acceptodds
Withdrawn from ICLR 2027

Process Memory Beyond Prediction: Auditing Algebraic Structure in Learned Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Does accurate prediction imply recovery of the process that generates the observations? We study this question in a controlled physical-representation benchmark with eight latent states and six primitive operations whose closure forms a 104-element transformation semigroup. The states are rendered as high-dimensional spatial fields under controlled carrier variation, letting us measure physical-state recoverability, predictive generalization, and process-structure recovery separately. The physical states are highly recoverable from observation, with a frozen diagnostic decoder reaching 98.31% mean validation accuracy, and learned models consistently outperform persistence across sealed distribution shifts. Yet the predictively trained Operation Channel Model (OCM), despite an explicit eight-state categorical representation and operation-specific transition channels, reaches only 15.72% state alignment and 13.89% sealed state-path accuracy against a 12.5% chance reference. It exactly recovers only 1 of 104 transformations and none of the nine held-out transformations, and it does not outperform the validation-selected B4 baseline in the frozen primary joint-OOD comparison. An earlier simpler observation regime provides a positive control in which OCM recovers the full finite transformation structure at low-to-moderate noise, showing that the broader process-memory analysis can register exact structural recovery under less demanding conditions. In Tier-C-v4, state recoverability and predictive learnability therefore do not, by themselves, guarantee recovery of the intended process algebra in the structured model tested here. When mechanistic or compositional structure is a modeling objective, process memory should be evaluated directly rather than inferred from forecasting accuracy.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, world models, process memory, compositional dynamics, dynamical systems, structured representations, algebraic structure, benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_G17VOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper examines whether accurate prediction leads to learning the underlying process structure. Using a synthetic benchmark, the authors find that OCM performs poorly in recovering latent states and transformations. However, the experiments do not clearly explain the reasons behind this failure.

Critical strengths
  1. The paper addresses an interesting question about the relationship between prediction and structural learning.

  2. The benchmark provides ground truth for evaluating structural recovery beyond prediction accuracy.

Critical weaknesses
  1. The novelty is limited. The main finding is closely related to existing identifiability problems, and the paper does not clearly establish what new insight it offers.

  2. OCM itself has relatively weak predictive performance, making it difficult to tell whether the failure comes from predictive learning or from the model and its optimization.

  3. Important baselines and ablations are missing. Structural consistency constraints, partial supervision, and structural probes on other predictive models would help clarify the findings.

  4. The experimental analysis is insufficient. The large gap between state recoverability and OCM alignment is not explained, and the reliance on hard transition maps may overlook useful structural information. The limited synthetic setting and the unexplained difference from the positive control also raise concerns about generalizability.

Questions for authors
  1. How can the authors distinguish a limitation of predictive learning from the weak predictive performance or optimization issues of OCM?

  2. What causes the large gap between 98.31% state recoverability and 15.72% OCM alignment?

  3. Would adding structural consistency constraints or partial state supervision improve structural recovery?

  4. Could soft transition metrics or structural probes on other predictive models lead to different conclusions?

Reviewer_ou5tOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduced a controlled benchmark Tier-C-v4 for the purpose of testing prediction model's ability to recover underlying dynamics. A frozen diagnostic decoder recovers states at 98.31%, and all learned models outperform persistence. However, the open channel OCM reaches only low state and state-path accuracy, and little process algebra, and does not outperform the best baseline model. The authors conclude that predictive success and state recoverability do not guarantee recovery of the intended finite process representation.

Critical strengths

None

Critical weaknesses

The authors brought up a meaningful topic, however, the actual investigation was far from "physical dynamics", instead, it is based on some discrete form automaton, sharing nothing with the governing principles or mathematical formulation of physical dynamic systems. It is based on language theory instead of physical mechanics.

Questions for authors

Investigate a real physical dynamic system

Reviewer_4HhgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work is an empirical study about the relationship between predictive accuracy and recovery of the underlying dynamical structure, eg whether accurate prediction of a dynamical system implies recovery of its underlying process structure. The authors introduce a benchmark with eight latent states and 104 transformations to evaluate predictive performance and structural recovery separately. Their experiments show that models can predict system behavior while failing to recover the underlying transformation algebra.

Critical strengths
  • The finite transformation semigroup provides a reference for evaluating structural recovery, it is a well defined ground truth.

  • The analysis is thorough it includes state alignment, state trajectories, exact transformations, information partitions, and information-order relations.

  • The sealed evaluation, multiple seeds, noise conditions, and distribution shifts improve the reliability of the reported measurements.

Critical weaknesses
  • The distinction between predictive equivalence and recovery of the generating mechanism is well established. The paper provides an empirical illustration of this distinction within a controlled finite-state benchmark. The main methodological contribution is the proposed structural evaluation framework. Its broader applicability remains to be demonstrated.

  • The experiments show weak algebraic recovery in OCM, but they do not identify the factors responsible for this result. The absence of targeted ablations makes it difficult to distinguish limitations of predictive training from limitations of the architecture or optimization procedure.

  • Structural recovery is evaluated only for OCM. The other predictive models are not examined using standardized post-hoc structural probes.

Questions for authors

Q1. What is the main methodological novelty of the proposed process-memory evaluation framework relative to existing approaches for assessing latent-state and transition-mechanism identifiability?

Q2. Can the authors identify the primary cause of OCM’s structural recovery failure through targeted ablations of the architecture, training objective, or observation-generation process?

Q3. How would the conclusions change if the continuous predictive baselines were evaluated using standardized post-hoc state and transition probes?

Q4. What evidence supports the interpretation of OCM’s predictive performance as sufficiently successful to establish a meaningful separation between predictive learning and structural recovery?

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

It is very difficult to understand what the paper actually describes, but abstractly, it deals with the discrepancy between the ground truth world model and the world model learned from observation. Beyond that, it is very difficult to precisely say what the topic of the paper is.

Critical strengths

I fail to identify the strengths in this paper.

Critical weaknesses

This paper is one of several in my reviewing batch which seems to be entirely LLM generated. That is by itself not a ground for rejection; a valuable contribution is a valuable contribution wherever it comes from, but I have to say that I fail to find the value in this work. It is very difficult to follow, not due to technical difficulty, but due to unclear and incomplete presentation. A few examples:

  1. Tier-C-v4 seems to be the name of an intermediate version of a benchmark suite. Why v4?
  2. Not a single one of the seven Tier-C-v4 model families are described in any detail.
  3. "Spatial persistence" is never defined, but is still used throughout the text.
  4. "OCM" is not described in sufficient detail; it is impossible to understand how it works from the given description.
  5. None of the six primitive operations in 3.1 are described in any detail.
  6. The carrier "c" in 3.2 was never defined.
  7. It is stated that "Carrier parameters vary morphology, including characteristic scale, interfaces, orientation statistics, and defect structure.". None of these properties are explained.

I believe that this suffices.

Questions for authors

I have several general questions which can hopefully increase the quality of the submission.

  1. How do you envision the impact of your method? Do you think it might have applicability in a more general setting, where the true underlying generating model is not accessible?
  2. Could you, without use of LLMs, summarise your paper for me? I would like to hear about the motivation behind it, the main findings, and what you believe would be interesting extensions.
  3. Can you provide a list of three main results which you think anyone should take away from your work? The list should be written without excessive use of technical concepts.

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

Related papers

Discussion (0)

Sign in to comment.