{"id":"8871f598-21a8-4ae2-b029-7cd7441b3d9d","arxiv_id":"2506.05286","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A vision transformer with a concept bottleneck and denoised diffusion smoothing is claimed to give stable concept explanations under input perturbations while keeping diagnostic accuracy.","lead":"This paper adds a concept-bottleneck layer to a vision transformer and then wraps it in a denoising-diffusion smoothing step, aiming to keep the model's explanation stable when the input image is perturbed. The authors claim this preserves accuracy while making the concept explanations more trustworthy for medical diagnosis.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's proof assumes fc(X) is a probability vector, but Eq. (1) defines fc as an unnormalized linear projection; D_α is undefined for negative components.","rationale":"The paper's central claim is that DDS converts VCT into an SVCT with provable top-k concept stability. This rests entirely on Theorem 2 and its proof in Appendix C. The proof chain is: Rényi divergence between smoothed inputs is bounded (via post-processing) by αR²/(2σ²); then Lemma 2 lower-bounds the Rényi divergence required to keep top-k overlap at least β; combining gives the σ² condition. Lemma 2, however, is stated only for vectors with unit ℓ1 norm (probability vectors), and the divergence D_α(w~, q) in its statement is the divergence of categorical distributions with probability vectors w~ and q. The actual w~ = fc(X^) from Eq. (1) and Algorithm 1 is a linear projection Wc f(X) trained for interpretability via CLIP-Dissect; there is no normalization step anywhere in the pipeline, and the projection can produce negative values because ViT features and Wc are real-valued. Thus the object D_α(w~, q) is not defined for the implemented w~, and the minimization in Lemma 2 does not correspond to anything computed by the method. The theorem therefore does not prove what it claims.\n\nI considered whether another gap—e.g., the unspecified denoiser or the mismatch between the theory's top-k overlap and the CFS/CPCS metrics—is more load-bearing. The denoiser gap is real but secondary: post-processing lets any deterministic T satisfy the inequality, though an imperfect denoiser would change the actual distribution; the paper gives no guarantee that T is accurate. The evaluation gap means the paper does not empirically test the quantity V_k it promises. But the normalization gap is the most load-bearing: even if the denoiser were perfect and the evaluation measured V_k, the proof of the stability guarantee would still be invalid because D_α is not defined for unnormalized, possibly negative concept vectors. I agree with the reader's assessment.\n\nThe paper's own limitations (Appendix J) acknowledge only Gaussian noise was tested and assert, without a proof, that 'our theory proved that Gaussian noise is near-optimal and gave the worst-case of perturbations.' This further supports the view that the theoretical apparatus is being over-claimed. The empirical trend—SVCT improving CFS/CPCS over baselines on four datasets—is plausible and worth reporting, but it does not rescue the theory.\n\nA single concrete check would settle the matter: verify whether any normalization is applied to fc before the divergence computations in the proof and the implementation. If none is, the theorem's premise fails, and the paper's strongest claim cannot stand as written.","tokens_in":24315,"tokens_out":7950,"duration_ms":75837,"concrete_test":"Check the proof's premise on a trained VCT: for test images from OCT2017, compute fc(X) from Eq. (1) before and after PGD perturbation (ρ_u = 8/255). Record the minimum component and the ℓ1 norm of fc(X); if any component is negative or the norm differs from 1, the probability-vector precondition of Lemma 2 fails. Then independently re-derive equations (4)-(5) of Lemma 2 using this unnormalized fc, computing Dα(fc(X), q) as in Appendix C; if the log term is undefined for any sample, Theorem 2's bound cannot be instantiated. Finally, if the implementation is amended to normalize fc (e.g., softmax), check whether the top-k indices of the normalized vector match those of the raw fc; if they differ, the theorem and the implemented concept vector are different objects.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central formal claim is Theorem 2: w~(X) = fc(T(X+S)) is an (R, Dα, γ, β, k, ∥·∥2)-stable concept module, with top-k indices stable under perturbations. The proof in Appendix C hinges on Lemma 2, which minimizes the Rényi divergence Dα(w~, q) over 'the set of all vectors with unit ℓ1-norm in R^T' — i.e., probability vectors. But Algorithm 1 returns w~ = fc(X^) with fc(X) = Wc f(X) (Eq. 1), a linear projection of ViT features trained by CLIP-Dissect for interpretability. The paper applies no normalization, and fc can take negative values and arbitrary ℓ1 norm. For such w~, Dα(w~, q) as used in Lemma 2 is either undefined (the log of a nonpositive argument for α > 1) or is not the categorical-divergence quantity being minimized. Consequently, the bound in Theorem 2 does not follow for the implemented method. The evaluation in Table 2 also measures CFS and CPCS, not the top-k overlap Vk that the theory promises; this is a separate gap, but the normalization issue alone invalidates the proof of the main claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Vision Concept Transformer (VCT), which adds a label-free concept bottleneck layer to a ViT and fuses concept features with backbone features for classification, and Stable VCT (SVCT), which applies Denoised Diffusion Smoothing to stabilize concept explanations under input perturbations. The authors claim a theoretical guarantee (Theorem 2) that SVCT's concept vector is stable in top-k indices under perturbations, and they present experiments on four medical datasets measuring accuracy, concept faithfulness (CFS), and concept perturbation cosine similarity (CPCS), along with ablations and concept-intervention examples.","tokens_in":24541,"tokens_out":10100,"duration_ms":102979,"significance":"If the theoretical guarantee were valid, the paper would make a useful contribution to interpretable medical image classification by combining concept bottleneck models with ViT feature fusion and diffusion-based smoothing. The empirical evaluation has strengths: four medical datasets, comparisons to standard and concept-based baselines, ablations of the DDS components, and details of concept generation and computational cost. However, the central formal claim is undermined by a mismatch between the theorem's assumptions (probability vectors) and the implemented unnormalized projection, and the experiments measure cosine/Euclidean stability rather than the top-k overlap ratio the theorem promises. The empirical results may still be of interest, but the paper does not currently deliver a valid stability certificate.","major_comments":[{"comment":"The proof of the top-k stability portion applies Lemma 2, which minimizes Rényi divergence over 'the set of all vectors with unit ℓ1-norm in R^T' (probability vectors), to the concept feature w~ = fc(X). However, Eq. (1) defines fc(X) = Wc f(X) as an unnormalized linear projection of ViT features, and Algorithm 1 returns this vector without any normalization; the components can be negative and need not sum to one. For such vectors, the Rényi divergence Dα used in Lemma 2 is undefined when α > 1 (the logarithm of a nonpositive argument), and the minimization in Lemma 2 does not apply. Consequently, the claimed bound on the top-k overlap ratio V_k does not follow for the implemented method.","section":"Section 3.2, Theorem 2 and Appendix C, Lemma 2"},{"comment":"The statement defines w~(X) = fc(T(X+S)) with S a random Gaussian variable, but Definition 2 defines a stable concept module as a deterministic function g. The theorem does not state whether the stability condition must hold for every realization of S or in probability, and the proof bounds the Rényi divergence between the distributions of w~(X) and w~(X′), not between deterministic outputs. In addition, the theorem's condition on σ² depends on the concept values w~_i through the sum over the index set S, making the condition data-dependent and not a verifiable a priori guarantee. Both issues need to be resolved for the theorem to be a meaningful stability certificate.","section":"Section 3.2, Theorem 2"},{"comment":"The empirical stability evaluation reports CFS (relative Euclidean distance between concept weight vectors) and CPCS (cosine similarity), but Definition 2 and Theorem 2 are about the top-k overlap ratio V_k(g(X′), g(X)). The reported metrics do not measure top-k index stability; high cosine similarity can coexist with different top-k sets, and large Euclidean changes can leave top-k sets unchanged. The paper should report V_k for the same perturbations and, ideally, check the radius condition of Theorem 2.","section":"Section 4.3, Table 2"},{"comment":"The experiments evaluate stability under PGD adversarial perturbations with ℓ∞ radius ρu, while the theory in Section 3.2 is developed for additive Gaussian noise S ~ N(0, σ²I) and an ℓ2 perturbation radius R; no mapping between ρu and (R, σ) is given. Moreover, the Limitations section states that the method was 'only tested in the case of Gaussian noise,' which is contradicted by the PGD-based experiments. These inconsistencies prevent the experimental section from testing the theoretical guarantee.","section":"Section 4.1 and Section J (Limitations)"}],"minor_comments":[{"comment":"The definition of T_k(x) mixes index sets: for x ∈ R^n it writes 'i ∈ [d]' and 'j ∈ [n]', which should be a single consistent index set of the same dimension.","section":"Definition 1"},{"comment":"Algorithm 1 takes an input standard deviation σ, but the experimental setup in Section 4.1 and Table 7 refers to a parameter S=8/255 without clarifying whether S is σ or the noise standard deviation; please use consistent notation.","section":"Algorithm 1 and Table 7"},{"comment":"The text states that 'All results are the average score running 10 times to reduce variance,' but Table 7 lists trial_num=5; please reconcile these numbers.","section":"Section 4.1"},{"comment":"The phrase 'while remaining interpretability' is grammatically incorrect and should be 'while remaining interpretable'.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The reference list contains a large share of self-citations from the same research group (e.g., refs. [13,15-23,33,52,59]); the editor may wish to verify that the novelty claims are adequately distinguished from those prior works. The central theoretical flaw (normalization of concept vectors) is fixable, but the revision will require rethinking the theorem and adding experiments that measure top-k overlap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: the paper's headline claim — that Denoised Diffusion Smoothing yields a provably stable concept module — is not supported as written. The proof of Theorem 2 treats the concept vector as a probability distribution, but the implementation just returns a linear projection of ViT features that is neither normalized nor nonnegative. That is the load-bearing step, and it fails.\n\nWhat it does well: the formal definition of an (R, D, γ, β, k, ∥·∥)-stable concept module (Definition 2) is a reasonable way to frame explanation stability, and the DDS idea — smooth the input, then denoise before extracting concepts — is genuinely worth testing. VCT (concatenating concept features with backbone features) is a routine extension of label-free CBM, but it is a sensible one, and the experiments on four medical datasets show it closes most of the accuracy gap between a pure CBM and the standard ViT. The ablation study separating denoising from smoothing is useful and shows both contribute.\n\nSoft spots, in order of severity. First, the theorem. Lemma 2 minimizes Rényi divergence over the simplex of unit-ℓ1 vectors, but Algorithm 1 outputs fc(X^) = Wc f(X^) with no normalization; components can be negative and the ℓ1 norm is arbitrary. For such vectors the Rényi divergence D_α is not a well-defined categorical divergence, so the bound in Theorem 2 does not follow for the implemented method. This is not a cosmetic gap; it is the central formal contribution. Second, the evaluation does not measure the promised top-k overlap ratio V_k. The reported CFS and CPCS metrics capture cosine similarity and relative Euclidean distance between concept weight vectors, which are related but not equivalent to the top-k index stability the theorem claims. The tables may still indicate a real empirical effect, but they do not test the theorem's conclusion. Third, the denoising diffusion model is unspecified — no architecture, training data, or step count — and no code is released, so nobody can reproduce the DDS part. Fourth, the limitation section claims 'our theory proved that Gaussian noise is near-optimal,' but I did not find that result anywhere in the paper; that overstatement should be removed or supported.\n\nThe empirical trend is plausible and the framework is coherent. But the paper currently asks the reader to accept a guarantee that the math does not deliver, and the evaluation is misaligned with that guarantee.\n\nWho this is for: people working on concept bottleneck models and explanation robustness will find the definition and the empirical comparison worth reading, mostly as a benchmark of what still needs fixing. It deserves referee time — the flaw is fixable (normalize fc, or prove a version for top-k of signed scores without the divergence machinery), and the problem is real. I would send it out, but the referee should insist on a corrected proof, a specified denoiser, and a direct measurement of top-k overlap before the claims are accepted.","headline":"The stability guarantee is a proof artifact: the theorem assumes a probability vector while the method outputs an unnormalized linear projection, and the experiments measure cosine similarity, not top-k overlap.","tokens_in":25101,"tokens_out":3300,"would_cite":false,"duration_ms":38055,"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":"This paper claims that adding denoised diffusion smoothing to a concept-aware vision transformer keeps the top-k concept explanations and predictions stable under input noise, while preserving classification accuracy on medical images.","keywords":["Explainable AI","Concept Bottleneck Models","Vision Transformer","Denoised Diffusion Smoothing","Stable explanations","Medical image classification","Rényi divergence","Randomized smoothing"],"falsifier":"On any image from the four medical datasets, compute fc(X)=Wc f(X) and check whether all entries are nonnegative and sum to 1; if they are not, the probability-vector assumption in Lemma 2 is violated and the certified stability bound of Theorem 2 does not apply. A direct experiment would be to run the SVCT pipeline with and without ℓ1-normalization of fc and compare the measured top-k overlap at the same perturbation radius and noise level.","tokens_in":24074,"feed_emoji":"🩺","tokens_out":3974,"duration_ms":49792,"temperature":0.7,"pith_summary":"The paper sets out to make concept-based explainable AI usable for medical diagnosis, where both accuracy and trust matter. It introduces the Vision Concept Transformer (VCT), which fuses a ViT backbone's image features with human-understandable concept features from a label-free concept bottleneck, and then the Stable Vision Concept Transformer (SVCT), which wraps VCT in a denoised diffusion smoothing step. The central claim is that with a sufficiently large smoothing noise level, SVCT satisfies a formal notion of stable explanation: the top-k important concepts stay nearly unchanged under small input perturbations, and the prediction distribution stays close in Rényi divergence. If true, this would give clinicians a model whose concept-level reasoning does not silently shift when an image is noisy or slightly perturbed, addressing a known weakness of prior concept bottleneck models.","feed_headline":"Concept explanations stay stable under medical-image noise","feed_subtitle":"SVCT wraps a vision transformer in denoised diffusion smoothing to keep top-k concepts and predictions faithful under perturbation.","key_machinery":"The load-bearing object is the stable concept module g(X)=fc(T(X+S)), where fc is the concept projection learned by CLIP-Dissect in the label-free CBM, T is a denoised diffusion probabilistic model, and S is Gaussian noise. The mechanism that carries the argument is the Rényi-divergence bound between two Gaussians with the same variance but shifted means, combined with the post-processing property of Rényi divergence: this gives D_α(g(X),g(X')) ≤ αR²/(2σ²) for any two inputs within radius R. The top-k overlap half of the proof rests on a lemma that computes the minimum Rényi divergence between two unit-ℓ1 probability vectors whose top-k sets overlap by at least a fraction β, which is then inverted to express the required noise level σ² in terms of k, α, and the concept weights.","core_discovery":"On its own terms, the paper discovers and proves a stability guarantee for an interpretable medical classifier. It defines a stable concept module as a function g whose top-k concept overlap between an input and any perturbed input within radius R is at least β, and whose prediction distributions are within γ in Rényi divergence. It then constructs such a module by taking the VCT's concept feature fc(X)=Wc f(X) and evaluating it on the denoised output of a diffusion model applied to a Gaussian-smoothed input, i.e. g(X)=fc(T(X+S)) with S~N(0,σ²I). Theorem 2 states a lower bound on σ² that, when satisfied, makes g a (R, D_α, γ, β, k, ℓ2)-stable concept module; the proof chains the post-processing inequality for Rényi divergence, a Gaussian divergence bound, and a new lemma characterizing the minimum Rényi divergence between two probability vectors that have at least β top-k overlap.","pith_inferences":["A natural implementation step the authors leave implicit is to ℓ1-normalize fc(X) before applying the stability theorem; doing so would make the probability-vector assumption explicit and testable.","The same diffusion-smoothing wrapper could be combined with other interpretable architectures, such as attention-based explainers, to transfer the top-k stability definition beyond concept bottlenecks.","The empirical evaluation only covers Gaussian noise; a matched experiment with non-Gaussian perturbations (e.g., uniform or adversarial-patch noise) would clarify how far the certified radius transfers to real clinical degradation.","Because Theorem 2's bound depends on the largest and second-largest concept weights, an interesting corollary is that nearly flat concept vectors require very large smoothing noise to stabilize, suggesting a testable trade-off between concept selectivity and explanation stability."],"forward_implications":["If the stability certificate holds, concept-bottleneck-style models can be deployed in noisy clinical settings while preserving human-readable reasoning.","The noise-level bound provides a concrete, tunable mechanism: raising σ² trades a bit of accuracy for a formally guaranteed explanation-stability radius.","Because the concept layer is label-free, the same recipe can be applied to any ViT backbone without gathering concept annotations, making faithful explanations scalable to new medical tasks.","Test-time concept intervention, where a clinician corrects a wrongly predicted concept, would remain reliable under input perturbation, supporting human-machine co-diagnosis.","The theoretical template extends beyond medical images to any high-stakes domain where both prediction and explanation must be certified against small input changes."],"supporting_citations":[{"why":"Supplies the label-free CBM transformation that produces the concept feature fc(X) without human-annotated concept labels.","marker":"[43]"},{"why":"Provides the CLIP-Dissect procedure used to learn the projection weights Wc that map backbone features to concept space.","marker":"[44]"},{"why":"Supplies the denoised diffusion probabilistic model that defines the denoising operator T used in the stable concept module.","marker":"[14]"},{"why":"Gives the Gaussian Rényi-divergence bound and the post-processing inequality that power the proof of Theorem 2.","marker":"[34]"},{"why":"Motivates the top-k overlap minimization technique that underlies Lemma 2's Rényi-divergence lower bound.","marker":"[36]"},{"why":"Provides the stability-of-interpretation perspective and the top-k overlap argument that Lemma 2 adapts to concept vectors.","marker":"[20]"},{"why":"Establishes Theorem 1, linking Rényi-divergence robustness of the prediction distribution to unchanged argmax predictions.","marker":"[59]"},{"why":"Defines the ViT backbone architecture on which the VCT and SVCT frameworks are built.","marker":"[5]"}],"fun_headline_variants":["Diffusion smoothing makes medical AI concepts noise-robust","Stable concept explanations survive medical image perturbation","Vision Concept Transformer with diffusion-smoothing stability","Guaranteed stable concepts for interpretable medical AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stability proof assumes the concept feature vector fc(X) is a probability vector—nonnegative and summing to one—but the method never normalizes this linear projection, so the claimed top-k overlap bound may not follow from the stated conditions.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion smoothing makes medical AI concepts noise-robust","Stable concept explanations survive medical image perturbation","Vision Concept Transformer with diffusion-smoothing stability","Guaranteed stable concepts for interpretable medical AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00052,"raw_usage":{"total_tokens":2546,"prompt_tokens":999,"completion_tokens":1547,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":1486}},"tokens_in":615,"tokens_out":1547,"duration_ms":15386,"temperature":1.0,"reasoning_tokens":1486,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:21:57.336371+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On any image from the four medical datasets, compute fc(X)=Wc f(X) and check whether all entries are nonnegative and sum to 1; if they are not, the probability-vector assumption in Lemma 2 is violated and the certified stability bound of Theorem 2 does not apply. A direct experiment would be to run the SVCT pipeline with and without ℓ1-normalization of fc and compare the measured top-k overlap at the same perturbation radius and noise level.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the label-free CBM transformation that produces the concept feature fc(X) without human-annotated concept labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CLIP-Dissect procedure used to learn the projection weights Wc that map backbone features to concept space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the Gaussian Rényi-divergence bound and the post-processing inequality that power the proof of Theorem 2."},{"cited_title":"Certifiably Robust Interpretation via Renyi Differential Privacy","cited_arxiv_id":"2107.01561","evidence_quote":"Motivates the top-k overlap minimization technique that underlies Lemma 2's Rényi-divergence lower bound."},{"cited_title":"In: Forty-first International Conference on Ma- chine Learning (2024)","cited_arxiv_id":null,"evidence_quote":"Provides the stability-of-interpretation perspective and the top-k overlap argument that Lemma 2 adapts to concept vectors."}],"review_version":1}