Adaptive Exploration in Lenia with Intrinsic Multi-Objective Ranking

Lee Spector; Niko Lorantos

REVIEW 4 major objections 4 minor 13 references

Adaptive Exploration in Lenia with Intrinsic Multi-Objective Ranking

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Ranking Lenia individuals on three intrinsic objectives—homeostasis, distinctiveness, and population sparsity—drives more adaptive exploration than a single homeostasis objective.

desk verdict An honest, incremental Lenia experiment whose small effects rest on pseudoreplicated t-tests and an ambiguous dominance definition. read the letter →

arxiv 2506.02990 v1 pith:IMGFHHBD submitted 2025-06-03 cs.NE

Niko Lorantos , Lee Spector This is my paper · ORCID

classification cs.NE

keywords Leniacontinuouscellularautomataartificiallifeopen-endedevolutionintrinsicfitnessmulti-objectiveoptimizationnoveltysearchquality-diversity

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Lenia is a family of continuous cellular automata used as a testbed for open-ended evolution in artificial life, but evolving populations tend to converge unless the fitness signal keeps pushing exploration. This paper tries to establish that ranking individuals against one another on three intrinsic objectives—homeostatic stability, behavioral distinctiveness, and occupancy of sparse regions of a learned behavioral-descriptor space—sustains more adaptive exploration than rewarding homeostasis alone. In 50 trials per condition, multi-objective ranking increased average mass by 0.73 percent and repertoire variance by 0.91 percent, while decreasing gzip complexity by 1.06 percent, with all differences reported as statistically significant. The authors interpret the simultaneous increase in mass and diversity with lower complexity as evidence of more modular internal structure and less trivial complexification, and as a step toward open-ended evolution.

What carries the argument

The carrying mechanism is a variational autoencoder that embeds Lenia phenotypes into a latent descriptor space, together with a domination-count ranking over three objectives defined in that space. Equations (1)–(3) express homeostasis, distinctiveness, and population sparsity as functions of latent vectors, and Equation (4) ranks each candidate by how few archive members dominate it. The ranking does the exploratory work: because fitness is relative and archive-relative, no fixed optimum exists, so the population is pushed toward new trade-offs and sparse regions of the descriptor space.

What would settle it

Replace the learned latent distances in Equations (1)–(3) with distances in a random projection of the raw grid states and rerun the 50-trial comparison; if the multi-objective advantage over homeostasis disappears, the reported exploration gain depends on the VAE embedding rather than on the ranking itself. A complementary check would measure final repertoire diversity with an independent behavioral descriptor not used by the VAE; if the multi-objective population is not more diverse under that measure, the variance increase is an artifact of the chosen descriptor space.

Watch

Extended reading notes

Core claim

The central claim is that intrinsic multi-objective ranking is a workable mechanism for adaptive exploration in evolving Lenia populations. Each individual is scored by domination count against the archive of previously seen patterns under three objectives: homeostasis (latent variance over time), distinctiveness (latent mean distance from average behavior), and population sparsity (negative radial-basis-function density in the descriptor space). Individuals dominated by fewer archive members are fitter, so selection continually favors novel lifeforms in underexplored parts of the behavioral space without prescribing a final target. The authors report that this mechanism yields final repertoires with greater mass, higher latent variance, and lower gzip complexity than the single-objective homeostasis baseline, and interpret these differences as enhanced exploration with preserved homeostatic regulation.

Load-bearing premise

The whole method assumes the learned latent space is a faithful behavioral-descriptor space, so that Euclidean distances between encoded patterns correspond to real phenotypic differences; if it is not, all three fitness objectives lose their intended meaning.

Editorial extensions

If this is right

Because selection is archive-relative, the evolutionary process does not converge to a fixed optimum; it keeps favoring individuals that explore new trade-offs among stability, novelty, and sparsity.
A practitioner can obtain broader behavioral coverage in Lenia by switching from single-objective homeostasis to the three-objective ranking, without sacrificing life content.
The decrease in gzip complexity, together with higher mass, is presented by the authors as evidence of more modular internal structure rather than trivial complexification.
The method removes the need for human-designed novelty metrics, since the learned descriptor space and the sparsity term supply the exploration signal automatically.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the learned latent space is behaviorally faithful, the same three-objective ranking should transfer to other evolved substrates with learned embeddings, such as neural morphologies or other cellular automata; the paper does not test this transfer.
The reported effect may depend on the RBF kernel width $\sigma$; an untested prediction is that adaptive kernel tuning or a learned density estimator would amplify the exploration advantage.
A more direct test of the modularity interpretation would measure compressibility per unit of mass or structural modularity directly, since the paper only reports aggregate gzip size.
Replacing the dominance-count ranking with a uniform random archive sample in the same VAE space would isolate whether the exploration gain comes from ranking or merely from the larger effective archive; this comparison is absent.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is an honest incremental test of a simple idea—rank Lenia individuals by domination count over three intrinsic objectives instead of single-objective homeostasis. The combination is new to Lenia, and the writing is clear about the limits: they call it a step toward open-endedness, not open-endedness itself. The empirical support, however, is shakier than the p-values suggest, and there's a real notation bug in the fitness definition.

What's good: The objectives are sensibly chosen (homeostasis for stability, distinctiveness and sparsity for exploration), the setup is standard Leniabreeder/AURORA, and the comparison against single-objective homeostasis is the right baseline. They also make the right call flagging that they haven't yet measured open-ended dynamics and point to Bedau and MODES as next steps. The paper is transparent enough that an interested reader could reimplement it.

The soft spots, in order of importance: 1. The statistics don't support the headline. The t-tests pool all individuals in final repertoires (50 trials × 1024 individuals ≈ 51,200 per condition) as if they were independent. Individuals from the same trial share the same VAE, selection history, and evolutionary trajectory; the effective replication is 50 trials, not 51,200 samples. The reported differences (0.73%, 0.91%, 1.06%) are small enough that a proper trial-level mixed model could easily put the confidence intervals across zero. Since the whole quantitative claim rests on these p-values, this needs to be fixed before the result can be trusted. 2. Equation 4 and the surrounding prose disagree on what ≺ means. If y≺x means 'y is dominated by x' as the prose says, then d(x) counts archive members that x dominates, but the text then says 'solutions dominated by fewer archive members are considered fitter'—which would be the opposite count. This makes the implemented ranking ambiguous; the authors need to say which one they actually coded. 3. The variance outcome is partly circular: repertoire variance is measured in the same VAE latent space that the distinctiveness objective rewards. Mass and complexity are independent and both move, but by tiny amounts, so the diversity result is less clean than it looks. 4. Missing hyperparameters (the RBF kernel width σ and the homeostasis window n) make reproduction harder. Minor, but should be reported.

Net: the idea is plausible, the execution is honest, and the flaws are typical of a workshop-level empirical note. It doesn't deserve desk rejection. Send it to review with a request for a trial-level analysis, a clarified dominance definition, and error bars. I would not cite it as evidence until the stats are fixed, but I'd happily read a revision.

Referee Report

4 major / 4 minor

Summary. The paper introduces a multi-objective intrinsic fitness ranking for evolving Lenia patterns in the Leniabreeder platform, combining homeostasis, latent-space distinctiveness, and population sparsity. It compares this mechanism with a single-objective homeostasis baseline over 50 trials and reports small increases in mass, VAE-latent variance, and compressibility, each claimed significant at p<0.001. The authors interpret these results as evidence of adaptive exploration and as a step toward open-ended evolution.

Significance. If the results were robust, the contribution would be a useful, practical demonstration that intrinsically motivated multi-objective ranking can improve exploration in a modern QD-based ALife system, and it would strengthen the case for abandoning hand-designed fitness functions. The motivation is clear and the combination of objectives is sensible. The paper also honestly acknowledges that it has not yet applied established OE metrics. However, the current empirical support is not yet convincing: the statistical significance rests on pseudoreplicated per-individual tests, the repertoire-variance metric is circular with respect to the objectives, and the reported effect sizes are very small. These issues are fixable and the central idea remains promising.

major comments (4)

[Results and Discussion (Table 1)] The claim that all differences in Table 1 are statistically significant at p<0.001 is based on pooled two-sample t-tests across all individuals in the final repertoires. Because the 1024 individuals within each of the 50 trials share the same VAE, selection history, and evolutionary trajectory, they are not independent observations; the effective replication is at most 50 per condition rather than roughly 51,200. A proper analysis should first aggregate each trial to a single value per metric (e.g., the per-repertoire mean or median) and then compare the 50 trial-level values between conditions, or use a mixed-effects model with trial as a random effect. The paper should also report 95% confidence intervals for the differences. Without this, the p<0.001 claim is pseudoreplication and does not support the stated significance.
[Methodology, Equation (4)] The definition of domination count is internally contradictory. Equation (4) sets d(x)=|{y∈A | y≺x}|, and the preceding sentence defines y≺x as 'y is dominated by x'. Thus d(x) counts the number of archive members that x dominates. The following sentence, however, states that 'solutions dominated by fewer archive members are considered fitter,' which would require d(x) to count archive members that dominate x (i.e., |{y∈A | x≺y}|). As written, a solution that dominates many archive members receives a very negative fitness, which is the opposite of the intended ranking. Please correct the notation or the equation and verify which version was actually implemented, because this is the core of the proposed mechanism.
[Results and Discussion, Table 1] The diversity measure called 'repertoire variance' is computed in the same VAE latent space that the distinctiveness objective (Eq. 2) and the sparsity reward (Eq. 3) explicitly optimize. The multi-objective algorithm is under selective pressure to increase spread in that space, so observing a higher latent variance under multi-objective ranking is expected by construction and is not an independent confirmation of greater diversity. The mass and complexity metrics are more independent, but their differences are only +0.73% and -1.06%, respectively; please report confidence intervals and discuss whether these effect sizes are of practical importance.
[Methodology, Equations (1) and (3)] Several parameters essential for reproduction are not specified: the homeostasis evaluation window n in Eq. (1), the RBF kernel width σ in Eq. (3), the VAE architecture and training procedure, and the exact rule for updating the archive. Without these details the experiments cannot be reproduced, and it is unclear whether the reported comparison is robust to reasonable choices of these parameters. Please add a supplementary or experimental setup section listing all hyperparameters.

minor comments (4)

[Results and Discussion] Throughout the text, 'increased compressibility' is used to describe the complexity change, but Table 1 shows complexity decreasing by 1.06%, which means compression size is smaller; please align the wording with the metric.
[Results and Discussion] The sentence 'attribute diversity in the homeostasis approach to the AURORA QD algorithm' appears incomplete; it should say 'we attribute' or similar.
[Figures] The figures, especially Figure 3, are not described in enough detail; the caption should state which phenotype is which and what 'columns 1-4' represent.
[Methodology] The paper would benefit from a table or text listing the VAE latent dimension, the number of timesteps used for encoding, and the update schedule of the archive, since these choices can affect the behavior of the QD algorithm.

Circularity Check

1 steps flagged · score 4.0 of 10

Repertoire variance is measured in the same VAE latent space that the distinctiveness and sparsity objectives optimize, so the headline diversity increase is partly by construction; mass and complexity are independent but shift by only ~1%.

self definitional [Methodology (Eq. 2 and Eq. 3), Results and Discussion (Table 1 and 'Repertoire variance' definition)]
"Distinctiveness encourages novelty relative to preexisting individuals. We calculate latent mean distance, measuring divergence from average behavior f2 =|¯z−E[¯z]| (2) ... Repertoire variance is calculated as the latent variance across solutions for the last timestep, giving a measure of phenotypic diversity."

Equation (2) is the selection objective for distinctiveness: it explicitly rewards an individual's latent encoding being far from the average latent encoding. Equation (3) likewise rewards positions in sparse regions of the same VAE descriptor space. The outcome reported as 'Variance' in Table 1 is the latent variance across solutions in that same VAE latent space. The multi-objective condition is therefore under direct selection to increase the very quantity used to measure diversity; the reported +0.91% variance increase is partly the objective being optimized rather than an independent confirmation of adaptive exploration. The paper's prose says the increase 'confirms greater diversity,' but the diversity metric is defined in the space the algorithm is told to spread out in.

full rationale

The only concrete circularity I can identify is that the repertoire-variance outcome is computed in the same VAE latent space that the distinctiveness and sparsity objectives explicitly optimize. Because f2 rewards distance from the average latent encoding and f3 rewards sparse regions of descriptor space, selecting for these objectives will tend to increase latent variance; reporting that latent variance increased is partly a restatement of the selection pressure rather than an independent discovery. The algorithm's multi-objective tradeoff with homeostasis and the external mass and gzip-complexity measures prevent this from being a complete self-reduction, but those external measures shift by less than about one percent and are not the main evidence for 'greater adaptive exploration.' I do not score the pseudoreplicated pooled t-tests as circularity, since that is a statistical robustness issue rather than a definitional reduction. The domination-count notation ambiguity in Equation (4) and the surrounding prose is also a consistency problem, not a circularity step. No load-bearing self-citation chain appears: the cited open-endedness definitions, including Banzhaf et al. 2016, are background framing and do not force the experimental results. Overall, the central diversity claim is partially circular, while the other metrics retain some independent content, giving a score of 4.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its main load-bearing assumptions are that the VAE latent space is behaviorally faithful, that Lenia is a good substrate for open-endedness research, and that domination count over three intrinsic objectives produces useful exploration. The two free parameters, the RBF kernel width and the homeostasis window, are not reported and materially affect the objective values.

free parameters (2)

RBF kernel width sigma
Controls the sparsity objective in Equation 3; the value is not reported, so it is effectively chosen by hand or inherited from an unstated default, and the resulting fitness landscape depends on it.
Homeostasis evaluation window n
Number of timesteps over which latent variance is averaged in Equation 1; not reported, and the homeostasis score depends on this choice.

assumptions (3)

domain assumption The VAE latent space provides a meaningful behavioral descriptor space for Lenia patterns.
All three objectives (Equations 1, 2, and 3) are computed from latent encodings, so the method assumes Euclidean distances in this space reflect behavioral differences.
domain assumption Lenia is a suitable model for investigating open-ended evolution.
The introduction calls Lenia the ideal testbed without comparing it to other artificial life substrates; the generality of the findings depends on this choice.
ad hoc to paper Domination count ranking with intrinsic objectives promotes exploration.
This is the proposed mechanism itself; the paper does not derive it from prior theory, and the definition in Equation 4 is inconsistent with the prose that follows.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptive Exploration in Lenia with Intrinsic Multi-Objective Ranking." pith.science (2026). https://pith.science/paper/IMGFHHBD

@misc{pith2026250602990,
  author       = {Pith},
  title        = {Pith review of: Adaptive Exploration in Lenia with Intrinsic Multi-Objective Ranking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IMGFHHBD}},
  note         = {Machine review of arXiv:2506.02990}
}

read the original abstract

Artificial life aims to understand the fundamental principles of biological life by creating computational models that exhibit life-like properties. Although artificial life systems show promise for simulating biological evolution, achieving open-endedness remains a central challenge. This work investigates mechanisms to promote exploration and unbounded innovation within evolving populations of Lenia continuous cellular automata by evaluating individuals against each other with respect to distinctiveness, population sparsity, and homeostatic regulation. Multi-objective ranking of these intrinsic fitness objectives encourages the perpetual selection of novel and explorative individuals in sparse regions of the descriptor space without restricting the scope of emergent behaviors. We present experiments demonstrating the effectiveness of our multi-objective approach and emphasize that intrinsic evolution allows diverse expressions of artificial life to emerge. We argue that adaptive exploration improves evolutionary dynamics and serves as an important step toward achieving open-ended evolution in artificial systems.

Figures

Figures reproduced from arXiv: 2506.02990 by the authors.

**Figure 2.** Temporal progression of three patterns start [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Columns 1-4 display patterns evolved through multi-objective ranking, while columns 5-8 display patterns evolved [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 11 canonical work pages

[1]

A., McMullin, B., De Melo, V

Banzhaf, W., Baumgaertner, B., Beslon, G., Doursat, R., Foster, J. A., McMullin, B., De Melo, V. V., Miconi, T., Spector, L., Stepney, S., and White, R. (2016). Defining and simulating open-ended novelty: requirements, guidelines, and challenges. Theory in Biosciences , 135:131--161

work page 2016
[2]

A., Snyder, E., Brown, C

Bedau, M. A., Snyder, E., Brown, C. T., Packard, N. H., et al. (1997). A comparison of evolutionary activity in artificial evolving systems and in the biosphere. In Proceedings of the fourth European conference on artificial life , pages 125--134. MIt Press Cambridge

work page 1997
[3]

Chan, B. W.-C. (2018). Lenia - biology of artificial life. arXiv preprint arXiv:1812.05433

arXiv 2018
[4]

Chan, B. W.-C. (2020). Lenia and expanded universe. In Artificial Life Conference Proceedings 32 , pages 221--229, Cambridge, MA, USA. MIT Press

work page 2020
[5]

L., Vostinar, A

Dolson, E. L., Vostinar, A. E., Wiser, M. J., and Ofria, C. (2019). The modes toolbox: Measurements of open-ended dynamics in evolving systems. Artificial life , 25(1):50--73

work page 2019
[6]

and Cully, A

Faldor, M. and Cully, A. (2024). Toward artificial open-ended evolution within lenia using quality-diversity. In ALIFE 2024: Proceedings of the 2024 Artificial Life Conference . MIT Press

work page 2024
[7]

and Cully, A

Grillotti, L. and Cully, A. (2022). Unsupervised behavior discovery with quality-diversity optimization. IEEE Transactions on Evolutionary Computation , 26(6):1539--1552

work page 2022
[8]

and Stanley, K

Lehman, J. and Stanley, K. O. (2011). Abandoning objectives: Evolution through the search for novelty alone. Evolutionary Computation , 19(2):189--223

2011

Show all 13 references

[9]

Plantec, E., Hamon, G., Etcheverry, M., Oudeyer, P.-Y., Moulin-Frier, C., and Chan, B. W.-C. (2023). Flow-lenia: Towards open-ended evolution in cellular automata through mass conservation and parameter localization. In Artificial Life Conference Proceedings 35 , volume 2023, ...

2023
[10]

Reinke, C., Etcheverry, M., and Oudeyer, P.-Y. (2019). Intrinsically motivated discovery of diverse patterns in self-organizing systems

2019
[11]

Stepney, S. (2021). Modelling and measuring open-endedness. Artificial Life , 25(1):9

2021
[12]

and Hickinbotham, S

Stepney, S. and Hickinbotham, S. (2024). On the open-endedness of detecting open-endedness. Artificial Life , 30(3):390--416

2024
[13]

Taylor, T., Bedau, M., Channon, A., Ackley, D., Banzhaf, W., Beslon, G., Dolson, E., Froese, T., Hickinbotham, S., Ikegami, T., and McMullin, B. (2016). Open-ended evolution: Perspectives from the oee workshop in york. Artificial Life , 22(3):408--423

2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

A., McMullin, B., De Melo, V

Banzhaf, W., Baumgaertner, B., Beslon, G., Doursat, R., Foster, J. A., McMullin, B., De Melo, V. V., Miconi, T., Spector, L., Stepney, S., and White, R. (2016). Defining and simulating open-ended novelty: requirements, guidelines, and challenges. Theory in Biosciences , 135:131--161

work page 2016

[2] [2]

A., Snyder, E., Brown, C

Bedau, M. A., Snyder, E., Brown, C. T., Packard, N. H., et al. (1997). A comparison of evolutionary activity in artificial evolving systems and in the biosphere. In Proceedings of the fourth European conference on artificial life , pages 125--134. MIt Press Cambridge

work page 1997

[3] [3]

Chan, B. W.-C. (2018). Lenia - biology of artificial life. arXiv preprint arXiv:1812.05433

arXiv 2018

[4] [4]

Chan, B. W.-C. (2020). Lenia and expanded universe. In Artificial Life Conference Proceedings 32 , pages 221--229, Cambridge, MA, USA. MIT Press

work page 2020

[5] [5]

L., Vostinar, A

Dolson, E. L., Vostinar, A. E., Wiser, M. J., and Ofria, C. (2019). The modes toolbox: Measurements of open-ended dynamics in evolving systems. Artificial life , 25(1):50--73

work page 2019

[6] [6]

and Cully, A

Faldor, M. and Cully, A. (2024). Toward artificial open-ended evolution within lenia using quality-diversity. In ALIFE 2024: Proceedings of the 2024 Artificial Life Conference . MIT Press

work page 2024

[7] [7]

and Cully, A

Grillotti, L. and Cully, A. (2022). Unsupervised behavior discovery with quality-diversity optimization. IEEE Transactions on Evolutionary Computation , 26(6):1539--1552

work page 2022

[8] [8]

and Stanley, K

Lehman, J. and Stanley, K. O. (2011). Abandoning objectives: Evolution through the search for novelty alone. Evolutionary Computation , 19(2):189--223

2011