REVIEW 4 major objections 2 minor 1 cited by
Reservoir Computing with Evolved Critical Neural Cellular Automata
T0 review · 4 major / 2 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read This paper claims that a neural cellular automaton evolved to exhibit power-law avalanches acts as a reservoir computer, scoring perfectly on a 5-bit memory task and matching or beating the best elementary cellular automata on…
desk verdict Evolved critical NCA as a reservoir is a fresh combination; the benchmarks look solid, but the criticality claim needs a control condition. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the neural cellular automaton (NCA): a lattice of cells whose next state is computed by a shared, small artificial neural network from the local neighborhood, so the same network is the update rule everywhere. Each cell also carries an integer counter that lets avalanches be defined as cascades of state changes; the evolution strategy maximizes a fitness signal derived from the size distribution of these avalanches being close to a power law. That power-law avalanche signature is the machinery's load-bearing feature: it is both the optimization target and, after evolution, the evidence that the substrate sits at criticality. The reservoir readout is a simple learned map from the final CA state to the task output, so the evolved dynamics do the computation and the readout only extracts it.
What would settle it
Fit competing models (power law versus log-normal or other finite-size distributions) to the avalanche sizes produced by the evolved NCA with a standard goodness-of-fit test; if the power law is not the best fit, the paper's evidence for criticality collapses, since the power-law statistic is both the fitness target and the claimed signature of criticality.
Extended reading notes
Core claim
The central claim is that criticality can be installed by evolution, not hand-designed; an evolution strategy tunes the weights of the neural network governing each cell of a cellular automaton until the activity in the lattice forms power-law distributed avalanches. The resulting critical neural cellular automaton is then used as a reservoir: it is driven by input, left to evolve, and read out by a trained layer. On the 5-bit memory task the readout achieves a perfect score, retaining all five bits, and on handwritten-digit image classification it matches or, in some cases, surpasses the performance of the best elementary cellular automaton used for the same task. The paper also claims the critical NCA is robust to extreme initial conditions, which it reads as evidence that the system may be operating as a self-organized critical system rather than a critically tuned one.
Load-bearing premise
The paper takes a power-law distribution of avalanche sizes as the definition of criticality, and that same power-law statistic is also what the evolutionary search is rewarded for; if that statistic is not a meaningful measure of criticality, the claim that the reservoir works because it is critical is not established.
Editorial extensions
If this is right
- An evolved critical NCA can hold all five bits of a test input in its dynamical state, so the memory task is passed at the maximum possible score.
- The same evolved substrate, with only a readout trained, classifies handwritten digits at least as well as the best elementary cellular automaton used for the task.
- Because the criticality is evolved rather than hand-designed, the method offers a general route to building reservoir computers from arbitrary local-update rules.
- If the substrate is genuinely self-organized critical, the reservoir will keep its computing behavior when initialized far from the usual starting conditions, without re-tuning parameters.
Reading between the lines
- Because the avalanche statistic appears both in the fitness function and in the evidence for criticality, the cleanest test of the interpretation is to evolve the same substrate with a fitness that does not reward power-law avalanches; if the benchmarks are unchanged, the computational results do not depend on criticality.
- The same evolution-to-criticality recipe could be applied to other local-update substrates, such as graph automata or continuous-state reaction-diffusion lattices, by defining an avalanche statistic for those systems; the paper does not test this, but nothing in its setup limits the idea to cellular automata.
- Since only the readout is task-specific, a single evolved critical NCA could plausibly be reused for several memory and classification tasks, though the paper reports each benchmark separately rather than a multi-task demonstration.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes to evolve a neural cellular automaton (NCA) toward criticality using evolution strategy, with the fitness function based on power-law avalanches, and then evaluate the evolved substrate on two reservoir-computing benchmarks: a 5-bit memory task and handwritten-digit image classification. The reported results are a perfect score on the 5-bit task and performance matching or exceeding the best elementary CA on the image task, with the additional claim that the critical NCA may operate as a self-organized critical system.
Significance. If the causal link between criticality and reservoir performance were established, this would be a valuable contribution, connecting self-organized criticality to practical computation. The use of evolution to search for critical substrates is a promising direction. However, the evidence presented in the abstract does not establish that link, because the criticality measure is both the optimization objective and the diagnostic, and no non-critical control is reported. The benchmark results themselves are positive existence claims but lack statistical detail.
major comments (4)
- [Abstract] The abstract states that the NCA is evolved "to achieve criticality, demonstrated by power law distributions in structures called avalanches," but if the same avalanche statistic is used as the fitness function, the power-law outcome is partly imposed by the search and does not independently confirm criticality. Please provide a criticality measure that is not the optimization target, such as branching ratio or susceptibility, or show that the avalanche distribution is a robust emergent property across different fitness definitions.
- [Abstract] The abstract reports no control condition. To attribute the 5-bit memory and image classification performance to criticality, the authors should compare against a non-critical NCA with the same architecture and readout but different dynamics, such as one evolved under a different fitness function or with different parameters. Without such a control, the benchmark scores are existence claims only and do not support the causal interpretation.
- [Abstract] Power-law fitting is known to be sensitive to fitting range, binning, and system size. The abstract provides no details of the fitting procedure, such as maximum-likelihood estimation, goodness-of-fit tests against alternative distributions, or error bars on the exponent. The claim of criticality is under-specified; please report the fitting methodology and finite-size analysis to rule out log-log artifacts.
- [Abstract] The benchmark results are presented without statistical context: no number of independent runs, variance, significance tests, or dataset partitioning details for the image classification task. The phrase "matched and sometimes surpassed" is not sufficient to assess reliability. Please provide standard deviations or confidence intervals and describe the readout training and evaluation protocol.
minor comments (2)
- [Abstract] The phrase "the system managed to remember all 5 bits" is informal; consider rephrasing to "the system correctly recalled all 5 bits" or similar.
- [Abstract] The sentence "Criticality is a behavioral state in dynamical systems that is known to present the highest computation capabilities" would benefit from a supporting citation or a more precise description of the computation-capability claim.
Circularity Check
Avalanche power-law statistic appears to be both the optimization target and the evidence for criticality; benchmark scores are independent.
-
self definitional
[Abstract (full text not available; abstract-only review)]
"We apply evolution strategy to optimize the NCA to achieve criticality, demonstrated by power law distributions in structures called avalanches."
The abstract states that the evolutionary strategy optimizes the NCA 'to achieve criticality' and that criticality is 'demonstrated by power law distributions in structures called avalanches.' If the fitness function rewards the same avalanche power-law statistic that is later reported as evidence of criticality, then observing power-law avalanches after evolution is forced by the selection criterion rather than being an independent discovery. The power-law finding is therefore partly tautological, because the optimization target and the verification metric coincide. However, this circularity affects only the criticality evidence, not the reservoir-computing benchmark scores, which are external tasks and remain independent of how criticality was measured.
full rationale
This is an abstract-only review, so the full derivation chain cannot be inspected. The reservoir-computing benchmarks (5-bit memory and image classification) are external and self-contained: a perfect 5-bit memory score and classification performance matching or surpassing the best elementary CA do not depend on the definition of criticality. Those claims are not circular. The only identifiable circular step is the criticality demonstration: the abstract says the NCA is evolved to achieve criticality and that criticality is demonstrated by power-law avalanche distributions. If the fitness function uses the same avalanche statistic, then the post-evolution power-law distribution is an expected outcome of optimization rather than an emergent property discovered after the fact. This is a mild case of self-definitional evidence. The lack of a non-critical control condition is a correctness concern but not a circularity concern. Because the central benchmark claims retain independent content and the circular step is limited to the internal criticality claim, a moderate score of 4 is appropriate.
Assumptions & free parameters
free parameters (2)
- Evolution strategy hyperparameters
- Avalanche power-law fit parameters
assumptions (3)
- domain assumption Power-law distributed avalanches imply criticality.
- domain assumption Critical dynamics are optimal for reservoir computing performance.
- domain assumption The benchmark protocols and comparison to the best elementary CA are fair and statistically adequate.
Cite this review
Pith. "Pith review of Reservoir Computing with Evolved Critical Neural Cellular Automata." pith.science (2026). https://pith.science/paper/TBDEUPAN
@misc{pith2026250802218,
author = {Pith},
title = {Pith review of: Reservoir Computing with Evolved Critical Neural Cellular Automata},
year = {2026},
howpublished = {\url{https://pith.science/paper/TBDEUPAN}},
note = {Machine review of arXiv:2508.02218}
}
read the original abstract
Criticality is a behavioral state in dynamical systems that is known to present the highest computation capabilities, i.e., information transmission, storage, and modification. Therefore, such systems are ideal candidates as a substrate for reservoir computing, a subfield in artificial intelligence. Our choice of a substrate is a cellular automaton (CA) governed by an artificial neural network, also known as neural cellular automaton (NCA). We apply evolution strategy to optimize the NCA to achieve criticality, demonstrated by power law distributions in structures called avalanches. With an evolved critical NCA, the substrate is tested for reservoir computing. Our evaluation of the substrate is performed with two benchmarks, 5-bit memory task and image classification of handwritten digits. The result of the 5-bit memory task achieved a perfect score and the system managed to remember all 5 bits. The result for the image classification task matched and sometimes surpassed the performance of the best elementary CA for this task. Moreover, the proposed critical NCA may operate as a self-organized critical system, due to its robustness to extreme initial conditions.
Forward citations
Cited by 1 Pith paper
-
CO-RFT: Efficient Fine-Tuning of Vision-Language-Action Models through Chunked Offline Reinforcement Learning
CO-RFT combines imitation learning with chunked offline reinforcement learning to fine-tune vision-language-action models from 30 to 60 demonstrations, reportedly improving success rate by 57% and cutting cycle time by 22.3%.
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.