{"id":"4476bdd8-bd74-4880-b87a-4c4be0d60d5d","arxiv_id":"2507.07532","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Neural Concept Verifier trains image classifiers so predictions must rely on small, verifiable subsets of extracted concepts rather than raw pixel masks.","lead":"This paper combines concept-based explanations with a prover-verifier game, so image classifiers justify decisions using a small set of high-level concepts instead of raw pixels. The reported model matches or beats pixel-based verifier baselines on several image benchmarks while keeping soundness high.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world completeness wins are confounded by representation: NCV uses dense CLIP-Sim while CBM baselines use sparse SpLiCE, so the prover-verifier game alone may not explain Table 1.","rationale":"Reader's verdict is high-confidence conditional, and my main concern is consistent with that rather than a reason to reject. The framework is interesting, code is public, limitations are unusually explicit, and the CLEVR-Hans NCB experiments do control the representation. But the strongest real-world evidence in Table 1 does not. The claim in Sec. 4.1 that baselines share the same concept extractor is contradicted by the appendix definitions of CLIP-Sim vs SpLiCE. Because CLIP-Sim's dense 10k-dimensional similarity vector contains strictly more per-image information than SpLiCE's optimized sparse embedding, any observed NCV advantage may be an artifact of representation, not of the prover-verifier game. The test I propose isolates the mechanism. This also means the abstention/robustness interpretation of soundness should be reported (fraction of ⊥ predictions), since soundness counts abstention as success and the paper's own theory (Sec. A.5) limits the adversary to the trained neural Morgana. Neither issue is fatal, but both need to be settled before the headline completeness/soundness claims are accepted as evidence for NCV's core mechanism.","tokens_in":30749,"tokens_out":7647,"duration_ms":87227,"concrete_test":"On CIFAR-100, rerun the Table 1 comparison with the nonlinear CBM baseline replaced by an MLP of the same architecture and training protocol as NCV's pretrained verifier (Suppl. E.1) but trained directly on the same dense CLIP-Sim activation vectors that NCV consumes, and report its completeness. If this same-representation nonlinear CBM reaches or exceeds NCV's 83.32%, the gains attributed to the NCV prover-verifier mechanism are not supported; if it does not, the Q2 claim narrows to the sparse-selection regime. As a secondary check, report the fraction of soundness-score examples that end in abstention for both NCV and Pixel-MAC.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 promises that NCV is compared against baselines sharing the same concept extractor, but this is not what Table 1 does for CIFAR-100, ImageNet-1k, or COCOLogic. The CBM baselines (Suppl. C.3) use per-sample optimized sparse SpLiCE encodings, while NCV's 'CLIP-Sim' is defined in Sec. 4.1 as retaining the full dense activation vector. The central Q1/Q2 completeness comparisons therefore conflate two changes: the Merlin-Arthur game and the choice of dense CLIP-Sim over sparse SpLiCE. Internal numbers support the confound: NCV's own pretrained verifier without any prover reaches 85.96% on CIFAR-100 (Table 8), already above the 79.29% nonlinear CBM, while NCV's game-based completeness is 83.32%. Without a nonlinear CBM trained on the identical CLIP-Sim features (or, conversely, NCV trained on SpLiCE), the paper's headline 'matches or outperforms' and its interpretability-accuracy-gap claim are not established for the real-world datasets. A reader also cannot tell how much of NCV's near-perfect soundness is abstention because rejection rates are never reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Neural Concept Verifier (NCV), which combines a pretrained concept extractor (NCB on CLEVR-Hans, CLIP-Sim on CIFAR-100, ImageNet-1k, and COCOLogic) with a Merlin-Arthur prover-verifier game played over sparse concept masks. A cooperative prover Merlin selects a small concept subset to support the true class, an adversarial prover Morgana selects misleading subsets, and a nonlinear verifier Arthur predicts from the masked concepts, with completeness and soundness as the headline metrics. The authors report that NCV scales prover-verifier games to high-dimensional images, narrows the interpretability-accuracy gap of linear concept bottleneck models, produces more interpretable explanations than pixel-level masks, and mitigates shortcut learning. The central empirical claim is that NCV matches or outperforms concept-based and pixel-based baselines in completeness on most real-world settings while retaining near-perfect soundness.","tokens_in":2104,"tokens_out":2311,"duration_ms":64356,"significance":"If the central claims were established, the contribution would be useful: NCV provides a clean way to enforce sparse, per-sample concept selection through an adversarial game while keeping a nonlinear classifier, and it is evaluated over a wide range of datasets (CLEVR-Hans3/7, CIFAR-100, ImageNet-1k, COCOLogic) with 10-20 seeds. The paper also ships public code, reports computational costs, includes ablations over mask size and the trade-off parameter gamma, and is unusually transparent about the limitations of the information-theoretic guarantees in App. A.5. These are genuine strengths. However, as submitted, the main empirical claims for the real-world datasets are not yet supported because the completeness comparisons conflate the prover-verifier game with a change of concept representation, and because the near-perfect soundness figures are reported without any information about how much of them is abstention.","major_comments":[{"comment":"Section 4.1 promises that within each dataset, NCV is compared against baselines that share the same concept extractor, but this is not what Table 1 does for CIFAR-100, ImageNet-1k, or COCOLogic. The CBM baselines use per-sample optimized sparse SpLiCE encodings (Suppl. C.3), while NCV uses CLIP-Sim, which Sec. 4.1 explicitly defines as retaining the full dense activation vector. The headline completeness comparison therefore conflates two changes: the Merlin-Arthur game and the choice of dense CLIP-Sim over sparse SpLiCE. To support the Q1/Q2 claims, the authors should add a nonlinear CBM trained on the identical dense CLIP-Sim features, or alternatively an NCV variant operating on SpLiCE encodings, so that the representation is held fixed.","section":"§4.1 and Table 1"},{"comment":"The claim that NCV matches or outperforms baselines in completeness on real-world datasets is undercut by the paper's own pretraining results. Table 8 shows that the pretrained verifier alone, without any prover, reaches 85.96% on CIFAR-100, 77.07% on ImageNet-1k, and 81.39% on COCOLogic, whereas the full NCV game achieves 83.32%, 67.04%, and 75.42%, respectively. Thus on every real-world dataset the prover-verifier game reduces completeness relative to its own pretrained verifier on the same CLIP-Sim features. The apparent advantage over the SpLiCE-based CBMs should therefore be attributed to the dense representation, not to the game, and the interpretability-accuracy gap reduction is not established without a nonlinear CBM on the same features.","section":"Table 8 vs. Table 1"},{"comment":"Soundness as defined in Eq. (2) treats the rejection class as success, and the Morgana loss in App. A.4 switches the target to the rejection class whenever Arthur already prefers abstention over the true class, in addition to a softplus penalty on the gap between the true-class and rejection logits. Under this training objective, near-perfect soundness can be achieved largely by abstaining. The paper never reports rejection or abstention rates for any model or dataset, so a reader cannot tell whether the reported 99.9% soundness means that Morgana fails to elicit wrong predictions or that Arthur simply declines to answer. The authors should report soundness conditional on making a class prediction, together with the overall rejection rate.","section":"§3.4, Eq. (2), and Suppl. A.4"},{"comment":"The Pixel-MAC comparison is not a controlled test of whether prover-verifier games scale to high-dimensional images. For CIFAR-100, Pixel-MAC is restricted to a mask of 32 pixels out of 32x32 pixels, and for ImageNet-1k and COCOLogic to 1000 pixels out of roughly 50k pixels, while NCV's 32 concepts are full-image CLIP similarity scores. The low Pixel-MAC completeness on these datasets is therefore confounded by an extremely small pixel budget and by the different information content of a pixel versus a global concept score. The authors should either give Pixel-MAC a comparable fraction of the input (e.g., by using a larger patch budget, or by reporting a Pixel-MAC variant with the same number of scalar features) or restrict the scaling claim to the CLEVR-Hans settings where the mask sizes are more comparable.","section":"§4.2 and Suppl. C.2"},{"comment":"The theoretical section is presented as providing information-theoretic grounding, but App. A.5 explicitly states that the mutual-information bounds are not estimated numerically, that the assumptions on asymmetric feature correlation and Morgana's relative success rate are not verified, and that the precision is only a population quantity that is not meaningfully estimable for dense continuous encodings. This is an honest and appropriate caveat, but it should be reflected in the abstract and introduction: the current wording, especially formal verifiability and strong soundness guarantees, overstates what the experiments establish, since the reported completeness and soundness are empirical proxies rather than certified guarantees.","section":"§3.5 and App. A.5"},{"comment":"Table 1's matches or outperforms summary is also inaccurate on ImageNet-1k, where NCV's completeness (67.04) is below both the linear CBM (68.59) and the nonlinear CBM (69.02). The paper's text does acknowledge this in §4.2, but the table caption and the Q2 discussion should state more precisely that the advantage is limited to CIFAR-100 and COCOLogic, and even there only against the SpLiCE-based CBMs.","section":"Table 1"}],"minor_comments":[{"comment":"The abstract's formal verifiability and the conclusion's verifibale explanations (typo) should be aligned with the limitations stated in App. A.5; I suggest replacing formal verifiability with something like verifiable, concept-level reasoning and correcting the typo.","section":"Abstract and §6"},{"comment":"The text says the three modules are updated jointly and then describes a two-phase alternating scheme with continuous masks for prover updates and hard top-m masks for Arthur's update; this is clear in App. A.4, but the main text would benefit from a pointer to the exact loss formulations for Merlin's and Morgana's updates, since the footnote referring to Sec. A.4 is easy to miss.","section":"§3.6"},{"comment":"The table header layout with three column blocks is hard to parse because the dataset names are separated from their columns by the shared completeness/soundness headers; adding explicit column groups for each dataset would improve readability.","section":"§4.1 and Table 1"},{"comment":"The comparison of selected concepts between NCV and a linear CBM is helpfully honest about being illustrative, but the sentence the verifier is therefore not simply handed the answer is a bit misleading because the concept vocabulary deliberately contains phrases that are surface forms of class names; I would acknowledge this more concretely in the main text when discussing explanation quality.","section":"Suppl. E.5"},{"comment":"The term CLIP-Sim is defined only in passing in Sec. 4.1; since it is the central representation for the real-world experiments, a short formal definition (e.g., c_j = cosine similarity between image embedding and text embedding of concept j) would help the reader.","section":"§4.1 and Suppl. C.3"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about representation confounding is real and load-bearing: the completeness comparisons in Table 1 for real-world datasets compare two different representations (dense CLIP-Sim vs. sparse SpLiCE), and the paper's own Table 8 shows that the prover-verifier game reduces accuracy relative to the pretrained verifier on those same features. The soundness-abstention issue compounds this. I recommend major revision rather than rejection because the missing controlled experiments (nonlinear CBM on identical CLIP-Sim, or NCV on SpLiCE, plus rejection-rate reporting) are feasible within the manuscript's scope and would allow the central claims to be evaluated fairly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Best quick take: this is a useful paper to know about, and I'd send it to reviewers, but the headline claims for real-world datasets do not yet isolate the NCV mechanism.\n\nWhat's actually new: shifting the prover-verifier game from pixels to concept encodings, using a frozen extractor (NCB or CLIP-Sim), a sparse-mask prover pair, and a nonlinear verifier. That combination is new and it does scale to CIFAR-100, ImageNet-1k, COCOLogic, where pixel-space MAC basically dies. The paper also ships code and data, includes ablations over mask size, gamma, prover architecture, and a DCR comparison; those are real assets. The authors are unusually upfront that the information-theoretic bounds from Wäldchen et al. are inherited, instantiated class-wise, and not numerically certified here; Supp. A.5 says exactly that. Credit where due.\n\nThe soft spots are real but fixable. First, the central real-world comparison is confounded. Section 4.1 promises baselines sharing the same concept extractor, but Table 1 compares NCV's dense CLIP-Sim against linear and nonlinear CBMs trained on sparse SpLiCE. Per-sample optimized sparse representations are a different feature space. Table 8 makes the problem concrete: the NCV verifier pretrained without any prover gets 85.96% on CIFAR-100, above the game-based 83.32% and far above the nonlinear CBM's 79.29%. So the 'interpretability-accuracy gap' claim rests on representation choice more than on the PVG. The fix is straightforward: train a nonlinear CBM on the same dense CLIP-Sim features, or run NCV on SpLiCE.\n\nSecond, soundness at 99.9x percent looks too good to interpret without knowing how often Arthur rejects. The definition counts abstention as sound, but rejection rates are never reported. Near-perfect soundness can be trivially achieved by always rejecting; the gamma sweep shows soundness collapses at gamma=0, which is useful context, but still no abstention numbers.\n\nThird, the Pixel-MAC on CIFAR-100 is restricted to 32 pixels per image (Supp. C.2) while NCV selects 32 full-image concept scores; that comparison mainly shows pixel-space MAC is brittle, not that concept encodings beat pixels fairly. The ImageNet pixel-MAC uses 1000 pixels, which is also tiny.\n\nWhat holds up: the CLEVR-Hans results with NCB are cleaner because extractor and baselines share representation; the shortcut experiments are honest, showing NCV does not eliminate shortcuts at 0% clean data, only improves with some clean samples. The qualitative explanations are genuinely more readable than pixel masks.\n\nWho benefits: people working on concept bottleneck models, verifiable explanations, or PVG scaling. I'd want the confounds resolved before citing the real-world numbers, but the framework and the honest limitations discussion deserve referee attention. Send it to review, ask for the same-representation CBM baseline, rejection rates, and a fairer pixel-MAC setup.","headline":"NCV is a sensible new combination of Merlin-Arthur games and concept encodings with honest limitations sections, but its headline real-world comparisons are confounded by representation choice and missing rejection rates.","tokens_in":31547,"tokens_out":6554,"would_cite":false,"duration_ms":59541,"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 argues that moving the prover-verifier proof game from pixels to compact concept encodings makes verifiable, interpretable classification practical on high-dimensional images such as ImageNet, with completeness matching or…","keywords":["prover-verifier games","Merlin-Arthur classifiers","concept bottleneck models","concept encodings","verifiable classification","completeness and soundness","shortcut learning","interpretable machine learning"],"falsifier":"Count how often Merlin's selected concept subsets actually recur across a held-out set in the CLIP-based regime: if the 32-concept masks are almost all unique per image, the containment event $\\{M(x')\\subseteq x\\}$ has near-zero empirical probability, so the average-precision and mutual-information quantities behind the faithfulness claim cannot be estimated from data and the certificate is vacuous there — the paper itself says the bound is most readily evaluated for discrete, low-dimensional encodings, not dense continuous ones such as CLIP. A second check is to search over concept subsets at inference time with an adversary that is not the trained Morgana; if soundness drops well below the reported ~99.9%, the guarantee is robustness against a specific trained adversary rather than protection against any misleading subset.","tokens_in":30509,"feed_emoji":"🧩","tokens_out":13690,"duration_ms":127380,"temperature":0.7,"pith_summary":"This paper claims that formal verifiability for nonlinear classifiers can be made to work on high-dimensional images by moving the proof game from raw pixels to compact concept encodings. It introduces the Neural Concept Verifier, in which a cooperative prover selects a sparse set of interpretable concepts, an adversarial prover selects deliberately misleading ones, and a nonlinear verifier must classify from those selections alone, never seeing the full image and with the option to abstain. If correct, the approach gives classifiers that are accurate and can justify every decision with a short, human-readable list of concepts, while closing the accuracy gap that has kept concept-based models tied to constrained linear heads. The paper reports that the scheme matches or outperforms pixel-space verifier baselines and opaque ResNets in completeness on most benchmarks while keeping soundness near perfect, and that it shrinks reliance on spurious correlations when even a few clean samples are available.","feed_headline":"Verifiable classifiers reach ImageNet via concept-level proofs","feed_subtitle":"Sparse concept subsets replace pixel masks, keeping accuracy high and the reasons checkable.","key_machinery":"The carrying mechanism is a prover-verifier game relocated into concept space. The named pieces are Merlin, the cooperative prover, and Morgana, the adversarial prover — two neural networks that output sparse binary masks over the concept encoding — and Arthur, the nonlinear verifier, which sees only the masked concept values and predicts into $K$ classes plus a rejection class $\\perp$. The two quantitative objects that carry the argument are completeness, $P(A(S)=y)$, and soundness, $P(A(\\hat{S})\\in\\{y,\\perp\\})$, where $S$ is the masked concept subset produced by Merlin and $\\hat{S}$ the one produced by Morgana. Training alternates between the provers and Arthur: the provers are updated through continuous soft masks so gradients flow, the masks are then rounded by a top-$m$ operator, and Arthur is trained on the hard-masked subsets with loss $(1-\\gamma)L_{\\text{Merlin}}+\\gamma L_{\\text{Morgana}}$, where Morgana's loss treats abstention as a sound outcome. The formal spine is a one-vs-rest instantiation of the Merlin-Arthur mutual-information bound: for each class $k$, the average precision of Merlin's concept features lower-bounds the mutual information $I(Y_k(y); M(x')\\subseteq x)$ through the binary entropy function, and the observable completeness and soundness errors jointly lower-bound that precision whenever the asymmetric feature correlation $\\kappa_k$, Morgana's relative success rate $\\alpha_k$, and the class imbalance $B_k$ are well-behaved.","core_discovery":"The central discovery is that the Merlin-Arthur prover-verifier game, which previously stalled on high-dimensional inputs because it argued over pixels, can be played in concept space without losing its formal guarantees. A frozen concept extractor maps each image to an encoding — object slots from the Neural Concept Binder for CLEVR-style scenes, or CLIP similarity scores against a 10,000-word caption vocabulary for natural images. Two differentiable provers, Merlin and Morgana, emit sparse binary masks selecting a fixed small number of concepts, and Arthur, a nonlinear verifier (a Set Transformer or two-layer MLP), predicts the label from the masked concepts alone, with an explicit rejection class it may use when uncertain. The paper defines completeness as the probability that Arthur recovers the true label from Merlin's subset and soundness as the probability that Arthur, on Morgana's adversarial subset, either stays correct or abstains; it then instantiates the Merlin-Arthur theory class-wise, showing that under bounded feature-correlation, class-imbalance, and relative-adversary-strength assumptions, high completeness and soundness imply that Merlin's sparse selections carry near-maximal mutual information about the label in each one-vs-rest subproblem. Across CLEVR-Hans3/7, CIFAR-100, ImageNet-1k, and COCOLogic, the paper reports that NCV matches or beats all baselines in completeness in most settings, achieves soundness at or near 100%, and narrows or closes the interpretability-accuracy gap of linear concept bottleneck models.","pith_inferences":["Soundness is reported against the specific Morgana network actually trained, so 'verifiable' here means robust to that adversary class; a natural stress test is an independent inference-time search over concept subsets, and the paper's own supplement concedes this dependence.","The mutual-information certificate needs selected concepts to recur across inputs, which holds for the discrete NCB encodings but not for dense CLIP similarities — so the strongest-sounding guarantee applies precisely in the regime where the explanation is already easiest to inspect by hand.","The shortcut results suggest Morgana's selections teach Arthur to distrust spurious attributes; one observable test is to record Morgana's masks early in training and check whether they preferentially mark shortcut concepts such as 'gray' on CLEVR-Hans.","The theory bounds each class separately and says nothing directly about the joint K-class argmax; deriving a genuine multiclass certificate, or testing whether the class-wise bounds compose, would be the next step toward an end-to-end guarantee."],"forward_implications":["Prover-verifier games leave the toy-data regime: the same game machinery runs on $224\\times224$ images with 1,000 classes, at a training cost the paper estimates as one to three orders of magnitude below pixel-space provers.","Concept bottleneck models keep their interpretable concept interface while gaining nonlinear heads: NCV matches or beats linear CBMs and opaque ResNets on CIFAR-100 and COCOLogic, so the interpretability-accuracy gap is not intrinsic to concept bottlenecks.","Verified decisions come with human-readable certificates: a class explanation is a short concept list such as 'small and metal and sphere' rather than a diffuse pixel mask, which is what pixel-space provers return.","The adversarial branch acts as robustness training against spurious correlations: with only 1-5% clean samples, NCV's validation-to-test gap under shortcuts drops faster than CBM baselines, indicating reduced shortcut reliance.","Near-perfect soundness means a misleading concept subset makes Arthur abstain rather than commit to an error, giving the model a built-in 'I don't know' response under adversarial selection."],"supporting_citations":[{"why":"Supplies the Merlin-Arthur classifier theory that NCV instantiates in concept space: completeness and soundness definitions, average-precision and mutual-information bounds, and the feature-correlation and relative-strength assumptions.","marker":"Wäldchen et al., 2024"},{"why":"Introduces prover-verifier games, the cooperative/adversarial training paradigm on which NCV is built.","marker":"Anil et al., 2021"},{"why":"Provides the Neural Concept Binder, the unsupervised slot-based concept extractor used for the CLEVR-Hans instantiations.","marker":"Stammer et al., 2024b"},{"why":"Provides SpLiCE and the CLIP-based sparse concept-embedding recipe behind NCV's CLIP-Sim extractor and its 10,000-concept LAION vocabulary.","marker":"Bhalla et al., 2024"},{"why":"Defines Concept Bottleneck Models, the reference point for the interpretability-accuracy gap NCV claims to narrow.","marker":"Koh et al., 2020"},{"why":"Supplies the CLEVR-Hans3 and CLEVR-Hans7 datasets, including the confounded splits used for shortcut experiments.","marker":"Stammer et al., 2021"},{"why":"Provides the COCOLogic benchmark of real-world images with compositional class rules, where NCV's completeness margin over baselines is largest.","marker":"Steinmann et al., 2025"},{"why":"Supplies CLIP, the vision-language backbone producing the concept similarities used on CIFAR-100, ImageNet-1k, and COCOLogic.","marker":"Radford et al., 2021"}],"fun_headline_variants":["Prover-verifier games scale to images via concept encodings","Concept-level proofs make verifiable AI practical on ImageNet","Sparse concepts let Merlin-Arthur verifiers handle high-dim data","NCV: verifiability meets interpretability on high-dimensional data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the frozen concept extractor's encodings contain enough task-relevant information, with consistent semantics across inputs, that a prover-verifier game played on top of them can both classify accurately and justify itself; if the concept space discards or entangles the information that separates classes, no game can recover it, and the paper's own Discussion concedes that effectiveness depends on the quality of the underlying concept extractor.","fun_headline_variants_meta":{"raw":{"variants":["Prover-verifier games scale to images via concept encodings","Concept-level proofs make verifiable AI practical on ImageNet","Sparse concepts let Merlin-Arthur verifiers handle high-dim data","NCV: verifiability meets interpretability on high-dimensional data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1612,"prompt_tokens":1058,"completion_tokens":554,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":674,"completion_tokens_details":{"reasoning_tokens":480}},"tokens_in":674,"tokens_out":554,"duration_ms":6005,"temperature":1.0,"reasoning_tokens":480,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:38:06.440877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count how often Merlin's selected concept subsets actually recur across a held-out set in the CLIP-based regime: if the 32-concept masks are almost all unique per image, the containment event $\\{M(x')\\subseteq x\\}$ has near-zero empirical probability, so the average-precision and mutual-information quantities behind the faithfulness claim cannot be estimated from data and the certificate is vacuous there — the paper itself says the bound is most readily evaluated for discrete, low-dimensional encodings, not dense continuous ones such as CLIP. A second check is to search over concept subsets at inference time with an adversary that is not the trained Morgana; if soundness drops well below the reported ~99.9%, the guarantee is robustness against a specific trained adversary rather than protection against any misleading subset.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides SpLiCE and the CLIP-based sparse concept-embedding recipe behind NCV's CLIP-Sim extractor and its 10,000-concept LAION vocabulary."},{"cited_title":"W., Nguyen, T., Tang, Y","cited_arxiv_id":null,"evidence_quote":"Defines Concept Bottleneck Models, the reference point for the interpretability-accuracy gap NCV claims to narrow."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CLEVR-Hans3 and CLEVR-Hans7 datasets, including the confounded splits used for shortcut experiments."}],"review_version":1}