acceptodds
Withdrawn from ICLR 2027

Learning Iterative Updates to Sampling Distributions from Density-Ratio Feedback

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative models enable inference from incomplete observations by sampling plausible underlying structures from learned distributions. However, a mismatch between the sampling distribution and the target posterior can leave regions with substantial posterior probability represented by only a few samples. Such mismatch can increase posterior-mean estimation error under a fixed number of target density evaluations. In this work, we introduce the Ratio-Guided Proposal Compiler (RGPC), a learned model that iteratively updates the sampling distribution to reduce distribution mismatch and posterior-mean estimation error. To guide these updates, RGPC compares the density of the target posterior with the current sampling density at trial samples, then combines the resulting density-ratio feedback with aggregate constraints to predict changes to the sampling distribution's parameters. Across 32 Public Use Microdata Areas (PUMAs), we evaluate population inference on development tasks defined by subsets of American Community Survey (ACS) aggregate statistics, with all methods approximating the same target posterior for each task. We measure distribution mismatch using squared maximum mean discrepancy (), which compares distributions through their samples. For posterior-mean estimation error, we use total variation distance (TVD) between the estimated and reference posterior means. RGPC achieves the lowest and TVD among all nine evaluated methods, including conditional generators and conventional sampling baselines. When using the same number of target density evaluations, RGPC reduces by 42.65% and TVD by 26.50% relative to the best-performing baseline. Code is available at https://anonymous.4open.science/r/RGPC-FD4B/.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Posterior approximation, Adaptive sampling, Density-ratio feedback, Population inference, Generative model

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_T7S3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Generative models enables population mean inference by sampling from the learnt distribution. The sampling distribution can differ from the target posterior. This paper proposes the Ratio-Guided Proposal Compiler (RGPC) method, which iteratively updates the sampling distribution toward the target posterior to address this problem.

Critical strengths

The proposed RGPC method achieves small distribution mismatch across 32 Public Use Microdata Areas and outperforms the chosen baselines.

Critical weaknesses
  1. My main concern is that the core research question appears to be a well-studied classical statistics problem. There have been extensive literature addressing the core research question, and the paper doesn't sufficiently clarify what knowledge gap it addresses. For example, in the related work section, the paper states that Frogner & Poggio (2019) uses a Bayesian inference method to compute a posterior mode, while RGPC computes the posterior mean. It is unclear to me what the significant gap is between computing population mean and population mode. Population mean appears to be a more regular and standard statistics than population mode. As far as I understand it, there are abundant literature on Bayesian inference that address the issue of population mean.
  2. The paper is poorly written and hard to follow
Questions for authors
  1. What is the novelty of the research problem?
  2. What is the novelty/advantages of the proposed method compared with the existing methods?
Reviewer_Kd6vOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers posterior inference over population tables. Each table is a vector p∈Δ95p \in \Delta^{95} of 96 cell proportions, constrained by observed marginals Aop=yr,oA_o p = y_{r,o} for region rr and observation pattern oo. The domain of the inference is the polytope (after removing cells forced to zero) Fr,o={,p∈Δ95:Aop=yr,o,}.\mathcal{F}{r,o} = \{, p \in \Delta^{95} : A_o p = y{r,o} ,\}.

A Gaussian prior ρr(p)\rho_r(p) over tables, fitted to census microdata (Public Use Microdata Sample, PUMS) of other regions (App. B.2), is restricted to it, π(p);∝;ρr(p),1[p∈Fr,o].\pi(p) ;\propto; \rho_r(p),\mathbf{1}{[p \in \mathcal{F}{r,o}]}. The posterior π\pi is thus a truncated Gaussian over tables. The target quantity is the posterior mean μ=Eπ[p]\mu = \mathbb{E}_\pi[p].

The authors propose the Ratio-Guided Proposal Compiler (RGPC), which places a Gaussian-mixture proposal qq plus a Student-tt tail on Fr,o\mathcal{F}_{r,o} and updates its parameters T=16T=16 times using a network trained across tasks and frozen at inference. Each update uses log-density ratios log⁡π(p)−log⁡q(p)\log\pi(p) - \log q(p) at 64 trial tables, i.e. 1,024 target density evaluations in total. The posterior mean is estimated by the mean of samples from the final proposal.

The method is evaluated on 192 tasks: 32 regions (US census Public Use Microdata Areas, PUMAs) ×\times 6 observation patterns. Both metrics are computed against a Hamiltonian Monte Carlo (HMC) reference: a Gaussian-kernel (empirical) MMD2\mathrm{MMD}^2 between the sampled tables and HMC samples, and the posterior-mean error 12∣μ^−μref∣1\tfrac12|\hat\mu - \mu^{\mathrm{ref}}|_1 (called TVD by the authors).

At this budget, the paper reports the lowest values on both metrics for RGPC on the development tasks, compared to other learned methods and conventional baselines, including its closest competitor, an adapted Distribution Transformer (PDT; Whittle et al., 2026). A hold-out on 32 further regions shows a similar gap to PDT.

Critical strengths

S1. Careful experiments. Within its scope, the evaluation appears well executed. The paper contains several ablations, results are reported across multiple seeds with paired bootstrap intervals, and a hold-out on unseen regions is included. The appendix also reports limitations (e.g. App. C.3, App. D.2), which I appreciate.

Critical weaknesses

W1. It is not clear to me that the proposal approximates the posterior. As far as I understand, the posterior mean is estimated as μ^=1N∑np’n\hat\mu = \tfrac{1}{N}\sum_{n} p’_n with fresh samples p’n∼q(T)p’n \sim q^{(T)} from the final proposal (Eq. 4), so E[μ^]=Eq(T)[p]\mathbb{E}[\hat\mu] = \mathbb{E}{q^{(T)}}[p], and the estimate is only as good as the match q(T)≈πq^{(T)} \approx \pi. The authors note themselves (Sec. 3.1) that increasing NN does not remove bias, and unlike MCMC or importance sampling, the mean thus neither converges to the target nor indicates when it is off. The importance correction of App. A.3 would provide such a diagnostic, but is not used in the main comparison. In the “reuse” setting the paper targets, a failure on a new task (i.e. a new π\pi) would therefore go unnoticed.

Moreover, the Student-tt component keeps its mass τ=0.5\tau = 0.5 and its parameters across all updates (Eq. 7, App. A.2), so half of the fresh samples come from a fixed component centred at the initial approximation:

Eq(T)[p];=;12,Emix(T)[p]+12,Etail[p].\mathbb{E}{q^{(T)}}[p] ;=; \tfrac12,\mathbb{E}{\mathrm{mix}^{(T)}}[p] + \tfrac12,\mathbb{E}_{\mathrm{tail}}[p].

For the unweighted mean to be accurate, the adapted mixture has to compensate for the fixed tail term. A fixed defensive component like this is natural when samples are reweighted, but the main estimator does not use importance weights. That this construction nevertheless yields the lowest TVD seems more consistent with the explanation in W2 than with q(T)≈πq^{(T)} \approx \pi.

W2. Missing baselines and narrow scope. The ablations (Sec. 4.2) perturb individual inputs, but as I see it they do not show how much of the gain comes from q(T)≈πq^{(T)} \approx \pi, and how much from a learned map ϕ(yr,o,Ao,ρr)↦Eq(T)[p]≈μ\phi(y_{r,o}, A_o, \rho_r) \mapsto \mathbb{E}{q^{(T)}}[p] \approx \mu from the task features, acquired offline from posterior samples of similar training tasks. A direct regressor ϕ(yr,o,Ao,ρr)↦μ\phi(y{r,o}, A_o, \rho_r) \mapsto \mu, the unweighted mean of samples from the initial proposal q0q_0, or a gradient-based sampler at equal budget would help to clarify this in my opinion. Also, the generality the authors claim is not clear to me. In particular, the target π\pi is always a truncated Gaussian, which the reference HMC seems to handle well, and only this target family is evaluated, although the title and introduction suggest a general-purpose sampler.

Questions for authors

Clarifications (affect assessment):

  1. How well does the final proposal of the main model match the posterior (e.g. ESS)? I could only find this for earlier checkpoints in the appendix.
  2. How does a simple regressor from the task features to μ\mu compare to RGPC? This would be needed in my opinion to understand how much the density-ratio feedback actually contributes.
  3. How does the total cost, including training, compare to just running HMC on each new task? App. B.5 states that the per-task evaluation counts “are not total-runtime comparisons”; an estimate of wall-clock cost per new task, including the amortized offline cost, would help.

Suggestions (do not affect score):

  • The abstract should say more about what the paper actually solves, which is inference over 96-cell population tables with exact margins, where the target π\pi is a truncated Gaussian. I suggest to either broaden the evaluation or adapt the title and abstract to this setting.

  • “Distribution” is used for pp, π\pi and qq, which I found confusing. Similarly, “TVD” is 12∣μ^−μref∣1\tfrac12|\hat\mu - \mu^{\mathrm{ref}}|_1 between two tables, so I would appreciate if you would say so.

  • The reported MMD2\mathrm{MMD}^2 (Eq. 20) also depends on the number of output samples and their weights, which may favour methods with many equally weighted samples.

  • Since TT is selected on the development tasks (App. B.3), the hold-out (App. B.7) would be the cleaner main result.

  • Some passages need editing, e.g. l. 89–91.

  • I did not check all references, but some citations in the main text do not seem to support the specific claims they are attached to. E.g., Blumenthal et al. (2026), a general review connecting control, transport, inference and thermodynamics, is cited for the mismatch between sampling distribution and posterior in population inference.

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

The paper proposes to reduce the posterior distribution mismatch and reduction of posterior estimation error with Ratio-Guided Proposal Compiler (RGPC). In applications like population inference where aggregated observations change the posterior distribution, RGPC iteratively updates the posterior estimation with the density ratio feedback and current-task constraints to update the posterior parameters according to the observation task. The empirical section reports improvement over multiple benchmarks in comparison to 8 baselines with limited target density evaluations.

Critical strengths

The method address the issue of posterior mismatch and iteratively updating a sampling distribution to approximate target posterior conditioned on the observed aggregate statistic. The updates are shown to reduce the posterior mismatch and improve posterior mean estimation under limited sampling and target evaluation budget.

Critical weaknesses

The provided scheme figures are overcrowded and difficult to follow. The blue legend in fig1 is not presented in the figure. It is difficult to follow the input arrows from all the blocks in fig2 e.g. in the chain Initial proposal -> current parameter state -> current proposal ... -> final proposal, defining 3 different blocks for the proposal seems unnecessary.

The model design is not well-motivated or consistently explained in section 3.2. Some details are missing or need further clarification.

1 - None of the components of eq (6) are defined (e.g. what is the dimension of ckjc_{kj}? is it real?

2 - The definition of eje_j is confusing. eje_j should constrain the direction of population change but not necessarily for retaining the zero set Sr,oS_{r,o} during the updates. So what constraints does eje_j enforce? Also it's addition in Eq (6) doesn't necessarily entail a linear update direction from embedding space to the parameter space.

2 - it is not clear how ukju_{kj} is used to the the Gaussian Mixture parameters, or why eq (6) is

3 - How is the log-density ratio in Eq (5) used to get the parameter updates? Why is objective (8) optimized.

4 - In the paper lines 91-94 and 144 suggest the density ratio is address lack of samples from the target posterior at inference. line 105 suggests that RGPC doesn't require retraining on the new tasks. line 290 suggests that objective (8), an expectation over the target posterior distribution, is optimized to train the posterior parameteres with gradient updates. This is one of the major claims of the paper and has to be written accurately throughout.

Minor:

The sentence "increasing N reduces sampling variability" is fundamentally wrong.

Line 244 base is not defined. is it the task or observation or region?

Questions for authors

1 - What is eje_j representing and how is it connected to the constraints imposed by Sr,oS_{r,o} and what are the example of these variables in the considered experiments.

2 - How complex is the prior in the considered experiments that requires the posterior estimation? Don't we have access to samples from the posterior if we are optimizing (8)?

3 - What is the connection of equation (5) to (6) and (7) and how the update rule is modifying the Gaussian mixture parameters and preserving the constraints?

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

Related papers

Discussion (0)

Sign in to comment.