acceptodds
Withdrawn from ICLR 2027

Learning Editable Predictive States for Ocean World Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurately modeling the coupled evolution of the ocean carbon–physics system and its accompanying events is a long-standing challenge in climate science. Conventional data-driven methods often rely on the regression capacity of neural networks to map past surface fields to future ones without a state into which a physical change could be written, or model individual events in isolation, neglecting the information that sparse event records carry about the continuous fields. In this paper, we propose the State-Perturbation Joint-Embedding Predictive Architecture (SP-JEPA), a world model that represents the environment by a compact state and advances this state autonomously. Central to our method is a basis-anchored state: each frame combines the coefficients of a fixed spatial basis with a learned residual representation, so that a perturbation written into the coefficients changes the decoded field exactly by the prescribed pattern, and a single predictor propagates both the natural and the perturbed futures without reading future observations. Historical fields and source-grounded event observations jointly initialize this state, from which future fields and events are read out. We evaluate natural prediction on three decades of global monthly surface carbon–physics fields with six event families, and response prediction in a controlled transport–exchange environment that provides the true future before and after each perturbation. On the validation splits, SP-JEPA reaches a normalized root mean squared error of 0.131, level with the best field forecaster and 18% below the best compared latent world model, and fusing historical events lowers event-prediction error by 9–11% against absent or mismatched event inputs without increasing field error. On mechanisms unseen during training, its rollout reproduces the mean response of each pulse and reaches a response skill of 0.575 and a latent counterfactual skill of 0.206, whereas compared latent world models score below zero.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
World models, Predictive state representations, State perturbations, Joint-embedding predictive architectures, Ocean forecasting

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_k68oOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper describes a model of the ocean carbon-physics system. For a non-expert in this area like me, it is incomprehensible. A lot of terms like “monthly surface fields”, “natural forecasting”, “physical edits”, and “localized pulses” are unexplained when used in the introduction. How should I understand a sentence like “In an environment with paired true futures, we separate the response error introduced by encoding and readout from the error added by the rollout, which gives a lower bound on response skill”, when terms like “paired true futures”, “response skill”, etc., are undefined in the context? The next part of the text is about “source-grounded event familes”, which leaves me clueless. The paper continues to be incomprehensible in the description of the model. For instance, I know what an action is in world models for robotics, but I need an explanation of what is meant in ocean modelling. I could have used AI to help me understand this paper, but then we end up with AI-written papers that only AI can read, so I have not done that. It is my impression that the authors have done a good job of testing their model against other models and so on, but in its present form, the paper would fit only in a very specialized journal or conference.

Critical strengths

Unknown

Critical weaknesses

Almost impossible to read if you do not have detailed knowledge of the narrow field of study

Questions for authors

Have you considered that the audience of general machine-learning scientists at ICLR should be able to at least understand the basic ideas in the paper?

Reviewer_qp5vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes State-Perturbation Joint-Embedding Predictive Architecture (SP-JEPA), a world model that represents the environment by a compact state and advances this state autonomously while also making them editable. The authors evaluate this architecture on natural ocean carbon–physics forecasting with fused event observations as well as on the response prediction in a synthetic paired transport–exchange setup.

Critical strengths

S1. The setup is well motivated since the paired environment supplies true counterfactual futures that the ocean record lacks.

Critical weaknesses

W1. My main concern is the limited evaluation with many dashes in tables 1 and 2. It’s unclear which ones are n/a and which ones don’t have runs computed. And if they aren’t computed, how do they affect the result once they are here?

Questions for authors

Q1. My primary question is the computations that aren’t run yet (the dashes on the tables). How do these runs affect the paper?

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

The paper proposes SP-JEPA, which separates predictive states into fixed-basis coefficients and a learned residual. An orthogonally constrained decoder permits exact edits within the basis span while holding the residual fixed. The method combines ocean forecasting and historical event observations with a controlled simulator evaluation of perturbation responses. It also introduces a diagnostic separating response capture from propagation. Validation results demonstrate competitive field prediction, while identifying substantial limitations in mechanism-dependent response prediction.

Critical strengths
  1. An explicit, independently editable state interface. The decoder guarantees that coefficient edits produce the corresponding basis-space change without interference from the residual. This is a concrete capability beyond merely interpreting latent coordinates after training.
  2. Useful diagnostic evaluation. The controlled simulator, capture–propagation decomposition, and action-mean baseline expose what the model learns and where it fails. This analysis provides scientific value beyond reporting aggregate forecasting accuracy.
Critical weaknesses
  1. Independent generalization remains unverified. The reported results use validation splits that also select checkpoints, and the ocean event gate is validation-selected. Results on the defined test partitions are necessary to establish that the reported improvements survive model selection.
  2. The response benchmark does not establish an advantage over a simple predictor. SP-JEPA achieves response skill of 0.575, below the stateless action-mean table’s 0.64, and does not recover the mechanism-dependent component. Its response-specific training objective also differs from the compared models’ objectives, limiting attribution of their performance gap to the state architecture.
Questions for authors
  1. Evidence needed for the current claims: What performance is obtained on the reserved ocean and simulator test partitions after freezing checkpoints, gates, and all selection choices?
  2. Clarification and supporting evidence: How much of the response advantage over other learned models comes from the state architecture versus response-specific supervision? Please provide a comparison with matched response supervision and training budgets.
  3. Evidence needed for the current claims: Can the model outperform the action-mean baseline on the mechanism-dependent response component? If not, how should the demonstrated intervention capability be scoped?

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

Related papers

Discussion (0)

Sign in to comment.