acceptodds
Under review as a conference paper at ICLR 2027

Why Do Conventional World Models Fail to Learn Cellular Automata?

Abstract

Although conventional world models — auto-regressive or diffusion models based on transformers or convolutional networks — may learn surface statistics of world dynamics, can they learn the exact world dynamics from its observed history? Leveraging cellular automata as a simple testbed, we find the answer to be no in many cases. Conventional architectures predict most pixels correctly yet rarely complete a rollout: a CNN predicts 96.3% of cells but completes 18.9% of rollouts; a joint diffusion model completes none. We trace the gap to three failure modes of these world models — namely, they fail to exactly capture spatial locality, temporal locality or temporal stability. Simple changes repair each: (1) for spatial locality, two-dimensional rotary positions lift a transformer from 39.1% to 100% on the Game of Life; (2) for temporal locality, handing each token its cell's previous-frame neighbourhood lifts the same transformer from 25.8% to 99.9% on unseen rules; (3) for temporal stability, causal freezing lifts the same diffusion weights from 42.2% to 99.9%. None of the three changes touches the architectural backbone; each only modifies the information flow within it. The same diagnosis applies to billiards and to a simulated Burgers equation.

Primary area:
interpretability and explainable AI
Keywords:
cellular automata, world models, inductive bias, interpretability, architecture design

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_qFvvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper asks whether conventional world models, i.e., autoregressive and diffusion models built on transformers or convolutional networks, can learn the exact rules of a world from its observed history. The testbed is binary outer-totalistic cellular automata on an 8×88\times 8 torus. A model observes 8 frames and must predict the next 8 exactly. Performance is measured by strict sequence accuracy: the fraction of rollouts in which all 512 predicted cells are correct. The authors attribute the failures of conventional models to three links. The first is spatial locality: identifying a cell's neighbourhood. The second is temporal locality: binding each observed situation to its outcome in the next frame. The third is temporal stability: settling each predicted frame before later frames are built on it. For each link, the authors propose a change that leaves the backbone untouched: two-dimensional rotary positions for spatial locality; a KV shift or a convolution over two adjacent frames for temporal locality; and causal freezing at sampling time for temporal stability. The paper also includes exploratory experiments on billiards and on several continuous dynamical systems.

Critical strengths

The authors use cellular automata as a testbed, a system whose rules are simple and fully known yet whose dynamics are rich. On this testbed they systematically test the mainstream world-model paradigms, namely autoregressive and diffusion models built on transformers or convolutional networks. Instead of relying on per-pixel accuracy, they require every cell of every predicted frame to be correct, which reveals where and how these models fail. The authors go beyond reporting failures and identify a common cause behind them: the models lack inductive biases matched to the task. They organize these inductive biases into three properties (spatial locality, temporal locality, and temporal stability) and give a corresponding fix for each. The controlled comparisons are also carefully designed; most of them change a single factor while holding everything else fixed.

Critical weaknesses

1. The writing and organization seriously hinder understanding. This is the main reason for my negative assessment.

(a) Most of the main text consists of self-contained paragraphs, each opening with a bold conclusion sentence, so the text effectively reads as a bulleted list. All four paragraphs of Section 3, most paragraphs of Sections 4.1–4.3, and all six paragraphs after the first in Section 5 open this way, and the introduction presents its contributions as three "Failure mode / Solution" boxes. This is more than a matter of style. Each paragraph states a conclusion and then its evidence, but the paper rarely says how one paragraph relates to the previous one: whether it adds to, qualifies, contradicts, or builds on it is left to the reader. As a result, the relations between the three properties, and between the experiments under each property, remain loose, and the paper reads as a list of findings rather than a coherent argument.

(b) Many concepts are used before they are defined, or are defined only in the appendix or in figure captions. For example, "situation" is already used in the introduction to state the temporal-locality failure but is only defined in Section 2. The levels L1–L4 appear in Figures 2 and 3 before they are defined in Section 2 and Table 1. The "structured diffusion model" is used to report results in Sections 4.2 and 4.3, but it is only defined at the end of Section 5. In addition, several terms are never explained in the main text, and the reader must look them up in the appendix or in figure captions: the "10% dose" in Section 4.3, the "canvas encoder" and "anchor frame" in Section 5, and width labels such as "d64" and "d128", which are used from Section 3 onward.

In my view, the paper needs substantial restructuring, not just local revision.

2. The paper adds limited new knowledge. The paper repeatedly emphasizes that adding the right inductive biases repairs the failures. However, the principle that architecture should align with the target computation is well known, and the authors themselves cite Battaglia et al. (2018) and Xu et al. (2020) on this point. The authors acknowledge that none of the devices they use (KV shift, causal freezing, 2D RoPE) is new. They position their contribution as providing a world in which the reason each device is needed can be stated in advance and tested. Yet the conclusions of these tests largely stay at the level of "whichever information pathway is missing, supply it through structure." After reading the paper, it is unclear to me what has been learned that was not known before and that transfers to other tasks.

The paper also does not place its experiments within the in-context learning framework. The contrast between L2 and L3 is essentially a distinction between in-weights and in-context learning. The task itself is associative recall, with the situation as the key and the next-frame state as the value. The authors cite in-context learning work such as Olsson et al. (2022) and Chan et al. (2022), but Chan et al. appears only in the L4 discussion of data dosage. Making this framing explicit would help readers see what the results add to existing understanding.

3. The paper offers no general methodology applicable to real-world world models. All three fixes rely on knowing the dependency structure of the rule in advance: the next state depends only on the 3×3 neighbourhood in the previous frame, and the rule is first-order in time. The checklist in Section 6 asks the reader to "write down which cells and frames must meet, and which parts of the future depend on which." For real-world world models, such as video generation or physical scene prediction, these dependencies are usually exactly what is unknown. Here the paper offers only a broad recommendation to attend to information and dependency structure. It does not give a concrete, actionable method for determining which information pathway needs to be supplied. The discussion of video models in Appendix H is likewise hypothetical. I understand that automatically discovering such pathways is beyond the scope of this paper. For exactly that reason, however, it is questionable how far its conclusions can guide the design of real-world world models.

Questions for authors
  1. The contrast between billiards and the Game of Life seems related to Wolfram's notion of computational irreducibility: the future positions of billiard balls have a closed-form solution, whereas the Game of Life must be evolved step by step. The authors give two explanations for this difference. First, the Game of Life cannot skip rule steps. Second, a later state can have several predecessors, so it cannot be used to verify an earlier guess. The first is close to irreducibility; the second is irreversibility. Could the authors compare joint sampling and causal freezing on rules with different degrees of reducibility and reversibility, to separate the contributions of these two factors?
  2. The held-out rules vary widely in difficulty. Under some rules, all cells quickly die or stay static; others are highly sensitive to perturbations. The paper reports only averages over all rules. Could the authors break down the results by rule type, for example by Wolfram class, Langton's λ parameter, or damage-spreading rate, to show whether the three properties play the same role across different kinds of rules?
Reviewer_aE22OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies how to learn the dynamics of cellular automata (CAs), focusing on the classical case of binary 2D grids with outer-totalistic (neighbor-counting) rules. The main insight is that long-horizon predictions work if spatio-temporal locality and symmetry (rule invariance) is baked into the architecture. In this case, results are perfect or near-perfect; when the spatial structure is ignored (like stacking time steps up as features in a CNN), performance degrades severely. A second source of degradation is running diffusion models in a temporal chain without ensuring convergence; the residual noise leads to accumulating drifts. Similarly, all architectures cannot generalize rules if no examples of local neighborhoods are present in the training data; they cannot guess the outer-totalistic nature of the rule (which, in principle, would be able to achieve generalization, if used as a prior).

Overall, the findings are straightforward and expected: By prescribing the problem structure to a point where only a small local rule has to be learned, performance reaches 100%; if the resulting patterns are put into higher-dimensional tokens or input vectors unrelated to the computational structure, accuracy drops dramatically. The findings on noise and missing training data seem similarly obvious.

Critical strengths

Understanding the ability of DNNs to understand and model evolving / emerging structures in dynamical systems is an important problem, and I do believe that the approach of creating toy model worlds by CAs is a promising avenue to improving out understanding. There is some, but not a lot, of related work in this area and I would find the general research direction interesting, providing still some potential for deeper insights.

Critical weaknesses

Unfortunately, in my understanding of the paper, it reaches no deeper, non-obvious insights. An outer-totalistic rule of a 3x3 neighborhood has a very small coding length; one could learn the rules with a hash-table if enough prior knowledge (symmetry and locality of the machine) was imposed. The CNNs that actually work exploit all invariances and locality constraints of the automata and only drop the "sum of active neighbors" restriction. It is not surprising that the CNN can learn the rule (there are 9 inputs, counting center + neighbors, i.e., 2^9 = 512 different input states that need to be assigned a zero or one for a concrete rule when ignoring the outer-totalistic construction; this means, 512 different examples fix the rule completely, and a small CNN can easily remembers these rules. Similarly, a transformer equipped with translational covariance and some locality prior should be able to represent the rules. On the other hand, dropping the symmetry prior makes the problem, obviously, very hard, as CAs can produce complex patterns that cover large areas in the space of boolean vectors when viewed as long random bit vectors). Similarly, the need for diffusion models to converge before taking the next step is not surprising, and the lack of generalization if examples of rules are missing (and critical information of the very specific and peculiar outer-totalistic bias is not available) is also not surprising.

In essence, the models do not learn which emergent patterns the dynamical system forms but just identify the rules from direct examples of their receptive fields and outputs. This problem is obviously easy and it is not surprising that this works. The counter-examples, where performance is low, are all obviously combinatorally challenging, working in a setting where one would not expect the learning algorithms to be able to find suitable structure. The challenge in this problem area is to find a suitable experimental setup that leaves aspects to be learned while remaining learnable; I found the proposed experimental design not convincing in this regard.

Overall, I think that, while the general direction is promising, the experimental scenario is too simplistic to gain significant non-obvious insights from the paper. More work in more complex scenarios, where actual computations of the CA-"machine" are predicted rather than just remembering the rules would be very interesting in my opinion (and there has been some related work already, with some surprising/non-obvious results). In this sense I would see the work in an too early stage for publication.

Questions for authors

I have no questions. If I am missing something important, or misunderstanding something, I am, of course, happy to be corrected.

Reviewer_27nXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Cellular automata is a computational model that evolves in discrete time based on a rule. This paper tries to answer the following question: why do the conventional world models fail to learn cellular automata? Two things must be learned: an initial configuration and the rule that cellular automata evolve under. This paper tries to answer this question by answering the following research question: Why can’t standard world models, autoregressive or diffusion models based on transformers or convolutional networks, infer an unseen cellular automaton’s rule from observed history?

Critical strengths
  1. Considering temporal and spatial locality, and also temporal stability.
  2. Studying these three features through experiments.
Critical weaknesses

It is not clear how the initial configuration or rule affects the patterns and the output of the model. It is not clear how the evaluation changes as the size of the cellular automata changes. The methodology used in this paper is not well-defined for cellular-automata community and I think the results obtained in this paper are not useful for this community.

Questions for authors
  1. What do you mean by “opposite centre cells”?
  2. Wolfram categorised cellular automata rules into distinct classes, each exhibiting characteristic behaviours that are now widely accepted by researchers. However, it remains unclear which of these behavioural categories the models’ findings correspond to. The authors should provide a detailed analysis framed around these categories, because each group represents a distinct dynamical regime; evaluating the results through this lens could yield significantly more insightful conclusions. What are the rules that a model cannot learn, or at least cannot learn well?
  3. The state space of samples (for example, two consecutive configurations of the cellular automata) is very large. The evaluation is not clear.
  4. How did you evaluate models? There is a large space; how did you generate samples?
  5. Can you find the initial configuration?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.