- Rating
- 2
- Confidence
- 1
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper studies amortized simulation based inference for hierarchical population models and proposes HAPERA, a flow matching method for estimating joint posteriors over global and unit level parameters. The authors first identify a bias caused by coupling observation windows with the true model parameters and propose a decoupled simulation design to address it. HAPERA then introduces a sensitivity based vector field and an allocation residual that adjusts global and local parameters while preserving unit level parameters to first order. The method is evaluated on synthetic degradation data, an LED dataset, and an indomethacin dataset, using NUTS as the reference posterior. The authors report that HAPERA outperforms existing methods on the synthetic benchmark and achieves near chance C2ST on the real datasets. Ablation studies show that coupled training introduces posterior bias, while the allocation residual provides a smaller improvement in posterior accuracy.
Critical strengths
-
The residual has a clear mathematical construction. Equations~5 to 7 establish the directional cancellation and normalize the correction across population sizes and scales.
-
The window design ablation supports Proposition 1. With matched pairs at three seeds, coupled training raises C2ST by 0.125 at f = 0.05 and by 0.068 when pooled over fractions.
Critical weaknesses
-
The theory is correct but thin. As the authors acknowledge, Proposition 1 applies the proposal prior identity of Papamakarios and Murray (2016) to the observation window. Corollary 1 follows directly from the fact that the ideal estimator equals the true conditional of its own training joint.
-
The paper does not report the simulation and tuning budgets for HNPE and TFMPE alongside those for HAPERA. This omission matters because HAPERA uses fresh simulations throughout training. It remains unclear whether its advantage would persist under matched budgets.
-
The paper does not show that the coupled design is common in practice. It says only that simulators "may" set windows this way, and cites no prior work that sets the window from the true lifetime, so it is hard to judge how often this problem arises. All evidence of harm comes from a simulator the authors built themselves, so the practical significance of the first contribution is unclear.
Questions for authors
-
Why are the prior anchor and design gate needed? Explaining their roles before introducing the equations would make the method easier to assess.
-
What accounts for HAPERA's large gain over HNPE and TFMPE? The methods share the same window design, and the residual ablation shows only a small improvement. What C2ST does the ported TFMPE architecture achieve under the same training pipeline? This comparison would help separate the contributions of the architecture and training procedure.
-
How many simulated populations and how much GPU time did each method use for training? I am also curious how long HAPERA and NUTS take per population at inference. NUTS seems to work well at the current sizes, so it is hard to see when amortization pays off.
-
Are the real data results stable across independent training runs when model selection uses only the fit cohort? Since lifetime predictions remain worse than the naive baseline, what practical benefit does the closer agreement with NUTS provide?
6 comments
Comment 2 of 5 : Response to Reviewer SqGj
On whether the coupled design arises in practice, the introduction of the revised paper (§1) cites published simulators in which the simulated parameters set what is observed, for example an amortized pharmacokinetic estimator trained on simulated studies whose sampling times depend on the peak time and half-life implied by each study's sampled rate constants, and evaluated on clinical studies with their own schedules (Ojeda et al., 2026), and the C-MAPSS engine trajectories, which end at simulated failure in training while test trajectories are cut short (Saxena et al., 2008). In the first, the sampling times play the role of our window and the clinical studies that of deployment: in training the times follow from the simulated parameters, and at evaluation they are the studies' own. The coupled design builds the same dependence into training, and a designer has a reason for it: a fixed window covers more of a fast population's lifetime than of a slow one's, whereas a random fraction of each population's own lifetime makes the observed stages comparable (§1; Figure 1a).
Because your point is that all evidence of harm comes from a simulator we built, an additional test for this response repeats the comparison of the two posteriors of Proposition 1 on a published drug model that is not ours, mrgsolve's pk1 one-compartment model, with a population prior we chose and windows set by the coupled design. No network is trained: for the same data, NUTS samples the posterior under the declared prior, which an estimator should return, and the posterior under the prior conditional on the window length, which the ideally trained estimator of the coupled design returns. At f = 0.05, a classifier separates the two posteriors with an accuracy of 0.723, against 0.500 when both sets of draws come from one posterior (two halves of one NUTS run), on average over the 188 of 200 populations whose NUTS runs passed their convergence thresholds.
The revised paper is also rearranged to read more easily, and every main-text symbol is explained where it first appears or earlier.
Our direct answers to your questions follow.
Q1. Why are the prior anchor and design gate needed, and can their roles be explained before the equations?
A1. The prior anchor lets the network learn only how the data move the posterior away from the prior. The design gate damps the learned output on coordinates the records say little about: those to which the simulated trajectories barely respond at the recorded times (§3.3).
HAPERA moves random starting draws along a vector field, a rule that gives each draw a direction of movement, and flow matching learns this rule (§2.2). The prior anchor is the fixed part of the rule that, by itself, carries standard normal starting draws to the prior. The design gate is computed from the simulator's sensitivity at the recorded times, not from the recorded values, and multiplies each coordinate of the learned output by a value between 0 and 1 before the output is added to the anchor (Eq. 4).
Both are there because a short window carries little information about some parameters, and where the records say little, the posterior stays close to the prior. The anchor builds the prior into the rule, so the network does not have to learn the prior as well. Without the gate, the learned output would act with full strength on every coordinate; with it, the flow stays close to the anchor on the coordinates the records say little about, unless the learned output is large. With no observation, the gate closes and the flow returns the prior (§1; §3.3–§3.4).
On the synthetic benchmark, removing the prior anchor with all else unchanged moves HAPERA (17.6M) farther from the NUTS reference at 3 of 3 training seeds: C2ST pooled over the window fractions increases by 0.0139 on average (Table 17 (c)). Removing the design gate moves it farther at 2 of 3 training seeds, by 0.0115 and 0.0497, and nearer at the first training seed, by 0.0178 (Table 17 (c); §4.4; App. E.6).
The introduction of the revised paper explains both roles before any numbered equation, and §3.3 introduces them in words before Eq. 4.
Comment 3 of 5 : Response to Reviewer SqGj
Q2. What accounts for HAPERA's large gain over HNPE and TFMPE, given the shared window design and the small effect of the residual ablation?
A2. The gain comes from the paper's second contribution, HAPERA's estimator itself, that is, its network architecture together with the three components of its vector field. The comparators share its window design and, under HAPERA's regime, its simulator, prior and training-budget rules, and still lie farther from the NUTS reference.
Three comparisons in the paper size this. The difference to HNPE is about half as large when HNPE's training procedure changes from its authors' recipe to HAPERA's regime: HAPERA's C2ST is lower than that of HNPE (19.8M) by 0.2271 under the recipe, as in the submitted paper, and by 0.1118 when both train under HAPERA's regime (Tables 12 (a), 1; App. E.3). Under one regime, the differences between complete estimators are this 0.1118 and, to TFMPE-Direct (17.8M), 0.0917 (Table 1), and the order holds in the smaller size class, where HAPERA (1.66M) reads 0.557 against 0.601 for HNPE (1.67M) and is nearer at 3 of 3 training seeds (Table 19 (a); App. E.7). Within HAPERA, Table 17 removes one component at a time with the other two in place, and pooled C2ST then changes by at most 0.0497, at one training seed (§4.4; Table 17 (a), (c)).
The differences of 0.1118 and 0.0917 are between the estimators as built and trained: each keeps its own architecture, objective, optimizer and batch size (App. D.3; Table 13 (a)). HAPERA's network is one joint transformer that reads the observations, the units' local offsets and the global parameters as tokens, with 12 layers at token width 256 (§2.2; Table 3). HNPE factors the posterior into a global and a local part, each a normalizing flow (App. D.3). TFMPE-Direct is a tokenized transformer with 5 layers at token width 768, trained by flow matching as HAPERA is (Table 7 (b); App. D.3).
The allocation residual is one of the three components of the estimator, and the three work together in one vector field: the prior anchor supplies the prior, the design gate scales the learned output, and the residual, a learned correction, makes explicit how each unit's parameter is split between population and unit (§3.3–§3.4). The residual head has 526 parameters in a network of 17.6M (Table 3), and on the synthetic benchmark removing it alone moves HAPERA farther from the reference at 3 of 3 training seeds: pooled C2ST increases by 0.0160, 0.0058 and 0.0025, each beyond twice its standard error (Table 17 (a)). On the population location, the only scored group of coordinates that the residual acts on, the coordinate-wise Wasserstein distance to the reference is lower with the residual at 3 of 3 training seeds, by 0.0314, 0.0048 and 0.0032 standard deviations of the NUTS draws (Table 17 (d)).
Q3. What C2ST does the ported TFMPE architecture achieve under the same training pipeline?
A3. The ported TFMPE architecture is scored on calibration, not on C2ST; the C2ST of TFMPE's own architecture under HAPERA's regime is 0.620, against 0.528 for HAPERA (Table 1). That value belongs to TFMPE-Direct (17.8M), TFMPE's posterior network from its authors' implementation, trained directly on simulator-generated populations under the same simulator, prior, window design and training-budget rules as HAPERA (§4.1). TFMPE's published network (821,473 parameters) trained the same way reads 0.615, against 0.557 for HAPERA (1.66M) (Tables 21 (a), 19 (a)).
The port is a different network. It is the port of App. D.5, built to test the window design outside HAPERA's architecture, with token width 256 and 22 encoder layers, where TFMPE-Direct has 768 and 5 (App. D.5; Table 7 (b)). It is trained once under each window design and then checked on windows drawn independently of the parameters: at f = 0.10, its 90% intervals for the first population-location coordinate cover the generating value at a rate of 0.820 under the coupled design, against 0.895 under the decoupled design, with 0.90 nominal (Table 14, rows 3 and 5).
Comment 1 of 5 : Response to Reviewer SqGj
Thank you for your comments on our paper. We have uploaded a revised PDF, which incorporates the comments of all reviewers, reviewed together.
Our benchmarks are models whose likelihood, the probability of the recorded data given all parameters, has a closed form. Their posterior can be sampled directly with the No-U-Turn sampler (NUTS), a Markov chain Monte Carlo method, and these samples, drawn under the same model, prior and window as every network, are the NUTS reference (§4.1). An amortized estimator such as HAPERA is trained on simulations alone and never evaluates the likelihood, so this kind of estimator also applies to simulators whose likelihood cannot be written down, where NUTS cannot be run; the closed-form benchmarks are what let us score it against a reference. The paper's main score is C2ST, the accuracy of a classifier that tries to separate a network's draws of the global parameters (the population-level parameters shared by all units) from the NUTS draws: it is 0.5 when the classifier cannot tell them apart, and higher the farther the network lies from the reference (§4.1). The window fraction f, from 0.05 to 0.20 on the synthetic benchmark, is the window's length as a fraction of a population lifetime drawn from the prior independently of the population observed.
The revised paper reports the simulation and training budgets side by side (Table 7 (a); Tables 13 and 20). Every amortized comparator of Tables 1–2 is now trained under HAPERA's regime: fresh simulations at every optimizer step, with HAPERA's per-epoch simulation budget, epoch cap and stopping rule (§4.1). The advantage persists under these rules. On the synthetic benchmark, HNPE (19.8M; M: millions of network parameters) reads 0.640 and TFMPE-Direct (17.8M; TFMPE without its surrogate stage) reads 0.620, against 0.528 for HAPERA (17.6M) (Table 1). HAPERA is nearer the reference in 24 of 24 cells, one per comparator, training run and window fraction (four values of f), each beyond twice its standard error (Table 9).
On tuning, HAPERA's optimization settings and network size (Table 3) were fixed in advance, and no run varied them to select them. Under HAPERA's regime, HNPE and TFMPE keep the optimizer, step size and batch size of their authors' recipes (Tables 13, 20), except that a short probe before training chose the batch size and step size of TFMPE at its published size on indomethacin (App. E.7).
On the theory, Proposition 1 starts from the identity of Papamakarios and Murray (2016) for training on parameters drawn from a proposal instead of the prior, and adds the case in which the window rule, not the parameter sampler, creates the proposal. Under the coupled design, which sets each training window to a fraction of the simulated population's own lifetime, the simulated parameters still follow the model's declared prior, so the standard correction, the ratio of the declared prior to their distribution, equals 1 and changes nothing (§3.2). Yet a long window signals a slow population, so an ideally trained estimator treats the window length as evidence. It then returns a different posterior from the one wanted at deployment, where the window is chosen independently of the parameters; HAPERA's decoupled design draws its training windows that way. The standard calibration check, a rank check on the training simulations, cannot detect the difference (Corollary 1).
The new Figure 2 shows both results on a scalar model: for the same observation, the means of the two designs' posteriors lie 0.36 apart, 0.81 standard deviations, and both rank histograms are uniform. On the synthetic benchmark, training under the coupled design, all else equal, places HAPERA farther from the NUTS reference at every window fraction and training seed, one seed per training run (Figure 3a); at the shortest window fraction, f = 0.05, its posterior of the first population-location coordinate lies farther by 0.55–0.60 standard deviations of the NUTS draws at 3 of 3 training seeds (§4.4; Table 16).