- Rating
- 4
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes a new method for inference-time prior adaptation of diffusion models for simulation-based inference. Prior work has achieved this with approximate corrections to the score (assuming Gaussian structure). This paper shows that for a class of prior ratio changes (isotropic Gaussian-like functions), the corrections can be obtained exactly. They extend this to ratios that can be represented by positively weighted sums of such functions by sampling separately with each function in proportion to weights that are estimated via sampling. The method improves over previous methods on a wide variety of benchmarks, with significant speedups over existing diffusion inference-time adaptation methods.
Critical strengths
- The paper is well written, and showcases it's core contributions clearly.
- The method is a natural improvement over prior approaches (PriorGuide), and improve over it significantly, especially in runtime and with narrow target priors. The exact single-factor transport and component-weight estimation for mixture prior ratios are clever and conceptually clean tricks, although the single-factor transport is limited to isotropic ratio changes.
Critical weaknesses
It would be appropriate to include citations to some recent papers that have noted mathematically similar results as the "exact single-factor transport" in this paper:
[1] Moitra, Ankur, Andrej Risteski, and Dhruv Rohatgi. "Steering diffusion models with quadratic rewards: a fine-grained analysis.", 2026
[2] McDonald, Curtis. "Technical Note on Relating Scores of Tilted Distributions.", 2026
Questions for authors
- How much upfront cost does the component weight estimation take (Eq. 16)? I would also like to see analysis on how the variance of the estimator changes with increasing problem dimensionality, given that importance sampling can become more difficult in that situation.
- Do you think that [1] or [2] could suggest interesting extensions of the method in this work?
4 comments
Response to Reviewer fF2R (2/2)
(2/2)
Estimator variance versus dimensionality
We agree that dimension matters, and we have added Appendix H.3 to address it. It computes the error of the path-space estimator exactly in a Gaussian setting and compares it with runs of the implemented estimator. In short, the error grows with dimension when every parameter receives the same factor, more reverse steps offset this growth, and the implemented estimator follows the exact error wherever that error is moderate.
For one estimator seed with proposal trajectories, the estimate is unbiased for the finite-chain normalizer . When its second moment is finite,
where is the proposal path law and is the base path law tilted by the factor at its endpoint. The square root of this ratio is the root-mean-square relative error of , which we call its relative error below.
We evaluate the divergence exactly for a base posterior and a zero-centered factor , with exact scores and the paper's noise schedule. The values and shrink the posterior variance by factors of 3 and 13. The table gives the relative error for and the population ESS benchmark at a budget of 4,096 trajectories, .
| Time points | Relative error () | Population ESS (of 4,096) | |||
|---|---|---|---|---|---|
| 40 | 20 | 400 | 0.217 | 0.73% | 3,364 |
| 40 | 64 | 400 | 0.877 | 1.5% | 2,182 |
| 240 | 20 | 400 | 1.54 | 1.9% | 1,612 |
| 240 | 64 | 400 | 18.8 | 6.8% | 207 |
| 240 | 64 | 1280 | 1.50 | 1.9% | 1,637 |
Both path laws factorize over the coordinates, so , where is the divergence for a single coordinate. With the same factor on every parameter, the error therefore grows with . More reverse steps reduce the mismatch per coordinate. With 20 time points per parameter, the error stays at 1.9% up to (last row), at a proportional increase in score evaluations. The growth reflects the total shift between component and base posterior. If is lowered with so that the KL divergence of the component from the base posterior stays fixed, the error falls slightly, from 1.9% at to 1.7% at . Direct estimation from base samples, the alternative estimator in the paper, has a far larger error at the same budget, which barely changes as reverse steps are added (Figure 5 in Appendix H.3).
We also ran the implemented estimator with 63 seeds in 18 configurations of this setting (Table 4 in Appendix H.3). Its error follows the same trends and agrees with the exact value to within 25% wherever the exact error is moderate, including the paper's 400 time points up to . When the exact error is large, the spread across seeds is usually smaller than the exact error and the empirical ESS is higher than the population benchmark, because most runs miss the rare trajectories with very large weights. The exact calculation therefore shows where a run's own diagnostics would mislead. For the component weights that Spectra uses, which average over three seeds, a two-component version with weights near 0.73 and 0.27 gives a mean weight error of at most 0.0091 for with 400 time points.
These results assume exact scores and zero-centered factors. They do not cover learned scores or the discretization gap between and .
We hope these revisions address the reviewer’s questions and would be glad to clarify any remaining points.
I thank the authors for the detailed reply! I am happy with the answers, and maintain my score.
Thank you for your time and constructive feedback. We are glad our responses addressed your questions.
Response to Reviewer fF2R (1/2)
(1/2)
We thank the reviewer for the positive assessment and for pointing out these relevant connections and practical questions.
Related work and possible extensions
We agree that the mathematical overlap with Moitra et al. [1] and McDonald [2] should be acknowledged explicitly. We have added both references in Section 3.3 and Appendix A.
Specifically, Moitra et al. [1] establish the exact score identity for linear exponential tilts, corresponding to the κ=0 case of our formulation. McDonald [2] extends this identity to isotropic exponential–quadratic tilts with nonnegative κ. After converting between variance-preserving and variance-exploding Gaussian-noising parameterizations, the latter identity is equivalent to the corresponding case of our single-factor score identity. Spectra applies this identity to prior adaptation of frozen amortized posteriors, extends it to κ<0 and positive mixtures, and characterizes the factors that admit exact transport (Theorem 2).
Regarding possible extensions, the Hubbard–Stratonovich construction of Moitra et al. [1] writes a low-rank positive-semidefinite quadratic tilt as a continuous mixture of linear tilts. Since each linear tilt is transportable, this suggests extending Spectra beyond the isotropic single-factor family. Appendix I now gives a closely related construction for target covariances that exceed an isotropic covariance along a few directions, whose number of components grows with the number of such directions and not with the dimension, and Appendix N.3 evaluates it in 10 and 20 dimensions. Their hardness result for rank-one negative-semidefinite tilts, which tighten the distribution along a single direction, also shows that the complementary case is intractable in the worst case when the tightening is very strong. The main practical challenges are discretizing the continuous mixture and reliably estimating its observation-specific component weights. We have noted this direction in Appendix A.
McDonald's result [2] also suggests extending the implementation to diffusion-SBI models trained with variance-preserving Gaussian noising.
Upfront cost of component-weight estimation (Eq. 16 → 17)
The path-space estimator is an upfront, observation-specific computation. Our practical configuration uses 4,096 proposal trajectories per component and estimator seed, three estimator seeds, 400 reverse time points, and no Langevin corrector. Each transition requires two frozen-score evaluations, amounting to approximately 19.6 million trajectory-wise score evaluations for K=2. For a fixed per-component budget, the weight-estimation cost scales linearly with K.
The table reports measured times in seconds for K=2 on one A100 GPU, excluding compilation, as medians over the two prior–observation pairs of Appendix M.2. Online sampling generates 1,000 posterior samples using (N,L)=(100,0). PriorGuide refers to the benchmarked implementation and includes its per-observation precomputation.
Weight estimation dominates first-use latency, making Spectra-PS more expensive than PriorGuide for the first 1,000 mixture samples. However, the estimated normalizers can be reused for a fixed observation, pretrained model, and factor dictionary when only the mixture coefficients change, without repeating weight estimation. Appendix M.2 now also reports the setup–sampling trade-off against PriorGuide (VJP), including the number of repeated batches after which Spectra has lower cumulative cost.