acceptodds
Withdrawn from ICLR 2027

PROVER: Precision-Indexed Risk Order via Euclidean Reachability in Gaussian-Affine Flow Decoding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

When does one source representation supply more useful information than another for velocity prediction? Classical comparison settles this when one representation is a random degradation of the other, but leaves many pairs unordered. We prove an exact order for flows that linearly mix a target with Gaussian noise. At each noise precision, a matrix records the posterior error removed by the representation. Matrix dominance at every precision is equivalent to no larger optimal Euclidean velocity risk on every complete flow. The new step is Euclidean reachability: any positive precision and error metric can be realized together on a normalized noise-to-data path. Complete flows therefore test the full Gaussian-affine quadratic prediction order. This comparison principle also orders frozen predictors with fitting error included and determines when a preferred system remains preferable throughout a prescribed path family.

Primary area:
learning theory
Keywords:
Flow Matching, Representation Comparison, Comparison of Statistical Experiments, Minimum Mean-Square Error, Precision-Indexed Risk Order, Euclidean Reachability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_DDy8OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how to compare source representations for velocity prediction in Gaussian-affine flow decoding. It introduces a precision-indexed matrix measure of the posterior error reduction provided by a representation and proves that dominance of this measure at every precision is equivalent to lower optimal Euclidean velocity risk over all complete Gaussian-affine flows. The key technical result is a Euclidean reachability theorem that realizes arbitrary positive precision and velocity-error metrics within a normalized complete flow. The framework is further extended to frozen fitted predictors and restricted families of flow directions, and is evaluated on image and video prediction tasks using several pretrained visual representations.

Critical strengths
  1. Exact and Non-trivial Theoretical Characterization.
    The paper develops a technically non-trivial and mathematically rigorous framework for comparing source representations across Gaussian-affine flow objectives. In particular, the Euclidean reachability result provides the key converse needed to show that local precision-indexed matrix dominance is equivalent to global velocity-risk dominance over the full class of complete flows. This gives the framework a clean and exact characterization rather than only a sufficient condition.

  2. Extension to Fitted Systems with Empirical Validation.
    The paper makes a meaningful effort to extend the theory beyond an idealized Bayes-optimal setting. The analysis is further developed for frozen fitted predictors, explicitly incorporating fitting error, and the empirical section evaluates the resulting preference behavior on image and video prediction tasks using several pretrained visual representations. This helps demonstrate that the proposed ordering is not limited to a purely synthetic setting.

Critical weaknesses
  1. The practical ML significance is not sufficiently established.
    The paper characterizes representation preference in terms of Euclidean velocity-prediction risk, but it does not convincingly show that this ordering leads to better generation quality, prompt alignment, downstream performance, or actionable model/schedule design. The experiments mainly verify the theoretical risk reversals rather than demonstrate a practically consequential benefit.

  2. The experimental setting is substantially removed from modern large-scale generative modeling.
    The evaluation uses frozen representation encoders and relatively small U-Net/Transformer prediction cores over VAE latents. It remains unclear whether the same preference reversals are meaningful in current large pretrained diffusion/flow models with complex conditioning mechanisms, guidance, and jointly adapted representations.

  3. The relevance of the constructed path family to realistic inference trajectories is insufficiently justified.
    The main empirical conclusions rely on a mathematically designed family of carrier rotations and directional perturbations. While this is appropriate for validating the theory, the paper does not establish that these perturbations correspond to practically used diffusion/flow schedules or that the observed fragility would occur under realistic inference choices.

  4. The writing structure and presentation choices do not communicate the paper’s contribution clearly.
    The paper is organized largely around a sequence of definitions, theorems, and increasingly general extensions, while the central ML motivation and practical consequence become difficult to track. The notation density, theorem-first organization, and presentation/layout choices make it hard to identify the main problem–insight–consequence chain. In particular, the paper would benefit from a clearer separation between the core reachability result, classical ingredients inherited from Gaussian estimation/comparison theory, and the concrete implication for generative modeling.

Questions for authors
  1. Can the proposed risk ordering be shown to predict actual generative performance rather than only velocity-prediction error?
    A key missing link for me is whether the differences identified by PROVER translate into practically meaningful differences in generation quality. For example, could the authors use PROVER to select between competing source representations and then evaluate whether this choice consistently improves final image/video quality, semantic alignment, or other downstream generation metrics on a pretrained flow/diffusion model? Such an experiment would substantially strengthen the practical motivation of the framework.

  2. Does the reported preference reversal persist in a modern large-scale pretrained generative model?
    The current controlled U-Net/Transformer experiments are useful for validating the theory, but they leave open whether the same phenomenon remains relevant in contemporary large pretrained flow/diffusion systems. I would encourage the authors to evaluate at least one realistic pretrained backbone, keeping the generation model fixed while varying the source representation or conditioning interface, and test whether the representation ordering predicted by PROVER agrees with both velocity risk and final generation performance.

I would also like to note that the current writing structure and presentation made the paper harder to follow than necessary. In particular, the organization, notation density, and layout sometimes obscured the central contribution and its practical significance. I believe clearer structuring and presentation would substantially improve the paper’s accessibility.

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

This paper investigate when one input representation (a “report”), such as CLIP or DINOv3 features, helps a flow model predict velocity better than another on every noise-to-data path. For each signal-to-noise level, the authors define a matrix that measures how much prediction error a report removes. Theorem 2.1 shows that report A has no larger optimal velocity error than report B on every path exactly when A’s matrix is at least as large as B’s in every direction, at every signal-to-noise level. The key new step, Euclidean reachability (Theorem 2.2), builds a valid path that reaches any signal-to-noise level and any error weighting at once. On CO3Dv2 new-view and SSv2 later-frame prediction, 9 of 48 comparisons keep the same winner across the family. Switching encoders on those 38 paths cuts error by 0.64% to 2.25% on new data.

Critical strengths

1. Math part is understandable and clear. The main theorem needs very few assumptions. Only the added noise must be Gaussian, and the target can follow any distribution with finite variance. The construction in Theorem 2.2 also handles the hard case where the signal-to-noise matrix and the error weighting point in different directions (they do not commute).

2. A clear experiment design. Every path in the rotation family has the same input distribution at each time: the same noisy state, target, context and reports, with the same clock and endpoints (Section 4). A change in the winner therefore comes from the velocity objective alone, not from a change in available information.

Critical weaknesses

1. The contribution is limited. The paper itself calls the easy direction of Theorem 2.1, equation (2.3), “the standard conditional-projection step.” The link to general prediction tasks (Proposition A.2) uses standard Gaussian regression. So the truly new part is the path-building result in Theorem 2.2. It combines a known matrix average (the geometric mean in A.4) with a smooth filling-in step. The paper never shows that the hardest case it solves, where the two matrices do not commute, actually occurs with real encoders.

2. Lack of image quality examination. Section 4 says the trained systems are ordered by theorem 2.3. But each network only sees four sine/cosine angle maps, so it only works when the signal-to-noise level is shared within four channel groups. Path choice therefore matters only through training error and step size, yet the paper reports only velocity error and never the quality of generated images.

Questions for authors
  1. Does the non-commuting case occur with real encoders? For the encoder pairs in Section 4, how often does the error-gap matrix fail the order in a direction that commuting paths cannot detect?

  2. Do the 38 witness-selected report switches produce better samples? I am curious about PSNR, LPIPS and FID for samples along the default, fixed-mix and witness paths.

  3. How does the number of ODE steps affect the comparison? Does the witness path need more steps than the default path to reach the same sample quality?

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

Prover introduces an order for ranking the usefulness of information for flow matching.

Critical strengths

None

Critical weaknesses

The presentation is extremely hard to follow. This is a niche mathematical field and the presentation is also not helping out either. E.g. the paragraph around l072 introduces H,S,Y, lambda and Z and only introduces these non standard objects in the next section l100.

The paper hardly makes the significance of its findings clear to me. The way it is written, I need to reread every sentence in the abstract several times to get a rough idea of what is actually being said.

I really cannot comment on the soundness of the paper, since I don't have the background necessary to evaluate that.

Questions for authors

Why is ICLR the right venue for this manuscript?

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

Related papers

Discussion (0)

Sign in to comment.