W1) The certified guarantee concerns a hand-constructed proxy SDE, not the system that generated the data. This is my main concern, and it is only visible in the appendix.
- The benchmark data are generated with additive Gaussian noise on every coordinate (Eq. (125)). The SDE that is actually verified has noise in the last coordinate only (Eq. (127)), which explains the zeros in \overline{G}_{\mathrm{diff}} in Example 1. The certified system is therefore NOT the data-generating system, and the certificate is not shown to apply to it.
- It remains unclear whether the authors' method actually accounts for diffusion-model estimation and the sampling errors.
- The paper itself states that "the reported probability bounds certify the fixed stochastic proxy used by each verifier" and promises that the statistical relationship between this proxy and the unknown disturbance "is discussed separately at the end of this section" (p. 27). I could not find this discussion anywhere in the paper.
Consequently, the "certified infinite-horizon safety lower bound" is a formal statement about an SDE that the authors construct, not about the unknown system. Calling this "formal safety verification" in the title and abstract, without stating any of the above in the main text, substantially overclaims what the method delivers.
W2) It is unclear whether the authors' method actually provides continuous-time or discrete-time guarantees.
The problem is initially posed for the Ito SDE (1), but Assumption 1 replaces it by a discrete-time model x_{t+1} = \Psi_{\Delta t}(x_t) + \delta_t with a residual whose conditional law depends on a trajectory history \mathcal{H}t. Everything that is learned is a one-step residual distribution. The continuous-time model is then recovered by the heuristic Q\theta = \Sigma_{\Delta x}/\Delta t (Eq. (105), Proposition 3). Proposition 3 only states that a positive semidefinite matrix has a square root; it does not establish any relation between the safety probability of the resulting SDE and that of the discrete-time process the data come from.
Several points remain unclear:
- How is \Psi_{\Delta t} obtained from f?. For data from a genuine continuous-time system, the residual would also contain discretization error of order \Delta t^2, which this construction cannot separate from the noise.
- Assumption 1 allows the residual to depend on the history (i.e., a non-Markovian perturbation), yet the verified SDE (17) is Markovian with a constant diffusion coefficient.
- Assumption 1 calls \mathcal{H}_t an "observable trajectory history", but it consists of full states, i.e., full state observability is assumed.
The paper should state precisely which guarantee holds for which object (e.g., discrete-time process or SDE), and in continuous time or not.
W3) The experiments do not support the claimed advantage of the diffusion model.
- The uncertainty in all benchmarks is additive, state-independent, and Gaussian (Eq. (125)). None of the "complex", "interaction-induced", or state-dependent disturbance patterns that motivate the paper (abstract and Section 1) is ever presented. The benchmarks are polynomial ODEs perturbed by additive Gaussian noise of constant amplitude, for which a diffusion model is unnecessary.
- The key ablation is missing: the same continuous-time SOS pipeline with \lambda_G (or with the true rate) instead of \lambda_{\mathrm{diff}}. Given Table 7, I expect this to perform at least as well as DiffSBC, which would mean that the diffusion component contributes nothing to the results.
- The state-dependent structure is discarded before verification. The covariance is collapsed to a single constant, and the residuals are symmetric as {r, -r} (Table 4), which forces a zero mean. The statement in Section 4.2 that "preserving the state-dependent stochastic structure learned by the diffusion model can lead to more informative certificates" is therefore not supported: no state-dependent structure is preserved in the verified model. If the only difference from the baselines is that the learned uncertainty is not reduced to an averaged additive disturbance, this should be stated plainly.
- The baselines are weak. "Additive+Neural" (called "Gauss.+Neural" in Appendix A.8) verifies a deterministic system, i.e., it is not a stochastic verification method at all. Table 5 also lists "Baseline CEGIS rounds: 1"; please clarify whether the baselines were given the same refinement budget as DiffSBC. There is no comparison with methods that make weaker assumptions (e.g., unknown drift, non-additive noise, or partial observability), nor with abstraction-based methods.
W4) The assumptions are substantially stronger than the introduction and abstract suggest.
The abstract and introduction present the method as learning data-driven stochastic behaviors "from observed trajectories" and adapting to "complex disturbance patterns encountered in real-world operation". In fact the method requires: (i) known polynomial nominal dynamics, stated only in a side sentence before Assumption 1 and in the "Require" line of Algorithm 1; (ii) full state observability; (iii) compact semi-algebraic state, initial, and unsafe sets; (iv) polynomial barrier functions of low degree (2 to 6); and (v) in all experiments, rank-one, constant, additive noise. Each of these is a substantial restriction, and together they should be stated up front. The introduction further claims that the prior data-driven methods it cites were developed "to relax these assumptions", but Mathiesen et al. (2026) assume additive noise and Salamati et al. (2024) assume Lipschitz continuity, so the contrast drawn with them is not accurate.
W5) Novelty and relevance for ICLR are low.
Theorem 1 restates the standard stochastic barrier certificate result of Prajna et al. (2007) (going all the way back to Kushner), Theorem 2 is Putinar's Positivstellensatz, and Theorem 4 seems to be a classical result for diffusions. Proposition 3 is the existence of a matrix square root. The remaining technical contribution is RPM, a quadratic local fit of a noise predictor plus a Gaussian moment closure, used to estimate the variance of a one-dimensional residual. For a one-dimensional variance, sampling from the diffusion model and computing the empirical variance would be simpler; the claim (Section A.2.4) that Monte Carlo estimation is impractical is not substantiated. It is also not clear whether the first contribution (a conditional diffusion model of residuals given known nominal dynamics) is novel.
More generally, many recent papers have already combined machine-learning tools with barrier certificates. Like much of that research line, this paper does not demonstrate relevance to real-world problems: all benchmarks are low-dimensional academic polynomial systems. SOS-based barrier certificates are inherently limited to such systems (the paper itself reports resource-limit rejections already at degree 4 for n = 7), and barrier certificates are known to be conservative in general, i.e., they cannot in general provide tight safety bounds (Laurenti & Lahijanian, 2025). The authors cite this work but do not discuss its implications for their method. Without a demonstration on a system of practical relevance, I do not see how the presented results are suitable for ICLR. The paper may be better suited to a formal-methods or control venue after the issues above are addressed, in particular after the gap between the verified proxy and the actual system is closed and a benchmark with genuinely non-additive, state-dependent noise is added.