel.","section":"Evaluation / Experimental Setup"},{"comment":"Results (and abstract): the superiority claim (highest balanced accuracy on 9/10 conditions) provides no statistical significance tests, run-to-run variance, or exact baseline implementation details (e.g., hyperparameter choices or training protocols for the dynamics-based comparators). This is load-bearing for verifying the performance edge, especially on imbalanced subsets where small differences could arise from implementation variance rather than the proposed kernel.","section":"Results"}],"minor_comments":[{"comment":"The single kernel hyperparameter is mentioned but its value, selection procedure, and sensitivity analysis are not detailed; adding this (with ablation) would improve reproducibility without altering the core contribution.","section":"Methods"},{"comment":"Notation for the two-term coverage kernel (direct vs. two-hop terms) could be clarified with an explicit equation reference to distinguish it from the full spectral heat kernel.","section":"Method"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. We address the major comments point by point below.","responses":[{"response":"We agree that the manuscript does not include explicit validation of manifold preservation in the embeddings. This is a valid observation. We will revise the evaluation section to add neighborhood preservation metrics (trustworthiness and continuity) on the MedMNIST embeddings and include an ablation against available domain-specific encoders to support the assumption.","revision_made":"yes","referee_comment":"[Evaluation / Experimental Setup] Evaluation section: no validation is provided of whether the generic pretrained embeddings preserve local manifold structure on the MedMNIST datasets (e.g., neighborhood preservation metrics such as trustworthiness or continuity, or ablation against domain-specific encoders). This assumption is load-bearing for the central claim, because the k-NN graph and two-term kernel selections are only useful downstream if the embedding geometry aligns with the data manifold; distortion would render the reported balanced-accuracy gains un-attributable to the coverage kernel."},{"response":"We agree that statistical tests, variance reporting, and detailed baseline protocols are necessary for robust claims. We will revise the results section and supplementary material to report means and standard deviations over multiple random seeds, add statistical significance tests (e.g., paired tests), and provide exact hyperparameter settings and training protocols for all baselines.","revision_made":"yes","referee_comment":"[Results] Results (and abstract): the superiority claim (highest balanced accuracy on 9/10 conditions) provides no statistical significance tests, run-to-run variance, or exact baseline implementation details (e.g., hyperparameter choices or training protocols for the dynamics-based comparators). This is load-bearing for verifying the performance edge, especially on imbalanced subsets where small differences could arise from implementation variance rather than the proposed kernel."}],"tokens_in":1470,"tokens_out":396,"duration_ms":37301,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper introduces a graph-based one-shot selection that constructs a k-NN graph on pretrained embeddings and derives a two-term coverage kernel from the heat diffusion kernel to run greedy facility location. This is a distinct technical step from plain pairwise distance methods, and it aims to cover both direct and two-hop neighborhoods with sparse operations and one hyperparameter.\n\nIt does a few things cleanly. The method stays training-free during selection, which matches the practical goal of cutting annotation cost in medical imaging. It tests on five MedMNIST datasets across modalities and shows the highest balanced accuracy on nine of ten dataset-ratio conditions, with bigger edges on the imbalanced subsets. Code is released, which lets others verify the kernel and the greedy step.\n\nThe main soft spot is the unexamined assumption that generic foundation model embeddings keep the local geometry of these medical datasets intact. Without neighborhood preservation metrics or a comparison to domain-specific encoders, it is hard to know whether the graph and kernel are actually doing the work or just riding whatever structure the embeddings happen to have. The abstract also gives no statistical significance numbers or sensitivity analysis for the kernel hyperparameter, so the performance edge is harder to trust at face value.\n\nThis is for people who need a simple, training-free baseline for data selection in medical classification. A reader working on annotation efficiency would find the kernel construction and the imbalance results useful to check. It deserves a serious referee because the method is well-defined, the code is public, and the empirical pattern on multiple datasets is worth closer inspection even if the embedding step needs more support.","headline":"The paper's new piece is the two-term coverage kernel from heat diffusion on k-NN graphs of frozen embeddings for one-shot facility location, with reported gains on MedMNIST but no checks on whether those embeddings preserve the needed manifold.","tokens_in":2417,"tokens_out":409,"would_cite":false,"duration_ms":18344,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A two-term coverage kernel on k-NN graphs from frozen embeddings selects subsets that maximize manifold coverage and raise balanced accuracy on medical image tasks without any training during selection.","keywords":["data selection","medical image classification","one-shot selection","k-nearest neighbor graph","heat diffusion kernel","facility location","class imbalance","balanced accuracy"],"falsifier":"On a held-out medical imaging dataset, the subsets chosen by the graph-coverage method produce classifiers whose balanced accuracy falls below that of random selection or the strongest baseline under the same annotation budget.","tokens_in":2669,"feed_emoji":"📊","tokens_out":688,"duration_ms":15198,"temperature":0.7,"pith_summary":"The paper aims to show that one-shot data selection for medical image classification can be performed by building a global k-nearest neighbor graph on pretrained embeddings and then applying a sparse two-term kernel derived from heat diffusion to guide greedy facility-location selection. This approach is meant to capture both direct and two-hop neighborhood relations across the entire dataset, including cross-class structure that per-class methods overlook. A reader would care because expert annotation of medical images is costly, so a method that picks high-value subsets in a single pass without repeated model training could cut labeling budgets while preserving classifier performance, especially on imbalanced data.","feed_headline":"Graph kernel on embeddings picks best medical image subsets","feed_subtitle":"Two-term coverage method beats baselines on nine of ten tests without training during selection, largest gains on imbalanced data.","key_machinery":"The two-term coverage kernel on the k-NN graph, which encodes direct and two-hop neighborhood relationships for facility-location selection.","core_discovery":"Given embeddings from a pretrained encoder, the method constructs a k-nearest neighbor graph over all training samples, derives a two-term coverage kernel from the heat diffusion kernel that approximates full spectral behavior through sparse matrix operations, and uses greedy facility location on this kernel to choose class-balanced subsets that maximize coverage of the data manifold; the resulting subsets produce the highest balanced accuracy on nine of ten dataset-ratio conditions across five MedMNIST datasets, with the largest gains on class-imbalanced collections.","pith_inferences":["If the same kernel construction were applied to embeddings from a domain-specific fine-tuned model, the coverage quality might increase further.","The approach could serve as an initialization step for subsequent active-learning rounds rather than a standalone one-shot solution.","Failure on a new dataset would most likely trace back to the pretrained embeddings failing to reflect the target manifold rather than to the kernel or selection algorithm."],"forward_implications":["The method requires no model training or oracle access during the selection step itself.","Global graph construction yields larger gains on class-imbalanced datasets than methods that operate per class.","The two-term kernel reproduces the selection behavior of the full heat kernel while remaining computationally cheap.","Performance is measured on five MedMNIST collections covering histopathology, radiology, and microscopy."],"fun_headline_variants":["KNN graph kernel selects medical image subsets one-shot","Two-term heat kernel selects balanced medical data subsets","Graph coverage kernel enables one-shot medical manifold selection","Graph method achieves highest accuracy on nine of ten medical tests"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Embeddings from a generic pretrained foundation model preserve enough local manifold structure of the specific medical datasets that selections based on the k-NN graph and two-term kernel will improve downstream classifier performance.","fun_headline_variants_meta":{"raw":{"variants":["KNN graph kernel selects medical image subsets one-shot","Two-term heat kernel selects balanced medical data subsets","Graph coverage kernel enables one-shot medical manifold selection","Graph method achieves highest accuracy on nine of ten medical tests"]},"model":"grok-4.3","cost_usd":0.007241,"raw_usage":{"total_tokens":3363,"prompt_tokens":718,"num_sources_used":0,"completion_tokens":60,"cost_in_usd_ticks":72412000,"prompt_tokens_details":{"text_tokens":718,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2585,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":718,"tokens_out":60,"duration_ms":13402,"temperature":1.0,"reasoning_tokens":2585,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T12:47:31.551849+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On a held-out medical imaging dataset, the subsets chosen by the graph-coverage method produce classifiers whose balanced accuracy falls below that of random selection or the strongest baseline under the same annotation budget.","supporting_citations":[],"review_version":1}