acceptodds
Under review as a conference paper at ICLR 2027

ENFR: Evidence Networks with Flow-Matched References for Amortized Simulation-Based Model Inference

Abstract

The evidence, , of a data-generating model is a powerful tool in Bayesian frameworks: it enables model inference, model misspecification tests, and goodness-of-fit assessment as an alternative to standard discrepancy statistics, e.g. . While versatile, the evidence is notoriously difficult to compute, even when the likelihood is available. In both likelihood-based and simulation-based inference, estimating provides an immediate solution to high-dimensional goodness-of-fit or model misspecification tests: using as the test statistic induces a minimum-volume level- test. Such tests are essential to assess model fidelity, which is necessary to trust simulation-based parameter inference. Existing learned approaches largely target Bayes factors, which only measure evidence ratios. Instead, we provide accurate, fast, and amortized estimates of the absolute evidence. We present ENFR: Evidence Networks with Flow-matched References. Building on the Evidence Network framework of Jeffrey & Wandelt (2024), the proposed method first learns Bayes factors with respect to a flow-matched reference, from which the absolute evidence is recovered. We show improvement in estimation of over the flow-based baseline in all applications considered and the best performance of all density estimation methods considered in our benchmarks. We furthermore introduce and provide the proof for an architecture that jointly learns the individual evidences of a set of models . We demonstrate the performance and scaling behavior of these approaches on data from a variety of scientific applications.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
simulation-based inference, Bayesian evidence, Bayesian model comparison

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_DzHiOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

ENFR (Evidence Networks with flow-matched references) combines flow matching with evidence networks to estimate the absolute evidence for likelihood-based and simulation-based inference techniques. Estimating the absolute evidence can address model misspecification and goodness-of-fit assessments. The paper proposes ENFR for single models (SingleENFR) and multiple models (MultiENFR). It also introduces a loss function to jointly estimate the Bayesian evidence and Bayesian factors for multiple models (MultiENFR).

Critical strengths
  • Overall, the work is interesting and a valuable contribution: the estimation of the absolute value of evidence has been largely underexplored in the literature, to my knowledge.
  • The work contains a wide range of tasks, building up from simple tasks with analytical solutions to complex neuroscience and cosmological problems.
Critical weaknesses

Figures

  • Figure 1: It would be good to adjust the colors and style of the first figure, since the colors strongly resemble what AI models typically generate (although I am not implying that the figure is directly LLM-generated).
  • Figures 3 and 4: Please add plot titles to the figures to increase clarity. For instance, for Figure 4, titles such as "Hodgkin-Huxley" and "Galaxy assembly bias" would make it visually clearer.

Language

  • Line 157: "We choose flow matching for this task". Do you mean throughout all the tasks/experiments in the paper?

Experiments

  • The baseline comparison with Flow matching and CTSM variants has not been shown in all experiments (e.g., no CTSM-SB + FlowRef in Figure 3). Please add these baseline comparisons or explain why you did not compare your approaches (SingleENFR/MultiENFR) to them.
  • A calibration test does not seem to me to be enough to validate accurate absolute evidence. Could you explain why you used this as the main conclusion that your method (MultiENFR) works on the Hodgkin-Huxley and Galaxy assembly tasks? Are there other ways to evaluate whether your method works on those tasks?

Reproducibility

  • Please share the codebase via an anonymous link so that the code and results can be verified.

Minor points

  • Line 045: "What we ultimately care about in science, however, is assessing the relative fits of models when confronted with data". It would be better to tone down this statement, since it is not the (only) ultimate goal of science.
  • There are some issues with the references in this paper. Could you double-check whether you have the same issues on your side?
    • Line 042: (e.g. von Wietersheim-Kramsta et al., 2025; Thomsen et al., 2026, …): These two citations are displayed differently from the others (not underlined/in blue).
    • Line 098: (Masnadi-Shirazi & Vasconcelos, 2008): the link to Masnadi-Shirazi is not clickable
    • Line 119: "While Rhodes et al. (2020)": "While" is a clickable link.
    • Line 1090: Adam Kingma & Ba (2014): Should be "(Adam; Kingma & Ba, 2024)"
  • Typos
    • Line 320: damped(β...): no space between "damped" and "("

Overall, I am open to considering a weak accept if the authors share the codebase to reproduce and confirm the results, and if they can provide a justifiable explanation of why the calibration test is sufficient to validate accurate absolute evidence.

Questions for authors
  • Out of curiosity, do you have an intuition for how your approach would work with other generative models (rather than flow matching)?

  • In the limitations section, Line 475, one mentions: "Encouragingly, the evidence-network correction is most effective precisely in the small data regime (Figure 2b)". However, in the figure caption, the simulation budget is fixed at 5 × 10610^6 (Figure 2a) or varies between 10510^5 and 10610^6, which does not seem to be a "small data regime". Could you clarify where this comment comes from?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 3

We thank the referee for their detailed feedback and for highlighting the value of our method for the simulation-based inference program.

W1: Colors in Figure 1

We have changed the colors to a palette that is also colorblind friendly.

W2: Plot titles

We have added plot titles to Figures 3 and 4.

W3: Flow matching language

Yes, in every application in this work we train flow matching. We have clarified the language in the main body.

W4: CTSM in other experiments

We do show CTSM-SB + FlowRef for the other likelihood-based experiments in the corresponding tables 5 and 7 in the appendix. We are not showing them in the same plot for readability. We are now explicitly referring to those tables in the figure caption. We are not showing CTSM-VP as its performance has been clearly worse than CTSM-SB + FlowRef already in the Gaussian example and is uncompetitive.

W5: Calibration test does not validate absolute evidences

We agree with the referee on this point. However, in the absence of a likelihood, there is no independent estimate to compare our simulation-based method to. In order to validate the absolute evidences of ENFR, we ran the likelihood-based benchmarks. We would argue that the simulation-based examples are not benchmarks of the absolute evidence, but research applications for which there is no known ground truth. As the absolute evidences cannot be tested, we instead test the calibration of the resulting evidence ratios.

W6: Publishing code

Please find the anonymized code at this repository: https://anonymous.4open.science/r/ENFR-F041/

W7 "minor points"

We have relaxed the statement about the ultimate goal of science and we corrected the two typos. In general, we could not reproduce the same problems with the pdf file in a standard Firefox browser.

Q1: Other generative models

Our approach is not restricted to flow matching. The essential requirements for the reference model are that we can sample from it and evaluate its density. Consequently, other generative models satisfying these requirements could be used, provided that they yield sufficient overlap with the target distribution.

Q2: Small data regime

For the GMM scaling tests, the smallest data set we have tested and shown in Figure 2b is 5×1045\times 10^4 simulations. We have softened the language on calling this the small data regime.

Thank you for your reply and for sharing the code. I could verify that the reported numbers are consistent with the code. Overall, I think this work presents an interesting and fresh idea for estimating the absolute evidence, which, to my knowledge, has been largely underexplored so far and deserves a place in the literature.

A small point of doubt came up regarding the code. I was wondering why the Flow, CTSM-VP, and CTSM-SB+FlowRef use a default 30-epoch cap in the GMM sweeps (MAX_EPOCHS="${GMM_DIM_MAX_EPOCHS:-30}"), while ENFR uses a default 100-epoch cap (EN_MAX_EPOCHS="${GMM_DIM_EN_MAX_EPOCHS:-100}"). Since the method uses early stopping, I would think that increasing the cap to a much higher value (especially for MAX_EPOCHS) and saving the best models would be a better practice, since this cap might have led to results that have not converged.

I change my score to a weak accept.

AuthorsOpenReview ↗

Response to Reviewer 3

We would like to thank the reviewer for explicitly checking our code and ensuring the reproducibility of our results.

We acknowledge the reviewer’s excellent point concerning the maximum epoch cap in our GMM experiments. We have rerun the flows and evidence networks for the GMM dimensionality scaling experiment, ensuring that every model is early stopped. We can report the following results with the newly trained flows (showing mean ±\pm standard deviation of RMSE/dimension×103\mathrm{RMSE} / \mathrm{dimension} \times 10^{3}):

dFlow (original)SingleENFR (original)Improvement (original)Flow (updated)SingleENFR (updated)Improvement (updated)
6411.93 ± 0.657.05 ± 0.6740.9%11.64 ± 1.196.94 ± 0.7040.4%
12818.00 ± 2.3210.33 ± 1.0842.6%15.92 ± 1.599.39 ± 0.2141.0%
25625.82 ± 0.5411.86 ± 0.6254.1%23.24 ± 0.4311.19 ± 0.6051.9%
51230.84 ± 0.6528.20 ± 2.288.6%29.81 ± 0.1224.31 ± 4.4118.5%
102434.17 ± 0.2332.65 ± 2.284.4%32.28 ± 0.4830.08 ± 1.556.8%

While the flow-only results slightly improve with longer training, the ENFR improvement remains consistent within a few percentage points up to 256 dimensions. At higher dimensions the improvements are slightly larger than in the original experiments. We can confirm that all CTSM-VP runs did in fact early stop.

We commit to including the updated GMM runs in the manuscript along with new CTSM-SB + FlowRef results using the updated flows by camera-ready, if accepted.

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

This paper introduces ENFR, a method for quickly estimating Bayesian evidence when researchers can simulate data from a model but cannot calculate its likelihood. ENFR combines flow matching with an evidence network that corrects errors in the flow's density estimate. The paper presents versions for both single and multiple models and shows that ENFR is generally more accurate than the other tested methods across applications in cosmology, neuroscience, and dynamical systems. It can also support model comparison and detect when a model does not fit observed data. However, the method still requires many simulations and becomes less effective for very high-dimensional data.

Critical strengths
  1. ENFR combines a flow-based density estimate with an evidence network that learns a correction to that estimate, enabling amortized absolute-evidence estimation when the likelihood cannot be evaluated. The paper also presents SingleENFR and MultiENFR variants for handling individual models and sets of competing models, respectively. For MultiENFR, the authors provide a theoretical result showing that the proposed loss recovers pairwise Bayes factors under idealized asymptotic conditions.
  2. The paper evaluates ENFR on synthetic benchmarks and scientific applications from cosmology, neuroscience, and dynamical systems. Where likelihood-based evidence estimates are available, ENFR generally improves on the flow baseline and the other tested estimators. In likelihood-free settings, the authors use calibration tests to assess the resulting model probabilities and demonstrate an application to model-misspecification testing.
Critical weaknesses
  1. As acknowledged by the authors, scientific experiments may require significant amount of training simulation runs per model, which is expensive, especially for slow simulators. The Gaussian-mixture results also show that the flow reference becomes less accurate at high dimensions (~500), where the evidence network provides only a small additional correction.
  2. The paper compares absolute evidence with analytic or nested-sampling estimates in the GMM, BAO, ODE, and drift-diffusion experiments. For the Hodgkin-Huxley and galaxy applications, it instead evaluates the calibration of model probabilities. These tests support the relative model comparisons, but they do not directly assess the absolute evidence because errors shared across models may cancel in the evidence ratios.
Questions for authors
  1. Most scientific applications use/require large training sets. How could ENFR be adapted to settings with much smaller simulation budgets, and which part of the method is likely to become the main limitation?
  2. Can the authors clarify how users should detect when the flow reference is not sufficiently accurate, especially in high-dimensional settings where the evidence-network correction may be limited?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 2

We would like to thank the reviewer for their thoughtful comments and highlighting our theoretical foundations and the performance of our method.

W1: High dimensionality and low training volume

It is true that the performance of our methods gets worse in higher dimensions. We view this decrease in performance as expected, as the simulation density will always be sparse in this regime. Concerning the lower simulation budgets we have tested, we have seen that our correction to the standard flow is even more significant (Figure 2b), making it even more valuable. At the same time we acknowledge that one might eventually encounter a regime in which the flow reference is no longer sufficient.

W2: Calibration does not test absolute evidence

We agree with the referee on this point. The coverage test does not validate the absolute evidences, but rather evidence ratios. However, we note that there is no independently available ground-truth evidence estimate for these applications. For this reason we present a large array of examples where the likelihood is known and the evidence can be estimated by alternative means in order to validate our approach.

Q1: Smaller simulation budgets and limitations

The referee is making an important point, that many simulators may be too expensive to run as many times. For these cases we see the flow as the most limiting part of our method. If the flow does not learn a good reference density, the classifier is limited by the density chasm problem. Possible solutions for this would therefore be to increase the performance and training efficiency of the underlying flow, for example with more intricate interpolants or the inclusion of cheap simulations through multifidelity training. We have added such a comment in the main body.

Q2: Detection of the failure of the evidence correction

The referee is making a valuable point for our manuscript. In general, we have seen that the evidence correction becomes ineffective when the classification accuracy is too high. This is due to an insufficient reference which is far away from the target distribution, and the density chasm problem arises. Overly high classification accuracy on a held-out test set could therefore be used as an indicator for an insufficient reference. We have added this point in the main body in the limitations section and thank the referee for this improvement of our submission.

The reviewer appreciated the authors' detailed response. The response addressed most of my concerns and the edits clarified the practical implications of the method (scalability at high dimension regime and real-sim gap / absolute evidence). My initial assessment already gave the highest score for technical soundness, therefore and I maintain my original score.

Reviewer_ttZJOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper develops a method to estimate the evidence of a scientific model under likelihood-free or simulation-based inference settings. The proposed method, ENFR, estimates the absolute evidence from the learned evidence ratios between proposed models and a flow matching-based reference through a classifier trained to discriminate between simulated vs. the flow-based reference samples. The method is evaluated and achieves good performance on a range of problems with and without likelihoods, the former providing ground-truth likelihood-based evidence estimates for validation, while the latter includes real scientific models such as the Hodgkin-Huxley model.

Critical strengths
  • tackles an important and broad problem, especially in the context of simulation-based inference, as having access to the evidence (or an accurate estimate of it) enables a range of model-critique applications
  • the method is evaluated on a wide range of problems, apparently achieve good performance across the board
  • the problem is well motivated and the introduction is clearly written,
Critical weaknesses
  • the paper is quite dense, especially in its exposition of the main method, and the schematic figure did not really help me with all the mathematical notations within
  • section 4 presents one application after another without any real explanation for the motivation of choosing those problems, other than some likelihood-based ones to obtain a ground-truth evidence for comparison
  • similarly, results are presented for each problem that demonstrates that the method does what it advertises, i.e., accurately estimates model evidence, e.g., for a variety of possible physical models, but due to how densely packed the paper is, I'm genuinely not sure what to take away from it nor can I properly contextualize how big of an improvement the proposed method is
  • more generally, I believe that this is a result of LLM usage, where text is comprehensible at the sentence-level without any strange phrases or common claude-isms, but the transition between subsections or even paragraphs are quite jarring.
Questions for authors
  • I'm a bit confused about the logic of the evaluation here: e.g., Table 1 reports RMSE between estimated and likelihood-based log-evidence (no error bars), while Table 2 reports the estimated log-evidence itself. Why?
  • More generally, if the comparisons for the first set of problems are against the "ground truth" likelihood-based estimates, while the latter can only be validated based on coverage, then why is the proposed method necessary or useful? In general I'm a bit confused about what is learned from using such a method.
  • This is not a specific question, but if it's possible in general to move a few of the problems to the appendix while describing the others (perhaps the more important or exemplary ones) in more detail, that would help understanding tremendously.
  • as noted above, the schematic figure 1 is very difficult to digest. It essentially re-writes all the key equations but in a figure (sorry this is not really a question).
2 comments
AuthorsOpenReview ↗

Response to Reviewer 1

We thank the referee for pointing out the importance of the problem we address and the breadth of our benchmarking efforts, as well as their helpful comments regarding the clarity of our manuscript.

W1: Dense structure and figure improvement

We have extended the discussion of our method and the flow matching approach. We have also restructured our figure, adding more explanatory text and splitting into a training and an inference part.

W2: Unclear motivation for applications

We have picked the GMM due to its analytically known evidence. It is furthermore very easy to adapt for testing the scaling behavior of our method.

We then consider an array of scientific problems for which we do have a likelihood. All three aim to identify models that describe nature: explanations of the accelerated expansion of the Universe, equations of motion of a particle, or information accumulation in the brain. In a Bayesian framework, these questions are addressed with the evidence.

We show two more simulation-based applications where there is no known ground truth evidence to compare to. Those applications also aim at answering scientific questions of model inference: Which ion channels are necessary to explain neuronal firing dynamics, and how do galaxies connect to dark matter?

We now state this distinction of analytic evidence, likelihood-based, and simulation-based applications more clearly in the text in Section 4 and have added more detailed motivations for each of the presented applications in their subsections.

W3: What is the advantage of our method and how big is the improvement?

Our method enables the estimation of the evidence without the need for an explicit likelihood. We show that ENFR achieves the best evidence-estimation accuracy in our benchmarks, outperforming flow matching and CTSM-SB + FlowRef, when we can compare to the likelihood-based estimate.

To demonstrate the usefulness of this increase in precision, we have now added a comparison for the BAO application: We count how often the evidence estimates strongly prefer a wrong model. While the likelihood-based evidence ratios strongly prefer the wrong model 12 times out of 500 test simulations, MultiENFR is comparable, preferring the wrong model 8 times. The flow without the ENFR correction strongly prefers the wrong model 49 times due to its less precise evidence estimates. Our method therefore reduces this error rate by a factor of 6 for this application, showing the necessity of highly precise evidences for reliable model inference and comparison. We have added this insight to the main body and thank the referee for improving the clarity of our manuscript.

W4: Writing style

We have revised the language to make the text more accessible.

Q1: Table 1 RMSE vs Table 2 log evidence

In Table 1, we report RMSE values calculated over 100 held-out simulated data realizations, without quoting additional uncertainty estimates on these RMSE values. Table 2 shows the log evidence estimate of our method on a single real observation. This is because there is only a single realization of this data vector in nature, and we can therefore not compute RMSE values here. We have clarified this in the text.

Q2: Usefulness and necessity of our method

We identify two advantages:

a) In our presented applications we are interested in answering scientific questions about neuronal dynamics and galaxy formation by identifying adequate physical models with the Bayesian evidence. As the likelihood is not tractable for these examples, we have to resort to a simulation-based estimator of the evidence. This is what our method provides. Since a simulation-based estimate is the only option for these scientific problems, we can't directly evaluate the absolute evidence estimates. To test the accuracy of our approach, we have therefore presented the likelihood-based examples. However, we can test the resulting evidence ratios with the coverage test that we present.

b) Even in the case where the likelihood is known, our method enables new analyses, such as the hypothesis test we present in Section 5. This test would be computationally much more demanding with a traditional analysis, as our neural networks are amortized and can provide extremely fast estimates of the evidence for each of the 1000 simulations.

Q3: Moving problems to appendix

We thank the reviewer for this suggestion. We have retained the applications in the main text as they illustrate different aspects of our method, but have expanded the introductory discussion in Section 4 and clarified the motivation for each application to improve readability.

Q4: Figure 1 is difficult to digest

We have restructured the figure and added more explanatory text.

I appreciate the the authors' constructive and clarifications. Just to clarify, I understand the problem and aim here in estimating the model evidence in a likelihood-free setting, as well as to compare with problems where the likelihood is available to benchmark the estimated evidence here. My main confusion was about each individual application, and its goal and evaluation metric or success criteria, due to the sheer number of problems presented across very different disciplines. I think presentation can still be improved, where less is potentially more, but I appreciate the importance and generality of the problem as well as the technicalities of the solution, and will increase my score to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.