acceptodds
Withdrawn from ICLR 2027

Bridging the Knowledge–Reasoning Gap in Natural Systems: A Benchmark for Inductive Rule and Noise Recovery in Cellular Automata

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Noise is an inherent part of any natural system. After considering the recent success of transformer-based models in natural system reasoning, this research explores the capability of large language models (LLMs) in addressing the question: given an example of a real-world natural system, can LLMs identify the presence of noise within it? Here, we consider the natural computing model cellular automata (CA) as a proxy of natural systems. Therefore, we formalize this question as the CA identification problem: given noisy evolution dynamics, can LLM recover both the underlying rule and the magnitude of noise? To explore this question, we construct four benchmark datasets (skewECA, αECA, stocECA, and tempECA) totalling 2,876,800 samples across four noise types, along with knowledge-check dataset. This question also captures the reasoning capabilities of LLMs, i.e. reasoning from observation. Evaluating Qwen, Llama, Mistral/Mixtral families (7-72B parameters), alongside GPT 5.1 and Claude Haiku 4.5 under zero-shot, few-shot, and fine-tuned settings (for open source models), we observe near-perfect knowledge accuracy but poor task performance. The rule prediction peaks at 14.60% and noise prediction at 32.50%. This gap is architectural. Next-token prediction does not equip models to extract the global and structural statistics required for this task, including change rates, spatial contiguity, and distributional variance. We therefore design four specialised transformer models (skewM, αM, stocM, and tempM) each aligned with the statistical signature of its corresponding noise process. These models achieve up to % rule exact match and % noise tolerance accuracy, outperforming the best LLM baseline by up to with at most M parameters. These findings emphasize the importance of structure-aware architectures for reasoning over spatiotemporal data, as exemplified by the identification problem. Code and Dataset are available at https://anonymous.4open.science/r/LLM_NOISE-9DC1/.

Primary area:
datasets and benchmarks
Keywords:
Large Language Models, Benchmarking, Identification Problem, Elementary Cellular Automata, Noise

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_wzc1OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

It introduces four benchmarks for recovering cellular automaton rules and noise parameters, evaluates LLMs, and proposes specialised transformers. The benchmark is potentially useful, but the evaluation protocol and overstated architectural conclusions prevent me from recommending acceptance.

Critical strengths

The controlled setting provides exact ground truth and supports analysis of identifiability. The specialised models perform strongly, and the algorithmic baselines help establish task difficulty.

Critical weaknesses
  1. It's such a disaster to read the whole paper, given the 117 pages. I highly suspect that most of them are generated by AI because it's not really readable for humans because of its verbose, repetitive, and defensive assertions. Especially for Figure 6, I retain high suspicion that it was completely generated by AI, which is not disclosed by the authors.

  2. The claimed relevance to natural systems remains speculative. Experiments assume known noise families, fully observed binary states, and synthetic one-dimensional dynamics. No real-system validation supports the broader framing.

  3. The central architectural conclusion exceeds the evidence. Limited LoRA adaptation is compared with extensive task-specific training, structured tokenisation, and mechanism-aware estimators. This does not isolate architecture as the cause of the performance gap.

  4. The test set is used for model selection. The paper explicitly selects checkpoints using repeated evaluation on the final test set. Disjoint training and test rules do not eliminate selection bias. The claim that online data generation makes an independent validation set impossible is untenable.

Questions for authors
  1. What justifies calling these results a held-out evaluation when the test set repeatedly determines checkpoint selection? Please report results on an untouched test set.

  2. What experiment isolates architecture from training budget, input representation, and analytical assistance? Without such a control, why claim a fundamental architectural limitation?

  3. What evidence supports transfer beyond this narrowly specified synthetic setting?

Reviewer_cHMFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces four benchmarks for recovering cellular automaton rules and noise parameters from observed evolution sequences. It compares general LLMs with four specialised transformer models that use local transition representations and analytical noise estimators. The specialised systems achieve substantially higher accuracy. The authors argue that LLMs know the relevant concepts but lack the architecture needed to apply them.

Critical strengths
  1. An interesting problem and a useful benchmark. Recovering cellular automaton rules and noise parameters from observed dynamics is an interesting task for studying how models infer underlying mechanisms from data.

  2. Useful diagnostic experiments. The paper includes fine-tuning, tool use, complete-input and compact-summary conditions, and baselines trained from scratch.

Critical weaknesses
  1. I am not convinced that architecture is the main reason for the baseline LLMs lower performance. The specialised models also use structured inputs, more task-specific training, and analytical noise estimators. The complete-input experiment does not fully resolve this: the best LLM results are 3.0% on stocECA and 6.5% on tempECA, but training uses only 500 examples instead of 5,000. Input completeness and training budget therefore change at the same time. Additionally, the abstract’s claim that LLM rule accuracy peaks at 14.6% also needs clarification. GPT-5.1 reaches 34% with reasoning and Python tools, and 60% with a compact transition summary. These experiments use different input conditions and only 50 test examples, but the stronger results should still be acknowledged.

  2. Test-set performance is used for checkpoint selection. The tempM and stocM results in Table 1 use checkpoints selected based on test-set performance, which may lead to overly optimistic accuracy estimates. Checkpoint selection should use a separate validation set, followed by evaluation on an untouched test set.

Questions for authors

Q1. How do the existing controls separate the effect of architecture from structured inputs and task-specific training?

Q2. Can you report the tempM and stocM results in Table 1 using checkpoints selected on a separate validation set and evaluated on an untouched test set?

Q3. Was generative AI used to create or modify Figure 6? If so, please describe this use in the AI Use Statement.

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

The paper proposes classification of noisy elementary cellular automata (ECA) as a stylized setup for noisy natural systems. They show that despite containing the knowledge of ECA rules and noise, LLMs perform poorly on the classification tasks. Vanilla models such as Transformers, CNNs, MLPs also underperform on these tasks. They design highly custom architectures with hand-tuned training methods which can perform well.

Critical strengths
  • The paper is well-written and easy to follow.

  • The paper is empirically sound. The authors have made efforts to make the experiments fair and robust, and ran useful evaluations and baseline comparisons.

Critical weaknesses

My main concern relates to the scope and generality of the papers findings.

  • Regarding the LLM evaluations, it is no surprising that understanding facts of ECA and noise does not translate to making predictions from bit strings. Comparing their performance with highly custom architectures and training objectives is very unfair. The link to reasoning is weak, and not made precise in the paper.

  • Their custom architecture and training is too hand-crafted to draw general conclusions from. To an extent, it is expected that if one divides a complex tasks into pieces and designs architectural components and training objectives for each, this would perform better than end-to-end models. With the right inductive bias (i.e. enough hand-holding) most task become easier, which is not a new insight. I fail to see what genuinely new insight their setup provides. (I am happy to hear counter-arguments to this from he authors.)

  • Claim that training vanilla architectures does not perform well on these tasks requires a scaling study with data. Equal-data comparison between vanilla and custom setup is not a like-to-like comparison because the design and training signal of the latter is much richer.

Questions for authors

Questions

  • The zero-training results suggest that the most important part of the custom methods is curating the architecture to the statistics of the data, rather than learning from data (albeit some noise cases fail in zero-training). Can the authors comment on why this is an interesting learning problem, especially given the emphasis on feeding the right biases and objectives rather than learning them from data?
  • The custom architectures are built assuming that the noise is perfectly understood. However, in natural systems, a precise statistical description of the noise may not be available. Can the authors address this shortcoming of the setup?

Suggestions

  • Figures 2 and 6 are too small to be comprehensible.
  • Are panels (c) and (d) in Fig.3 swapped?

Comment

My assessment lies somewhere between reject and accept, with inclination towards reject. However, if the authors satisfactorily address at least part of my concerns, I am willing to change my evaluation.

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

Related papers

Discussion (0)

Sign in to comment.