- limited novelty.
My main concern is the lack of substantial novelty. The proposed method is essentially a combination of existing QuDDPM/MSQuDDPM architectures, information-based noise scheduling, and established quantum recovery theory. Information-guided diffusion scheduling has already been studied in classical diffusion models. Theorem 1 is a rather straightforward consequence of distributing a fixed total information loss equally across timesteps. The recovery bounds largely follow from existing universal recovery results and continuity inequalities. Similarly, the observation that local reconstruction accuracy does not guarantee distribution matching is already well known in generative modeling, and QuDDPM already considers distribution-level objectives.
Although the paper provides quantum-specific formulations, I struggle to identify a fundamentally new theoretical or algorithmic insight. The actual method seems to be little more than modifying the noise schedule and adding a recovery regularization term to MSQuDDPM. I do not think combining these existing ideas constitutes a sufficiently meaningful contribution.
- Limited practical significance and questionable theoretical justification.
The theoretical results do not convincingly justify the proposed training strategy. The main recovery guarantees concern the existence of an ideal CPTP recovery channel, but they do not establish that the parameterized reverse models can achieve these guarantees. More seriously, the proposed recovery constraints are not necessarily compatible with correct distribution generation.
For example, consider two equally likely states, ρ0=diag(0.99,0.01) and ρ1=diag(0.01,0.99), under complete depolarization. With γ=0.01, a generator that exactly reproduces the target distribution has recovery loss approximately 1.614, exceeding the information budget of 0.637. In contrast, a generator that always outputs I/2 satisfies the budget with loss approximately 0.512, despite producing an incorrect distribution. Thus, the proposed constraint can exclude a correct generator while allowing a collapsed one.
Similarly, equal-information scheduling does not necessarily minimize actual recovery difficulty. For two orthogonal qubit states, a linear two-step schedule achieves a maximum optimal recovery error of 0.25, whereas equal-information scheduling gives approximately 0.390. This directly challenges the interpretation that equalizing information loss balances recovery difficulty.
The endpoint Wasserstein bound is also practically uninformative in the reported TFIM setting, where even under ideal calibration it exceeds the trivial upper bound of 1. Overall, I find the theoretical analysis largely disconnected from the actual learning objective and generation performance.
- Very weak experimental evaluation and inadequate baselines.
The experimental evaluation is extremely limited and, in my view, provides little meaningful evidence of the proposed method's usefulness. The main experiments involve only a one-qubit mixed-state ring and a four-qubit TFIM system. A four-qubit system has a Hilbert-space dimension of only 16, with density matrices of size 16 × 16. Such a small system is readily tractable using classical simulation, and even relatively simple neural networks or QNNs could serve as meaningful baselines. I fail to see what these experiments demonstrate beyond improvements on a very small and relatively simple benchmark.
More importantly, recent quantum-state generative modeling work has already explored substantially larger systems, including 10–14-qubit settings (e.g., arXiv:2605.03573 and arXiv:2606.21256), as well as more diverse benchmarks and stronger generative baselines. In comparison, the experimental scope here is far below what is needed to establish competitiveness or scalability. There is also no real quantum hardware experiment, no realistic measurement or noise analysis, and no evidence of practical quantum deployment advantages. The experiments therefore provide little insight into either larger-scale quantum generative modeling or actual quantum implementation. For an ICLR 2027 submission, I find this level of experimental validation seriously inadequate.
Most comparisons are simply different scheduling and training variants of the same MSQuDDPM architecture. I do not consider QuDDPM/MSQuDDPM alone to be sufficiently strong baselines. A different neural architecture, improved optimization, or even a conventional generative model could potentially achieve similar or better results. The authors should consider stronger alternatives, such as stochastic Schrödinger/Riemannian score-based diffusion (arXiv:2605.03573), intrinsic quantum-state flow matching (arXiv:2606.21256), measurement-based quantum diffusion, and other quantum generative approaches.
Furthermore, since the goal is to generate distributions of quantum states, the evaluation should more thoroughly examine diversity, mode coverage, observable distributions, and entanglement statistics, rather than relying mainly on aggregate Wasserstein distances and local fidelity measurements.
Even within the limited experiments, the improvements are not particularly convincing. On the ring dataset, the best method only reduces the endpoint Wasserstein error from 0.0779 to 0.0757 compared with cosine scheduling without calibration. On TFIM, adding recovery calibration barely changes the maximum budget violation (0.1638 vs. 0.1640). Enforcing the constraints more strongly significantly worsens generation quality, and none of the tested joint configurations satisfies all recovery budgets.
Computing the Holevo-information schedule also requires information about the quantum-state ensemble, yet the authors do not demonstrate that this is practical for larger systems or realistic measurement budgets.
Overall, I find both the theoretical novelty and empirical contribution seriously insufficient. The paper mainly combines existing techniques without introducing a convincing new generative mechanism, the theoretical results do not adequately justify the proposed training strategy, and the experiments are far too limited to establish a meaningful advance over existing methods. I do not see a sufficiently strong contribution to justify acceptance.