{"id":"a9797a96-0119-4f9c-aa9d-d650358a5ea5","arxiv_id":"2508.13880","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"A concept-activation-vector regularizer improves robustness of medical-image CNNs to spurious correlations and out-of-distribution shifts without concept labels in the main training set.","lead":"This paper introduces LCRReg, a training-time regularizer that steers a CNN's internal representations toward concept directions learned from a small auxiliary dataset. It offers a lightweight, label-light route to making medical image classifiers more robust to spurious correlations and distribution shifts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The robustness claim depends on concept directions from a small auxiliary set transferring to the main and OOD domains; the abstract asserts this without stability or domain-match evidence.","rationale":"The reader's weakest assumption is exactly the transferability and quality of the small auxiliary concept dataset, and that is where my concern lands. I do not identify an internal inconsistency; the issue is that the strongest empirical claim depends on an unshown premise. The proposed check is feasible with the public code and directly tests whether auxiliary-to-main concept transfer is stable and whether the regularization's benefit is actually concept-specific. Since the paper is already UNVERDICTED and I cannot verify the method from the garbled full text, my read does not change that verdict; it sharpens the reason for withholding verification and specifies the experiment that would resolve it.","tokens_in":30607,"tokens_out":5334,"duration_ms":60909,"concrete_test":"Run the released code on the diabetic retinopathy task with CAVs computed from two disjoint 50% splits of the auxiliary concept dataset, keeping the main training and OOD test splits fixed, and compare OOD AUC. Add a control using auxiliary concept labels randomly permuted (or random directions) as the regularizer. If the two split-based runs differ by more than about one AUC point, or if the random-direction control retains the same OOD gain, then concept-direction stability and transfer are not established and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LCRReg improves OOD and spurious-perturbation robustness by encouraging a CNN to activate in concept-defined latent subspaces. For that to be true, the LCRs (e.g., CAVs) extracted from the small auxiliary concept dataset must be stable enough to estimate, must align with the same features in the main training distribution, and must remain aligned under the OOD shifts tested. The abstract states this premise as an assumption ('high-quality, disentangled concept examples') and reports no control showing that the gains are mediated by concept-direction transfer rather than by generic regularization. There is also a risk that the synthetic spurious perturbations and the auxiliary concept examples share the same hand-defined features, which would make the reported transfer favorable by construction. Because the full text is corrupted, I cannot check whether such controls exist in the paper; the abstract alone provides no evidence for the decisive premise. This is an empirical premise, not an internal inconsistency, but it is load-bearing: if the concept directions are unstable or domain-specific, LCRReg is not validated as concept grounding even if the reported numbers are correct.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LCRReg, a regularisation method for convolutional neural networks in medical imaging. The method first extracts Latent Concept Representations (LCRs), such as Concept Activation Vectors (CAVs), from a small auxiliary dataset that has concept labels, and then adds a regularisation term that encourages the main network to activate within latent subspaces associated with these concepts. The authors claim that LCRReg improves robustness to synthetic spurious correlations and to out-of-distribution (OOD) shifts on a controlled toy dataset and on a diabetic retinopathy binary classification task, in comparison with baselines such as multitask learning, linear probing, and post-hoc concept-based models, without requiring concept labels in the main training set. A code repository link is provided in the abstract.","tokens_in":30833,"tokens_out":4086,"duration_ms":44920,"significance":"If the empirical claims hold, LCRReg would be a practically useful addition to the robustness toolbox: it is architecture-agnostic, requires only a small concept-labelled auxiliary set, and is compared against several relevant baselines. The provision of a public code link is a reproducibility strength, and the design gives rise to a falsifiable prediction, namely that concept-aligned regularisation, rather than generic regularisation, is responsible for the reported gains. However, the central evidence is currently inaccessible: the supplied full text is severely corrupted, so no experimental protocol, numeric table, or equation can be checked. The abstract alone reports qualitative claims without confidence intervals or protocol details. The decisive transfer premise of the method is asserted but not demonstrated in the available material.","major_comments":[{"comment":"The full text supplied to the referee is undecodable mojibake; no section, equation, table, or experimental protocol can be checked. The abstract reports qualitative claims (\"significantly improves robustness\", \"enhances performance\") without numeric results, confidence intervals, dataset splits, or baseline specifications. Because the paper's central claim is empirical, the unavailability of the body text is a load-bearing gap: I cannot verify that the experiments support the stated conclusion.","section":"Full text / Abstract"},{"comment":"The method's success depends on the premise that LCRs estimated from a small auxiliary concept dataset remain stable and aligned with the corresponding features in the main training distribution and under the OOD shifts tested. The abstract asserts this premise through the phrase \"high-quality, disentangled concept examples\" but provides no supporting evidence, such as a measure of CAV stability across domains, a comparison with random concept directions, or a control with shuffled concept labels. Without such a control, the reported gains could be attributable to generic regularisation rather than to semantically grounded concept alignment.","section":"Abstract, transfer premise"},{"comment":"The abstract does not report the regularisation weight or the dimensionality of the concept subspaces, and the visible fragments of the manuscript do not reveal how these were selected. If these hyperparameters were tuned on the OOD test sets, the claimed robustness improvement would be weakened by selection on the test distribution. The authors should state the validation-based selection procedure and, preferably, provide a sensitivity analysis over the regularisation weight and subspace dimension.","section":"Hyperparameters / experimental protocol"}],"minor_comments":[{"comment":"The acronym CAV is defined in the abstract, but LCR is not; please define the acronym at first use and state explicitly the form of the regularisation term (e.g., projection loss, cosine-similarity penalty, or CAV-alignment loss) in the main text.","section":"Abstract / Notation"},{"comment":"Please state the size of the auxiliary concept dataset, the number and type of concepts used, and any measures taken to assess domain mismatch between the auxiliary dataset and the main training set; this information is essential for assessing the transfer claim.","section":"Experimental description"},{"comment":"The code repository link should include a version or commit identifier and a license statement; this would make the reported experiments reproducible from the archival record.","section":"Reproducibility"},{"comment":"The corrupted text includes unreadable table and figure fragments; the final PDF must be re-rendered so that tables show units, sample sizes, and confidence intervals or error bars, and so that figure axes and legends are legible.","section":"Figures/Tables"}],"recommendation":"uncertain","confidential_remarks":"The version of the manuscript provided to me is not readable beyond the abstract; I cannot perform a substantive technical review of the experiments. If a clean PDF or a decodable source text is available, the manuscript should ideally be re-reviewed. The archival integrity of the arXiv source may be worth checking, since the current corruption prevents verification of the central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea here is a legitimately useful combination: take Concept Activation Vectors from a small auxiliary dataset and use them as a training-time regularizer so the main model learns to activate along concept subspaces. That is a clean extension of post-hoc concept interpretability into the training loop, and it sidesteps the usual need for dense concept annotations on the main training set. If the reported robustness gains are real, it is a subfield-level tool for medical imaging, not a breakthrough, but a solid one.\n\nWhat I can actually verify is close to nothing, because the full text I received is garbled—the equations, tables, and most of the prose are unreadable. The abstract alone says LCRReg improves robustness on a toy task and on diabetic retinopathy classification under synthetic spurious perturbations and OOD shifts, compared to multitask learning, linear probing, and post-hoc concept models. That is a sensible set of baselines, and the authors include a code link. But the abstract contains no numbers, no confidence intervals, no hyperparameter details, and no ablation showing the gain is mediated by concept-direction transfer rather than generic regularization.\n\nThe stress-test concern is on point and load-bearing: the whole method assumes that concept directions estimated from a small auxiliary set transfer to the main training distribution and remain aligned under OOD shifts. The abstract states this as an assumption, not a demonstrated property. There is also a real risk that the synthetic spurious perturbations and the auxiliary concept examples share hand-defined features, which would make the transfer favorable by construction. Whether the full paper contains controls, domain-mismatch checks, or stability analyses, I cannot tell from what I have.\n\nNone of this is an internal contradiction. The paper is coherent on its own terms as far as the abstract goes, and the framing is honest about requiring only a small auxiliary set. But I cannot in good conscience say the claims are supported based on what I can read.\n\nRecommendation: get a clean PDF and send it to peer review. The idea is worth a serious referee, and the authors have asked the right questions. If the full experiments hold up, this is a useful paper for people working on robustness in medical imaging and concept-based interpretability. My own verdict is unverified, not negative.","headline":"The idea—using CAVs as a training-time regularizer without concept labels—is plausible and worth checking, but the version I have is corrupted, so I can only judge the abstract and the claims remain unverified.","tokens_in":31296,"tokens_out":1635,"would_cite":false,"duration_ms":19475,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A concept-direction regularizer steers medical-image CNNs toward clinically meaningful features, improving accuracy when the input distribution shifts.","keywords":["concept activation vectors","latent concept representations","distribution shift","spurious correlations","diabetic retinopathy","regularization","medical imaging","out-of-distribution generalization"],"falsifier":"Train LCRReg on diabetic retinopathy with a small auxiliary concept set drawn from a different camera or population than the main training set, then measure accuracy on the out-of-distribution test shift; if the regularized model is no better than the unregularized baseline, or worse, concept-direction transfer is the point of failure.","tokens_in":30488,"feed_emoji":"🩺","tokens_out":4956,"duration_ms":52567,"temperature":0.7,"pith_summary":"LCRReg is a regularizer designed to stop medical-image classifiers from succeeding for the wrong reasons. It uses a small auxiliary set of concept examples to locate directions in the network's latent space that correspond to clinically meaningful features, then adds a loss term encouraging the model to activate along those directions. The main training set does not need concept labels. The paper reports that this improves resistance to injected spurious correlations on a controlled toy task and improves diabetic retinopathy classification under synthetic perturbations and out-of-distribution shifts. If the approach works as claimed, it offers a lightweight way to inject clinical knowledge into any CNN without dense annotation.","feed_headline":"Regularizing CNNs toward concept directions improves medical imaging","feed_subtitle":"A small auxiliary set supplies clinically meaningful directions, improving diabetic retinopathy models on shifted and spuriously perturbed…","key_machinery":"The load-bearing mechanism is the LCR: a direction in the network's latent feature space that represents a clinically relevant concept. A small auxiliary dataset provides positive and negative examples of each concept, and a Concept Activation Vector-style linear probe identifies those directions. The training objective adds a regularization term that encourages the model's activations to stay in the subspaces spanned by the concept directions. This is what lets the main model use concepts without ever seeing concept labels in its main training set.","core_discovery":"The paper's central claim is that deep learning models in medical imaging can be made to rely on clinically meaningful features rather than spurious correlations by adding a regularization term computed from Latent Concept Representations (LCRs). LCRReg first extracts concept directions, for example Concept Activation Vectors, from a small auxiliary dataset with concept examples, then trains a CNN on the main dataset with no concept labels while guiding activations to lie in the latent subspaces associated with those concepts. The authors report that this improves robustness against injected spurious correlations on a toy dataset and improves performance on diabetic retinopathy binary classification under synthetic spurious perturbations and out-of-distribution generalization, compared with baselines including multitask learning, linear probing, and post-hoc concept-based models.","pith_inferences":["Inference: The same auxiliary concept set could be reused as an audit probe at deployment time, by projecting test activations onto the LCRs to check whether the model still relies on the intended concepts when the input shifts.","Inference: The method's dependence on one small auxiliary set suggests a practical recipe: a shared concept set per imaging modality could inject domain knowledge into many downstream models without re-annotating every dataset.","Inference: A natural next test is whether the same latent-direction steering transfers to vision transformer architectures, which have comparable latent spaces but were not evaluated here.","Inference: Deliberately omitting a known confounder from the concept set should make LCRReg fail to protect against that confounder, providing a direct way to measure how much concept coverage the method needs."],"forward_implications":["LCRReg should let medical-image models keep higher accuracy under input shifts, such as a new scanner or different population, than the same architecture trained with a plain classification loss.","The main training set does not need dense concept labels; only a small auxiliary set with concept examples is required.","The same architecture can be used unchanged because the regularizer is a single additional term, and the paper shows it works in multi-concept and multiclass settings.","The method gives a way to compare against stronger baselines such as multitask learning, linear probing, and post-hoc concept-based models without adding heavy supervision.","If the concept directions are well chosen, the resulting representations are semantically grounded rather than driven purely by spurious correlations."],"supporting_citations":[],"fun_headline_variants":["Concept-guided regularization boosts OOD performance in medical imaging","LCRReg: steering CNNs to clinical features reduces spurious reliance","Small auxiliary set guides CNNs to clinically meaningful features","Concept directions make CNNs robust to shifts in medical imaging","No concept labels needed for concept-based regularization in imaging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The small auxiliary concept set has to produce directions that still point at the same clinically meaningful features in the main training data and in the shifted test data; if those directions do not transfer, the regularizer steers the model toward features that are irrelevant or misleading.","fun_headline_variants_meta":{"raw":{"variants":["Concept-guided regularization boosts OOD performance in medical imaging","LCRReg: steering CNNs to clinical features reduces spurious reliance","Small auxiliary set guides CNNs to clinically meaningful features","Concept directions make CNNs robust to shifts in medical imaging","No concept labels needed for concept-based regularization in imaging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1305,"prompt_tokens":944,"completion_tokens":361,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":280}},"tokens_in":560,"tokens_out":361,"duration_ms":4194,"temperature":1.0,"reasoning_tokens":280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:14:20.463047+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train LCRReg on diabetic retinopathy with a small auxiliary concept set drawn from a different camera or population than the main training set, then measure accuracy on the out-of-distribution test shift; if the regularized model is no better than the unregularized baseline, or worse, concept-direction transfer is the point of failure.","supporting_citations":[],"review_version":2}