acceptodds
Under review as a conference paper at ICLR 2027

MA-SBI: Computable Ceilings Decide Which Misspecification Corrector to Fit

Abstract

Simulation-based inference degrades when the simulator misspecifies the process that produced the real observations, and the existing repairs ask for data the practitioner rarely has, such as ground-truth parameters paired with real observations, a functional form for the noise, or a credible alternative prior. What practitioners often do have is a written record of the conditions under which each observation was collected, such as the instruction text given to participants or the policy bulletin in force that week. We show that the share of the simulator's error this side information can remove has a ceiling that is computable in advance from held-out residuals, without evaluating any posterior, and that a second ceiling says whether the correction also needs to see the observation. We introduce misspecification-aware simulation-based inference (MA-SBI), a family of observation-space correctors that the two ceilings select between, fitted in front of a frozen amortised posterior. In the Gaussian-linear model the conditional mean provably attains both ceilings, and the fitted correction is identically zero when the simulator is well specified, however the prior behaves. A second result separates the estimators by sample complexity, since the structured one pays for the rank of the embedded design where an unstructured one pays for the width of the encoder. Across four benchmarks and three physical datasets the ceilings are tight, with every arm landing between 0.99 and 1.0 of its own, so the number tells a practitioner what an estimator will achieve before one is fitted. On a wind-tunnel benchmark a corrector that uses no parameter labels matches a calibration-based method that uses them, and we report the negative results too, including an epidemiological dataset whose apparent coverage gain does not survive our pre-registered test.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
simulation-based inference, model misspecification, neural posterior estimation, amortized inference, side information, calibration, uncertainty quantification

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_WjwQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes context-conditioned observation corrections for a neural posterior estimator under simulator misspecification. Two population residual-prediction ceilings are intended to guide the choice between context-only and observation-dependent correctors. The paper also develops pairing-free variants and evaluates the approach on synthetic benchmarks, physical datasets, and negative controls.

Critical strengths
  1. The practical motivation is relevant. Recorded operating conditions may help explain systematic simulator discrepancies, and correcting inputs while reusing a pretrained posterior estimator is a convenient workflow.

  2. The paper reports substantial failures and limitations. The negative physical and epidemiological results, together with the Gaussian-linear analysis of bootstrap attenuation, provide useful information about the method’s scope.

Critical weaknesses
  1. The proposed ceilings do not adequately justify the corrector selection rule. The ceilings quantify optimal residual prediction in observation space, but their relevance to posterior inference is only theoretically established under restrictive Gaussian-linear assumptions. It remains unclear whether a larger ceiling translates into better posterior accuracy or calibration for general nonlinear NPE models. Furthermore, the estimated ceilings depend on fitted regressors and may underestimate population predictability due to fitting error. Finally, Algorithm 1 in Appendix B.4 stops when Cz≈0C_z \approx 0, even though Cyz>0C_{yz} > 0 may still make observation-dependent correction useful.

  2. Observation-space correction does not establish reliable posterior repair. The proposed correctors minimize residual prediction error in observation space, but the experiments do not sufficiently demonstrate that this translates into reliable posterior inference. Real-data improvements are limited and inconsistent, while reductions in posterior-mean error do not establish calibrated uncertainty. The reported coverage results raise further concerns about whether the method reliably improves the full posterior distribution.

Questions for authors
  1. Why use ceiling-based selection instead of directly comparing the two correctors on held-out calibration data? Since estimating the ceilings already requires fitting residual regressors, it is unclear what advantage the proposed selection rule provides over standard validation-based selection. Does it improve selection accuracy or reduce computational cost?

  2. Does the proposed correction reliably improve posterior calibration rather than only point estimation? The real-data experiments emphasize posterior-mean error, while empirical coverage remains poor or inconsistent in several settings. Can the authors clarify whether the corrected posteriors consistently improve coverage over the uncorrected NPE, and explain the remaining calibration failures?

  3. Is the necessity condition in Theorem 5 correct? Consider a parameter space containing only two values and two contexts with positive probability. Suppose the two parameters induce identical conditional distributions in one context but distinct distributions in the other. The joint distributions of ((y,z)) are then distinguishable, although slice-wise injectivity fails on a positive-measure set of contexts. Does this contradict the claimed equivalence?

Reviewer_9zn4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors proposed to correct for misspecification in SBI, when such misspecification is in the form of an additive residual from real observation process and simulation, by predicting that residual using side information.

Critical strengths
  • It makes use of additional information in a way that can correct at least a large family of misspecifications.
Critical weaknesses
  • I am unsure I understand where the supervision for training the residual correction model is from. Which seems to require access to the ground-truth simulation-observation residual, but if we know which parameter would generate the observation, there is no need for SBI, if we do not, then the "ground truth" residual depends on the parameters used to generate simulated data. It might be just me not understanding the setup. In that case, it is easily fixable by some writing changes.
Questions for authors
  • Reference on SPIN (Anonymous, 2026) need to be corrected. The given arxiv number is pointed to https://arxiv.org/abs/2605.05652, which does have SPIN method so I would say it is legit but need correction.
  • I don't think I fully understand how δ\delta is modeled. It is said as a more general"misspecification" but seems more like a residual term. So is the idea that we learn predicting a "clean" simulated data by subtracting δ^(z)\hat{\delta}(z)? This is only one notion of misspecification. Is it possible for some misspecification to have the form that the projection between θ\theta and yy is off that cannot be written as an additive residual.
  • For linked triplet zi,yobs,i,ysim,iz_i, y_{obs,i}, y_{sim,i}, what θ\theta is used to simulate y∼,iy_{\sim,i}? This seems important as that θ\theta would be treated as the ground truth parameter associated with yobs,iy_{obs,i} no?
    • A related question is where the ground truth δ\delta used in training come from? If I have a way to know the difference between yobsy_{obs} and ysimy_{sim}, should I absorb δ\delta into my simulator and do usual SBI?
  • Does the residual δ\delta depend on yobsy_{obs}? In eq 1, δ∗\delta^* is only a function of zz but at the start of sec.4, it is also a function of yobsy_{obs}.
Reviewer_cNPzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Misspecification-Aware Simulation-Based Inference (MA-SBI), which uses auxiliary contextual information, such as experimental protocol text, to correct observations before passing them to a frozen SBI posterior estimator. Its central contribution is a pair of conditional-mean-based ceilings on residual error reduction for context-only and observation-and-context correctors, together with a rule for selecting between them. The theoretical analysis relates these ceilings to parameter estimation risk under Gaussian-linear assumptions and examines sample efficiency and correction without paired observations.

Validation includes synthetic misspecification on four SBI benchmarks, generalization to unseen textual formulations, ablations, and evaluations on wind-tunnel, light-tunnel, and epidemiological data. The authors report near-ceiling residual correction in controlled settings with sufficient calibration data, while real-data results are mixed and depend on the dataset, corrector, and supervision available. The method targets recovery of the simulator posterior at an uncorrupted simulated input, rather than guaranteeing the correct posterior under the real data-generating process.

Critical strengths
  • S1: The paper offers an interpretable framework for distinguishing information limits from estimation error in discrepancy correction. Its two residual correction ceilings characterize the best attainable performance from context alone and from observations together with context. This distinction helps explain when limited gains reflect insufficient predictive information rather than an inadequately learned corrector, providing a useful analytical perspective on simulator misspecification.

  • S2: The empirical analysis provides meaningful evidence about the method’s applicability and limitations. The evaluation extends beyond controlled SBI benchmarks to textual generalization and real physical and epidemiological datasets, explicitly documenting failures under unseen conditions, limited real-data gains, and poor posterior coverage despite improved parameter estimates. This transparency makes the empirical claims easier to assess and helps establish where the method may be useful without overstating its reliability.

Critical weaknesses
  • W1: The novelty and practical value of the ceiling-based selection framework remain insufficiently demonstrated. The core method performs conditional residual regression before applying a frozen posterior estimator, and its theoretical characterization builds largely on conditional-mean projection identities and linear regression risk analysis. These results are informative, but the paper does not establish what the ceiling-based framework adds over standard residual correction with held-out model selection. Since estimating the ceilings itself requires residual regression, a direct comparison is needed to show whether the proposed rule selects correctors more reliably or efficiently. Without this evidence, the substantive benefit of the paper’s central contribution remains unclear.

  • W2: The independent contribution of natural-language semantics is not adequately isolated. The synthetic experiments explicitly associate contexts with discrepancy mechanisms, allowing a model to succeed by identifying the context and learning its corresponding residual. Although the paraphrase experiments demonstrate linguistic generalization, the true-context one-hot control achieves 100% gap closure, compared with 34.4% for the default sentence encoder with ten training paraphrases per context (Table 14). This does not negate the usefulness of text when context labels are unavailable, but it leaves unclear whether language contributes beyond recovering context identity. Matched comparisons with structured experimental metadata are needed to establish the language representation’s independent value and clarify the source of the reported gains.

Questions for authors
  • Q1: The evidence for reliable posterior inference is weaker than the evidence for residual correction. MA-SBI minimizes residual squared error and passes a point-corrected observation to a frozen posterior estimator, without explicitly propagating uncertainty in the correction. The exact connection to parameter estimation risk relies on Gaussian-linear assumptions and does not establish posterior calibration for general nonlinear models. Thus, approaching a residual correction ceiling or improving parameter estimates is insufficient to demonstrate reliable posterior uncertainty. Although the authors acknowledge this limitation, it materially restricts the significance of the results for SBI, where uncertainty quantification is a central objective.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.