acceptodds
Under review as a conference paper at ICLR 2027

Learning to Test: Latent Representation Learning for Dynamic Instability Detection

Abstract

Stability assessment is central to safety-critical systems governed by differential–algebraic equations (DAEs), whose dynamics are constrained by physical laws and admissibility conditions. These systems operate under stochastically varying environmental inputs, so stability must be reassessed as the context distribution shifts, yet repeated large-scale DAE simulation is prohibitive in high-dimensional or real-time settings. We propose a test-oriented learning framework that tests a trajectory-level chance constraint from deployment contexts alone. Rather than re-estimating physical parameters or re-solving the DAE, we learn a latent representation from baseline context–trajectory data that retains stability-relevant information and is regularized toward a tractable reference distribution, and calibrate a latent discrepancy test over the composite safety null. We establish exact validity for oracle calibration and asymptotic validity for its empirical implementation under class-conditional latent invariance, and show that the guarantee degrades gracefully when this invariance holds only approximately. On simulated dynamical systems and a power-system real-world example, our approach strikes a balance between validity and power, whereas other baseline methods each fail in at least one regime. An ablation that injects controlled violations of the invariance assumption shows that decisions stay correct while the violation is below the safety margin, as the theory predicts.

Primary area:
learning theory
Keywords:
representation learning, hypothesis testing, dynamical systems, distribution shift

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yz4NOpenReview ↗
Rating
2
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a learning-to-test framework to verify the stability of dynamical systems. The goal is to remove the expensive re-simulations and provide efficient stability criteria. A theoretical framework is also provided for the statistical test in the latent representation space.

Critical strengths

The idea is well-motivated to take the place of the expensive DAE simulations for stability analysis. The theory is also provided. The whole framework seems doable in some test cases.

Critical weaknesses

The first concern is about the connection to the classic stability definition in dynamical systems. The author's definition is not well recognized in the dynamical system domain. The second concern is about the effectiveness of uniform regularization and the guarantee of the invariance assumption.

Questions for authors
  1. In Fig. 1, the author said that under stable regime, the learned latent representation is approximately uniform. Any explanations?
  2. One concern is that the authors should first explain how their stability definition relates to classical stability theory for dynamical systems, including Lyapunov and asymptotic stability. Definitions 1 and 2 characterize finite-horizon chance-constraint satisfaction, which is not generally equivalent to these classical notions. Without a clearer connection, it is difficult to see what the proposed theory contributes to dynamical-system stability analysis.
  3. The quality of Fig. 2 needs to be improved. Some texts/formula are grey and hard to clearly identify. It’s also not easy to capture the main flow.
  4. The role of uniformity regularization is questionable since it does not guarantee uniformity always holds. The implemented test uses an empirical reference.
  5. The validity guarantee relies on class-conditional latent invariance, which the training objective does not enforce.
  6. The tests systems are insufficient. More realistic cases may need to be considered. Sensitivity with respect to the hyper-parameters should also be considered.
Reviewer_35RROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a test-oriented learning framework for stability assessment that tests against a trajectory-level chance constraint only using deployment context rather than re-estimating or re-solving the differential-algebraic equation (DAE). They establish validity for oracle calibration and asymptotic validity under class-conditional invariance. Experiments on synthetic and real-world data demonstrate a balanced trade-off between validity and power.

Critical strengths

The proposed learning to test framework is well articulated and justified, while circumventing the need for re-estimation or simulation.

Critical weaknesses

The theoretical results are welcome; however, these seem disconnected from the experiments.

The experimental results are convincing from a conceptual perspective, however, there are not enough details in the main body of the paper or direct references to the appendix that describe the details of the methodology and metrics used in the experiments. For example, in Figure 4 it is not clear why the ground-truth rejection rate is not presented given that the data is synthetic so the true rejection rate is known given the magnitude of the shift (as shown in Figure 7) and how exactly are the type-I and -II errors obtained in Table 1.

Questions for authors

Why is h_{k+1} in (7) not explicitly adding to h_k as in (6)?

How is omega in the weighting term set?

Figure 3 is not sufficiently explained or justified.

Consider moving the details of the model in Section 3.1 to the appendix to make room for more details of the testing procedure, experiments and connecting the theoretical results with the empirical findings.

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

This paper addresses a timely problem and presents a general framework that could be particularly useful for emerging trajectory-level anomaly and distribution-shift detection problems. A key strength is that the proposed method requires only in-distribution baseline context–trajectory data for training, without relying on explicitly collected OOD samples. The framework learns a stability-relevant latent representation from baseline data and performs a calibrated distributional test using deployment contexts alone, thereby avoiding repeated trajectory simulation or re-estimation of the underlying physical dynamics. In addition, the paper provides theoretical guarantees for the validity of the proposed test and analyzes how these guarantees degrade when the latent invariance assumption is only approximately satisfied.

Critical strengths

S1. The problem is well posed, and the paper is clearly written and easy to follow. I found the proposed formulation particularly valuable as a principled template for trajectory-level anomaly and safety detection in learning-based systems. Rather than introducing a task-specific detector, the paper connects representation learning, latent distribution testing, and trajectory-level safety guarantees within a unified framework.

S2. I also find the direction timely from an architectural perspective. For emerging predictive or representation-based models, such as JEPA-style architectures, scalability alone is not sufficient; one would also like a principled way to justify whether the learned representation is suitable for downstream safety or anomaly detection. This work provides a useful framework for making such a connection through trajectory-informed representation learning and calibrated testing.

S3. Another strength is that, if an appropriate context variable XX can be defined in advance, the framework enables safety assessment using deployment contexts alone, without requiring the corresponding trajectories to be executed or simulated. In particular, the formulation suggests a useful perspective in which one first identifies a safety-relevant context XX, learns a representation that preserves its relationship to trajectory-level behavior, and then tests whether the deployment distribution remains within the calibrated safe regime. This makes the setup practically appealing and suggests several natural follow-up directions, including how to design XX, how to preserve the relevant safety information in the latent representation, and how to integrate such tests into larger predictive or control architectures.

Critical weaknesses

W1. The test statistic appears to require a sufficiently large number of context samples for reliable calibration. This may limit applicability to settings where such context sampling is feasible, especially in high-dimensional or data-scarce regimes.

W2. Some parts of the learning pipeline still feel somewhat intuitive or architecture-dependent. In particular, there seems to be substantial room for improving the encoder, decoder, and update network, and the paper could benefit from a clearer characterization of which representation properties are theoretically essential for the test.

Questions for authors

Q1. Is the assumption that one can sample the context variable XX common in related literature? I am curious about which application domains naturally lead to this problem formulation. The motivating examples are helpful, but a more detailed discussion would improve the paper.

Q2. It would also be helpful to discuss possible architectural extensions of the framework. For example, are there directions involving more expressive encoders, decoders, or update networks that the authors view as promising future work?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.