{"id":"c7b4146c-41ab-43f6-9288-5b1f5dc0438f","arxiv_id":"2508.02431","paper_version":3,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors propose an Asymmetric Transformer Decoder that uses tissue type and patch embeddings to predict six actionable lung cancer mutations from H&E slides, but the uploaded full text is a different document.","lead":"This paper claims a new machine-learning model can spot six cancer-driving gene mutations in lung tumor images, outperforming existing methods by 3% on average. The supplied manuscript text, however, is an unrelated appendix about measuring innovation, so the results could not be checked.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central 3% performance claim is unverifiable: the submitted full text is an unrelated appendix and the cited repository points to a different project, so no model, data, or evaluation is present to check.","rationale":"The reader's verdict of UNVERDICTED is appropriate because the central numerical claim cannot be checked. My stress-test finds the same core problem, but I locate it even more broadly than the reader did: not only is tissue-type handling unverifiable, the entire method and evaluation are absent from the supplied full text, and the companion repository belongs to a different research topic. This is not an accusation of misconduct; it is a straightforward statement that the evidence required to evaluate the strongest claim is missing. I considered REJECT, but reject would require a positive basis to disbelieve the 3% figure, which we do not have. The honest status is unverified. The reader's weakest_assumption about tissue annotations captures one possible deployment risk, but the more load-bearing concern is the absence of any methods or results section at all. A single concrete check, retrieving the arXiv source and locating the experimental description, would settle whether this is a formatting artifact or a genuine gap. Until that check is done, no score can be assigned, so the verdict remains UNVERDICTED.","tokens_in":819,"tokens_out":2043,"duration_ms":26845,"concrete_test":"Download the actual arXiv source for 2508.02431 from arXiv.org and compare it with the text supplied to reviewers; then inspect the Methods and Experiments sections to confirm presence of the Asymmetric Transformer Decoder, tissue-type labeling, dataset splits, and baseline tables. If those sections are absent, obtain the correct project repository and run the released inference code on the same mutation-prediction task; if no lung-cancer code exists, the headline improvement is unsupported and the verdict should remain UNVERDICTED.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is the reported 3–4% average AUROC improvement over top MIL models for six actionable NSCLC mutations. To hold, the manuscript must contain a specification of the Asymmetric Transformer Decoder, the tissue-type input construction, the dataset split, and the baseline comparisons. The supplied full text, however, is an online appendix on 'AI-Based Measurement of Innovation' with a companion repository (github.com/robi979/AI-Innovation-Measurement) unrelated to lung cancer or computational pathology. None of the components needed to verify the headline result are present. The reader's weakest assumption about noisy or unavailable tissue annotations is therefore secondary: even the existence of a tissue-type channel that can be added to the model cannot be confirmed from the submission. This is a missing-evidence problem rather than an internal inconsistency, so it does not by itself falsify the 3% advantage, but it prevents any reliability assessment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript under review is arXiv:2508.02431 (eess.IV), titled \"Identifying actionable driver mutations in lung cancer using an efficient Asymmetric Transformer Decoder.\" The abstract claims that a new Multiple Instance Learning (MIL) architecture, the Asymmetric Transformer Decoder, combined with a method to directly use tissue type, outperforms top MIL models by an average of 3% and by over 4% for rare mutations (ERBB2, BRAF) across six actionable NSCLC driver mutations (ALK, BRAF, EGFR, ERBB2, KRAS, MET ex14). However, the supplied full text is not the paper described in the abstract. It is an online appendix for an unrelated paper titled \"AI-Based Measurement of Innovation: Mapping Expert Insight into Large Language Model Applications,\" containing a single figure (Figure C1) on training data distribution and a companion repository link (github.com/robi979/AI-Innovation-Measurement) that points to a different project. No model specification, tissue-type construction, dataset description, experimental protocol, cross-validation scheme, baseline identities, or result tables are present in the submission.","tokens_in":987,"tokens_out":2645,"duration_ms":30784,"significance":"If the claimed results were substantiated, the paper would be a clinically relevant contribution to computational pathology, particularly because it addresses multiple actionable mutations and includes rare ones. The proposed tissue-type integration would also be a methodological novelty that could improve MIL interpretability and performance. However, the submitted manuscript provides no evidence whatsoever for these claims. The absence of the actual paper text makes any assessment of scientific significance impossible: there is no architecture to evaluate, no dataset to scrutinize, and no experimental design to check for bias or circularity. The central claim therefore remains entirely unverified.","major_comments":[{"comment":"The body of the submission is an unrelated online appendix. The title of the supplied full text is \"AI-Based Measurement of Innovation: Mapping Expert Insight into Large Language Model Applications,\" and its companion repository (https://github.com/robi979/AI-Innovation-Measurement) is not about lung cancer, MIL, or computational pathology. None of the components needed to verify the headline claim are present: the Asymmetric Transformer Decoder architecture, the tissue-type input method, the dataset description, the cross-validation scheme, the baseline MIL models, and the per-mutation AUROC results for ALK, BRAF, EGFR, ERBB2, KRAS, and MET ex14 are all absent. This is a load-bearing omission that prevents any assessment of the paper's central claim and cannot be fixed by a local revision.","section":"Full Text"},{"comment":"Even taken in isolation, the abstract's quantitative claim is not backed by any experimental data in the submission. The abstract states an average 3% improvement and over 4% for ERBB2 and BRAF, but provides no confidence intervals, number of patients or whole-slide images, evaluation metric definition, baseline identities, or statistical significance. The only figure in the supplied text (Figure C1) pertains to TF-IDF and GloVe embeddings for ML classifiers in an innovation-measurement context, which is unrelated to mutation prediction. Consequently, the performance claim as stated cannot be checked or reproduced.","section":"Abstract"}],"minor_comments":[{"comment":"The manuscript's title and abstract are incompatible with its body; the authors should verify that the correct file was uploaded. If the submission is a placeholder or submission error, it should be corrected before any further review is attempted.","section":"General"}],"recommendation":"reject","confidential_remarks":"This submission appears to be an erroneous upload: the full text is an appendix from an unrelated paper. There is no technical content to review, and the central claim is not supported by any accessible evidence. I recommend rejection in the current form. If the authors later submit the correct manuscript with full methodological and experimental details, it could be considered de novo."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: you cannot review this paper, because there is no paper. The supplied full text is an online appendix about measuring innovation with LLMs, and the companion repository points to a different project entirely. The abstract is all that exists, and abstracts aren't enough to verify a load-bearing performance claim.\n\nWhat is genuinely new and worth acknowledging: the clinical target. Detecting six actionable NSCLC mutations from slides, including rare ones like ERBB2 and BRAF, addresses a real gap — most prior work stops at one or two mutations. The Asymmetric Transformer Decoder idea, using low-dimensional queries with higher-dimensional keys/values, is a plausible architectural variation, and feeding tissue type into a MIL model is a reasonable attempt to add biological signal. On the merits, the abstract describes a sensible study.\n\nBut here is the soft spot, and it is not minor: nothing in the manuscript supports the headline. There are no methods, no dataset description, no split protocol, no baseline names, no confidence intervals, no ablation, no error analysis. The repository cited in the appendix has no connection to lung cancer or pathology. I cannot even confirm that the tissue-type channel exists, much less that it doesn't leak genetic-testing outcome. The stress-test note is right: this is a missing-evidence problem, not an internal contradiction. The results could be perfectly fine — but they are not auditable.\n\nMy recommendation: desk reject this version. Do not send it to referees, because they would have nothing to evaluate. If the authors have a real manuscript, let them resubmit it. The abstract alone suggests a study that might deserve peer review once the actual methods and data are on the table, but as submitted, this is not a paper.","headline":"The abstract promises a useful clinical tool, but the submission contains no paper—only an unrelated appendix—so the 3–4% claim is uncheckable and the work isn't ready for review.","tokens_in":1458,"tokens_out":1837,"would_cite":false,"duration_ms":24532,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An Asymmetric Transformer Decoder predicts six actionable lung cancer driver mutations from tissue slides, beating top MIL models by 3% on average and over 4% on rare mutations.","keywords":["non-small cell lung cancer","driver mutations","multiple instance learning","asymmetric transformer decoder","whole-slide images","computational pathology","weakly supervised learning"],"falsifier":"Train the Asymmetric Transformer Decoder on the same six-mutation task with and without tissue-type input, and evaluate both versions on an independent multi-site cohort; if the tissue-type-free version matches or beats the full model, or if the 3% average and over-4% rare-mutation margins do not reproduce, the paper's central claim is refuted.","tokens_in":677,"feed_emoji":"🧬","tokens_out":8484,"duration_ms":89261,"temperature":0.7,"pith_summary":"This paper aims to establish that a machine-learning model can read routine pathology slides and detect six actionable driver mutations in non-small cell lung cancer at once, rather than focusing on only one or two common alterations. The authors introduce an Asymmetric Transformer Decoder for multiple-instance learning (MIL), where query vectors stay low-dimensional while key and value vectors are larger, an arrangement they argue extracts information from many small image patches without overfitting. They also feed tissue-type information directly into the model, addressing a typical MIL limitation of treating all regions alike and neglecting biological relevance. The reported result is that the model outperforms top MIL models by an average of 3%, and by over 4% on rare mutations such as ERBB2 and BRAF, moving ML-based slide tests closer to practical complements to standard genetic testing.","feed_headline":"Transformer reads tissue slides for six lung cancer mutations","feed_subtitle":"Asymmetric decoder beats standard slide-analysis models by 3% on average and 4% on rare mutations.","key_machinery":"The load-bearing object is the Asymmetric Transformer Decoder, a transformer decoder in which query vectors and key/value vectors have different dimensions: the query side stays narrow while keys and values are wider. This lets the decoder process the very large number of patch embeddings in a whole-slide image at low query dimensionality, which the paper argues reduces overfitting and fits the multiple-instance learning setting naturally. The second piece of machinery is tissue-type conditioning: each patch is associated with a tissue-type label and that information is injected into the model, countering the MIL tendency to analyze all regions equally and neglecting biological relevance. The decoder's job is to turn patch embeddings into a slide-level prediction for each of the six mutations tested.","core_discovery":"The central claim is that an asymmetric design for a transformer decoder makes multiple-instance learning work well for a six-way lung cancer mutation-prediction task on whole-slide images. In the model, queries use fewer dimensions than keys and values, so the decoder can interrogate a large bank of patch embeddings while keeping parameter count low and limiting overfitting. The paper further claims that supplying the model with tissue-type labels for patches gives it biological context that standard MIL approaches lack, and that these choices together improve average performance over top MIL models by 3% and rare-mutation performance by more than 4% for ERBB2 and BRAF. The intended consequence is a practical slide-based screening tool for actionable mutations in non-small cell lung cancer.","pith_inferences":["If tissue-type annotations are what carry the improvement, the method's deployability depends on those annotations being available and accurate at inference time; a clean ablation withholding tissue-type input would settle how much of the 3% margin is architectural versus biological context.","The asymmetric query/key-value split is not obviously pathology-specific, so the same decoder shape could be tried in other weakly supervised tasks with enormous instance sets, such as satellite imagery or video-level prediction.","Because the rare-mutation gains are the largest, one testable implication is that jointly training on all six mutations shares morphology features that help rare classes; comparing the six-way model against six separately trained one-vs-rest models would isolate that multi-task benefit."],"forward_implications":["If the results hold, slide-based screening for the full six-mutation panel becomes a realistic triage step before genetic testing, widening the pool of NSCLC patients who get mutation information.","The tissue-type input offers a general fix for a known MIL weakness, so it could be reused for other whole-slide biomarker prediction tasks beyond these six mutations.","The low-dimensional query design means the decoder can handle very large slide representations without a proportional increase in parameters, which should help it scale to bigger cohorts.","The larger gains on rare mutations (ERBB2 and BRAF) suggest the model is not merely re-learning common mutation patterns, but is capturing morphology that matters exactly where existing tests are weakest."],"supporting_citations":[],"fun_headline_variants":["Asymmetric transformer screens six lung cancer mutations","Tissue-aware model beats MIL on rare mutations","Efficient decoder improves lung cancer mutation prediction","Slide AI uses tissue context to spot six mutations","Asymmetric decoder wins on rare lung cancer mutations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that reliable tissue-type labels can be assigned to every patch and that these labels do not leak the genetic-testing outcome; if tissue annotations are noisy or unavailable at deployment, the reported 3% and 4% advantages may disappear.","fun_headline_variants_meta":{"raw":{"variants":["Asymmetric transformer screens six lung cancer mutations","Tissue-aware model beats MIL on rare mutations","Efficient decoder improves lung cancer mutation prediction","Slide AI uses tissue context to spot six mutations","Asymmetric decoder wins on rare lung cancer mutations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1595,"prompt_tokens":926,"completion_tokens":669,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":600}},"tokens_in":542,"tokens_out":669,"duration_ms":8061,"temperature":1.0,"reasoning_tokens":600,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:56:28.414751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the Asymmetric Transformer Decoder on the same six-mutation task with and without tissue-type input, and evaluate both versions on an independent multi-site cohort; if the tissue-type-free version matches or beats the full model, or if the 3% average and over-4% rare-mutation margins do not reproduce, the paper's central claim is refuted.","supporting_citations":[],"review_version":1}