acceptodds
Under review as a conference paper at ICLR 2027

Conditional Mean Dependence Governs One-Step Endpoint Regression

Abstract

One-step generative policies may match iterative samplers or collapse to a conditional mean. We characterize this outcome for endpoint regression. Let denote the ratio of target within-condition variance reproduced by a one-step map. For the optimal map, , the squared correlation ratio between source noise and target given context. Collapse occurs exactly when the target's conditional mean is independent of the source noise; full statistical independence is sufficient but unnecessary. Moreover, depends on the training pairing, not on the target marginal alone. A fixed-marginal coupling family makes span , giving every target-marginal-only estimator worst-case absolute error at least ; paired training samples permit estimation under regularity assumptions. Synthetic experiments recover the predicted curve. A cost-matched construction further shows that quadratic transport cost does not identify endpoint retention. An area-downsampled full-image MNIST check recovers the predicted ordering but exposes the failure of a generic high-dimensional estimator. An exploratory OGBench test recovers the map-spread ordering yet refutes the behavioral prediction: the collapsed policy succeeds more often. Thus the criterion diagnoses endpoint-map spread, not task utility.

Primary area:
reinforcement learning
Keywords:
one-step generative policy, conditional spread, correlation ratio, conditional mean dependence, flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_jYYfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper primarily studies the effects of couplings on generation in one-step generative models under ( L2L_2 ) endpoint regression. In addition, the authors provide experiments involving non-( L2L_2 ) endpoint regression.

Critical strengths

This paper is easy to follow, and the theory is mostly solid under the constraints and settings imposed by the authors.

Critical weaknesses

Novelty

The effect of couplings on one-step generation has already been studied in various papers. For example, ρ=0\rho = 0 for independent couplings is well known, and previous papers have also identified this collapse. Similarly, it is straightforward from optimal transport and diffusion distillation that ρ=1\rho = 1 for deterministic couplings. However, mixtures of these couplings are not commonly studied.

The main result in C1 appears to follow directly from classical L2L^2 regression theory and the law of total variance. The characterization of collapse through conditional mean independence is closely related to the established correlation ratio (Rényi, 1959; Sampson, 1984). Furthermore, Shou (2026, Theorems 1–2) already establishes mean collapse under independent endpoint pairing and characterizes the optimal flow map using conditional expectations. The deterministic-coupling result also follows immediately from this characterization and is consistent with earlier rectified-flow and distillation work (Liu et al., 2022). While the distinction between conditional mean independence and full independence is useful, it is unclear what fundamentally new theoretical insight C1 provides beyond these established results.

The applicability of the theory to batch OT remains unclear, as batch OT does not necessarily satisfy the assumption that the coupling is a linear combination of deterministic and independent couplings.

The value of ρ\rho depends on the determinism of the coupling. Two arbitrary couplings can have ρ=1\rho = 1 and ρ=0\rho = 0 while having different transport costs. However, at the infimum of the transport cost the transport becomes deterministic, and therefore ρ=1\rho = 1. This does not imply a general relationship between ρ\rho and the transport cost. The authors appear to conflate the relationship between coupling distance and ρ\rho while attempting to establish that no such relationship exists. Although the counterexample in Figure 1 illustrates this distinction, its necessity and additional contribution are unclear.

The authors should clarify what they intend to establish by demonstrating that $ \rho^* $ spans the entire interval [0,1][0,1], given that the two extreme cases (ρ∗=0\rho^* = 0 and ρ∗=1\rho^* = 1) are already sufficient to establish non-identifiability and the minimax lower bound. Furthermore, the practical relevance of the proposed Bernoulli-mixture coupling, which randomly switches between independent and deterministic source–target pairings, remains unclear. Is such a coupling used in existing one-step generative modeling methods, or is it purely a theoretical construction? What additional theoretical insight or practical significance does this interpolation provide beyond the two extreme cases?

Finally, the authors provide experiments involving objectives other than L2L^2 endpoint regression, but little theoretical analysis or explanation of these results is provided.

Writing

The authors use inconsistent notation in multiple sections of the paper. Random variables are sometimes represented by lowercase letters, making the paper difficult to follow.

L17: "sufficient but unnecessary" is not standard mathematical terminology. I assume the authors meant "sufficient but not necessary."

References

  1. Shou (2026), ODE-free Neural Flow Matching for One-Step Generative Modeling.
  2. Liu et al. (2022), Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow.
  3. Rényi (1959), On Measures of Dependence.
  4. Sampson (1984), An Inequality for Correlation Ratios.
Questions for authors

The authors have developed a theoretical characterization of these coupling combinations. However, how this theory contributes to designing better couplings or extends beyond L2L^2 endpoint regression remains unanswered.

Reviewer_4YWDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper explains detection of one step generative policy samples' to mean regression using a measure introduced in equation (1). To the best of my understanding there is an effort to motivate the use of this measure. And derive some coupling identifiablity properties.

Critical strengths

Empirical results accompany theoretical material

Critical weaknesses

The text is not easy to follow. Starting from the abstract the sentences appear to be independent from each other, lacking the flow of the logically linked and structured story. Some sentences read unnaturally e.g. "For the optimal map, ρ=ρ∗\rho = \rho^*, the squared correlation ratio between source noise and target given context." (line 14) or are too vague "paired training samples permit estimation under regularity assumptions." (line 19) or "The two couplings that matter in practice are..." (line 94) these couplings matter in what sense?

The main notation used to explain the contributions (cc, x0x_0 and x1x_1 in eq (1)) is not introduced until Section 2. This renders the motivation and contributions incomprehensible. Figure 1 did not help with the situation which is reference in line 42 and explained in proposition 1. I was not able to verify the validity of the claims in introduction.

The theorems are formalized without proper motivation and discussion. Most of the theoretical work appears to have unnecessary conditions and is limited to special cases without a clear justification for their introduction (Thm 2, 3 and Proposition 1, Col 1 in closed form).

Section titles don't describe the content and need revision. as an example "5.3 WHAT DOES DECIDE : THE METHOD MAP"

The novelty or application is not apparent in the current form of the paper.

Questions for authors

Proposition 1. Why is the assumption dd even relevant to the proposition?

Why is the family in section 3.1 introduced? and how does the results in Thm 3 generalized to any other coupling? How consequential is the choice of Bernoulli distribution over B to the proof?

Where is the target-only metric used in Figure 2?

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

The paper studies when a one-step generator trained with squared error produces diverse outputs and when it predicts only the conditional mean (collapse). The paper identifies that the most important factor is how source noise is paired with training targets. The authors show that collapse occurs when knowing the source noise does not change the expected target. They also show that the target distribution alone cannot predict this outcome. Experiments examine different pairings and loss functions on synthetic data and MNIST. An OGBench experiment confirms the predicted differences in output spread, but the collapsed policy performs better.

Critical strengths

None.

Critical weaknesses
  • Most importantly, the motivation of the paper is not established. The paper did not cite or include any concrete example of an existing one-step generative method whose failure is explained by the proposed criterion. This makes the contribution of the paper unclear.
  • The downstream experiment does not support the paper's claim that avoiding mean collapse improves policy performance. In the OGBench experiment, the collapsed policy substantially outperforms the spread-preserving policy.
  • The paper is very hard to read. A lot of terms are not clearly introduced (e.g., D^\hat{D} in Line 137, "sorted-quantile pairing" in Section 5.6 ). And it has some suspected AI traces that the authors did not clean up before the submission. For example, at the beginning of Section 5, the paper writes, "All five pre-registered criteria pass (verdict PASS)." What is this supposed to mean?
Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.