Pith. sign in

REVIEW 4 major objections 4 minor 20 references

Exploring Flow-Lenia Universes with a Curiosity-driven AI Scientist: Discovering Diverse Ecosystem Dynamics

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 The paper shows that setting goals in simulation-wide metric space lets an intrinsically motivated exploration algorithm find far more diverse ecosystem-level dynamics in Flow-Lenia than random search does.

desk verdict A useful extension of IMGEP to ecosystem-level metrics in Flow-Lenia, but the headline coverage numbers rest on an evolutionary-activity metric that is not computable as written, and the abstract promises a scaling study the body never delivers. read the letter →

arxiv 2505.15998 v4 pith:AQ45GZBG submitted 2025-05-21 cs.AI

Thomas Michel , Marko Cvjetko , Gautier Hamon , Pierre-Yves Oudeyer , Clément Moulin-Frier This is my paper · ORCID
classification cs.AI
keywords Flow-Leniacellularautomataopen-endedevolutionintrinsicallymotivatedgoalexplorationdiversitysearchevolutionaryactivitymulti-scaleentropyartificiallife
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

This paper tries to establish that a curiosity-driven goal exploration process (IMGEP) can systematically reveal diverse ecosystem-scale dynamics in Flow-Lenia, a continuous cellular automaton where matter flows and carries localized parameters. The key move is to define goals as simulation-wide metrics—evolutionary activity, compression-based complexity, and multi-scale entropy—rather than individual patterns. Quantitatively, the method achieves an average pairwise goal-space distance of 0.874 versus 0.476 for random search, and covers 576 non-empty metric bins versus 205. If the claim holds, this gives a cheap, principled way to find promising regions of parameter space in complex systems before running more expensive experiments.

What carries the argument

The engine is the IMGEP loop paired with a goal space of global metrics. IMGEP repeatedly samples a goal vector in metric space, finds the closest previously explored goal, mutates its Flow-Lenia parameters, runs the simulation, and archives the reached goal. The load-bearing metrics are non-neutral evolutionary activity (summing positive changes in the proportion of mass carrying each localized parameter type), compression-based complexity estimated by MP4 file size, and entropy of matter distribution at several downscaled resolutions. The Flow-Lenia setup adds a negotiation mixing rule and Gaussian parameter mutation so distinct patterns governed by different local rules can coexist and interact.

What would settle it

Re-run the two experiments after quantizing the parameter map into explicit bins before computing $p_i(t)$; if the IMGEP advantage in pairwise distance and coverage shrinks to random-search levels, the central discovery collapses.

Watch

Extended reading notes

Core claim

The central discovery is that goal-conditioned exploration with system-level metrics locates a broader, qualitatively richer set of ecosystem-scale behaviors in Flow-Lenia than random parameter sampling, in the first application of this kind to large multi-pattern cellular-automaton environments. The authors adapt IMGEP by sampling goals in a space of global descriptors and then selecting and mutating Flow-Lenia parameters closest to those goals. Across two experiments, the reached-goal archive is roughly twice as spread out and twice as coverage-rich as random search, and inspection of the archive shows structures that resemble colonies, feeding, and allopatric speciation. A scaling study across six spatial scales and seven time horizons reports macro-scale organization with no analogue at the base scale, and the paper argues this loop of cheap exploration, inspection, and targeted follow-up is the methodological contribution.

Load-bearing premise

The headline numbers depend on counting identical localized parameter types, while mutations add continuous Gaussian noise, so the evolutionary-activity metric needs an explicit discretization or binning to be well defined.

Editorial extensions

If this is right

  • IMGEP with system-level goals reaches an average pairwise goal-space distance of 0.874 and 576 non-empty metric bins, roughly twice the coverage of random search (0.476 and 205 bins).
  • The discovered archive contains self-organized behaviors analogous to colonies, feeding, and allopatric speciation that random search did not surface.
  • When matter starts in one corner of an obstacle maze, the same goal-driven loop discovers movement modes including fragmenting, turning, and directed travel through corridors.
  • Scaling up simulations to six spatial scales and seven time horizons reveals macro-scale organization that has no analogue at the base scale.
  • The workflow can act as a cheap scaffold: large-scale diversity search proposes regions of parameter space for follow-up, more expensive experiments.

Reading between the lines

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

  • A natural extension the authors leave implicit: the same archived goal space could seed targeted follow-up experiments by letting a human select interesting regions, then restarting IMGEP there at higher resolution.
  • If the balancing observation generalizes, the most interesting configurations should concentrate at intermediate values of multi-scale entropy and sustained non-zero evolutionary activity, a testable prediction for new metric spaces.
  • The compression-based complexity proxy could likely be replaced by a faster online statistic such as local entropy change, so the exploration loop can flag promising simulations in real time.
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.

Referee Report

4 major / 4 minor

Summary. The paper adapts Intrinsically Motivated Goal Exploration Processes (IMGEP) to Flow-Lenia, a continuous cellular automaton with mass conservation and parameter localization, in order to discover system-level ecosystem-like dynamics rather than individual patterns. The authors define three simulation-wide metrics (non-neutral evolutionary activity, MP4 compression size, and multi-scale entropy), use them as IMGEP goal spaces, and compare IMGEP against random search in two experiments: one targeting ecosystem dynamics and one targeting matter movement through obstacles. They report that IMGEP achieves an average pairwise goal-space distance of 0.874 versus 0.476 for random search, and a coverage of 576 non-empty bins versus 205, and they present qualitative behaviors such as feeding, colony-like clumps, and allopatric-speciation-like divergence. The abstract additionally promises a scaling study across six spatial scales and seven time horizons, but no such study appears in the full text. The paper also describes an interactive exploration tool and states an intention to release code.

Significance. If the quantitative claims were sound, the paper would demonstrate a genuinely useful extension of autotelic exploration from individual pattern discovery to ecosystem-level dynamics in a continuous cellular automaton, and the qualitative catalogue of emergent behaviors together with the interactive tool would be valuable resources for the artificial-life and complex-systems communities. The choice of an external random-search baseline is a reasonable starting point for comparison. However, the central quantitative evidence is currently compromised: the evolutionary-activity metric is undefined under the stated continuous mutation dynamics, the promised scaling study is absent, and the headline comparisons come from single runs without error bars. These issues must be addressed before the paper's main claim about systematically illuminating ecosystem dynamics can be accepted.

major comments (4)
  1. [§Methods, 'System-level Metrics', Eqs. (7)–(9)] The non-neutral evolutionary activity metric is not well defined under the stated dynamics. Equation (7) applies continuous Gaussian noise to the parameter map; after the first mutation every cell's parameter vector is distinct with probability one. Equation (9) defines p_i(t) as the proportion of mass carrying 'identical local parameter' i, so the sum in Eq. (8) degenerates into a per-cell redistribution measure rather than a measure of changes in a discrete genome population. The authors need to specify an explicit discretization or binning of the parameter space, or a discrete mutation operator, before the EA values in Table 1 and Figure 4 can be interpreted. The Discussion's concession that the method cannot 'definitively distinguish meaningful adaptations from random variation' reinforces rather than resolves this problem.
  2. [Abstract and §Experiments] The abstract promises a scaling study across six spatial scales and seven time horizons that 'uncovers macro-scale organization with no analogue at the base scale,' but no such experiment appears anywhere in the full text. Neither the Experiments section nor the Discussion reports results at different grid sizes or durations. This missing study is load-bearing for the claimed iterative scaffold for experiment design; either add the study with its numbers and analysis or remove the claim from the abstract.
  3. [Table 1 and Figures 4–5] The headline quantitative comparison rests on single runs. Table 1 reports average pairwise distance and coverage values, and Figure 5 shows curves for IMGEP versus random search, but no error bars, confidence intervals, or numbers of seeds are given. Because both algorithms contain stochastic components (random goals and mutations), 'significantly higher' needs a distribution over runs; otherwise the 0.874 versus 0.476 and 576 versus 205 comparisons could be within run-to-run noise. Multiple independent runs with variance reporting should accompany all coverage and distance statistics.
  4. [§Methods, 'Compression-based Complexity Metric'] The MP4-file-size complexity metric is under-specified: the text does not state the encoder settings, frame sampling rate, color quantization, or normalization used when converting a 256×256 Flow-Lenia state to video. Since compressed size is one of the three goal dimensions in Experiment 1, the metric's scaling properties directly affect which regions of goal space are reachable. The authors should provide the exact encoding protocol and, ideally, a stability check (e.g., file size versus encoder parameters) to show the proxy tracks simulation complexity rather than encoder artifacts.
minor comments (4)
  1. [Throughout] There are several typos and formatting slips: 'exibit' in the Experiments section, 'it's volume' in the Figure 7 caption, 'denots' in Eq. (4), and 'Frdric Kaplan' in the References should be 'Frédéric Kaplan'.
  2. [Algorithm 1 and §IMGEP Implementation] The goal sampling step ('Sample random goal in system-level metrics space') does not specify the sampling distribution or the bounds and normalization of the three metrics; without this information the IMGEP variant is not fully reproducible.
  3. [Discussion and Conclusion] The statement that 'we will open source the full codebase' is future tense; for a scientific discovery pipeline, releasing code and data (or a permanent artifact DOI) is important for reproducibility. The reader should not have to rely only on the companion visualization website.
  4. [Figures 6 and 9] The qualitative showcases would benefit from an explicit statement that these are selected examples and from an indication of how many of the archived simulations exhibit each qualitative category (feeding, colony-like clumps, allopatric-like divergence).

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the goal metric space doubles as the evaluation space, but a random-search comparator keeps the central claim externally grounded.

full rationale

The paper's derivation chain is not circular in any load-bearing sense. IMGEP's goals are defined in the same system-level metric space used for evaluation (Algorithm 1, Figure 4, Table 1), so the reported coverage and average-pairwise-distance gains partly reflect the algorithm's own objective. However, the values are empirical outcomes of parameter mutations followed by simulation and metric computation, not fitted inputs, and random search is scored on the identical metrics, providing an external baseline. No equation in the System-level Metrics section is constructed from the outputs it is said to predict, and no fitted parameter is renamed as a prediction. The repeated citations to overlapping-author work (Flow-Lenia, IMGEP, prior Lenia diversity search) provide substrate and algorithmic background; they are not invoked as a uniqueness theorem or as the sole justification for the central claim. The undefined status of p_i(t) under continuous Gaussian mutation (Eqs. 7-9) is a correctness and robustness concern, not a circularity concern. Overall, the only mild issue is the goal-space/evaluation-space overlap plus non-load-bearing self-citations, which warrants a low score rather than a finding of significant circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. The main burden comes from hand-chosen simulation and algorithm parameters whose values are not reported, plus assumptions about how continuous parameter maps can serve as discrete genomes and how compression and entropy metrics reflect meaningful complexity. The most fragile is the genome assumption underlying the evolutionary-activity metric.

free parameters (7)
  • Negotiation-rule inverse temperature beta
    Eq (5) controls selection pressure in the new mixing rule; its value is not given, yet it determines coexistence and diversity.
  • Mutation frequency and covariance Sigma
    Eq (7); controls rate at which new parameter variants appear; value not reported, critical for evolutionary-activity measure.
  • Flow and diffusion parameters alpha, theta_A
    Eq (2); strengths of attraction and diffusion, not reported.
  • IMGEP bootstrapping length N and goal sampling distribution
    Algorithm 1; shapes exploration and comparison; not specified.
  • Metric normalization and binning = 5 bins per dimension
    Coverage comparison uses five bins per goal dimension; normalization of metrics before distance computations is not described, and scales differ by many orders of magnitude.
  • MP4 encoder settings
    Compressed-size complexity metric depends on encoding parameters, grid, frame rate; none reported.
  • Grid size and simulation steps = 256x256, 10,000 steps
    Chosen to fit computational constraints; behavior may change at other scales.
assumptions (5)
  • domain assumption Flow-Lenia with mass conservation and parameter localization is an appropriate substrate for studying emergent ecosystem dynamics.
    The entire study relies on this model, cited from Plantec et al. 2023, with no independent validation of its ecological relevance.
  • ad hoc to paper Local parameter vectors can be treated as discrete 'genomes' for the evolutionary activity metric.
    Eq (8)-(9) require counting proportions of mass carrying identical parameters; continuous mutation in Eq (7) makes exact equality measure-zero without discretization.
  • domain assumption MP4 compressed file size approximates Kolmogorov complexity of the simulation.
    Section 'System-level Metrics'; no validation or comparison against other complexity measures.
  • domain assumption Multi-scale entropy of downscaled representations captures hierarchical spatial organization.
    Section 'System-level Metrics'; downscaling method is not specified and no justification links these entropies to organizational hierarchy.
  • ad hoc to paper The 'negotiation' mixing rule (Eq 5-6) encourages coexistence of multiple species; results depend on it.
    A new mixing rule introduced without comparison to the original stochastic selection, yet it is central to observed multi-pattern coexistence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Flow-Lenia Universes with a Curiosity-driven AI Scientist: Discovering Diverse Ecosystem Dynamics." pith.science (2026). https://pith.science/paper/AQ45GZBG

@misc{pith2026250515998,
  author       = {Pith},
  title        = {Pith review of: Exploring Flow-Lenia Universes with a Curiosity-driven AI Scientist: Discovering Diverse Ecosystem Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AQ45GZBG}},
  note         = {Machine review of arXiv:2505.15998}
}
read the original abstract

We present a curiosity-driven AI scientist method for discovering system-level dynamics in Flow-Lenia, a continuous cellular automaton (CA) with mass conservation and parameter localization. Building on prior work that uses diversity search in Lenia to find individual self-organized patterns, we adapt Intrinsically Motivated Goal Exploration Processes (IMGEPs) to large environments of interacting patterns, using simulation-wide metrics such as evolutionary activity, compression ratio, and multi-scale matter distribution. We apply IMGEP in two exploration experiments: one targeting ecosystem-level dynamics, the other matter movement through obstacle-laden environments. In both, IMGEP illuminates significantly more of the metric space than random search and reveals self-organized behaviors qualitatively resembling many biological phenomena. Leveraging the resulting archive, we then run a scaling study across six spatial scales and seven time horizons, uncovering macro-scale organization with no analogue at the base scale and characterizing how goal-space metrics behave at scale. This illustrates a strength of our approach: a relatively cheap large-scale diversity search can act as a principled scaffold for designing subsequent, more expensive experiments, enabling an iterative loop of experiment design, inspection, and redesign, supported by an interactive exploration tool that keeps scientists in the loop. Though demonstrated with Flow-Lenia, this approach potentially applies to other parameterizable complex systems where studying bottom-up collective behavior is of interest.

Figures

Figures reproduced from arXiv: 2505.15998 by the authors.

Figure 1
Figure 1. Snapshot of an advanced state of Flow Lenia. Each color represents a different set of localized parameters, determining how [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. IMGEP approach for exploring a diversity of self-organization in Flow Lenia. The algorithm samples new goals in a system-wide [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Examples of environments with multiple parameters co [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Evolution of exploration metrics over time. The con [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 8
Figure 8. Figure 8: An example of the matter distribution experimental [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 7
Figure 7. Figure 7: An example of dynamics resembling feeding. The [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 10
Figure 10. Figure 10: Comparison of reached goals between (left) and cu [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 11 canonical work pages

  1. [7]

    Omni-epic: Open-endedness via models of human no- tions of interestingness with environments programmed in code.arXiv preprint arXiv:2405.15568,

    Maxence Faldor, Jenny Zhang, Antoine Cully, and Jeff Clune. Omni-epic: Open-endedness via models of human no- tions of interestingness with environments programmed in code.arXiv preprint arXiv:2405.15568,

    arXiv

  2. [10]

    Conservation of matter increases evolutionary activity

    Simon John Hickinbotham and Susan Stepney. Conservation of matter increases evolutionary activity. InEuropean Conference on Artificial Life 2015, pages 98–105. MIT Press,

    work page 2015

  3. [15]

    Pierre-Yves Oudeyer, Frdric Kaplan, and Verena V Hafner

    Accessed: 2025-05-05. Pierre-Yves Oudeyer, Frdric Kaplan, and Verena V Hafner. Intrinsic motivation systems for autonomous mental development.IEEE transactions on evolutionary com- putation, 11(2):265–286,

    work page 2025

  4. [17]

    doi: 10.1002/cplx.20060

    ISSN 1099-0526. doi: 10.1002/cplx.20060. Hiroki Sayama. Toward the realization of an evolving ecosys- tem on cellular automata.Proc. Fourth Int. Symp. Artifi- cial Life and Robotics, pages 254–257,

    work page doi:10.1002/cplx.20060

  5. [18]

    Requirements for Open-Ended Evolution in Natural and Artificial Systems

    Tim Taylor. Requirements for open-ended evolution in natural and artificial systems.arXiv preprint arXiv:1507.07403,

    work page Pith review arXiv

  6. [19]

    V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291,

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291,

    arXiv

  7. [1984]

    Two-dimensional kolmogorov complexity and an empirical validation of the coding theorem method by compressibility.PeerJ Computer Science, 1:e23, 2015

    Hector Zenil, Fernando Soler-Toscano, Jean-Paul Delahaye, and Nicolas Gauvrit. Two-dimensional kolmogorov complexity and an empirical validation of the coding theorem method by compressibility.PeerJ Computer Science, 1:e23, 2015

    work page 2015

  8. [1988]

    The ai scientist: Towards fully automated open-ended scientific discovery.arXiv preprint arXiv:2408.06292,

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The ai scientist: Towards fully automated open-ended scientific discovery.arXiv preprint arXiv:2408.06292,

    arXiv

Show all 20 references
  1. [2004]

    doi: 10.1162/1064546041255539

    ISSN 1064-5462. doi: 10.1162/1064546041255539. Randall D. Beer. The cognitive domain of a glider in the game of life.Artificial Life, 20(2):183–206, 04

    doi

  2. [2007]

    Flow-lenia: Towards open-ended evolu- tion in cellular automata through mass conservation and parameter localization

    Erwan Plantec, Gautier Hamon, Mayalen Etcheverry, Pierre- Yves Oudeyer, Clément Moulin-Frier, and Bert Wang- Chak Chan. Flow-lenia: Towards open-ended evolu- tion in cellular automata through mass conservation and parameter localization. InArtificial Life Conference Proceeding...

    2023

  3. [2010]

    Mykhailo Moroz

    doi: 10.1063/1.3326990. Mykhailo Moroz. Reintegration tracking. https://michaelmoroz.github.io/ Reintegration-Tracking/,

    doi

  4. [2011]

    doi: 10.1007/978-1-4614-1770-5_3

    ISBN 978-1-4614-1770-5. doi: 10.1007/978-1-4614-1770-5_3. Seth Lloyd and Heinz Pagels. Complexity as thermodynamic depth.Annals of physics, 188(1):186–213,

    doi

  5. [2012]

    Mayalen Etcheverry.Curiosity-driven AI for Science: Au- tomated Discovery of Self-Organized Structures

    doi: 10.1162/978-0-262-31050-5-ch007. Mayalen Etcheverry.Curiosity-driven AI for Science: Au- tomated Discovery of Self-Organized Structures. PhD thesis, Université de Bordeaux,

    doi

  6. [2014]

    doi: 10.1162/ARTL_a_00125

    ISSN 1064-5462. doi: 10.1162/ARTL_a_00125. Bert Wang-Chak Chan. Lenia-biology of artificial life.arXiv preprint arXiv:1812.05433,

    arXiv

  7. [2015]

    Automat- ing the search for artificial life with foundation models

    Akarsh Kumar, Chris Lu, Louis Kirsch, Yujin Tang, Ken- neth O Stanley, Phillip Isola, and David Ha. Automat- ing the search for artificial life with foundation models. arXiv preprint arXiv:2412.17799,

    arXiv

  8. [2018]

    Towards Large-Scale Simulations of Open-Ended Evolution in Continuous Cellular Au- tomata.arXiv preprint arXiv:2304.05639,

    Bert Wang-Chak Chan. Towards Large-Scale Simulations of Open-Ended Evolution in Continuous Cellular Au- tomata.arXiv preprint arXiv:2304.05639,

    arXiv

  9. [2019]

    A quantitative mea- sure of non-neutral evolutionary activity for systems that exhibit intrinsic fitness

    Alastair Droop and Simon Hickinbotham. A quantitative mea- sure of non-neutral evolutionary activity for systems that exhibit intrinsic fitness. InALIFE 2012: The Thirteenth International Conference on the Synthesis and Simula- tion of Living Systems, pages 45–52. MIT Press, July

    2012

  10. [2020]

    Dis- covering sensorimotor agency in cellular automata us- ing diversity search.arXiv preprint arXiv:2402.10236,

    Gautier Hamon, Mayalen Etcheverry, Bert Wang-Chak Chan, Clément Moulin-Frier, and Pierre-Yves Oudeyer. Dis- covering sensorimotor agency in cellular automata us- ing diversity search.arXiv preprint arXiv:2402.10236,

    arXiv

  11. [2022]

    Towards an ai co-scientist.arXiv preprint arXiv:2502.18864,

    Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Anil Palepu, Petar Sirkovic, Artiom Myaskovsky, Felix Weissenberger, Keran Rong, Ryutaro Tanno, et al. Towards an ai co-scientist.arXiv preprint arXiv:2502.18864,

    arXiv

  12. [2024]

    Toward artificial open- ended evolution within lenia using quality-diversity

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

    2024

Pith tools

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