acceptodds
Under review as a conference paper at ICLR 2027

Your World Model Is Not Blind, Your Detector Is: Native Prediction Errors for Regime Detection

Abstract

A world model that has learned how a system evolves should be the first to notice when that system starts to evolve differently. Yet learned representations are routinely judged blind to regime change: a predictive representation that recovers the hidden state of a dynamical system still lets standard sequential change-point detectors miss most changes. We show that this blindness belongs to the detector, not the representation. While standard detectors assume latent states follow a linear process, we instead query the world model directly-scoring each transition by its one-step predictive error normalized by covariance fitted on nominal (calm) data. This native-residual detector requires no retraining and elevates detection rates from a small minority of shifts to near the exact-likelihood ceiling. We then prove why this gap exists. The standard detector’s variance channel can only gather evidence above a threshold fixed by its alternative hypothesis grid, which no grid refinement resolves. Furthermore, we prove an absorption principle: a flexible predictor trained on calm data systematically explains away the component of a regime shift aligned with its learned dynamics, while a learned variance head re-weights the residual back toward nominal behavior. This reveals why standardizing by model uncertainty is inherently deceptive. Across convolutional, transformer, and period-folding backbones, native-residual detection alarms roughly three times faster than classical baselines, fails predictably only when an encoder discards one-step temporal order, and outperforms recent detector families, calm-fitted forecasters, and a zero-shot foundation model.

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

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TBuiOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a compelling diagnosis of detector–representation mismatch in learned world models, supported by an elegant residual-based intervention and useful theoretical identities.

Critical strengths
  1. The paper raises a compelling research question: when a world model performs poorly at detecting regime changes, is it because the learned representations lack information about such changes, or because the downstream detector fails to effectively utilize that information? This perspective offers a valuable lens for re-examining existing representation evaluation protocols and holds significant research merit.
  2. The proposed Native-Residual Detection (NRD) method does not require model retraining; instead, it performs detection by directly utilizing the one-step prediction residuals generated by the world model itself.
Critical weaknesses
  1. The core operation of NRD involves replacing the traditional linear VAR predictor with a predictor derived from a "World Model," while retaining existing covariance-based scoring and e-process mechanisms. Although the problem analysis offers a fresh perspective, the actual methodology closely resembles established paradigms for forecasting-based anomaly detection; thus, the level of independent innovation at the algorithmic level remains limited.
  2. The paper’s most notable performance gains are derived primarily from synthetically constructed benchmarks, with a lack of validation on real-world datasets. In particular, the SWIRL benchmark is designed as a challenging scenario featuring matched marginal distributions and non-linear dynamic shifts.
  3. NRD does not consistently outperform traditional detectors. On the SWIRL-LIN benchmark, its detection rate drops from 0.963 (achieved by a frozen detector) to 0.002; a significant decline is also observed on the ROTOR benchmark. Furthermore, the method proves almost entirely ineffective at detecting shifts when applied to PatchTST. These findings indicate that the method is highly dependent on the relationship between the regime shift and the predictor, as well as on whether the learned representation preserves one-step temporal information.
Questions for authors

Why are the unique properties of "co-adapted predictors" a widespread phenomenon that has not been sufficiently explored in existing forecast-based anomaly detection methods?

3 comments
AuthorsOpenReview ↗

Response to Reviewer TBui

Thank you for the review.

W1: NRD "closely resembles established paradigms for forecasting-based anomaly detection."

The pipeline is shared on purpose, so that the predictor is the only variable (C1; Section 2, p.3, l.138-143). The comparison shows two things.

  • Which predictor matters. On the same latent and pipeline, post-hoc predictors stay near the frozen level: LSTM 0.157 (Table 20, p.30), and a calm-fitted MLP 0.304 on Swirl and 0.023 on Sectors (new, Table 21). The co-adapted head reaches 0.951 and 0.946.
  • Why. The theory explains the frozen detector's miss (C2) and when the model's own predictor fails (C3).

W2: "gains are derived primarily from synthetically constructed benchmarks."

See General Response 3.

W3: "On the SWIRL-LIN benchmark, its detection rate drops … a significant decline is also observed on the ROTOR benchmark."

These drops belong to Var-JEPA. They are the designed controls of Section 5.2. JEPA keeps 0.951 and 0.943 (General Response 1), and the min-p rule recovers 0.947 on Rotor with Var-JEPA (General Response 2).

W3: "ineffective … to PatchTST … dependent on the relationship between the regime shift and the predictor … one-step temporal information."

Yes, and both are results of the paper: Theorem 2 (p.5), the precondition "stated and tested" (C4, p.3; p.8, l.408-411). On Swirl, the frozen detector also fails on PatchTST (0.001). On Rotor, NRD on PatchTST reaches 0.929.

Q: why the unique properties of "co-adapted predictors" are "a widespread phenomenon that has not been sufficiently explored" in forecast-based anomaly detection.

We claim it for the backbones we test, not as widespread. In forecast-based anomaly detection the forecaster is the model, so no separate representation is confounded. The confound appears when a self-supervised latent is frozen and scored by a fixed detector, and the predictor it was trained with is thrown away. The new post-hoc MLP rows (Table 21) show that, on Swirl and Sectors, only that predictor reads the change.

Although the view of correcting errors is interesting, the method doesn't bring new method to solve this problem. In addition, existing experiments are not enough to support my positive score.

Thank you for reading our responses and for taking the time to follow up. Thanks a lot !

Reviewer_epN7OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a method for regime change in sequence predictor: it is basically normalized predictor error. The paper compares this method to linear and various nonlinear detectors built on top of the sequence model. The motivation is from world models, although all experiments are synthetic time series.

Critical strengths

The paper seems to find a useful observation regarding regime change detection for world models.

Critical weaknesses

The method proposed is a straightforward application for prediction error for detection, I'd rank the novelty as low. It is also not clear if the findings will extend to world models beyond synthetic data.

Questions for authors

I don't have any particular question, unless there is something I missed in the review?

3 comments
AuthorsOpenReview ↗

Response to Reviewer epN7

Thank you for the review.

"a straightforward application for prediction error for detection … novelty as low."

The detector is simple on purpose; C1 is that it is drop-in. Prediction error alone does not give the result. With the same pipeline:

predictor on the same latentSwirlSectors
frozen linear null0.2230.029
calm-fitted MLP (post hoc)0.3040.023
the model's own head (NRD)0.9510.946

(7 seeds, Table 21; also LSTM 0.047 on the observations, 0.157 post hoc on the latent, and TimesFM 0.000 in Section 5.3.) The theory says why the standard detector misses (C2) and when the model's own predictor fails (C3), and the controls confirm it.

"not clear if the findings will extend to world models beyond synthetic data."

We do not claim that. Section 6 (p.9) limits the claim to synthetic benchmarks; we make no real-data claim. On NAB the frozen detector already sees the changes, and it stays best when both detectors run (Table 24).

"unless there is something I missed in the review?"

  • The table above.
  • The controls that Theorem 2 predicts (Table 2).
  • The standardization trap: NRD-σ^\hat\sigma gives 0.000 (Table 1); in the sweep of Table 18 it gives 0.109 against a fixed metric and 0.878 once the calm metric is re-fitted (Theorem 3).
Reviewer_epN7OpenReview ↗

I have read your response

Thank you for your reply and extra information. I agree to some of the points you made, and I agree that simplicity is a feature.

Thanks for the response

Reviewer_Yi1rOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper studies whether the poor regime change detection is from the learned representation of from the detector used to evaluate it. The authors proposed NRD which uses the model's own prediction errors instead of errors from a separately fitted linear model. Then it examines whether those errors depart from their normal behavior. On the synthetic SWIRL benchmark, this change increases JEPA's detection rate from 23.1% to 95% at a 5% false alarm probability. The paper also studies why some changes remain difficult to detect even with the model's own predictor.

Critical strengths
  1. The paper presents a clear example of how detector selection can affect conclusions about the learned representation quality. With the same JEPA encoder, detection improves from 23.1% to 95%. This supports the main argument that a poor detection result does not necessarily mean the representation has lost useful information. Improvements with Transformer-JEPA and a TimesNet-style encoder also support this finding.
  2. The paper examines when the proposed method fails, rather than reporting only the favorable results. Its analysis explains how the direction of a change can affect the magnitude of the prediction error. The experiments include scenarios where NRD performs much worse than the linear detector, which also help readers recognize the limitations of the method.
Critical weaknesses
  1. The experiments use synthetic data and it is unclear if the method would help in practice. Specifically, the strongest result are based on a synthetic benchmark designed to be difficult for the linear detector, and the paper does not show an advantage on real world datasets. Also, on other synthetic settings, the results are different. For Var-JEPA, detection decreases from 93.3% to 16.% on ROTOR and from 96.3% to 0.2% on SWIRL-LIN. And the paper does not make it clear that which detector should be used before the change occurs.
Questions for authors
  1. Var-JEPA's detection rate changes from 36.6% in the main results to 83.7% in another run with the same seeds and settings. What is the reason for this difference and which findings remain stable across repeated runs?
  2. The theory assumes that the latent state distribution is unchanged across regimes. How did authors check this? Because matching the distribution of individual hidden states does not necessarily match the distribution of the entire encoded windows.
  3. NRD uses a predictor trained jointly with the encoder,
  4. Is there a way to decide whether NRD or the linear detector is more suitable using the data only available before a change? This is more meaningful for practical use.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Yi1r

Thank you for a careful review. The new results are in Appendix L.

W1: "synthetic data … the paper does not show an advantage on real world datasets."

Correct, and the paper says so. Section 6 (p.9, l.457-465) and Appendix F (p.21, l.1113-1126) explain it: NAB's changes are already visible to the frozen detector (median Rlin=1.91>τGR_{\rm lin}=1.91\gt\tau_\mathcal G). The new NAB run with both detectors confirms that the frozen detector stays best there (General Response 3). We claim NRD's advantage only for changes that the linear alternative class cannot see.

W1: "For Var-JEPA, detection decreases from 93.3% to 16.% on ROTOR and from 96.3% to 0.2% on SWIRL-LIN."

These drops belong to Var-JEPA. They are the controls of Section 5.2, the analysis you list as Strength 2. In new runs, JEPA keeps 0.943 on Rotor and 0.951 on Swirl-Lin, and reaches 0.946 on Sectors, where the frozen detector reaches 0.029 (General Response 1).

W1/Q4: "which detector should be used before the change occurs?"

Run both; this was a fair gap, and we added it (General Response 2). The min-p rule stays within 0.045 of the better detector on synthetic data, and costs detection on NAB.

Q1: "36.6% … to 83.7% … which findings remain stable across repeated runs?"

JEPA's result is stable; Var-JEPA's NRD level is not. Var-JEPA samples latents during training, so each run is a different draw (Appendix H.7, p.30). We added three draws per model (General Response 4).

  • Stable: JEPA's near-oracle NRD (0.950-0.959 over five runs); Var-JEPA's frozen detection (0.049-0.090); and NRD > frozen in every draw.
  • Not stable: Var-JEPA's NRD level (0.366-0.837). RrawR_{\rm raw} alone does not order the draws, and we corrected Appendix H.7, which had suggested that it does.

Q2: "The theory assumes that the latent state distribution is unchanged across regimes. How did authors check this? … individual hidden states [vs] the entire encoded windows."

You are right. Matching holds by construction for the hidden state, not for the encoded window, and the "within 1%" sentence had no released check behind it. What we changed:

  • Theorem 2 without (M). Appendix D (proof of Theorem 2, p.19, l.972-973) already had the general form Ilin−Imodel=2EQ⟨c,δ⟩+ΔMI_{\rm lin}-I_{\rm model}=2 \mathbb E_Q\langle c,\delta\rangle+\Delta_M, with ΔM=(EQ−EP)[∥c∥2+2⟨a,c⟩]\Delta_M=(\mathbb E_Q-\mathbb E_P)[\Vert c\Vert^2+2\langle a,c\rangle]. The theorem now states it first (proof via Lemma 5, App. D), and Lean checks every line of the proof (App. E.1).
  • Measurement (Table 22). We compare latents from windows entirely after the change with calm windows: covariance traces, an episode-block MMD test, and ΔM\Delta_M.
    • JEPA on Swirl and Sectors: the covariance changes by at most 4.1%, and ΔM\Delta_M stays within a few thousandths of the calm energy. (M) is a good approximation there.
    • Var-JEPA: the latent contracts (trace ratio 0.861 on Swirl), and on Swirl the estimated ΔM\Delta_M exceeds the removed inflation on every seed of the new CUDA draws (1.009 vs 0.079 of the calm energy, Table 22). So the submitted reading of Rraw<1R_{\rm raw}\lt 1 as alignment (Corollary 1) does not hold for Var-JEPA: its absorption happens in the encoder. We corrected Corollary 1. Its conclusion for the frozen and standardized views, that they cannot use the variance channel, is unchanged.

Q3: "NRD uses a predictor trained jointly with the encoder," (cut off)

  • Leakage: none. "No threshold ever sees a test episode or a label" (p.6, l.308; Appendix F).
  • Is joint training needed? Yes (Section 5.3, "Which null?"; the design rule of C3). A calm-fitted MLP in place of the model's own head, on the same latent, stays near the frozen level: 0.304 vs 0.951 on Swirl, and 0.023 vs 0.946 on Sectors (7 seeds, Table 21).

We hope our answers help clarify our contributions and address your concerns.

Thank you for your time and consideration.

Reviewer_Yi1rOpenReview ↗

Response

Thank you for the additional experiments. They addressed most of my theoretical concerns and I have raised my rating from 2 to 3, although my major concern about the practical benefit on real world cases has not been addressed.

Thank you so much for your responses

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

We uploaded a revised PDF. Changes are in blue; the title is unchanged. Page (p.) and line (l.) numbers in our replies refer to the original submission. The new experiments test the submitted claims and support them, with two corrections (Corollary 1; Var-JEPA's re-runs in §5.1, §6 and App. H.7). Details: App. L. The new supplementary adds App. L's code and results and updates the Lean (App. E).

Where each concern is answered

  • Yi1r W1, TBui W3 (drops on Rotor and Swirl-Lin) → §5.2, "JEPA in the same regimes"; App. L, Table 21, Fig. 13.
  • Yi1r W1/Q4, TBui W3 (which detector to run before the change) → §5.3, "Which detector, before the change?"; App. L, Table 21, min-p column.
  • Yi1r Q1 (36.6% vs 83.7%) → §5.1; App. L, Table 23; App. H.7.
  • Yi1r Q2 (assumption (M)) → §4.1, item (M), Theorem 2 and Corollary 1; App. D; App. L, Table 22.
  • Yi1r Q3, TBui Q, epN7 (is the jointly trained predictor needed?) → §5.3, "Which null?"; App. L, Table 21, post-hoc MLP column.
  • All reviewers (real data) → §6; App. L, Table 24 (NAB with both detectors).

Every other change, by section

  • Abstract: the submitted sentences, in the same order. "On a JEPA latent" is added to the near-oracle claim. "Roughly three times faster" became "sooner than the frozen detector", and the comparison with other detectors is now stated for JEPA latents. "Helps where the frozen detector is blind and" is added before "fails predictably", and "only when" became "when". "Is inherently deceptive" became "without refitting the calm metric is deceptive".
  • §1: VAR(1) and e-process are glossed. C3 notes the measured marginal term of Theorem 2 where (M) fails. C4 gives the new theorem count. The contributions are labelled C1-C4. C1 says that the gain, not near-oracle detection, holds across encoders. C2's last clause is reworded. C3 glosses "co-adapted" and says "the measured energy ratios" where it said "every measured energy ratio". C4 gains pointers. C1's sensitivity range over recent detector families is recomputed from the result files: 2.9-214×, where it said 3-240×. The Var-JEPA gain in §1 is marked as one training draw. Fig. 1 is redrawn.
  • §2: ELBO and EMA are spelled out.
  • Throughout: long sentences and paragraphs are split; the claims are unchanged.
  • §3: a short problem formulation; the four regimes in one paragraph; one plain sentence above each result. The Lean sentence now says what is checked (proof steps except those marked, for finitely supported laws). PSR, the restart discount ρ and the drift prior variance are named where first used. Fig. 2 says "no growth" where it said "closed", and "decreases to 1" corrects a direction word. The dilution sentence gives the detection after the switch (0.158, Table 5) where it said "collapses". The grid-density sentence now quotes Var-JEPA's frozen detection, whose inflation it discusses (it quoted JEPA's).
  • §4: one plain sentence above each result; (M) is stated for the latent and measured; Theorem 2 in its general form; Corollary 1 narrowed for Var-JEPA, and its variance-channel conclusion to the frozen and standardized views; the drift-channel expectation follows as plain text. (O)'s example says "at a calm-risk optimum". Fig. 3 is redrawn.
  • §5: one opening sentence and one takeaway per subsection. Sentences that only repeated earlier results (in "Controls and ablations" and two one-line summaries) are removed; their content stays in C3, §5.2 and Table 1. The min-p rule is stated as chosen after the e-value average fell short. For Var-JEPA's re-runs, only the frozen energy ratio is stable, not all ratios; the same correction is made in §6. Figs. 4 and 5 print their key values. Seed counts other than 7 are stated (Var-JEPA's inflation: 3 seeds), "eight architectural variants" is scoped to the small tier, PatchTST reaches the ceiling on Rotor "like the other deterministic backbones", and the marginal-family bound is printed as ≤0.222. The largest gap above the per-seed oracle is printed to four digits (0.0035), since the rounded table entries suggest 0.004. The captions of Figs. 1, 2, 4, 5 and Tables 1-3 open with their takeaway.
  • §7: the near-oracle claim is stated for JEPA.
  • Appendices C and D: proofs rewritten for readability: proof ideas first, standard results cited, routine algebra grouped; Lean mapping in App. E.1; the rewrite changes no statement. App. E: 148 theorems, all audited. Proposition 1 is stated for the expected log-evidence, with crossings "above about 1.169" (1.255 for the grid); App. C.2's Lean sentence says ḡ in [1, 1.125]; App. C lists (A1)-(A3) and App. D positivity (P); Theorem 2's readings state their assumptions. App. A lists App. L and the new symbols. App. F points to the NAB runs of App. L and states the series filter. App. G names the checks fixed before training "design checks". Sentences in App. G, H.1, H.2, H.4, H.7, I, J and K, and the captions of Table 9 and Figs. 10 and 11, are corrected to match the tables and result files. App. L is new.
AuthorsOpenReview ↗

General Response to All Reviewers

Thank you for the reviews. Several concerns are answered in the submission, so this response restates its claims (p.2-3) before the new results.

The claims.

  • (C1) A world model's own one-step predictor, read in a calm-fitted metric, is a drop-in detector. On the same frozen JEPA latent it raises detection from 0.231 to 0.950 (oracle 0.965; Table 1, p.7).
  • (C2) The frozen detector misses the change because its variance channel gains evidence only above τG=1.2164\tau_\mathcal G=1.2164. A denser grid does not help (Table 16, p.28).
  • (C3) A calm-trained predictor absorbs changes it can express (Theorem 2). Scaling by its own variance fixes the null at one (Theorem 3).
  • (C4) The theorems are checked in Lean, and the method's precondition is stated and tested (p.3).

1. Where NRD loses (Yi1r, TBui). The losses are specific to Var-JEPA's latent; JEPA's NRD does not collapse there (0.943 and 0.951). Swirl-Lin and Rotor are designed controls (p.7, l.363-364: "NRD helps exactly where the change is not expressible by the calm predictor, and stops helping where it is"). We ran JEPA in all four regimes (7 seeds, Table 21):

regimeJEPA frozen → NRDVar-JEPA frozen → NRD
Swirl0.223 → 0.9510.065 → 0.765
Sectors0.029 → 0.946 (oracle 0.941; detection saturates near 0.95, App. L)0.017 → 0.257
Rotor0.956 → 0.9430.955 → 0.220
Swirl-Lin0.955 → 0.9510.952 → 0.532

These are new CUDA training draws (Table 21); the Table 1 run gives Var-JEPA 0.090 → 0.366 on Swirl (see item 4).

Where Var-JEPA's NRD fails, its latent contracts under stress (covariance-trace ratio 0.187 on Rotor, 0.150 on Swirl-Lin). PatchTST on Swirl is the stated precondition (p.8, l.408-411), and the frozen detector fails there too (0.001).

2. Which detector to run before the change (Yi1r Q4, TBui W3). On our synthetic regimes, run both. Calm data alone cannot choose, because absorption depends on the unknown shift (Theorem 2).

  • We alarm when either detector is extreme for its own calm law (min-p, same calibration).
  • This rule is within 0.045 of the better detector in all eight cells.
  • On Rotor with Var-JEPA it gives 0.947, where NRD alone gives 0.220.
  • Averaging the two e-values falls short there (0.670).

3. Real data (all). We make no real-data claim, and Section 6 (p.9, l.457-465) explains why with the theory.

  • On NAB the frozen detector's inflation is above τG\tau_\mathcal G on 9 of 10 series, so the frozen detector already sees the change.
  • With both detectors on NAB (3 seeds; Table 24), the frozen detector stays best (JEPA 0.522, Var-JEPA 0.477), and min-p costs detection (0.352 and 0.347).
  • The benchmark is synthetic because the question needs a change that linear statistics cannot see, and an exact-likelihood oracle.

4. Training draws (Yi1r Q1). JEPA is stable; Var-JEPA's level is not. JEPA's NRD gives 0.950, 0.950, 0.951, 0.959 and 0.956 over five runs (the first two are the same draw). Var-JEPA's ranges from 0.366 to 0.837, always far above its frozen detection (0.049-0.090) (Table 23).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.