acceptodds
Under review as a conference paper at ICLR 2027

Uncertainty-Aware Sequential Monte Carlo for Hard-Constrained Diffusion

Abstract

Deploying diffusion models often requires strict constraints, yet existing methods have distinct drawbacks: projection-based approaches rely on restrictive assumptions and can distort the constrained distribution, while particle methods construct intermediate potentials from point predictions of the terminal state, ignoring posterior uncertainty. Moreover, many practical constraints are non-differentiable or black-box, so gradients are unavailable. We introduce Uncertainty-Aware Sequential Monte Carlo (UA-SMC), a projection-free sampler for hard-constrained diffusion. UA-SMC directly approximates the probability of eventual constraint satisfaction and uses it as a Feynman–Kac potential. A practical time-dependent schedule captures contracting terminal uncertainty while avoiding state-dependent gradients. We analyze how approximation error in these potentials propagates to the intermediate targets and derive a finite-particle error bound. Across porous material design, rare-event sampling for nonlinear dynamics, and copyright-safe text-to-image generation, UA-SMC consistently attains high constraint satisfaction and distributional fidelity for non-differentiable and black-box constraints.

Primary area:
generative models
Keywords:
diffusion, sequential monte carlo, hard constraints

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_D3yCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Uncertainty-Aware Sequential Monte Carlo (UA-SMC), a projection-free framework for sampling from diffusion models under hard constraints. The authors observe that existing particle-based methods often rely on point estimates of the terminal state to construct intermediate feasibility potentials, potentially discarding particles that still have a substantial probability of satisfying the constraint. To address this limitation, UA-SMC interprets the exact intermediate potential as the conditional probability of eventual constraint satisfaction and approximates it using an uncertainty-calibrated sigmoid function. A practical time-dependent sharpness schedule is introduced to avoid estimating state-dependent uncertainty or differentiating the constraint function. The paper also provides theoretical bounds on potential approximation error, its effect on intermediate target distributions, and finite-particle sampling error. Experiments on an analytically tractable distribution, nonlinear dynamical systems, porous material generation, and text-to-image generation demonstrate improvements in constraint satisfaction and task-specific distributional fidelity over several existing methods.

Critical strengths

(1) A practically useful approach to handling uncertainty in constrained diffusion sampling. The paper identifies an important limitation of conventional point-estimate-based feasibility potentials: intermediate predictions may incorrectly eliminate particles that retain a meaningful probability of satisfying the terminal constraint. Interpreting the exact potential as a conditional feasibility probability provides a principled motivation for uncertainty-aware particle selection. The proposed sigmoid approximation and time-dependent schedule offer a computationally convenient implementation that does not require differentiating the constraint function, making the method applicable to non-differentiable and black-box constraints.

(2) Promising empirical results across diverse constrained-generation tasks. The paper evaluates UA-SMC on both analytically tractable distributions and more complex applications involving physical systems, material generation, and image synthesis. The controlled experiment provides a meaningful assessment of distributional accuracy against a known constrained target, rather than relying solely on constraint satisfaction. The broader experiments further demonstrate that the proposed approach can achieve favorable constraint satisfaction while maintaining task-specific fidelity. These results support the practical relevance of soft feasibility-based particle selection across different domains.

Critical weaknesses

(1) The relationship between the practical sampler and the exact hard-constrained target is not sufficiently established. Although the paper aims to sample from a hard-constrained distribution, the proposed algorithm uses a finite-sharpness sigmoid potential even at the terminal sampling step. Consequently, the limiting target of the implemented SMC procedure is generally a softly reweighted distribution rather than the exact distribution conditioned on feasibility. Increasing the number of particles does not eliminate this approximation bias. Furthermore, the theoretical justification of the Gaussian approximation relies on regularity and nondegeneracy conditions that are not sufficiently characterized. In particular, the first-order approximation can become degenerate when the constraint function has a vanishing gradient, while the paper emphasizes applicability to general non-differentiable and black-box constraints. These issues weaken the connection between the theoretical analysis and the stated hard-constraint sampling objective.

(2) The experiments do not sufficiently isolate the benefits of uncertainty-aware calibration from those of soft potential selection. The practical method replaces state-dependent uncertainty with a manually tuned time-dependent sharpness schedule, making it unclear how accurately the resulting potentials reflect the conditional uncertainty of individual particles. Moreover, a constant-sharpness configuration is used in one of the main application settings, suggesting that some of the observed improvements may arise from the sigmoid potential itself rather than uncertainty-dependent calibration. Although the paper compares against existing particle-based methods, more controlled comparisons against equally tuned constant-sharpness potentials are needed to establish the contribution of the proposed uncertainty mechanism. Additionally, distributional fidelity in several complex experiments is evaluated primarily through the marginal distribution of the constraint value, which does not necessarily establish fidelity of the full constrained sample distribution.

Questions for authors

(1) Exactness of the terminal constrained distribution. Does the implementation apply an exact feasibility indicator, a terminal importance correction, or an additional filtering step after the final sigmoid weighting? As described, the terminal potential remains h~0(x0)=σ(α0(V(x0)−V^))\tilde{h}_0(x_0)=\sigma(\alpha_0(V(x_0)-\hat{V})), where σ\sigma denotes the sigmoid function and V^\hat{V} is the constraint threshold. For finite α0\alpha_0, this assigns positive weight to infeasible samples and therefore does not recover the exact hard-constrained target even with infinitely many particles. Could the authors clarify the actual terminal distribution and quantify the resulting approximation bias?

(2) Validity and scope of the Gaussian approximation. Could the authors clarify the regularity and nondegeneracy conditions required for the Gaussian approximation error to vanish as the posterior uncertainty contracts? In particular, when ∇V(x^0(xt))=0\nabla V(\hat{x}_0(x_t))=0, the variance of the first-order approximation can vanish even though the true constraint-value error remains nontrivial. How does the theoretical analysis handle such cases, and to what extent do the stated guarantees apply when the constraint is non-differentiable or available only through black-box evaluations?

(3) Isolating the contribution of uncertainty-aware calibration. Could the authors provide controlled comparisons between the proposed time-dependent sharpness schedule and an equally tuned constant-sharpness sigmoid potential under identical computational budgets? Since a constant-sharpness configuration is already used in one of the application settings, it is difficult to determine whether the performance gains are attributable to uncertainty calibration or to the use of a smooth feasibility potential. Additionally, could the authors provide stronger distributional-fidelity evaluations beyond the marginal constraint-value distribution to establish that the generated samples approximate the intended constrained distribution?

5 comments
AuthorsOpenReview ↗

Response to Reviewer D3yC (1/3)

We thank the reviewer for their thoughtful comments about our paper and address several of their points below.


1. Exact Hard-Constrained Terminal Distribution (W1, Q1)\Large\text{1. Exact Hard-Constrained Terminal Distribution (W1, Q1)}

We agree that a finite-sharpness sigmoid at the terminal step assigns nonzero probability to infeasible samples, and therefore does not recover the exact hard-constrained target even in the infinite-particle limit. To address this issue, we have revised the UA-SMC algorithm to use the exact feasibility indicator at the terminal step, while retaining the uncertainty-aware sigmoid potentials at intermediate denoising steps:

\begin{cases} \sigma(\alpha_t[V(\hat x_0(x_t))-\hat V]), & t>0,\ \mathbf 1_{C}(x_0), & t=0. \end{cases}$$ This modification restores the exact hard-constrained terminal target in the infinite-particle limit. In particular, because the incremental importance weights are constructed from ratios of consecutive potentials, the intermediate potentials telescope, yielding the terminal distribution $$\pi_0(x_{T:0})\propto p_\theta(x_{T:0})\mathbf 1_{C}(x_0).$$ Thus, under standard SMC consistency conditions, the particle approximation converges to the desired hard-constrained distribution as the particle population increases. We note the exact terminal correction has one failure mode: with finite particles, a run may contain no feasible terminal particles. In this case, every terminal weight is zero and the normalized terminal distribution is undefined. If at least one terminal particle is feasible, the indicator assigns zero weight to every infeasible particle; consequently, every sample drawn from the normalized terminal particle measure is feasible. If no terminal particle is feasible, all terminal weights vanish, and we record the run as a failure rather than defining a normalized output distribution. **Empirical validation.** We repeated our text-to-image experiments with the corrected terminal potential. We report the feasibility rate of the final selected output. Under the exact terminal indicator, whenever the terminal particle population contains at least one feasible particle, terminal resampling returns a hard-constrained output by construction. If no feasible terminal particle is present, all terminal weights are zero and the run is recorded as a failure. For Squirtle, the hard-terminal correction increases the final-output constraint-satisfaction rate from $92\%$ to $94\%$ while TOPIQ was essentially unchanged (0.7660 versus 0.7650). This improvement occurs because for two runs, the original finite-sharpness terminal potential selected constraint-violating outputs despite the presence of feasible particles. The indicator potential instead assigns these infeasible particles zero weight and selects only feasible ones. The remaining $6\%$ of runs contain no feasible terminal particles. For Hulk, the final-output constraint-satisfaction rate remains $97\%$, with no decrease in TOPIQ. We have updated Algorithm 1 to make the terminal correction and its implications explicit. ---
AuthorsOpenReview ↗

Response to Reviewer D3yC (2/3)

2. W2, Q3: Isolating the contribution of uncertainty-aware calibration.\Large\text{2. W2, Q3: Isolating the contribution of uncertainty-aware calibration.}

To address the reviewer’s concern about the contribution of the time-varying schedule, we conducted an additional ablation on the analytical experiment presented in Appendix B.2. We compared 27 logarithmically spaced constant-sharpness configurations against the original 27 time-dependent schedules. Apart from the sharpness parameterization, all configurations use the same sigmoid intermediate feasibility potential, terminal correction, diffusion model, number of diffusion steps, particle count, sample count, and resampling criterion. Additionally, this setting permits direct evaluation of total variation (TV) distance to the full two-dimensional hard-constrained target distribution, rather than relying only on the marginal distribution of constraint values.

MethodSharpnessCS ↑\uparrowTV ↓\downarrow
Minimum-TV constantα≈0.005247\alpha \approx 0.0052470.628±0.0100.628 \pm 0.0100.406±0.0090.406 \pm 0.009
Perfect-CS constantα≈63.096\alpha \approx 63.0961.000±0.000\mathbf{1.000 \pm 0.000}0.473±0.0040.473 \pm 0.004
Tuned schedule0→1000\mathbin{\to}100, p=10p=101.000±0.000\mathbf{1.000 \pm 0.000}0.227±0.0150.227 \pm 0.015
Oracle sharpnessState-Dependent αt⋆\alpha_t^\star1.000±0.000\mathbf{1.000 \pm 0.000}0.192±0.010\mathbf{0.192 \pm 0.010}
The constant-sharpness configurations exhibit a clear constraint-satisfaction--fidelity trade-off. The minimum-TV constant achieves a mean TV of 0.4060.406 but only 62.862.8 %\% constraint satisfaction; conversely, the displayed constant-sharpness configuration achieving perfect constraint satisfaction has a mean TV 0.4730.473. The practical time-dependent schedule however, attains both 100%100\% constraint satisfaction and a mean TV of 0.2270.227, only slightly worse than the oracle state-dependent sharpness which achieves a mean TV of 0.1920.192, without requiring derivatives or state-dependent uncertainty estimation. These results show that the improvement is not attributable solely to sigmoid smoothing; the time-dependent schedule achieves a substantially better feasibility--full-distribution-fidelity trade-off than any of the constant schedules.

AuthorsOpenReview ↗

Response to Reviewer D3yC (3/3)

3. Regularity Conditions for Gaussian Error.

We appreciate the opportunity to clarify this point. To contextualize the new assumption, recall that our Gaussian approximation is based on the Taylor expansion

V(X0)=V(x^0(xt))+(∇V(x^0(xt)))⊤(X0−x^0(xt))+Rt(X0)V(\mathbf{X}_0) = V(\hat{x}_0(x_t)) + \big(\nabla V(\hat{x}_0(x_t))\big)^\top (\mathbf{X}_0 - \hat{x}_0(x_t)) + R_t(\mathbf{X}_0) where Rt(X0)R_t(\mathbf{X}_0) collects the terms that are nonlinear in Tr⁡Σt(xt)\sqrt{\operatorname{Tr}\Sigma_t(x_t)}.

To state the regularity condition needed for EtGauss(xt)→0\mathcal{E}_t^{\mathrm{Gauss}}(x_t) \to 0 as t→0t\to 0, we define νt(xt)=Rˉt(xt)τt(xt)\nu_t(x_t)=\frac{\bar{R}_t(x_t)}{\tau_t(x_t)} where Rˉt(xt)=E[∣Rt(X0)∣∣Xt=xt]\bar{R}_t(x_t)=\mathbb{E}\left[\left|R_t(\mathbf{X}_0)\right|\mid \mathbf{X}_t=x_t\right] is the expected magnitude of the nonlinear remainder and τt2(xt)=(∇V(x^0(xt)))⊤Σt(xt)∇V(x^0(xt))\tau_t^2(x_t)=\big(\nabla V(\hat{x}_0(x_t))\big)^\top\Sigma_t(x_t)\nabla V(\hat{x}_0(x_t)) is the variance of the linearized constraint-score error.

We revisited the derivation of EtGauss(xt)\mathcal{E}_t^{\mathrm{Gauss}}(x_t), which now shows that this error is controlled by νt(xt)\sqrt{\nu_t(x_t)}; see the blue text in Appendix C. We therefore assume νt(xt)=O(TrΣt(xt)).\nu_t(x_t)=\mathcal{O}\left(\sqrt{\text{Tr}\Sigma_t(x_t)}\right). As TrΣt(xt)→0\mathrm{Tr}\Sigma_t(x_t)\to 0 when t→0t\to 0, this condition implies that EtGauss(xt)→0\mathcal{E}_t^{\mathrm{Gauss}}(x_t)\to 0.

This assumption lets us state the required condition directly in terms of the relative magnitude of the nonlinear Taylor remainder, rather than separately imposing first-order nondegeneracy. The latter setting is studied in [1, 2], which show that more refined techniques are generally needed when the first-order term is degenerate.

We also note that Algorithm 1 itself requires only evaluations of VV and is therefore applicable to black-box or non-differentiable constraints. The regularity condition above only makes explicit the local regime under which our Gaussian approximation admits a vanishing-error guarantee.


We hope these new results and revisions address the reviewer's concerns, and we would be grateful if they could be considered in the final assessment. We are happy to address your further comments.

Reviewer_6Ad2OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper use uncertainty to weight SMC in diffusion sampling for constrained generation.

Critical strengths

The paper is easy to follow.

Critical weaknesses

See questions.

Questions for authors

The uncertainty-awared propery should be regarded as a main contribution. The formulation of (3) is noval to me. However, the implementation later seems to be much more trivial than (3). In real implementation, the author applies a Gaussian/sigmoid estimation of uncertainty of x_t. Although this is intuitive, this approximation looks coarse. In (3), there is a \epsilon_t, which is good because it means in high t, the point-mass estimation may not satisfy constraint but the final sampled x_0 may do. But in real implementation, this term is completely ignored. Can the author explain how the algorithm estimate \epsilon_t?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6Ad2

We thank the reviewer for acknowledging the novelty of the uncertainty-aware potential and appreciate the opportunity to clarify how the prediction error ϵt\epsilon_t is incorporated into the practical algorithm.

Our method does not ignore ϵt\epsilon_t. We approximate its conditional distribution ϵt∣Xt=xt\epsilon_t|X_t = x_t to derive an oracle potential whose particle-dependent sharpness reflects conditional constraint-score uncertainty. The practical algorithm then amortizes this state-dependent sharpness with a time-dependent schedule.

To derive the oracle potential, recall from Eq. (3) that the exact intermediate potential is given by ht(xt)=Pr⁡(ϵt≤rt(xt)∣Xt=xt),h_t(x_t)=\Pr(\epsilon_t\leq r_t(x_t)\mid X_t=x_t),

where rt(xt)=V(x^0(xt))−V^r_t(x_t)=V(\hat{x}_0(x_t))-\hat V is the predicted constraint margin and ϵt∣Xt=xt\epsilon_t|X_t = x_t is the conditional constraint-score prediction error.

Under a Gaussian approximation to the posterior X0∣Xt=xt≈N(x^0(xt),Σt(xt)),X_0| X_t=x_t\approx\mathcal{N}(\hat x_0(x_t),\Sigma_t(x_t)), A first-order linearization argument allows ϵt\epsilon_t to be approximated as a Gaussian ϵt∣Xt=xt≈N(0,τt2(xt))τt2(xt)=∇V(x^0(xt))TΣt(xt)∇V(x^0(xt))\epsilon_t|X_t=x_t \approx \mathcal{N}(0,\tau_t^2(x_t)) \quad \tau_t^2(x_t) = \nabla V(\hat x_0(x_t))^T \Sigma_t(x_t) \nabla V(\hat x_0(x_t)) where τt(xt)\tau_t(x_t) characterizes the conditional uncertainty in the constraint prediction. Under this approximation, the potential in Eq. (3) can be expressed as a Gaussian CDF ht(xt)≈Φ(rt(xt)τt(xt)).h_t(x_t)\approx \Phi\left(\frac{r_t(x_t)}{\tau_t(x_t)}\right). Approximating the Gaussian CDF with a sigmoid yields the oracle potential

\quad \alpha_t^\star(x_t)=\frac{c}{\tau_t(x_t)}.$$ As presented in [1], with $c=1.702$, this replacement is uniformly accurate: $$\sup_{u\in\mathbb R}\left|\Phi(u)-\sigma(cu)\right|\approx 0.00944.$$ Thus, with $\alpha_t^\star(x_t)=1.702/\tau_t(x_t)$, the oracle sigmoid differs from the Gaussian-CDF potential by at most $0.0095$ pointwise. The remaining practical approximation replaces this unavailable state-dependent sharpness with a time-dependent schedule. Hence, $\epsilon_t$ affects the potential through its conditional uncertainty, rather than through explicit samples of the prediction error. While the oracle potential requires estimating state-dependent uncertainty, our practical algorithm replaces it with a time-dependent schedule $(\alpha_t)_{t=0}^T$ which serves as a computationally inexpensive proxy for the typical contraction of uncertainty during denoising. More specifically, at early denoising stages, when the conditional uncertainty of $\epsilon_t$ may be large, smaller values of $\alpha_t$ yield softer potentials, preserving particles whose predicted terminal states may currently violate the constraint but could ultimately become feasible. As denoising progresses and uncertainty contracts, larger values of $\alpha_t$ yield more selective potentials that increasingly favor constraint satisfaction. Consequently, the practical algorithm retains the uncertainty-aware motivation of Eq. (3) through a time-dependent approximation to the evolution of prediction uncertainty, without requiring explicit sampling or estimation of $\epsilon_t$ at each denoising step. Moreover, as presented in Appendix B.2, for the analytically tractable Gaussian-mixture experiment, the practical schedule attains TV $0.227 \pm 0.015$, compared with $0.192 \pm 0.010$ for the oracle calibration, with both attaining $100\%$ constraint satisfaction. This ablation shows that the time-dependent schedule recovers much of the performance of the oracle sharpness, without requiring state-dependent uncertainty estimation. [1] Camilli, G. Origin of the Scaling Constant d=1.7 in Item Response Theory. Journal of Educational and Behavioral Statistics, 19(3):293–295, 1994. --- We hope these new results and revisions address the reviewer's concerns, and we would be grateful if they could be considered in the final assessment. We are happy to address your further comments.

We thank the reviewer for their thoughtful feedback on our paper. We hope that our rebuttal response addressed your concerns. If we have addressed them, we would appreciate it if you would consider raising your score. Otherwise, please let us know what your objections are at your earliest convenience so we can address them before the rebuttal deadline.

Reviewer_ZWi1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a new sequential monte carlo (SMC) to sample from a prior diffusion model given hard constraints, which, in particular, can be written in the form of V(x) > \hat V where x is the generated data, V is a given function, and \hat V is a pre-specified value. The core idea is to construct a sigmoid approximation to the intermediate SMC target. This sigmoid approximation accounts for variance, which prior SMC works do not usually consider. However, the practical algorithm relies on heuristic, diffusion state and time-dependent schedule for the approximation, and the experiment mainly tune the hyperparameter using a grid search. The paper also presents theoretical error control analysis. Experiments include toy Gaussian mixtures, trajectory sampling, porous material generation and image generation. The main metrics considered are constraint satisfication rate and the TV distance in the constrained value space (i.e. V(x) instead of x), where the proposed method demonstrates advantages in most of cases.

Critical strengths
  • I found the method well-motivated to incorporate uncertainty to improve Tweedie based estimator in SMC.

  • The proposed algorithm is simple to implement.

Critical weaknesses

While I found incorporating the uncertainty in designing intermediate target meaningful (as has been explored in prior works, e.g. [1]), there are three main concerns of the paper:

(1) Limited novelty and scope. The proposed method is largely based on the diffusion SMC literature and the biggest methodological novelty is consider a different potential function that accounts for the uncertainty, while the construction itself is quite heuristic. In addition, the paper also considers a specific type of hard constraint.

(2) Empirical results. The proposed method involves choosing 3 hyperpameters; it seems like these hyperaprameters are chose by gird search to optimize constraint safistifications and TV (both are test metrics, and TV would require oracle distribution which may not be available in practice). There is not additional validation split. In comparison, other methods do not seem to have a related hyperparameter search; moreover the added hyperparameter tuning increases the compute cost of the proposed method to baselines.

(3) Presentation. I found there is significant room for improving the presentation clarify in particular for the core construction of the method. E.g. τt\tau_t the assumed variance component in line 219 is not clearly defined in the main text. And all practical approximations are deferred to the appendix.

(40 Less informative theoretical results. While the theory shows a more-or-less standard results on error control, it is not established on the practical algorithm. So I found it hard to extract meaningful takeaways from it.

[1] Tweedie Moment Projected Diffusions For Inverse Problems. Benjamin Boys, Mark Girolami, Jakiw Pidstrigach, Sebastian Reich, Alan Mosca, O. Deniz Akyildiz

Questions for authors
  • Can you explain how you define τt\tau_t and in practice how you choose αt\alpha_t when you don't have an oracle metric?

  • Why do you consider a sigmoid approximation, instead of a Gaussian approximation directly?

  • Why do you only consider the constrain generation problem in the form of V(x) < \hat V? Is there any part of the method that prevents it from extending to general hard constraints?

7 comments
AuthorsOpenReview ↗

Response to Reviewer ZW1i (1/3)

We thank the reviewer for the feedback. We respectfully clarify several points where we believe the review mischaracterizes the method, its scope, and its theoretical connection to the practical algorithm. We believe these misunderstandings lead to your low rating on our paper. Below, we clarify these key points and address your detailed comments.


1. Novelty (W1)\Large\text{1. Novelty (W1)}

We respectfully disagree with the characterization of UA-SMC as merely an application of prior Tweedie-based SMC. Prior diffusion-SMC methods evaluate on a Tweedie point estimate, 1C(x^0(xt))\mathbf{1}_{\mathcal C}(\hat{\bf x}_0(\bf x_t)). In contrast, UA-SMC starts from the exact intermediate potential ht(xt)=Pr⁡(X0∈C∣Xt=xt),h_t(\bf x_t)=\Pr(X_0\in\mathcal C\mid X_t=\bf x_t), the conditional probability that the eventual terminal sample is feasible. Our method then derives an uncertainty-calibrated approximation to this probability through the conditional uncertainty of the scalar constraint score. Thus, UA-SMC reweights particles according to their probability of eventual feasibility, rather than according to whether a point prediction is feasible. This distinction is precisely what prevents premature elimination of particles when the terminal state remains uncertain. Further, the practical algorithm amortizes the unavailable state-dependent calibration into a time-dependent schedule and, at sampling time, requires only evaluations of VV. It therefore applies to non-differentiable and black-box constraints, unlike gradient-guided approaches that require differentiating through the constraint or a differentiable surrogate. We make this distinction from [1] explicit in the main text.

[1] B. Boys et al. Tweedie Moment Projected Diffusions For Inverse Problems.


2. Generality of Hard Constraint (W1, Q3)\Large\text{2. Generality of Hard Constraint (W1, Q3)}

We respectfully clarify that the reviewer conflates the algorithmic scope with the regularity assumptions used only for the local Gaussian analysis. As illustrated in Appendix A, any feasible set C\mathcal C can be represented by a binary membership oracle V(x)=1C(x)V(x)=\mathbf{1}_{\mathcal C}(x) with threshold V^=1/2\hat V= 1/2 . Thus, Algorithm 1 is applicable to any hard constraint for which a membership oracle can be evaluated, including non-convex, non-differentiable, and black-box constraints.

The Gaussian approximation analysis is more restricted: it assumes regularity of VV and hence does not apply to a discontinuous binary membership oracle. This does not limit the algorithmic applicability of UA-SMC; it delineates the setting in which our approximation bound applies. We will state this distinction more explicitly.


AuthorsOpenReview ↗

Response to Reviewer ZW1i (2/3)

3. Hyperparameter Search and Baseline Fairness (W2)\Large\text{3. Hyperparameter Search and Baseline Fairness (W2)}

Tuning without access to oracle. While the analytical experiments permit evaluation of TV distance against the exact constrained distribution, this oracle information is not required for hyperparameter selection in practical applications. As stated in Appendix B.1, we select configurations according to “constraint satisfaction and the task-specific fidelity metric.” For example, in our text-to-image experiments, we selected the sharpness schedule using constraint-satisfaction rates along with the TOPIQ fidelity metric. We clarified this procedure in the appendix.

Parameter sweep protocol. We agree that schedule selection and evaluation should be performed on separate validation and evaluation seeds and have updated our schedule selection protocol accordingly. We reran the schedule selection ablation for the analytical example (Appendix B.2) using disjoint validation and test seeds and the optimal selected schedule remained (0,100,10)(0,100,10). Thus, the selected schedule and qualitative conclusion are unchanged under a fully held-out evaluation protocol.

Baseline fairness. The hyperparameter tuning was not exclusive to UA-SMC. In particular, for the dynamical systems experiment, we conducted additional sweeps over FK Steering's potential formulations and ensemble size to identify competitive configurations. Thus, our baseline comparisons do not rely on arbitrarily selected or untuned configurations; rather, we made a deliberate effort to optimize the performance of each baseline under comparable computational budgets.

Comparison to existing methods. We acknowledge that this tuning incurs an additional one-time offline cost beyond the matched inference-time NFE budgets. However, this type of task-specific offline hyperparameter selection is not unique to UA-SMC. For example, DAS [2] (ICLR 2025 Spotlight) uses a two-stage offline grid search to select its KL coefficient defining its target distribution for sampling.

[2] S. Kim et al. Test-time Alignment of Diffusion Models without Reward Over-optimization.


4. Uncertainty Aware Potential Structure (W2, Q1)\Large\text{4. Uncertainty Aware Potential Structure (W2, Q1)}

We now clarify the uncertainty-aware potential and practical schedule. Let μt(xt)=x^0(xt)=E[X0∣Xt=xt]\mu_t(x_t)=\hat{x}_0(x_t)=\mathbb{E}[X_0|X_t=x_t]

denote the posterior mean, and define the predicted constraint margin and terminal-score error as rt(xt)=V(μt(xt))−V^, ϵt=V(μt(xt))−V(X0).r_t(x_t)=V(\mu_t(x_t))-\hat V, \ \epsilon_t=V(\mu_t(x_t))-V(X_0). Under our inequality-constraint representation, terminal feasibility is equivalent to V(X0)≥V^⟺ϵt≤rt(xt).V(X_0)\geq\hat V\quad\Longleftrightarrow\quad\epsilon_t\leq r_t(x_t). Therefore, the exact intermediate potential is ht(xt)=Pr⁡(ϵt≤rt(xt)∣Xt=xt).h_t(x_t)=\Pr(\epsilon_t\leq r_t(x_t)|X_t=x_t ). Using a local Gaussian approximation, X0∣Xt=xt≈N(μt(xt),Σt(xt)),X_0|X_t=x_t\approx\mathcal{N}(\mu_t(x_t),\Sigma_t(x_t)), and linearizing VV at μt(xt)\mu_t(x_t), V(X0)≈V(μt(xt))+∇V(μt(xt))⊤(X0−μt(xt)).V(X_0)\approx V(\mu_t(x_t)) + \nabla V(\mu_t(x_t))^\top \bigl(X_0-\mu_t(x_t)\bigr). gives ϵt∣Xt=xt≈N(0,τt2(xt)),\epsilon_t|X_t=x_t\approx\mathcal{N}(0,\tau_t^2(x_t)), where τt2(xt)=∇V(μt(xt))⊤Σt(xt)∇V(μt(xt)).\tau_t^2(x_t)=\nabla V(\mu_t(x_t))^\top\Sigma_t(x_t)\nabla V(\mu_t(x_t)). Thus, τt(xt)\tau_t(x_t) is the local posterior standard deviation of the scalar constraint score, giving ht(xt)≈Φ(rt(xt)τt(xt)).h_t(x_t)\approx\Phi(\frac{r_t(x_t)}{\tau_t(x_t)}). We then approximate this Gaussian CDF with a scaled sigmoid, yielding the oracle-sharpness potential h~t⋆(xt)=σ(αt⋆(xt)rt(xt)),αt⋆(xt)=cτt(xt),\widetilde h_t^\star(x_t)=\sigma(\alpha_t^\star(x_t)r_t(x_t)),\quad \alpha_t^\star(x_t)=\frac{c}{\tau_t(x_t)}, Our analytical example in Section 5.1 permits computation of this oracle sharpness and demonstrates improvements over SOTA projective diffusion approaches.

For black-box or non-differentiable constraints, directly estimating ∇V(μt(xt))\nabla V(\mu_t(x_t)) is infeasible. We therefore replace the state-dependent oracle sharpness with a practical time-only schedule: αt=αmin⁡+(1−tT)p(αmax⁡−αmin⁡),t>0,\alpha_t=\alpha_{\min}+\left(1-\frac{t}{T}\right)^p(\alpha_{\max}-\alpha_{\min}),\qquad t>0, This schedule, parameterized by a minimum sharpness, maximum sharpness, and power, mirrors the typical contraction of uncertainty. We select (αmin⁡,αmax⁡,p)(\alpha_{\min},\alpha_{\max},p) by grid search on held-out validation seeds using constraint satisfaction and task-specific fidelity. Crucially, after this schedule selection, the practical potential h~t(xt)\widetilde h_t(x_t) only requires evaluations of V(μt(xt))V(\mu_t(x_t)), making UA-SMC directly applicable to hard constraints specified by non-differentiable or black-box functions, for which the state-dependent Gaussian approximation and gradient-based guidance methods are not directly available. We have moved this derivation and the definition of τt\tau_t into the main text to make the practical construction clearer.


AuthorsOpenReview ↗

Response to Reviewer ZW1i (3/3)

5. Theoretical Contribution (W4)\Large\text{5. Theoretical Contribution (W4)}

We respectfully disagree that the theory is disconnected from the practical algorithm. Proposition 1 directly bounds the discrepancy between the exact conditional-feasibility potential ht(xt)h_t(x_t) and the practical potential h~t(xt)\tilde h_t(x_t) used in Algorithm 1. The bound explicitly separates the error from the local Gaussian and first-order score approximation, and error from replacing the resulting state-dependent sharpness by the sigmoid/time-dependent calibration.

The subsequent results propagate this potential error to the intermediate Feynman--Kac targets and combine it with a finite-particle SMC bound. Thus, the analysis characterizes both approximation error in the practical UA-SMC potentials and the additional Monte Carlo error incurred by finite-particle sampling.


6. Why not use the Gaussian CDF directly? (Q2)\Large\text{6. Why not use the Gaussian CDF directly? (Q2)}

We agree that the Gaussian CDF is the direct potential obtained under the local Gaussian and first-order approximation, however we note that the Gaussian CDF can be well approximated with a sigmoid. In particular, [2] presents a uniform approximation bound for our chosen constant c=1.702c=1.702

sup⁡u∈R∣Φ(u)−σ(cu)∣≈0.00944.\sup_{u \in \mathbb{R}} |\Phi(u) - \sigma(c u)| \approx 0.00944.

Hence, with αt⋆(xt)=1.702/τt(xt)\alpha_t^\star(x_t)=1.702/\tau_t(x_t), the sigmoid oracle potential differs from the Gaussian-CDF approximation by at most 0.00950.0095 pointwise. The sigmoid also provides the convenient sharpness parameterization σ(αtrt)\sigma(\alpha_t r_t), which lets us replace the unavailable state-dependent inverse uncertainty 1/τt(xt)1/\tau_t(x_t) with a simple time-dependent schedule. Our theoretical bound explicitly includes this small link-approximation error.

[2] Camilli, G. Origin of the Scaling Constant d=1.7 in Item Response Theory. Journal of Educational and Behavioral Statistics, 19(3):293–295, 1994.


We hope these new results and revisions address the reviewer's concerns, and we would be grateful if they could be considered in the final assessment. We are happy to address your further comments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.