acceptodds
Under review as a conference paper at ICLR 2027

Conditional Flow Models Convert Epistemic Uncertainty into Apparent Aleatoric Variance

Abstract

Conditional Flow Models (CFMs) are increasingly being used in place of deterministic predictors across machine learning, including for tasks that were classically formulated as regression. Given their success in such tasks, we ask the question: what useful information do the conditional distributions produced by CFMs carry beyond what a point predictor would provide? We identify a systematic phenomenon: in many natural regimes, CFM samples routinely have larger variance than the true conditional distribution, even in nearly deterministic regression problems. This effect is most pronounced where the model has not resolved the dependence on the conditioning variable, and thus the additional variance can be viewed as converting epistemic uncertainty in fitting the ground truth conditional distribution into apparent aleatoric variance in the generated samples. Through controlled synthetic and real-data experiments, and comparisons with explicit smoothing estimators, we show that the induced variance tracks model underfitting of complex or hard-to-fit targets, and decreases with additional training or data. We provide a theoretical account of the phenomenon in a simplified setting and demonstrate a consequent practical use-case of conditional flow-matching: designing high-coverage predictive sets that cover the ground-truth regression target.

Primary area:
generative models
Keywords:
Flow Matching, Generative Modeling

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_f8pMOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies conditional flow-matching models (CFMs) in regression-type settings and identifies a variance inflation phenomenon: when the learned velocity field underfits its dependence on the conditioning variable, the generated samples have larger conditional variance than the true conditional distribution. The authors attribute this to a smoothing bias in vector-field learning, support it with synthetic experiments (early stopping, explicit roughness penalty, kernel ridge regression) and an offline RL benchmark, and provide an idealized theoretical model based on an equivalent-kernel approximation to regularized flow matching, which yields a bias–variance inequality (Theorem 1). Finally, they show empirically that predictive sets built directly from CFM samples are approximately calibrated in the underfitting regime on policy evaluation, cell perturbation and behavior cloning tasks.
I find the phenomenon itself clearly presented and the synthetic experiments well designed. However, in my opinion the contribution is overstated: the effect is neither specific to flow matching nor surprising from a probabilistic modeling perspective, the theoretical model relies on explicit regularization rather than on the underfitting/early-stopping regime the paper actually argues about, and the experimental section lacks numerical results, comparisons to other probabilistic methods and implementation details. The practical use-case (coverage without conformal correction) is only verified empirically and trades off predictive accuracy of the mean.

Critical strengths

I like the clear identification and illustration of the variance inflation effect, in particular the controlled experiments in Figure 2 and Appendix C that vary three different smoothing mechanisms and show the same qualitative behavior. The idealized model in Section 5 is simple and gives a clean interpretation of the inflated variance as a local mixture of true conditionals, and the correlation between squared bias and excess variance in Figure 18 is a nice empirical check of the theory.

Critical weaknesses

1. Motivation and contribution The main claim of the paper is, in my view, not specific to conditional flow models and not particularly surprising. Any probabilistic model minimizes some divergence or metric between its predictive distribution and the true conditional distribution. If the model class is constrained, e.g. by regularization or limited capacity, one expects it to learn a projection onto that class, which naturally absorbs the unresolved (epistemic) information into the predictive uncertainty. I verified that the same qualitative effect appears for a simple MLP trained with a Gaussian NLL under the same smoothing regimes, so the selling point is really about smoothing of probabilistic models in general rather than about flow matching. The paper should either sharpen what is specific to CFMs or reframe the contribution accordingly.
Related to this, the theoretical model uses an explicit regularization term in (5.1) and a smoothing bias, whereas the paper argues that the same holds for early stopping and the underfitting regime. In practice, one is interested in the optimal point just before overfitting (selected by validation loss), and it is unclear whether early stopping at this point induces a bias strong enough for the effect to be noticeable. Figure 1 with proper early stopping would likely look quite different, and with a proper training protocol one would not stop as early as in the shown examples. The paper also claims that “any” inductive bias producing smoothing has the same effect (l. 83ff.), but only three mechanisms are tested.
Finally, the practical motivation is not clear to me. The proposed use-case sits between underfitting (inflated variance, badly fitted mean) and overfitting (loss of coverage). Reliable coverage is obtained at the cost of predictive accuracy of the mean function, whereas in practice one wants a good mean function together with reliable uncertainty. The coverage analysis is purely empirical, which the authors acknowledge, and the decomposition (A.15) essentially already contains the variance inflation statement, so Theorem 1 adds comparatively little beyond the local-mixture representation.

2. Baselines and experiments The experimental section is in its current form not sufficient to assess the method as an uncertainty quantification approach. There are no numerical results beyond the coverage and width curves, e.g. no accuracy or other metrics and no comparison to other probabilistic methods (apart from DER and HGR at a single checkpoint). Several strong claims rely on visual inspection of figures, e.g. l. 169 (“establish that it is driven by smoothing”), l. 223 (“reveals that the phenomenon is not tied to a single estimator”) and l. 229 (“as can be observed from the figures”), and should be backed by quantitative evidence.
Further, no code is provided, and the synthetic experiments lack a description of the actual implementation, i.e. the MLP architecture and the functions used in Figures 1, 2 and 15–17. Several benchmarks are cited only via preprints.

3. Minor points and style

  • l. 38: "numerous other scientific estimation and time-series problems" is backed by a single reference, I would expect more citations here.
  • Definition 3.1: vv should take three arguments; the notation mixes vt(x∣c)v_t(x \mid c), vt(x,c)v_t(x, c) and vt(c,x)v_t(c, x) throughout the paper. Also, the domain/codomain in the definition appear to be swapped (Rdc\mathbb{R}^{d_c} vs. Rdx\mathbb{R}^{d_x}).
  • Eq. (5.1): the space V\mathcal{V} is never defined.
  • l. 324: the whole motivation of Assumption 5.1 is moved to the appendix; at least 1–2 sentences should remain in the main text.
  • l. 681: missing reference, indicated by a question mark.
  • Assumption A.1: ξ\xi and μ∗\mu^{\ast} are not defined before use.
  • Why is "apparent aleatoric variance" used in the title and throughout? It should be clear that the model's predictive uncertainty is always influenced by the estimation error, so the terminology seems somewhat misleading. This might be worth a discussion with respect to literature on aleatoric/epistemic uncertainty in general.
Questions for authors
  1. In Figures 5–7 and 11–14, how is the end of the underfitting regime (the point where train and validation losses diverge) selected? Is this done qualitatively, or is there a specific mechanism or heuristic behind the shaded region?
  2. Can the authors comment on what is specific to conditional flow models in the variance inflation phenomenon, as opposed to any probabilistic model under a smoothing bias, for example, a proper scoring rule based network (e.g. a Gaussian NLL network), a BNN or a diffusion model?
  3. Does the effect persist with proper early stopping based on validation loss, i.e. at the training point one would actually use in practice, and how large is the resulting inflation in that case?
Reviewer_oXygOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether Conditional Flow Models (CFMs) incorporate useful information that would complement the information provided by a point-driven predictor. It also explores whether CFMs can convert epistemic uncertainty into aleatoric variance. To this end, by assuming an equivalent-kernel approximation, the authors prove that expected sample variance bounds true variance plus the squared bias of the mean. The experiments carried out examine this theory using synthetic regression, three policy-evaluation environments, three cell-perturbation datasets, and PushT behaviour cloning, demonstrating interesting outcomes.

Critical strengths
  • The paper studies a timely and challenging problem in analysing and assessing the uncertainty of AI models.

  • The empirical analysis carried out is well executed. The key observation that similar behaviour occurs across mechanisms provides substantially stronger evidence for the smoothing explanation than a single training trajectory would.

  • The theoretical explanation for the density-weighted mixture construction connects velocity-field approximation to the resulting endpoint distribution. Coupled with the variance decomposition, it gives interesting insights regarding why sample diversity can reflect unresolved conditional structure.

Critical weaknesses
  • Novelty: Although the paper investigates variance inflation specifically for conditional flow matching in point-based predictors, the core insight, i.e., that underfitting leads to sample over-dispersion, is strongly linked to prior work regarding selective underfitting and score-matching inductive biases (e.g., Appendix A - Yoon et al., 2023). While the novelty aspect is present, the distinction with respect to this highly relevant related work is not sufficiently substantiated.

  • Soundness: The core theoretical result (Theorem 1) relies heavily on Assumption 5.1, which assumes a smoothing kernel factorisation process. However, as acknowledged in the text, this approximation weakens significantly in low-noise situations where the conditional distribution localises correctly around the target level sets. Thus, the theoretical assumption does not hold strictly across the full time domain. The strength of those assumptions underpinning the Theorem should be justified further.

  • Rigour: Although the theoretical framework is solid, it is based on the assumption of a population-level regularised objective driven by an explicit linear smoothing operator. In practice, however, the training process of CFMs, typically based on stochastic gradient descent, does not match closely the linear kernel operation underpinning the mathematical formulation. The authors should explain better how closely this matching is.

  • Evaluation: The nominal coverage throughout the underfitting regime deserves a closer look. For instance, in Figure 7, the validation coverage for nominal 75% sets falls below nominal before the shaded underfitting regime ends. Likewise, Figure 12 shows further departures, which entail that smaller sets cannot establish improved efficiency when their achieved coverage is lower.

  • Presentation: The paper is, in general, adequately written. However, it is quite dense in its mathematical formulation, making it difficult for readers to follow the narrative easily. Some equations (e.g., in Section 5) should be further elaborated. The paper also contains some syntactical/grammatical errors (l. 267) and missing references (l. 680).

Questions for authors

(1) How strong are the assumptions (especially Assumption 1) underpinning Theorem 1?

(2) What guarantees that the normalised equivalent kernel in Lemma A.1 is nonnegative? Is an additional assumption needed?

(3) How were the underfitting cutoffs selected?

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

This paper studies how underfitting affects the distributions generated by conditional flow models (CFMs). The authors propose that smoothing across conditioning inputs can cause a CFM to mix probability paths under different conditions. The mixing could carry differences between conditional means into the variance of the generated distribution. With additional assumptions on the smoothing kernel, the paper relates average variance inflation to squared bias in the predictive mean. Experiments on synthetic data, policy evaluation, cellular perturbation prediction, and behavior cloning investigate how predictive variance and the coverage of prediction sets evolve during training.

Critical strengths
  1. The comparisons across early stopping, explicit smoothing penalties, and kernel ridge regression provide sufficient evidence that the observed variance inflation is associated with smoothing.
  2. The interpreation about the local mixture of true conditionals gives an intuitive explanation and a concrete variance–bias relation.
  3. Experiments across policy evaluation, cell perturbation prediction and behavior cloning establish relevance beyond synthetic regression.
Critical weaknesses

The paper reads like overclaiming and the connection between underfitting and the proposed mechanism is only partially established. The first two contribution statements claim variance inflation “whenever” a CFM underfits and under “any inductive bias” that smooths its conditional field. However, Theorem 1 requires a specific density-weighted kernel representation, with a kernel shared across flow time and state, together with stationarity, reversibility, and positive semidefiniteness.

Questions for authors
  1. Is the smoothing bias a well-established concept in flow matching? I would appreciate a dedicated background discuss around it to maker the paper easier to understand.
  2. Related to the Weaknesses, how strong is the assumption for Theorem 1 in practice?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.