acceptodds
Under review as a conference paper at ICLR 2027

When the Change Carries No Energy: A Detectable Regime Change That No Learned Monitor Detects

Abstract

Latent world models need reliable regime detection, yet existing monitors track only residual energy-shifts in the mean or variance of prediction error. When an environment switches between noise laws of identical variance, every prediction error keeps its size and only its shape changes, so an energy-based monitor, sensitive only to the second moment, is provably blind to the shift, a claim we verify-along with our other guarantees-in a machine-checked proof. We introduce a generative world model that pairs a joint-embedding predictive architecture with a conditional flow-matching predictor over future latent trajectories (F-JEPA-RD), which scores the full predictive density rather than its energy. An analytic Gaussian base coupled with a learned residual yields exact trajectory density ratios while preserving the representation structure the joint-embedding design protects. An audit of sequential detection benchmarks reveals a pervasive validity failure: standard protocols mistake a detection statistic's natural drift during calm periods for genuine detection, creating an illusion of success across existing methods. Under a drift-robust protocol with null controls, an oracle given the true noise law detects variance-matched shifts almost perfectly, while every learned monitor we test-including our own model's readouts-collapses to chance, on both synthetic benchmarks and a four-year real-world recording. A diagnostic locates the loss: the encoder preserves and even sharpens the shape signature, which is then diluted by the multi-lag residual the readouts consume. Applying the same shape core to a single-step innovation residual instead recovers part of the lost power in a single-seed pilot, indicating the information is still present in the representation. When regime changes instead alter the underlying dynamics, F-JEPA-RD's density ratio retains strong power while staying silent under null controls, exactly as the theory requires. We release the benchmark, proofs, and scoring code for independent re-scoring.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
World Model, JEPA, Flow Matching, Out of Distribution, Deep Learning Theory

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_qq4KOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper argues that there are regime changes that standard monitors working with means and variances do not detect. Then, the authors propose a methodology and show that it does not work.

Critical strengths

None.

Critical weaknesses

This paper is exceptionally hard to follow. Is the paper trying to argue that current monitors are insufficient because there are meaningful situations in which regime changes are not detected? If so, the paper needs a substantial narrative change, focus on diagnosing why and evaluating it on a large selection of such monitors. If the paper is instead proposing a new methodology (F-JEPA-RD?), it should show its effectiveness. However, by the authors' own admission, this method does not work. So why should a reader care about it?

Overall, this manuscript feels like an AI generated running draft/experimental notebook (Tables 32,33 are really emblematic) and it does not meet the bar for publication, especially in a venue like ICLR.

Questions for authors

None.

2 comments
AuthorsOpenReview ↗

Response to Reviewer qq4K

Insufficient monitors, or a new method? Both, as the contribution list states (p.3 l.110-129). The paper argues that monitors watching residual energy miss a class of changes. It then evaluates nine monitors under a corrected protocol, and diagnoses why they fail:

  • Insufficient monitors (C2, C5). Proposition 3 proves that energy statistics are blind to variance-matched changes; the exact oracle detects them at 0.972.
  • Many monitors, corrected protocol (C3, C4). Three readouts and six independent detectors run on identical latents (BOCPD: same-seed retrained model) with a calm test; Tables 1 and 4 show all of them at chance on this change. The revision adds a neural likelihood-ratio statistic that learns the alternative; it is at chance too (Table 38).
  • Why (C5). The encoder keeps the change, the predictor's one-step and multi-lag residuals dilute it (Tables 28 and 35, Appendix P), and BOCPD sees the mixture on raw data but not on codes (Appendix F).
  • The method (C1). It works where its alternative class applies: RATIO 0.626 on the dynamics change, and the strongest learned detector at lags (1, 2, 4) (0.818 vs. 0.682, horizon chosen on the swirl, Table 26). It is silent on the linear control (0.000). The analytic base is what keeps the code usable by standard detectors (Table 29).

The revision adds the multi-seed evidence for C5 (Appendix Z). On the same one-step code innovation, the energy statistic stays at 0.117 and the shape statistic of Prop. 4 reaches 0.375. The paper's own readouts sit at chance because the predictor's residuals they read dilute the fourth-moment signature.

Readability and Tables 32/33. We agree the manuscript was hard to read; see the Summary of Revisions. Table 33 lists the six design predictions, each with its outcome; it is kept because the drift-robust rules of the main text are post hoc (Section 5; Appendix T, p.34 l.1805-1810). Table 32 is a reported negative result on public data. Both are condensed in the revision.

I appreciate the authors' efforts to make the manuscript more readable. I am now more convinced that the general direction has some merit.

However, in the current form, I don't think that the work is ready for publication. Despite the efforts, I still find it very difficult to read, and it still feels like an AI-generated manuscript with occasional clarity.

I encourage the authors to consider a rewrite focusing on the main points that they would like a reader to take home, and to simplify the main text and the appendix, without the usual <statement><result><caveat> form, as it often obscures the actual point.

Reviewer_p9rxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates sequential regime detection when distributional changes alter the shape of prediction errors without changing their mean or variance. The authors propose F-JEPA-RD, a generative latent world model combining a joint-embedding predictive architecture (JEPA) with conditional flow matching and an analytic Gaussian base. The framework supports three detection statistics: a one-step residual score, a source-shape score, and a trajectory density ratio.

The paper also provides machine-checked theoretical results and identifies an important flaw in conventional detection evaluation protocols, where the natural drift of monitoring statistics can be incorrectly interpreted as successful change detection. The authors introduce drift-robust evaluation protocols and compare their approach against several classical and learned detectors.

Experiments on synthetic benchmarks and real-world data with injected changes show that an exact-likelihood oracle can reliably detect variance-matched distribution shifts, while the evaluated learned detectors, including F-JEPA-RD, largely fail under corrected evaluation protocols.

Critical strengths
  1. One of the contributions is the identification of a flaw in the evaluation protocol used in this study, where the natural drift of detection statistics can be mistaken for successful regime detection. The authors demonstrate this issue through controlled simulations and null-control experiments, showing that detectors can achieve seemingly high detection rates even when no distributional change occurs. The proposed drift-robust evaluation protocols substantially change the experimental conclusions for both the proposed method and several baselines. This finding highlights the importance of distinguishing genuine detection capability from artifacts introduced by evaluation procedures, potentially benefiting research beyond the specific framework studied in this paper.

  2. The paper provides a systematic theoretical analysis of variance-matched distribution shifts, supported by machine-checked proofs. The use of an exact-likelihood oracle establishes that the constructed distribution changes are detectable in principle, while the evaluated learned detectors largely fail to detect them under the corrected protocols. Rather than presenting misleading positive results, the authors explicitly acknowledge these failures and investigate their underlying causes. In particular, the analysis suggests that the latent encoder preserves distributional shape information, which is subsequently diluted in the prediction residuals. These findings provide insights into the limitations of the evaluated latent prediction and detection pipelines.

Critical weaknesses
  1. The paper motivates F-JEPA-RD as a density-aware extension of JEPA for detecting distributional changes that conventional prediction-error-based methods may miss. However, the proposed framework fails on the central shape-switch benchmark. More importantly, the experiments do not establish whether the proposed flow-matching architecture is necessary for addressing the motivating problem. Consequently, the paper does not sufficiently demonstrate the practical advantage or necessity of the proposed architectural complexity for its central research objective.

  2. In Appendix P, the authors attribute the failure primarily to the dilution of fourth-moment information in multi-lag prediction residuals. However, the main diagnostic relies on a single trained model, and the proposed one-step innovation-based shape detector is evaluated using only a single seed. Also, the main experiments rely on carefully constructed synthetic changes, while the real-world evaluation uses artificially injected distribution shifts rather than naturally occurring regime changes. These limitations make it difficult to determine whether the observed failures reflect a broader limitation of learned regime detectors or arise from the particular model architectures, residual definitions, and experimental settings considered. Therefore, the empirical evidence is not yet sufficient to establish the generality of the proposed failure mechanism or its implications for learned latent world models more broadly.

Questions for authors
  1. Could the authors provide a direct comparison between deterministic JEPA and F-JEPA-RD on the shape-switch task under the corrected evaluation protocol? In particular, would applying the same shape-sensitive detector to a deterministic JEPA's one-step innovations yield comparable performance?

  2. Appendix P suggests that distributional shape information is preserved in the latent representations but diluted in the multi-lag prediction residuals. However, the analysis uses a single trained model, and it evaluates the improved one-step innovation-based detector on only one seed. Does this finding hold across the seven seeds used in the main experiments? Additionally, how does the modified detector perform on the nonlinear swirl and linear-rotation controls?

  3. Proposition 3 establishes identical expected evidence for statistics that are affine functions of squared residuals under variance-matched distributions. However, nonlinear functions of residual energy, such as fourth-moment statistics, may still distinguish distributions with identical variances. Could the authors clarify the precise class of detectors covered by the theoretical blindness result? In particular, how does this result relate to shape-sensitive statistics constructed directly from squared residuals?

2 comments
AuthorsOpenReview ↗

Response to Reviewer p9rx

Q3 (scope of Prop. 3). The class is narrower than one Section 1 sentence implied. Proposition 3 is stated for statistics affine in the squared residual, α+βe2\alpha + \beta e^2 (p.5 l.253-258). This class includes the log-likelihood ratio between two zero-mean Gaussian laws, such as the variance grid of the frozen e-process; an elementary extension (Section 3.3, Scope) covers any fixed Gaussian pair.

The shape statistic from squared residuals that you describe is Proposition 4. There, log⁡tν(u)−log⁡φ(u)=cν+gν(u2)\log t_\nu(u) - \log\varphi(u) = c_\nu + g_\nu(u^2) is a nonlinear function of u2u^2, and its expected gain under any two-scale variance-matched mixture is non-negative (0.034 nats per coordinate for our mixture, p.5 l.259-267).

The two propositions mark the boundary you ask about. GLR statistics that re-estimate a covariance, and segmentation models such as BOCPD, are outside Prop. 3 (Appendix O, p.31 l.1655-1657; Appendix F shows BOCPD responds to the mixture on raw data).

One sentence in Section 1 (p.1 l.049-051) said "function of the residual energy"; it now says "affine in", and Section 3.3 states the class.

W1/Q1 (is the flow head necessary; deterministic JEPA). Not for the energy-free change: deterministic-JEPA codes do as well (0.433 vs. 0.375, Appendix Z).

The submission already compares the two representations, on the swirl with the frozen VAR e-process under the calm boundary (Table 33, P1, p.44): 0.719 vs. 0.736, paired p=0.393p = 0.393 over 7 seeds, and higher linear predictability for our codes (R2R^2 0.701 vs. 0.675, p=0.008p = 0.008). It also shows what the analytic base buys: without it, VAR R2R^2 falls from 0.691 to 0.064 and the frozen e-process from 0.500 to 0.002 (Section 4.4, p.8 l.386-389; Table 29).

On the shape switch this comparison was missing; we ran it (Appendix Z, 7 seeds, calm boundary, same statistic on the one-step innovation of a calm linear model):

  • Shape core: F-JEPA-RD codes 0.375 (calm 0.035), deterministic-JEPA codes 0.433 (calm 0.032), raw observations 0.421 (calm 0.037).
  • Energy e-process on the same innovations: 0.117, 0.080 and 0.068.
  • F-JEPA-RD vs. JEPA: paired Δ=−0.058\Delta=-0.058, Wilcoxon p=0.375p=0.375.

So on this change the flow head is not what makes the information reachable (revised Section 5). It contributes RATIO on dynamics changes and, via the analytic base, a code usable by standard detectors. The result also sharpens C5: the representation was not the binding constraint on this benchmark; the residual and the statistic are.

W2/Q2 (one model, one seed; swirl and linear rotation). Both were disclosed as limits (Table 28 caption, "one trained model"; Appendix P, p.32 l.1695: "a pilot rather than as a result"); both now use 7 seeds (Appendix Z):

  • Kurtosis diagnostic. The code innovation gains +0.396 excess kurtosis (range +0.303 to +0.491), more than the observation innovation (+0.235) and the flow residual (+0.085) on 7/7 seeds. The predicted scale is unchanged (post/pre 0.9993-1.0003).
  • Shape detector on the code innovation. It reaches 0.375 on the shape switch (calm 0.035) and 0.109 on the swirl (calm 0.025), where the energy statistic reaches 0.711. On the linear rotation it reaches 0.973 and energy 0.984, because the rotation raises the calm linear model's innovation energy. Config B gives the same pattern.
  • The model's own residuals. The same statistic reaches 0.067 on the one-step residual, 0.067 on the multi-lag residual and 0.008 at the inverse base point. So the loss occurs in the learned predictor's residuals, one-step and multi-lag, as the revision now states.
  • Device. The new runs use CUDA; seed 0 of the pilot gives 0.387 there (0.297 on MPS, Appendix P).

Broader limitation, or our architecture? We added the family Appendix O names as the next test, a neural likelihood-ratio statistic that learns the alternative online (Table 38). It detects the linear rotation (0.956) but stays at chance on the shape switch on every residual (0.012-0.031), including the code innovation where the shape core reaches 0.375. The failure is not specific to our readouts.

W2 (injected rather than natural changes). Only injection certifies that the change carries no innovation energy under the fitted VAR(1) and fixes its time. Appendix R enforces realized innovation energy equal to 1.000000 on a four-year recording (p.32-33 l.1716-1739).

Over 7 seeds (Appendix Z), every learned readout stays at chance, while the injection-aware reference reaches 0.666. On the code innovation the reference reaches 0.148: on this heavy-tailed background the code keeps part of the change, less than on synthetic data.

A natural change cannot be certified energy-free; our attempt with annotated TCPD series is in Appendix S. The negative result is stated for the monitors evaluated (Appendix O, p.31 l.1657-1658; Section 5, p.9 l.439-440).

Thank you for the detailed response and substantial revisions. However, my main concerns remain. In particular, the additional comparison with deterministic JEPA does not clearly demonstrate an advantage or the necessity of the proposed flow-matching architecture for the central shape-switch task. I appreciate the authors' efforts, but I do not find the additional evidence sufficient to change my overall assessment. Therefore, I will maintain my original score.

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

This work focuses on reliable regime detection, a task that loosely resembles to anomaly detection, in which a monitor is tasked to emit a signal when a stream of sequential regime states changes. In particular, the authors focus on states represented by (latent) world models, and seek to overcome the limitations of energy-based monitoring methods, which can be blind to changes when states abide to noise laws of identical variance, as these methods are sensitive to second moments only.

To attack the problem, this work introduces a generative world model that combines traditional joint-embedding predictive architectures with conditional flow-matching. The proposed method, together with alternatives from the literature, is evaluated according to a sequential detection benchmark, which indicates failure modes: the natural drift of distributions in "calm" or "no change" scenarios are misdetected as changes, which is not correctly captured by success metrics. Instead, in clear "change" scenarios, the proposed method successfully detect changes.

Critical strengths
  • Let me begin with an apology: I usually encourage authors with positive remarks, whatever the standards of the submission are, even if I find mistakes. Even when I am not familiar with the topic I do my best to read the literature to gain more confidence in my assessment. Unfortunately, for this paper, I could not parse its contents. I could have asked an LLM to help "translate" this paper for me, but I didn't do it.
  • In the distant past, I worked on traditional change detection methods applied to telecommunication networks. This is a very important topic, and I am convinced the authors genuinely worked hard to build on very recent literature (i.e., "Flow-JEPA: Flow matching for robust latent dynamics in JEPA world models") and adapt it to change detection. The idea I matured after several attempts at "digesting" this work suggests that combining reliable embeddings with conditional flow-matching is interesting, in my opinion. As such, I strongly invite the authors to keep pushing their ideas, and improve the exposition to make sure human reviewers can properly assess the quality and the implications of their work.
Critical weaknesses
  • The main methodology of this work, in section 3.1, is barely readable. As a reader, the whole section 3 is extremely hard to parse, the mathematical notation is not properly explained (Table in Section A does not help), the theoretical analysis in section 3.2 is only proved in Lean, so it is difficult to assess and follow.
  • The synthetic benchmark definition is not properly motivated. No standard benchmarks for change detection is discussed in the main paper, such as TCPD, or Ruptures, but are mentioned in the appendix (e.g. appendix S) with no explicit reference and accompanying text. Overall, the take home message from section 4 is really hard to evince. Is the proposed method competitive?
  • Overall, this work is barely readable, from abstract to conclusion. Appendices kept comments from an LLM/authors about moving content from main to appendix (See Appendix E).
Questions for authors
  1. I have read the AI statement from the authors, but this article uses terminology and constructions that are LLM jargon. The issues that have been amply reported by the practitioners communities about "unreadable English" permeate this paper: sentences lack a proper noun-verb-subject construction, use colloquial terms and are in general very confusing. There is no logical structure, there is a mix of high-level and extremely detailed exposition that hinders the task of an appropriate understanding. Do you confirm that this article has been carefully reviewed by humans who could understand the English used to expose the ideas of this work?
  2. Why the evaluation protocol is discussed in the related work section?
  3. In section 3.1:
  • you mention a "context", which has not been introduced before. Can you clarify?
  • you mention "a moment head reads the code": what does it mean?
  • you introduce mtm_t and sts_t, but don't define them
  • you mention that a residual network "corrects it": what is "it"? What is τ\tau?
  • what are "per-frame trunk features"?
  • is the residual network a transformer?
  • what do the green chekmarks and keywords embedded in the main text mean?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Jpyw

The revision defines each term you list where it first appears. Several answers were already in the submission, too far from where the question arises.

Q2, Section 3.1.

  • Context: the window Xt=xt−63:tX_t = x_{t-63:t} of 64 frames (p.4 l.179; Appendix A, p.15: "context window of 64 frames"). The revision names it at its definition.
  • Moment head, mtm_t, sts_t: (mt,log⁡st2)=hψ(zt)(m_t, \log s_t^2) = h_\psi(z_t) (p.4 l.186), an MLP 8→64→64 (Table 3, p.18). Its outputs are the mean and per-coordinate standard deviation of the Gaussian N(mt,diag⁡st2)\mathcal{N}(m_t, \operatorname{diag} s_t^2) over the future trajectory yty_t (same sentence; Appendix A).
  • "Corrects it": rϕr_\phi adds a learned correction to the base velocity vt⋆v^\star_t of Eq. 2; vt=vt⋆+rϕv_t = v^\star_t + r_\phi. τ∈[0,1]\tau \in [0,1] is flow time, the interpolation variable between noise and target. It is not stream time tt. The submission gave only τ∼U[0,1]\tau \sim U[0,1] (l.198), and the revision defines it.
  • Per-frame trunk features: the 64 outputs (one per frame) of the encoder's last convolutional block. ztz_t is a linear readout of the last one (Table 3: "linear readout of the last position"; "cross-attention over 64 trunk tokens").
  • Transformer? No: one 4-head cross-attention layer followed by an MLP (x,τ,c)→64→64→32(x,\tau,c)\to64\to64\to32, 44.4k parameters (Table 3, p.18 l.937-939).
  • Green check marks: names of the Lean theorems proving each statement. Section 3.3 explains this (p.5 l.238-240), but they first appear in Section 3.1. The revision moves them to Table 2.

Theory "only proved in Lean". Appendix B gives each argument in prose (p.15 l.797-798: "below we give the mathematical argument behind each proposition"; Base flow and Props. 1-5 on p.15-16).

The revision adds a one-line proof idea after each proposition, and Appendix B.1 now gives each proof in full, with standard results cited by name. Appendix B.2 maps each proof line to the Lean lemmas that check it and lists what is argued on paper only; each proposition is proved in Lean by composing these lemmas, with the ODE change-of-variables formula (A3) as the one assumed input (lean/FJEPA/Steps.lean; lake build passes with no sorry or new axiom).

Benchmark motivation and standard benchmarks. The synthetic benchmark provides the two things the claims need: an exact-likelihood oracle, and a change with provably no energy (Section 5, p.8 l.408-410; Appendix X, p.36).

Public benchmarks are cited in Section 2 (p.3 l.140-141: Truong et al., 2020, the ruptures review; van den Burg & Williams, 2020, TCPD). The TCPD result is in the main text (Section 4.5, p.8 l.396-397; Section 5, p.8 l.411-412), with the citation in Appendix S (p.33 l.1773). The revision names TCPD and ruptures in Section 2.

Is the method competitive? Partly. On the dynamics change, RATIO reaches 0.626, behind the frozen VAR e-process on the same codes (0.715/0.757) and BOCPD (0.655), and ahead of Scan-B, GLR-CUSUM and the diffusion detector in config A (Tables 1, 4 and 37). At lags (1, 2, 4), a horizon chosen on the swirl, it is the strongest learned detector, 0.818 against 0.682 (Table 26).

On the energy-free change no learned detector is above chance, ours included. That is contribution C5, and the paper locates where the information is lost (Appendix P). Over 7 seeds, a shape statistic on the code innovation reaches 0.375, against 0.117 for the energy statistic on the same innovation (Appendix Z). This is the readout the conclusion of the paper points to.

Q1. The protocol paragraph describes the rule the literature uses, which Section 4.2 then audits. It is now in Section 4.1.

Q0. We agree that the text was still hard to read; the revision rewrites Section 3 so that the terms you listed are defined where they appear, and edits the abstract and every other main-text section for readability.

Reviewer_JpywOpenReview ↗

Thank you for the rebuttal

Dear Authors, thank you for the rebuttal and your work to improve your submitted paper.

I particularly appreciate the effort you put to clarify theoretical claims and their proofs. There is a lot of new material to read, especially in the appendix, but you also put some cues in the main text that help in getting a gist of the proofs.

I read the revised work (mainly the main paper) and I still find it extremely difficult to parse. To be honest, what helped most was reading other reviewers' comments, summaries and strengths which highlight the main contribution of this work: identifying flaws in current evaluation protocols for change detection. In the proposed revised protocol, however, no method appears to be a "silver bullet" and the authors point at the latent encoder as the main culprit for poor performance.

In the current state, I still think this work is not ready for publication: in my humble opinion, it still reads as an AI generated work that should be parsed using AI to fully extract meaning out of it. I strongly encourage the authors to pursue their goals, as I think this paper has some potential, and invite them to consider a complete human-made rewrite of the paper to address human readers.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

We uploaded a revised PDF; every change is in blue. Sections and Tables 1-33 are numbered the same in both versions; Appendix Z (Tables 34-38, Figure 14) is new. Line numbers refer to the original submission.

  1. Section 3 rewritten; abstract and Sections 1, 2 and 4-6 edited for readability (Jpyw, qq4K).
    • The edits: the class of Proposition 3 (statistics affine in the squared prediction error); the loss located in the residuals of the learned predictor (one-step and multi-lag); the single-seed pilot replaced by the seven-seed result (Appendix Z); the abstract's audit and chance-level sentences are narrowed to the monitors they cover, "exact" density ratios become "exact-trace", and "strong power" becomes "among the strongest learned detectors", and "existing monitors" becomes "most existing monitors"; C2 states what Proposition 5 proves, and C4 and C5 are narrowed to the detectors of Table 1 and the monitors of the main evaluation; Section 1 points to the seven-seed real-recording runs; Section 5 adds that the flow head is not what makes the energy-free change reachable; the linear-control claim no longer says "exactly as the theory requires"; and a short pointer to the neural likelihood-ratio statistic.
    • Section 3 opens with a Problem setup paragraph that defines the stream, the change time, the residual and its energy, and how a monitor is scored.
    • It continues with a plain-language overview; context window, moment head, mtm_t, sts_t, flow time τ\tau and per-frame features HtH_t are defined where they appear; Lean theorem names moved from the running text to Table 2; a one-line proof idea and a plain-language sentence for each proposition.
    • Innovations, excess kurtosis and Lean are glossed at first use; the notation table (Appendix A) covers every main-text symbol.
  2. Scope of Proposition 3 stated in the main text (p9rx Q3), and one overstated sentence in Section 1 corrected.
  3. Evaluation protocol moved from Section 2 to Section 4.1, with a forward reference to the drift-robust rules (Jpyw Q1). The contributions are labelled C1-C5; each experimental subsection opens with the contribution it tests and ends with a takeaway.
  4. New Appendix Z with multi-seed results (p9rx Q1/Q2):
    • Energy vs. shape statistic on the same one-step innovation for F-JEPA-RD codes, deterministic-JEPA codes and raw observations (7 seeds, three settings, both configurations; per seed in Figure 14).
    • The shape statistic on the model's own residuals.
    • The kurtosis diagnostic over 7 seeds.
    • The full protocol re-run on a second device with both configurations: it fills Table 1's two empty cells and tests Appendix H's joint VAR null.
    • The real recording over 7 seeds.
    • An online neural likelihood-ratio statistic that learns the alternative, the family Appendix O named as the next test (Table 38).
  5. Editing leftovers and small errors fixed: seven "(moved from Section …)" labels; captions of every standard-rule table and figure now say so; Table 25 caption, Appendix E aggregation label and the Appendix Y table references corrected; Tables 32 and 33 condensed (qq4K); in Appendix X the W2W_2-optimal map needs S1/2S^{1/2}, and "freeze the moment head" becomes "clamp its outputs".
  6. Figures: Figure 1 redrawn as an overview of the paper's structure; Figures 2-13 re-rendered from the same results at print size with value labels; Figure 13(a) now uses the main text's calm-boundary rule, not the standard rule.
  7. Wording: the rule the literature uses is named the standard alarm rule throughout, and the warm-up bookkeeping without alignment the unaligned protocol. Section 5 states that the drift-robust rules and the aligned protocol are post hoc; Appendix T points to it. Long sentences are split; many captions lead with their result. Appendix P points to Appendix Z's seven-seed runs. Section 4.1 notes that BOCPD is not run on the same latents (Appendix F: a same-seed retrained model).
  8. Appendix proofs rewritten for readability: proof ideas first, standard results cited, routine algebra grouped; Lean mapping in App. B.2; statements unchanged except that Props. 1, 4, 5 name side conditions (A1), (A5), (A6) and Lean already use. After Prop. 4, centre and tails are rewarded relative to the shoulders.

Where each concern is addressed in the revised PDF:

  • Jpyw, Sec. 3.1 and notation → §3 "Problem setup", §3.1; App. A.
  • Jpyw, theory only in Lean → §3.3 proof ideas; App. B.
  • Jpyw Q1, protocol placement → §4.1. Jpyw, standard benchmarks → §2; App. S.
  • p9rx Q1, flow head and deterministic JEPA → §4.3, §5; App. Z (Table 34).
  • p9rx Q2, seeds, swirl and linear rotation → App. Z (Tables 34-37).
  • p9rx, broader limitation or our architecture → App. Z (Table 38); App. O.
  • p9rx Q3, scope of Prop. 3 → §1; §3.3 "Scope".
  • p9rx, injected vs. natural changes → App. R; App. Z.
  • qq4K, why care and readability → §1 contributions, §3, §4 takeaways; Tables 32-33 (App. S, T).
AuthorsOpenReview ↗

General Response to All Reviewers and the AC

Two reviews read the submission as a method that fails. It reports a limit, locates it, and fixes a protocol.

What the paper contributes (p.3 l.110-129; C5 now names the one-step and multi-lag residuals):

  • C1. F-JEPA-RD, a flow-matching JEPA with an analytic Gaussian base that yields three detectors from one model, with an exact base inverse, an exact base density and an exact-trace density (Section 3).
  • C2. Machine-checked theory: 123 Lean theorems with no sorry. Energy statistics are blind to variance-matched changes (Prop. 3), and the shape score is not (Prop. 4).
  • C3. A validity failure in the standard detection-rate protocol, and a fix. The usual rule credits calm drift as detection: on simulated paths with no model at all it reports 0.952 on episodes containing no change (Table 6). Our own readouts and a concurrent diffusion detector reported 0.77-0.96 on calm episodes under that rule (Table 7). We add a calm test and two drift-robust rules.
  • C4. A controlled head-to-head: three readouts and six independent detectors on identical latents (BOCPD: a same-seed retrained model, App. F), over 7 seeds, 2 configurations and 3 alarm rules, with raw monitoring paths released.
  • C5. A limit on predictive embeddings, with the loss located. An energy-free change that an exact oracle detects at 0.972 stays at chance for every learned monitor of Table 1 (0.003-0.105). The encoder keeps the change: the one-step code innovation gains +0.361 excess kurtosis (Table 28). The multi-lag residual the readouts consume dilutes it: +0.052.

The method is not reported as failing where the theory says it should work.

  • On a change in the dynamics, RATIO reaches 0.626 (Table 1), and 0.818 against 0.682 for the best baseline at lags (1, 2, 4), a horizon chosen on the swirl (Table 26).
  • On the linear control, which its null explains, it is silent under both drift-robust rules (0.000, calm test 0.012).
  • On the energy-free change the paper reports a limit, measures where it arises, and shows with a pilot that the information is reachable (Appendix P; seven seeds in Appendix Z).

What the revision adds (Appendix Z, 7 seeds, calm boundary; logs and per-seed outputs are in the revised supplementary material):

  1. Presentation, per the Summary of Revisions.
  2. C2/C5 measured on learned codes.
    • On the one-step innovation of a linear model fitted on F-JEPA-RD codes, the energy e-process reaches 0.117 and the shape core of Prop. 4 reaches 0.375 (calm 0.038 / 0.035). The shape core is higher on 7/7 seeds (Wilcoxon p=0.016p=0.016).
    • On the swirl, whose innovations keep their shape, it stays at 0.109 while energy reaches 0.711.
    • On the model's own one-step and multi-lag residuals it stays at 0.008-0.067.
    • The code innovation gains +0.396 excess kurtosis (range +0.303 to +0.491) against +0.085 for the flow residual, on 7/7 seeds.
  3. C4 completed. RATIO in config B: 0.580 on the swirl. Diffusion in config A: 0.299. Both are at chance on the shape switch. The config-A cells replicate on CUDA: RATIO 0.630 vs. 0.626.
  4. Appendix H's joint VAR null. It removes RATIO's calm drift (+4.0 → −0.04 nats/step) but, unlike what Appendix H expected, changes what RATIO detects: 0.178 on the swirl and 0.920 on the linear control. We keep the composite null and report both.
  5. Real recording, 7 seeds. Every learned readout is at chance (≤0.058), while the injection-aware reference reaches 0.666. The same reference on the code innovation reaches 0.148.

Corrections.

  • (a) Section 1, p.1 l.049-051, said "every statistic that is a function of the residual energy". Proposition 3 is stated for statistics affine in e2e^2, and the sentence now says so.
  • (b) A leftover sentence, "The first prediction holds." (p.6 l.279), contradicted the sentence before it and is removed.
  • (c) Equation 7 now states its hypothesis (the flow equals the true conditional law). Prop. 1 escapes it because the base flow is a product over lags (not because Φ\Phi is diagonal), and the joint lag covariance gains its transpose.
  • (d) Under the calm boundary, the calm-test rate is a first-crossing rate after ν\nu and is expected at or below 0.05; three captions said "near 0.05".
  • (e) Appendix N said the oracle is "constant at 0.904" across horizons; 0.904 is RATIO's best value, and the oracle's is 0.992 (Table 27).
  • (f) Section 3.1 said "no residual" is realizable; it is, but not optimal under independent coupling, and Proposition 1 now states ∣Φii∣<1|\Phi_{ii}|\lt 1.
  • (g) The Prop. 5 sketch lacked a factor ½; Lean has 123 theorems (78 plus 45 step lemmas).

Apart from (e) and (g), no previously reported number changes; Figure 13(a) is re-plotted under the calm boundary.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.