acceptodds
Under review as a conference paper at ICLR 2027

Geometry of Phase Ordering in Diffusion and Flow

Abstract

We study same-state conditional prediction contrasts as descriptive measurements, separating the temporal concentration of prediction disagreement from terminal intervention response. The primary statistic is the appearance-minus-structural phase-center gap within a trained evaluating network. For three four-factor diffusion checkpoints, archived model-level summaries report a mean gap of (sample standard deviation ) on common trained trajectories, versus for untrained evaluators on the same states; this contrast is descriptive, not a causal estimate of training. A separate Gaussian-block calculation derives variance-dependent phase ordering for unnormalized population contrasts; in a learned-predictor test with five variance ratios, the reported non-null gaps have the predicted signs, whereas their magnitudes, the equal-variance control, and the archived peak positions disagree with the prediction. In SD3.5 Medium, static directions attain the same reported classification accuracy as aligned updates, and text-only features classify 24/24 queries versus 23/24. In an onset audit, eight of 18 cases pass the archived primary gate flags, the geometry-peak error is 1.375 steps versus 1.25 for a fixed-step predictor, and a routing pilot does not establish an advantage. We therefore distinguish the reported phase organization from mechanistic identification, calibrated onset prediction, and verified-budget control utility.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion models, Flow 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_JHJcOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates same-state conditional prediction contrasts as descriptive measurements in diffusion and rectified-flow models, separating the temporal concentration of prediction disagreement from terminal intervention response. It examines whether conditional update geometry can diagnose functional regimes and restrict policy spaces, while explicitly defining the limits of its control and causal interpretations.

Critical strengths

The paper maintains exceptional scientific honesty by explicitly separating descriptive measurements of phase organization from unverified causal assumptions, mechanistic attributions, and control utility. By providing a detailed claim ledger and conducting rigorous audits, the authors transparently acknowledge the limits of their findings. This is highly consequential because it prevents over-interpretation and provides a reliable, self-audited foundation for future research.

The introduction of the same-state conditional prediction contrast framework allows for the precise isolation of local field differences by evaluating different prompt conditions at a shared trajectory state. This methodological design effectively avoids the confounding effects of accumulated state drift that plague independent trajectory comparisons, establishing a clean, reproducible baseline to analyze how prediction disagreement is temporally concentrated across features.

Critical weaknesses

The paper's empirical audits show that the proposed geometric diagnostics fail to yield tangible operational benefits. Specifically, in the causal-onset audit, the geometry-based peak predictor achieves a mean absolute error (MAE) of 1.375, which fails to outperform a naive, fixed-step-4 predictor (MAE 1.25). Furthermore, the routing pilot does not establish an advantage over the best fixed policy, and static/text-only controls match or exceed the classification accuracy of aligned updates. This leaves the practical utility and incremental value of the proposed framework unverified.

While the idealized Gaussian-block calculation (Proposition 1) predicts the correct signs for non-null ratios, it fails to provide a quantitative bridge for the actual renderer. The post-hoc stability certificate (Lemma 1) fails in all non-null models (0/200/20), and the observed magnitudes, equal-variance controls, and peak positions directly disagree with the theoretical predictions. Consequently, the theoretical framework does not rigorously certify or explain the empirical normalized-LL ordering observed in the neural networks.

Questions for authors

The paper notes that the Gaussian-block model's stability certificate (Lemma 1) fails in 0/200/20 non-null models, and that the predicted magnitudes, equal-variance controls, and peak positions directly disagree with empirical observations. Furthermore, population raw contrasts QfQ_f and separately normalized empirical diagnostics LfL_f represent different mathematical objects.

Question: Without a rigorous quantitative bridge between the unnormalized population model and the normalized renderer, how should readers interpret the theoretical alignment of signs beyond a qualitative coincidence?

The empirical audits reveal that a naive, fixed-step-4 predictor achieves a better Mean Absolute Error (MAE of 1.25) than the geometry-peak predictor (1.375) on primary-valid cases, and that static/text-only controls match or exceed the classification performance of temporally aligned updates.

Question: Given these findings, under what specific operational regimes or task distributions does the same-state conditional contrast provide a tangible advantage that outweighs its additional computational cost?

1 comment

Thank you for your careful review and constructive comments.

We accept both weaknesses: operational value is not established, and the Gaussian theory does not explain the normalized measurement. Blue marks substantive changes only.

Q1. Sign agreement without a quantitative bridge. Only a qualitative coincidence. The linking certificate held in 0/20 models and is removed (Appendix C). Proposition 1 (Sec. 3.1): the contrast keeps the angle and discards norms, so a phase centre is not a commitment time. It does not explain why the order appears; finite-switch assumptions are unverified (App. D).

Q2. Regime where it outweighs its cost. We found none. The SD3.5 onset predictor lost to a fixed-step rule and the routing pilot showed no advantage; both are withdrawn. The phase-centre order needs no interventions; full switches confirm it in 4/4 models (Table 3), also on unused label combinations (Table 7). But a post hoc constant locates horizons better (Sec. 5.2), immediate switching beats phase-gated for 3 of 4 factors, at ρ=0.5 the order holds in 1/4 models (Table 5), and calibration costs extra evaluations (Sec. 5.2). Large-model transfer is untested; the classification comparison is removed.

Reviewer_gEseOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The submission examines temporal patterns in conditional predictions from diffusion and flow models, focusing on phase ordering and intervention timing. It presents an idealized Gaussian analysis, controlled diffusion experiments, and additional studies of larger generative models.

Critical strengths

None

Critical weaknesses
  1. The submission does not provide a coherent, finalized account of a scientific study. Editorial notes, execution records, references to superseded material, and unresolved experimental status appear throughout the manuscript. Their prevalence shows a fundamental lack of consolidation and makes the document difficult to assess as a finished research contribution.
  2. The manuscript’s definitions, experimental descriptions, figures, and conclusions are not consistently aligned. Changes to measurement definitions and the status of reported evidence have not been incorporated throughout the document, leaving incompatible descriptions and outdated interpretations in place. These inconsistencies compromise the reader’s ability to establish what the reported comparisons actually demonstrate.
  3. The experimental account does not consistently connect the stated procedures to the retained evidence. Completed results are interwoven with withdrawn analyses and unexecuted revisions, while the retained onset analysis lacks information needed to reconstruct its response measurements and quality gates. The reader is left to reconstruct the basis of the study from fragmented records and qualifications that should have been resolved before submission.
Questions for authors
  1. What is the final set of completed experiments supporting the conclusions, and which measurement definitions and evaluation procedures apply to each?
  2. How can the retained results be verified and reconciled with the manuscript’s own statements that relevant measurements, comparisons, or selection procedures remain unverified?
  3. Why does the manuscript omit the AI Use Statement explicitly required by the ICLR 2027 author guidelines? Please explain this omission and provide the required disclosure.
1 comment

Thank you for your careful review and constructive comments.

We accept all three weaknesses; the PDF is rewritten, with editorial notes and unverifiable results removed. Blue marks substantive changes only.

Q1. Final experiments. Each states estimand, units and baseline: same-state gap (Sec. 4, Table 2); full switch ρ=1, primary, 4 models × 200 seeds (Sec. 5, Tables 3–4); ρ=0.5, secondary (Table 5); unused label combinations (Table 7); data-only predictors (Table 9). Held-out units are noise seeds, not unseen combinations (Appendix A).

Q2. Verification. The "unverified" statements concerned removed material, except finite-switch assumptions (App. D). For retained results, protocol and criteria were fixed before the full run, all models are reported without exclusions, and a second implementation reproduced Table 7 exactly. The initial caption was right that no equivalence test was run; the contrary statement elsewhere was wrong, and the tests in Table 2 are new. Not done: Table 9 has no independent recomputation. Weakening results: at ρ=0.5 the order holds in 1/4 models (Table 5), and the criterion measures source suppression (Sec. 5.1).

Q3. AI Use Statement. Our omission; it is now on p. 10.

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

It appears as though the paper is investigating when and whether conditional predictions disagree during denoising. There are a few "contributions" which is the empirical phase ordering with some preliminary metrics like the phase-center gap and between-model measures.

There seems to be "some" validation of what they've been doing. The experiments do support what the paper seems to be trying to parse but the presentation is horrid to say the least. The theory does make sense and seems to be legitimate but it doesn't necessarily get validated in the experiments. It is clearly LLM rewritten and this is most clear in the abstract.

There is no clear motivation for why they would like to study this. It reads more like a standard report than a specific question (and motivation) that they're interested in solving. Although, I admit this might just be my own biased view on how to write papers today.

The theory that they provide tends to be relatively disconnected from their measurement.

That being said, it is understandable but you will need to actually read through it very carefully.

Critical strengths

I cannot identify any clear critical strengths with the way it's written. It needs to be rewritten in a way where the motivations for the question at hand and their experimental design to answer these questions is clear. However, I do believe that LLMs might be able to partially ingest this information (after filtering out inconsistent use of terminology). It is just not written for a human properly.

Critical weaknesses

The title claims some formalization of the geometry of phase ordering but it is heavily misleading. There is no geometry in this paper. The theory and experiments seems to be disconnected. However, the experiments that were carried out are carried out with relatively strong validations but it doesn't predict anything in particular that appears to be novel. The point of the paper seems to be that there are theory contributions that are supported by the experiments but there seems to be no support of the theory. The mathematics does check out but it doesn't lead to any insight. Another key problem is that there is no clear motivators for the entire paper.

Questions for authors

Are you able to make clear the reasoning and motivation for why you have decided to investigate the scientific questions proposed in the paper?

More importantly, let's assume that there's a strong motivation for this. It is currently not obviously clear what your proposed measurement tells us that existing methods do not, which result in the paper clearly demonstrates this?

1 comment

Thank you for your time for reviewing our manuscript. Sec. 1 now opens with the question; terms are fixed in Table 1. Blue marks substantive changes only.

Q1. Motivation. Editing methods switch conditioning at a hand-set time that treats all attributes alike; Sec. 1 (p. 1) asks where single-factor prediction differences concentrate and whether that order matches the endpoint response.

Q2. What is new. An attribute order read from the network without interventions: gap 0.225–0.243, positive on 20/20 trajectories per model; untrained evaluator ≈ 0 (Table 2). Switch horizons agree in 4/4 models (Table 7); in one coupled model the margin interval includes zero. Limits: conditional prediction differences are not new; we claim no better editor; and a post hoc constant locates horizons better (Sec. 5.2). Not done: independent recomputation of Table 9; closed-form Gaussian-score explanations are unevaluated (Sec. 6).

"No geometry." Correct; the abstract states the scope.

"Theory disconnected." Agreed. The certificate held in 0/20 models and is removed; the Gaussian calculation matches in sign only (Appendix C).

Presentation, LLM use. The AI Use Statement is on p. 10.

Comments on the paper

2 comments

Hi authors,

If you are planning to respond to the reviews, I am happy to receive your responses to my concerns in stages. This would also allow us to engage in further discussion and clarification, given that only 1 day remain before the discussion phase ends.

Thanks!

We thank all reviewers for their careful and constructive reviews.

We accept the main criticisms (motivation, theory gap, value, AI statement). PDF rewritten around one claim; blue text marks substantive changes only.

Claim. Trained evaluators place appearance later than structure in same-state prediction contrasts (gap 0.232, 20/20 trajectories per model; untrained ≈ 0, Table 2), and full switches show the same order in 4/4 models (Table 7).

Withdrawn. Geometry/flow claims; theory as evidence (Appendix C); horizon localisation and timing advantage (Table 4); target-correct editing; large-model audit.

Corrections. The criterion measures source suppression (Sec. 5.1); held-out units are noise seeds (Appendix A). Not done: independent recomputation of Table 9. An earlier small run had undefined horizons (App. A).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.