acceptodds
Under review as a conference paper at ICLR 2027

The ε-Bridge: Dimension-Dependent Bridge-Noise Selection for Stochastic Interpolants

Abstract

Choosing between deterministic Flow Matching (FM) and stochastic bridge-based methods is a basic design decision in continuous-time generative modeling, and the two perform differently across tasks. Stochastic interpolants place both in a common family, yet there is little quantitative guidance on when, and how much, bridge noise helps. We study this question through the -bridge, a one-parameter Brownian-bridge stochastic-interpolant family that recovers FM at . We show that the interpolant, conditional velocity, and fixed-predictor objective converge to their FM counterparts as (Theorem A); that the raw bridge correction energy grows linearly in the ambient dimension (Proposition B); and that, under explicit dimension-uniform posterior-moment and contraction conditions, the marginal-velocity regularity bound contains no trace factor (Theorem C). The asymmetry between these two scalings motivates a two-term error model whose minimizer yields a dimension-dependent noise-selection rule (Model D). Under a fixed training and evaluation protocol on Gaussian mixtures, the selected noise decreases with dimension and reaches the FM endpoint at a crossover that is reproduced on an independent target family, and it moves toward zero as the inference budget grows. PCA-CIFAR, pixel-space image, tabular, molecular, robotics, and point-cloud experiments show the same ordering along dimension, but no single calibrated boundary transfers between families. The crossover instead tracks the fit quality of the FM baseline: bridge noise helps essentially only where that baseline is well fit, and changing model capacity at fixed dimension moves the boundary accordingly. Dimension is therefore a search coordinate, and the FM baseline is an operational diagnostic of whether the search is worth running.

Primary area:
generative models
Keywords:
flow matching, stochastic interpolants, diffusion bridge, Schrödinger bridge, bridge matching, generative models

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

The paper studies how to choose a single scalar noise parameter ε in stochastic interpolants, with ε=0 corresponding to flow matching. It proposes a dimension-dependent selection rule and evaluates noise choices through grid searches across several tasks, finding that the benefit of noise also depends on model capacity and the experimental setup.

Critical strengths

The paper aims to explain sensitivity to ε and finds that it depends on the input space, model capacity, and dataset type. We learn little, if anything, beyond this broad dependence.

Critical weaknesses

No theoretical novelty: the construction and FM connection underlying Theorem A are already established in [1] (§§4.2–4.3), and the theorem mainly formalizes a straightforward limit.

The preferred ε depends on the input space, model capacity, dataset type, and sampling budget. Despite the focus on selecting one scalar, we still end up doing a grid search, with no demonstrated advantage over a small search. I see this primarily as a negative result about transferable ε selection, while the abstract’s framing suggests more practical progress than is established.

[1] Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. "Stochastic Interpolants: A Unifying Framework for Flows and Diffusions" Journal of Machine Learning Research, 2025.

Questions for authors

What evidence shows that the proposed ε-selection rule improves performance or reduces tuning cost compared with a small grid search (common practice)?

Reviewer_FRg6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the ε-bridge stochastic interpolation framework and investigates the relationships among bridge noise intensity, data dimensionality, and model fitting capacity through theoretical analysis and experiments across multiple tasks. Its main contribution is the identification of a model-capacity-dependent dimensional transition in the benefits of bridge noise. However, it has yet to establish a noise selection rule with reliable predictive capability across tasks.

Critical strengths

The theoretical analysis provides a clear framework for studying the problem of noise selection. Theorem A, Proposition B, and Theorem C address the FM limit, the O(dε)O(d\varepsilon) growth in the energy of the bridge correction term, and conditional regularity bounds for the velocity field, respectively. Although these results are insufficient to directly derive sampling error bounds, they clearly distinguish the costs introduced by noise from its potential smoothing effects, providing a useful analytical foundation for future research.

Critical weaknesses
  1. In the nine experiments with predictions recorded in advance in Table 6, Model D correctly predicts only 5 out of 9 cases, performing even worse than the simple baseline of always selecting FM (6/9). Moreover, its original fixed-parameter calibration could not be reproduced. More importantly, Eq. (9) predicts greater benefits from noise as the fitting-error coefficient AA increases, whereas the experiments show that these benefits disappear when model capacity is insufficient. Thus, while the paper establishes an empirical phenomenon, it has not yet developed a practical selection rule capable of explaining the key capacity-dependent effects.
  2. Proposition B establishes that the energy of the unconditioned regression-target correction grows as O(dε)O(d\varepsilon), rather than providing a bound on the actual regression error or SW2 sampling error. Appendix AM further shows that part of the target variance can be eliminated without changing the marginal velocity field. The regularity bounds in Theorem C rely on dimension-uniform posterior moment conditions, which are currently supported primarily by numerical checks on a finite GMM grid and do not directly imply the fitting benefits assumed by Model D. Therefore, the central error model remains an empirical hypothesis rather than a theoretically justified rule for noise selection.
  3. With a fixed GMM network, the normalized FM error deteriorates substantially as dimensionality increases, suggesting that the preference for FM in high dimensions may also reflect a model capacity bottleneck. Although the existing capacity interventions demonstrate that model configurations can change the conclusions, they do not match fitting quality across dimensions to isolate the effect of dimensionality itself. Furthermore, the PCA-CIFAR and tabular experiments use training data as the reference for distribution matching, and Table 32 does not report paired gains over FM on an independent test set. These limitations weaken the evidence for the generalizability of the observed improvements.
Questions for authors
  1. Could the authors calibrate a selection rule that incorporates model capacity and sampling budget, without tuning on the test tasks, and report its prediction accuracy on previously unseen datasets?
  2. Could the authors adjust network width and training steps across different GMM dimensions to achieve comparable levels of normalized FM error, and then re-evaluate the optimal value of ε\varepsilon?
  3. Appendix AM proposes a conditional-mean regression target that preserves the same marginal velocity field. Could the authors compare the original and alternative targets in terms of SW2, gradient variance, and optimal ε\varepsilon under the same training budget?
Reviewer_j2yuOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the choice of the additional noise amplitude in the stochastic interpolant X_t=(1-t)X_0+tX_1+\sqrt{\varepsilon t(1-t)}Z, where \varepsilon=0 corresponds to straight-line flow matching. The authors examine the trade-off between smoothing the velocity field and increasing the variability of the training target, and propose a phenomenological model predicting that optimal \varepsilon decreases as dimensionality increases. Experiments cover Gaussian mixtures, images, tabular data, point clouds, and robotic tasks. The main contribution is an empirical investigation of when additional noise helps training. However, the results do not yet establish dimensionality as the determining factor or sufficiently substantiate the proposed noise-selection rule. The theoretical statements are largely correct under their assumptions, but the connection to the quality of the learned generator remains incomplete.

Critical strengths
  1. A practically relevant question. The paper examines noise-level selection under computational constraints and shows that its effect depends on the number of sampling steps and model capacity. This is more informative than comparing methods at a single, arbitrarily chosen \varepsilon.
  2. Substantive controlled experiments on Gaussian mixtures. The authors investigate different dimensions, distribution families, integration budgets, and model capacities. Particularly useful are interventions where increasing network width restores the benefit of positive noise at a fixed dimension. The main synthetic protocol separates hyperparameter selection from final evaluation.
  3. Transparent discussion of limitations. The authors explicitly discuss the phenomenological nature of the error model, the sensitivity of the results to the experimental protocol, and issues with earlier experiments. Additional checks of endpoint conditions and an analysis of an alternative training-target parameterization help distinguish the observed effects from potential artifacts.
Critical weaknesses
  1. The novelty lies primarily in the empirical analysis rather than in a new methodological construction. The interpolant with a noise term of the form \sqrt{\varepsilon t(1-t)}Z and the corresponding velocity-field training objective fall within the existing stochastic interpolants framework. The limiting connection to flow matching and the computation of the energy of the additional term do not, by themselves, constitute substantial methodological novelty. The paper should distinguish its contribution more clearly from existing studies of interpolation schedules, velocity-field regularity, and the dependence of discretization errors on dimensionality.
  2. The theory does not derive the claimed dependence of quality on dimensionality. The linear growth of the energy of the additional term with d characterizes the training target, but does not provide an error bound for the learned generator. The Jacobian bound requires uniform conditional covariance bounds, whose constants may depend on dimensionality and the noise level. A decrease in an upper bound also does not establish a decrease in the actual complexity of the field. Finally, the relationship \mathcal E^2(\varepsilon)=\frac{A}{1+\beta_0\varepsilon}+d\widehat c_0\varepsilon is postulated rather than derived from an analysis of training and sampling. The resulting dimensionality threshold is therefore a consequence of the assumed model rather than an established property of the method.
  3. The authors’ own analysis admits an alternative explanation of the main mechanism. Appendix AM constructs the target w_t=\mathbb E[v_t\mid X_0,X_t,t], which induces the same marginal velocity field. The original population loss differs from the loss associated with w_t only by an additive constant independent of the network parameters. Consequently, part of the target variability under discussion may affect stochastic optimization, but does not change the optimal predictor within a fixed model class. Without an experiment using this alternative target, it is not possible to establish whether the observed effect arises from the geometry of the interpolant or from the particular training procedure.
Questions for authors
  1. Do the benefits of positive \varepsilon and the transition to zero noise persist when training with the target from Appendix AM, which removes part of the conditional variability without changing the marginal velocity field?
  2. Does the dependence of the optimal noise level on dimensionality persist when model capacity and training budget are scaled to maintain comparable FM baseline quality? How can the influence of dimensionality be separated from insufficient training or limited model expressivity?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.