acceptodds
Withdrawn from ICLR 2027

High-Dimensional Simulation-Based Inference in Latent Spaces

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural simulation-based inference (SBI) has been widely successful in inferring a relatively small number of interpretable parameters from potentially high-dimensional observations, such as images or time series. Accordingly, representation learning in SBI has focused almost exclusively on compressing the observations used to condition the posterior. More recently, however, SBI has begun to target increasingly high-dimensional parameter spaces, raising the complementary question of whether the inference target itself should be compressed. Our answer is a practical merger of SBI and latent generative modeling, which learns a low-dimensional representation of the simulator parameters, performs posterior inference directly in this latent space, and maps posterior samples back to the original parameter space. We characterize the conditions under which latent-space inference recovers the desired target posterior and systematically study its empirical trade-offs. Across four case studies and three generative families, we compare latent and standard estimators while controlling for network capacity, regularization, optimization, and training compute. At matched training compute, latent-space inference achieves accuracy and marginal calibration comparable to direct target-space inference while sampling up to more than an order of magnitude faster.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
latent generative models, simulation-based inference, amortized Bayesian inference, neural posterior estimation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_X2KrOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors of this paper investigate neural simulation-based inference and the tradeoffs between flow/diffusion-based inference in data vs. latent space. They translate the language of latent diffusion models into an SBI framework and demonstrate its efficacy on four tasks, primarily comparing SBI in data space to SBI in latent space. They visually assess conditional data reconstructions as well as quantitatively assessing calibration of their models.

Critical strengths

The motivations behind this paper are relatively clear, and there is in my opinion a large visual improvement in the conditional latent diffusion to generate satellite images from maps (fig. 4).

Critical weaknesses

The primary weakness of this paper is that it is a straightforward application of conditional latent diffusion/flow-matching models, rather than a new method/promising set of experiments/theoretical result. The theoretical analyses are a reframing of prior data compression results in the authors' mathematical language, but do not provide additional insight into the "simulation-based inference" aspect of the approach.

Additionally, the rationale for some of the authors' choices is sometimes hard to follow. For instance, the first stage is framed as an InfoVAE but the encoders trained are all β\beta-VAEs. The InfoVAE aspect doesn't seem important for the actual implementation, but is the focus of much of section 2.

Questions for authors

Just after Eq.2, there's a p(θ∣x)p(\theta|x) that shows up. Where does this come from? Should it be qξ(θ∣x)q_\xi(\theta|x)?

Reviewer_kjWGOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies simulation-based inference with high-dimensional parameter spaces. It first trains an InfoVAE on prior samples, freezes the autoencoder, and learns a conditional generative model in the latent space. Posterior samples are then decoded into the original parameter space. The paper provides a KL error decomposition and evaluates four case studies using flow matching, diffusion models, and normalizing flows. Its main claim is that latent inference offers substantially faster sampling while maintaining comparable accuracy and marginal calibration.

Critical strengths
  1. The paper addresses a relevant practical problem: reducing the computational cost of posterior inference over high-dimensional targets.
  2. The training budget and model-size comparisons include the autoencoder. Accounting for this additional component makes the comparison more informative.
  3. The experiments include exact posterior comparisons, reconstruction controls, and latent-space C2ST diagnostics. These provide useful evidence about the limitations introduced by the autoencoder.
Critical weaknesses
  1. The theory does not directly cover the implemented sampler.

(a) Proposition 1 analyzes a stochastic Gaussian decoder, whereas the experiments use deterministic decoding. These define different posterior approximations, and the stated KL decomposition and bound do not automatically transfer to the implemented sampler.

(b) Proposition 2 also requires sufficient decoder flexibility to eliminate the decoder term. A fixed-variance Gaussian decoder generally cannot satisfy this requirement merely by increasing the capacity of its mean network, as the proof itself acknowledges.

The main text should clearly distinguish these idealized results from guarantees for the proposed method.

  1. The speedup in the experiment remains questionable. Except for the correlated Gaussian experiment, Table 1 combines accuracy obtained with adaptive solvers and speedups measured on untrained networks with fixed solver steps. Actual sampling time depends on both the cost of each network evaluation and the number of evaluations required. Therefore, the reported speedups do not establish how much faster latent inference achieves the reported posterior accuracy. The authors should measure accuracy and sampling time using the same trained models and solver settings.

  2. The calibration comparison is problematic for the correlated Gaussian experiment. The element-wise metrics include 1,792 structurally zero coordinates out of 2,080. Their exact posterior marginals are point masses, whose central credible intervals have coverage one rather than the nominal level. Under ordinary central credible intervals, these point-mass marginals have coverage one at every nominal level. The stated coverage-error metric therefore do not necessarily need to be minimized by the exact posterior. Unless degenerate marginals are handled separately, which is not described, Table 1 cannot reliably support the claimed calibration improvement.

  3. Experiment now suggests that the posterior quality is impaired for latent inference. From standard inference to latent inference, C2ST is strontly increasing in Fashion-MNIST and Gaussian random fields; see Table 1. The authors acknowledge these discrepancies and attribute much of the error to the autoencoder. However, the decoder is part of the posterior estimator, so this attribution does not remove the error. Comparable marginal metrics therefore do not justify the broader conclusion that posterior quality is retained or improved.

  4. The paper is poorly organized and contains inconsistent presentations. Table 1 and Section 4.3 do not make clear that posterior accuracy and sampling speed are measured using different sampling protocols; this distinction appears only in Appendix E. Compression ratios also use different normalizations without explanation: Figure 5 defines the ratio as 64/d, whereas Table 3 reports a 16-dimensional code with a ratio of 130:1 for the 2,080-dimensional target. These evaluation details and definitions should be clarified alongside the relevant results.

Questions for authors
  1. What are the actual sampling times and numbers of network evaluations for the trained models and adaptive solvers used to obtain the posterior-accuracy results in Table 1? Do the reported speedups still hold?

  2. What calibration error does the exact posterior obtain under the Table 1 evaluation protocol? Can the authors report coverage and calibration error on the eight block variances, excluding the structurally zero coordinates?

  3. Which theoretical guarantees apply to the deterministic decoder actually used in the experiments? What additional assumptions are needed for the decoder term in Proposition 2 to vanish?

Reviewer_EW9nOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies latent-space amortized simulation-based inference (SBI). It first trains an InfoVAE on prior samples of high-dimensional targets to encode them into compressed latent representations, and then trains a conditional generative model in the latent space. Posterior samples are mapped back through a deterministic decoder. The paper presents a posterior/code/decoder error decomposition and compares latent and target-space estimators across four case studies under matched training budgets and total parameter counts.

Critical strengths
  1. The proposed approach is practically motivated, and the empirical evaluation is well executed.

  2. The proposed approach can accommodate different generative models, making it easy to apply to existing SBI frameworks.

Critical weaknesses
  1. The studied SBI problems appear artificial. It would have been more interesting to see how the proposed method performs in more complex or real-world high-dimensional SBI problems.

  2. The compression ratio appears a free hyperparameter while strongly influence the inferential quality, which makes it hard for practioners to use when facing diverse concrete problems.

  3. The results confirm preservation of selected marginal statistics, but show substantial discrepancies in the final conditional distributions, e.g. in the Fashion-MNIST and the Gaussian random fields experiments.

Questions for authors
  1. What information can/must the latent representation actually retain?

  2. How does one decide the acceptable compression level when facing different problems?

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

Related papers

Discussion (0)

Sign in to comment.