{"id":"9deec178-0ab9-4a9f-85e5-aea542c26f8f","arxiv_id":"2607.16293","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SLT refines the label-noise transition matrix only when a QNN's predictive entropy hits a new low, improving noisy-label medical image classification.","lead":"This paper proposes SLT, which updates a noise-correction matrix in quantum neural networks whenever prediction confidence reaches a new minimum, improving noisy-label medical image classification. The approach reports gains on five MedMNIST datasets and a CheXpert subset, but its convergence proof has a gap.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's convergence proof is invalid: Δ(t)→0 only makes entropy-record updates sparse; it does not make {T(t)} Cauchy, so the advertised fixed-point guarantee is unsupported.","rationale":"The paper's central empirical claim—that SLT improves QNN classification under label noise—is supported by extensive tables and ablations, and I do not see a reason to reject it outright. However, the paper also stakes part of its novelty on a convergence guarantee: 'the proposed transition-refinement process reaches a steady state.' That theoretical claim is the least secure point. The reader's rationale correctly flagged the invalid Cauchy step, but the reader's stated weakest_assumption focuses on Eq. (4)'s reliance on argmax predictions as proxies for clean labels. That is a legitimate empirical risk, but it is shared by essentially all anchor-free transition-estimation methods and is not the point where the argument most clearly fails. Even if predictions were perfectly reliable, the supplied proof of Theorem 2 does not establish convergence. The missing fixed-schedule ablation is also important, but it is secondary: the theorem's failure undercuts the advertised stability guarantee itself, not just the incremental benefit of the supermartingale framing. I recommend keeping the CONDITIONAL verdict: the authors could repair the paper by removing or weakening Theorem 2 (e.g., proving only bounded oscillation or adding explicit assumptions on T′), or by adding the missing ablation and code release. Those are addressable changes, so REJECT is not warranted.","tokens_in":27631,"tokens_out":7148,"duration_ms":72242,"concrete_test":"Construct the counterexample process: let s(t) hit running minima at t_k=2^k with S(t_k)=S(t_{k-1})−1/k^2 (so Δ_k→0), and at each record set T′_k alternately to the identity and a fixed permutation before applying Eq. (5) with η=0.5; leave T unchanged between records. Compute ‖T(t_m)−T(t_n)‖∞ as m,n→∞. If the sequence has two accumulation points (distance bounded away from 0), the Cauchy assertion in Appendix A is refuted. A complementary empirical check: in existing synthetic-noise runs, record T′ and the prediction changes at every entropy-record update, and test whether ‖T′−T‖ stays bounded away from zero while Δ(t)→0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SLT's advertised theoretical foundation is the almost-sure convergence of {T(t)} to a fixed point T* (Theorem 2). The proof in Appendix A derives lim Δ(t)=0 from Doob's theorem, notes updates occur only when Δ(t)>0, and then asserts that {T(t)} is Cauchy. That inference is invalid. Δ(t)→0 bounds how much the running entropy minimum drops, and hence how frequently updates occur, but it does not bound the size or direction of the transition-matrix jumps. Eq. (5) updates T←(1−η)T+ηT′, where T′ is recomputed from current argmax predictions and can change discontinuously. A new entropy record of size 1/k^2 can occur at times t_k=2^k while T′ alternates between two matrices; the update rule then yields two accumulation points and no limit. Nothing in the paper's assumptions prevents this: model predictions can change without decreasing entropy, and T′ is not shown to converge to T. Thus the claimed 'fixed-point' convergence and the 'stability criterion' are unsupported. The empirical gains may survive, but the supermartingale convergence guarantee—a stated contribution—does not follow from the supplied argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SLT, a label-transition loss-correction method for QNNs under noisy labels. SLT tracks the historical minimum of predictive entropy, treats it as a supermartingale, and triggers an update of the noise transition matrix whenever a new entropy minimum is reached; the matrix is updated by a moving average of the empirical co-occurrence between predicted and noisy labels. The paper claims a convergence theorem (Theorem 2) that the transition sequence almost surely reaches a fixed point, and supports the method with experiments on five MedMNIST datasets and CheXpert under synthetic and real noise.","tokens_in":27930,"tokens_out":6239,"duration_ms":56698,"significance":"If the claims were supported, the paper would make a useful contribution: it provides a simple anchor-free mechanism tailored to QNN properties, broad empirical evaluation with multiple baselines, ablations over backbone, hyperparameters, and a real-world dataset. The method is parameter-efficient and the empirical gains over CE and Forward are consistent, especially on the small datasets. However, the theoretical foundation — advertised as a key contribution — is not established: Lemma 1 is tautological and Theorem 2's proof has a clear logical gap. The empirical claims appear plausible and the work is likely salvageable by weakening the theory or providing a real proof, but in its current form the paper overclaims.","major_comments":[{"comment":"The proof asserts that lim_{t→∞} Δ(t)=0 makes {T(t)} Cauchy. This is invalid: update events can occur infinitely often with summable Δ(t) (e.g., at t_k=2^k with Δ(t_k)=1/k^2). The update rule (5) with T′ recomputed from current argmax predictions can then produce two accumulation points if T′ alternates between two matrices; nothing in the paper rules this out. The claimed almost-sure convergence to a fixed point T* is therefore unsupported.","section":"Appendix A, proof of Theorem 2"},{"comment":"Lemma 1 is a tautology: since S(t)=min_{τ≤t} s(τ), S(t)≤S(t−1) with probability 1, so E[S(t)|F(t−1)]≤S(t−1) holds for any s(t). The 'supermartingale' property carries no information about prediction reliability or about the transition matrix. Doob's theorem only yields convergence of the running minimum, which already follows from monotonicity and boundedness. Thus the theorem does not provide the advertised stability guarantee for transition refinement.","section":"Section 3, Lemma 1 / Eq. (3)–(5)"},{"comment":"The transition estimator T′ treats current argmax predictions as proxies for latent clean labels. If the QNN is confidently wrong (or has memorized noise), T′ is biased and the corrected loss reinforces the error. The paper invokes 'natural smoothness' to justify this, but offers no empirical or formal evidence that entropy-record events coincide with increasing label-correct confidence. The ablation in Fig. 3 (classical backbones degrade under SLT) is consistent with this concern; a direct diagnostic (e.g., comparing T′ to the true noise matrix on synthetic noise) is needed to support the method's central mechanism.","section":"Eq. (4) and Fig. 3"},{"comment":"The comparison on CheXpert may not be apples-to-apples. The implementation section states that all baselines use the same backbone QNN, but methods such as VisualCheXbert and BoMD include text/other modalities; it is unclear whether these were re-implemented on the QNN backbone or run in their native classical form. The paper should state the backbone for each CheXpert baseline; as written, the 'state-of-the-art' claim in Table 5 is not verifiable.","section":"Table 5, CheXpert experiments, Appendix B"}],"minor_comments":[{"comment":"'del' and 'pat' are used in the main text (Fig. 4) without definition; they are only explained in Appendix B. Please define or forward-reference them.","section":"Algorithm 1 / Section 3"},{"comment":"The section promises an introduction to barren plateaus, but no discussion of barren plateaus follows. Either add the material or remove the phrase.","section":"Section 2"},{"comment":"Several formatting issues: 'V olMinNet' contains a spurious space; 'T revision' and the reference [11] author name have similar spacing problems. Please proofread.","section":"Typos"}],"recommendation":"major_revision","confidential_remarks":"The theoretical gap is serious and the paper should be revised to either prove Theorem 2 under stated assumptions or remove the convergence claim and reposition the paper as empirical. The empirical section is broad, but the CheXpert comparison fairness should be clarified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the empirical part of this paper is worth your time; the advertised theory is not. The SLT recipe — update the noise transition matrix whenever the model's predictive entropy sets a new historical minimum — beats the listed baselines across several small medical datasets, including a CheXpert subset. That is a useful result for the QML community. But the 'supermartingale' framing is a wrapper around a running-minimum process, and the convergence guarantee in Theorem 2 does not follow from the supplied argument.\n\nWhat is genuinely new and done well: this is the first application I'm aware of that couples anchor-free transition-matrix correction with QNN backbones for noisy-label medical imaging. The evaluation is thorough: five MedMNIST datasets, three synthetic noise types at three rates, instance-dependent noise, CheXpert, ablations on backbone, eta, qubit/layer counts, update strategies, and a plausible narrative about QNN smoothness mitigating overconfidence. The authors also report failure cases in the qubit trade-off analysis, which is honest.\n\nWhere it gets soft. Lemma 1 is tautological: for any sequence, the running minimum is a supermartingale because S(t) = S(t-1) - Delta(t) with Delta(t) >= 0. It says nothing about entropy. Theorem 2's proof has a real gap: lim Delta(t) = 0 means the entropy records become sparse, not that the transition-matrix updates stop or shrink. T' is recomputed from current predictions and can jump discontinuously; with eta fixed, infinitely many updates can make T oscillate without converging. The space of stochastic matrices being complete only helps if you've already shown the sequence is Cauchy, which the proof doesn't. So the headline convergence guarantee is unsupported. There is also a mismatch between the analyzed rule (update if s(t) < S(t-1)) and the actual implementation (NDU+PTU with delay and patience). And the paper never compares entropy-triggered updates against a fixed or random schedule, so the specific benefit of the supermartingale trigger isn't isolated. Code isn't available yet.\n\nThese are fixable. The empirical claim can survive with the theory either repaired or explicitly downgraded to a heuristic; the missing fixed-schedule ablation is straightforward; code release would help. But as written, I wouldn't bet on the convergence theorem.\n\nFor peer review: send it out, but with referees who can check the martingale argument and push for the fixed-schedule ablation. The empirical work deserves to be seen; the theory as stated should not be.","headline":"Good empirical recipe buried under a convergence theorem that doesn't follow; the \"supermartingale\" is a relabeled running minimum.","tokens_in":28423,"tokens_out":4045,"would_cite":false,"duration_ms":36994,"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":"A supermartingale view of predictive entropy lets quantum neural networks correct noisy labels in small medical images without anchor points, and the transition-matrix updates converge to a fixed point.","keywords":["noisy-label learning","quantum neural networks","loss correction","noise transition matrix","supermartingale","entropy","anchor-free","medical image classification"],"falsifier":"Construct a small medical dataset with instance-dependent noise that makes the network confidently wrong about a recognizable subset from the first epochs. If SLT's entropy-triggered updates then lower test F1 relative to a fixed-transition baseline while the recorded entropy minimum keeps decreasing — that is, if the selected update epochs coincide with falling clean-label agreement on held-out data — the central claim that falling entropy tracks reliable confidence would be refuted.","tokens_in":27477,"feed_emoji":"⚛️","tokens_out":6914,"duration_ms":69384,"temperature":0.7,"pith_summary":"SLT claims that the smoothness of quantum neural networks, usually a drawback for noisy-label learning, can be used as a stabilizing signal. The method tracks the historical minimum of predictive entropy across training, models that running minimum as a supermartingale, and updates the noise-transition matrix only when that minimum is beaten. The paper proves that this update process converges to a fixed matrix, and reports that the resulting anchor-free loss correction improves classification on five small medical image datasets under several synthetic noise types and on a real-world chest X-ray set. The reason to care is that small medical datasets with unreliable annotations are common, and SLT offers a stable correction mechanism that does not need hard-to-find anchor points.","feed_headline":"Entropy-record updates make quantum classifiers robust to noisy labels","feed_subtitle":"Updating the noise model only at entropy record lows stabilizes quantum training on small medical image sets.","key_machinery":"The central object is the noise-transition matrix T, a row-stochastic K-by-K matrix whose entry T[j,k] estimates the probability that an instance predicted as class j carries noisy label k; it is used for forward loss correction. The trigger mechanism is the running minimum S(t) of normalized predictive entropy. Because S(t) is a monotone, bounded sequence, the paper models it as a supermartingale and invokes a standard martingale convergence theorem to argue that S(t) converges and that updates of T become asymptotically infrequent. Each update constructs an empirical co-occurrence matrix T' from current argmax predictions and noisy labels, then applies a moving-average update T ← (1-η)T +","core_discovery":"On its own terms, the paper's discovery is that the historical minimum of the model's normalized predictive entropy is a supermartingale: at every step, the expected next running minimum is no larger than the current one. Because the sequence is bounded in [0,1], a standard martingale convergence theorem gives a finite almost-sure limit, and the paper argues from this that the update rule for the noise-transition matrix — which is revised only in epochs where the entropy score beats all previous scores — becomes asymptotically infrequent and converges almost surely to a fixed matrix T*. The refinement itself is anchor-free: each entry is the empirical co-occurrence of predicted class j with","pith_inferences":["The convergence proof establishes that updates become rare, but not that the fixed matrix equals the true noise transition; under instance-dependent or feature-dependent noise the class-conditional matrix is misspecified, so the steady state could be stable yet biased.","The same entropy-record trigger could be tested on calibrated classical networks: if calibration removes early overconfidence, the performance gap between classical and quantum backbones under SLT should shrink — a testable prediction the paper does not run.","One could gate the trigger on an external signal such as validation accuracy or agreement with a small clean set; if new entropy minima still occur while clean-label accuracy drops, that would expose the assumption that confidence growth equals correctness growth.","Adapting the moving-average step η to the observed rate of entropy decrease might give a better stability-adaptability balance than a fixed η tuned per dataset."],"forward_implications":["If the entropy-minimum process is indeed a supermartingale, the transition-refinement process has a guaranteed steady state, so an SLT-trained QNN does not chase oscillating label-noise estimates in late training.","Because the transition matrix is built from predicted-versus-noisy co-occurrences rather than anchor points, the method applies to small medical datasets where clean anchors are unavailable.","The reported experiments support that SLT improves QNN classification under uniform, cyclic, custom-mapping, and instance-dependent label noise, with gains that persist at high noise ratios.","Temperature scaling of QNN outputs does not improve SLT, consistent with the claim that the intrinsic smoothness of quantum measurements already provides the right confidence-growth profile.","On a sampled real-world chest X-ray set with automatically extracted labels, the entropy-triggered correction is reported to yield higher AUC and F1 than several generic noisy-label and domain-specific baselines."],"fun_headline_variants":["Quantum nets tame noisy labels via entropy-record updates","Supermartingale trick steadies quantum classifiers on noisy data","Label noise beaten by entropy-minimum quantum training","Quantum AI copes with bad labels using entropy record lows","Noisy-label quantum learning anchored by entropy minima"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method's safety rests on the assumption that the model's most-confident predictions at entropy minima are reliable proxies for the latent clean labels; if the QNN is confidently wrong, whether by memorizing corrupted labels or by systematic bias, the co-occurrence counts poison the transition matrix and the corrected loss amplifies the mistake.","fun_headline_variants_meta":{"raw":{"variants":["Quantum nets tame noisy labels via entropy-record updates","Supermartingale trick steadies quantum classifiers on noisy data","Label noise beaten by entropy-minimum quantum training","Quantum AI copes with bad labels using entropy record lows","Noisy-label quantum learning anchored by entropy minima"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000164,"raw_usage":{"total_tokens":1070,"prompt_tokens":716,"completion_tokens":354,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":278}},"tokens_in":460,"tokens_out":354,"duration_ms":3888,"temperature":1.0,"reasoning_tokens":278,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:59:59.247355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a small medical dataset with instance-dependent noise that makes the network confidently wrong about a recognizable subset from the first epochs. If SLT's entropy-triggered updates then lower test F1 relative to a fixed-transition baseline while the recorded entropy minimum keeps decreasing — that is, if the selected update epochs coincide with falling clean-label agreement on held-out data — the central claim that falling entropy tracks reliable confidence would be refuted.","supporting_citations":[],"review_version":1}