{"id":"25385638-59ac-47d6-bca3-ca4392e3a50f","arxiv_id":"2506.08569","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A mass-conserving extension of Lenia yields stable localized patterns, supports easy optimization of behaviors, and produces multispecies simulations with measurable evolutionary dynamics.","lead":"Flow-Lenia is a modified version of the Lenia cellular automaton where total mass is conserved, causing simple blobs of matter to form stable, life-like creatures instead of exploding or dying. The paper shows these creatures can be optimized for tasks and that embedding their rules in the moving matter enables multi-species worlds with measurable evolutionary dynamics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Intrinsic-evolution claim lacks a neutral control: stochastic parameter sampling plus continuous mutations can produce the reported saturation, activity curves, and PCA branching without selection on the embedded parameters.","rationale":"The paper's strongest and most novel claim is that Flow-Lenia's parameter-embedding mechanism supports genuine emergent evolutionary dynamics, evidenced by evolutionary trees, diversity saturation, and evolutionary-activity measurements. The reader's conditional verdict already hinges on the absence of a neutral control and the provisional species definition. My stress-test concurs and sharpens the concern: because species are continuous, real-valued parameter points, every mutation defines a new 'species', and the stochastic softmax copying rule of Eq. 8 is itself a neutral genealogical process if parameters do not affect fitness. The reported sublinear parameter-count saturation, linear EA growth, power-law decline with pmut, and PCA branching patterns are all qualitative signatures that a neutral model with finite grid capacity would also be expected to produce. The absence of a control in which parameters are heritable but phenotypically inert leaves the selection claim underdetermined. I do not find a more serious internal inconsistency: the mass-conservation optimization results are well supported, the dissipative/food mass-bias reversal is honestly reported, and the reproducible JAX notebook is a genuine asset. The missing neutral baseline does not warrant rejection, but it does justify keeping the conditional verdict until the control is run.","tokens_in":15903,"tokens_out":3837,"duration_ms":52635,"concrete_test":"Run a neutral-control simulation identical to the vanilla intrinsic-evolution setting (Sec. 4.3.1) except that P is decoupled from dynamics: compute U_t in Eq. 7 with a fixed, species-independent weighting vector h while still transporting P with matter (Eq. 6), applying mutation beams at the same pmut values, and sampling parameter inheritance via Eq. 8. Then compare (i) number of distinct parameters vs time and vs pmut, (ii) EA^C and EA^N curves and their power-law fits, and (iii) PCA evolutionary-tree branch structure against Figs. 8-10. If the neutral control reproduces the key qualitative patterns (saturation, linear EA growth, power-law pmut decay, branching trees), the claim that these metrics evidence selection on embedded parameters fails; if the neutral curves are flat or lower and trees do not form persistent branches, the original interpretation is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Sec. 5.3 is that the measured evolutionary activity, diversity saturation, and PCA branching trees reveal an intrinsic fitness landscape and emergent evolutionary dynamics. The load-bearing premise is that the stochastic parameter inheritance rule (Eq. 8) plus mutation beams (Sec. 4.3.1) create selective propagation of heritable parameter vectors. The paper provides no neutral baseline in which the parameter map is transported and mutated but does not affect the affinity U_t in Eq. 7. This matters because 'species' are defined as unique real-valued points in a continuous parameter space; every mutation beam creates a new species, and Eq. 8 is a softmax copying process that, under neutrality alone, produces coalescent-like genealogies, random-walk PCA trajectories, and competitive saturation because the grid has finite carrying capacity. The sublinear |P|-vs-pmut curve and the power-law decay of EA with pmut are consistent with neutral drift plus bounded space. EAN increments on upward proportion changes, so it also rewards variants that drift upward by chance. The authors' own Sec. 6 caveat that the species definition is provisional and that coherently moving clusters may be the true units reinforces that point-based metrics may be measuring genealogy, not adaptation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Flow-Lenia, a mass-conservative continuous cellular automaton extending Lenia. It claims that mass conservation makes spatially localized patterns (SLPs) the generic outcome of random parameter sampling, that update-rule parameters can be optimized with plain evolution strategies for directed motion, angular motion, obstacle navigation, and chemotaxis, and that embedding parameters within the dynamics enables multispecies simulations in which evolutionary activity, diversity, and parameter-space branching reveal emergent intrinsic evolutionary dynamics. The paper presents three model variants (vanilla, dissipative, and food) and analyzes them with count-based and non-neutral evolutionary activity metrics, with accompanying videos, a companion website, and open-source JAX code.","tokens_in":16135,"tokens_out":7073,"duration_ms":90932,"significance":"If the main claims hold, Flow-Lenia is a substantial contribution to artificial life: it addresses a known difficulty in Lenia, namely the scarcity of stable spatially localized patterns, by construction rather than by search, and it offers a concrete mechanism for local rule parameters to coexist and compete, which is a prerequisite for studying open-ended evolution in continuous cellular automata. The random-search and evolution-strategy results are convincing as qualitative evidence, and the paper's reproducibility assets (code, notebook, videos) are valuable. However, the intrinsic-evolution claim currently rests on metrics and species definitions that may not distinguish selection from neutral drift; the absence of a neutral control is the main threat to the central interpretation.","major_comments":[{"comment":"The comparison between the three model variants is not robust because total mass is not conserved in the dissipative and food settings. The text first reports that these settings have significantly higher count-based and non-neutral activity (p < 10^-5), then acknowledges that dividing by total mass reverses the ordering. This is more than a nuisance: the corrected and uncorrected measures support opposite conclusions, and the Discussion's wording ('the EA of the dissipative and food conditions are lower than in the vanilla systems') appears to rely on the corrected measure while the Results headline relies on the uncorrected one. The paper should state one primary measure, justify why it is appropriate, and report both consistently when summarizing the findings.","section":"§5.3, Fig. 10"}],"minor_comments":[{"comment":"The abstract contains two typos: 'phenomenons' should be 'phenomena' and 'spatially-localized patters' should be 'spatially-localized patterns'.","section":"Abstract"},{"comment":"The claim that most Flow-Lenia patterns are SLPs and that the same parameters mostly lead to empty or exploding patterns in Lenia would be stronger with a quantitative classification (for example, the fraction of runs yielding non-empty, bounded patterns) rather than visual inspection of 105 instances.","section":"§5.1, Fig. 3"},{"comment":"The exact mass-conservation argument applies to the full sum over all cells L, but the implemented computation restricts source cells to a Chebyshev distance of less than 5; the authors should either state a boundedness assumption on dt·F that guarantees no mass is lost, or provide a numerical check that total mass drift is negligible over 500,000 steps.","section":"§3, Eq. (6)"},{"comment":"For the directed-motion comparison with original Lenia, only a qualitative statement is given ('yellow curve', 'every Lenia optimized patterns are exploding ones'); reporting final fitness distributions or learning curves for both systems would make the optimization claim more precise and easier to verify.","section":"§5.2, Fig. 5"},{"comment":"The power-law fits and Mann-Whitney tests should specify the unit of analysis (per seed, per time step, or aggregated) and should report confidence intervals or effect sizes; with five seeds and long time series, the reported R² values alone are hard to evaluate.","section":"§5.3, Figs. 9–10"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has already been accepted for publication in Artificial Life according to the footnote; if this report is for the journal, the neutral-control issue should be addressed in a revision or explicitly reframed as a known limitation. The model and qualitative results are a useful contribution, and the missing control is fixable within the paper's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Flow-Lenia is a real step up for continuous CA as an experimental platform: the mass-conservation reformulation makes spatially localized patterns the default outcome, and the evolutionary-strategy results for directed motion, chemotaxis, and obstacle navigation are convincing. That part of the paper is solid, and the authors are honest that the core mechanism dates to their ALIFE 2023 paper; the genuinely new material here is the large-scale multispecies simulations, the dissipative and food model variants, and the evolutionary-activity/diversity analyses.\n\nThe weak spine is the intrinsic-evolution claim in Section 5.3. The metrics define a species as a unique point in parameter space, and the authors themselves flag in Section 6 that coherently moving clusters may be the right units. More importantly, the stochastic parameter-copying rule (Eq. 8) plus mutation beams will, under pure neutrality, produce the exact qualitative signatures they report: sublinear growth of unique parameter counts with mutation rate, saturation after a rapid initial phase, and PCA trajectories that look like a branching tree. The grid has finite carrying capacity, so any copying process with mutation and spatial competition will generate coalescent-like genealogies and apparent speciation. Without a control in which the embedded parameters are transported and mutated but do not affect the affinity map U_t, the measured EA activity cannot be attributed to selection rather than drift. The sublinear |P| vs pmut curve is not evidence for competition by itself; neutral drift in a bounded world gives the same scaling.\n\nThis is a fixable gap, not a fatal one. The model is a good tool, and the optimization results stand independent of the evolutionary-dynamics interpretation. The mass-bias issue in the dissipative/food comparison is minor because they report the corrected values and discuss the reversal. Minor reproducibility gap: the specific pmut values used in Figures 8 and 9 are not stated in the text.\n\nI'd send this to a serious referee and ask for either a neutral control or a much more careful argument that the observed dynamics exceed a null model of neutral copying. If that control lands, the paper becomes a strong contribution to the open-endedness literature; if it doesn't, those evolutionary-dynamics claims need to be scaled back to genealogical trees rather than fitness landscapes.","headline":"Worth publishing as a model paper, but the evolutionary-dynamics claims need a neutral control before they should be trusted.","tokens_in":16685,"tokens_out":3811,"would_cite":true,"duration_ms":42459,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37B15","68Q80"],"pacs":[],"model":"deepseek-v4-flash","headline":"One conserved quantity makes Lenia creatures evolve together","keywords":["Flow-Lenia","mass conservation","continuous cellular automata","Lenia","intrinsic evolution","evolutionary activity","multi-species simulation","open-ended evolution"],"falsifier":"Re-run the multispecies simulations with the same mass dynamics and mutation beams but replace the mass-proportional softmax sampling of Eq. 8 with a uniform random choice among the parameter vectors arriving in each cell. If evolutionary-activity measures and branching trees persist, selection on parameters is not needed for the observed dynamics; if they collapse to linear parameter growth, the mass-weighted competition is the causal mechanism claimed.","tokens_in":15725,"feed_emoji":"🧬","tokens_out":8177,"duration_ms":91551,"temperature":0.7,"pith_summary":"The paper introduces Flow-Lenia, a continuous cellular automaton that extends Lenia by conserving the total mass of matter in the world. It argues that mass conservation is the key ingredient that makes life-like structures easy to find: almost all randomly sampled rules yield spatially localized patterns, while the same rules mostly explode or vanish in Lenia. A second consequence is that the parameters of each local update rule can be carried by the matter itself, so creatures with different rules can occupy one world, compete for mass, and pass their parameters on. Using evolutionary-activity measures on very long simulations, the paper claims to observe emergent evolutionary dynamics, including branching parameter trajectories and systematic responses to mutation rate and to dissipative and food-based environment variants.","feed_headline":"One conserved quantity makes Lenia creatures evolve together","feed_subtitle":"Flow-Lenia keeps total matter fixed, so patterns stay localized and competing rule parameters form branching evolutionary trees.","key_machinery":"The carrying object is the parameter map $\\mathcal{P}_t: L \\to \\Theta$, which attaches a local update-rule weight vector to every cell of matter and is transported with the matter during each reintegration-tracking step. Affinity scores are computed with the localized parameters via Eq. 7, and the stochastic mixing rule (Eq. 8) samples one incoming parameter set per cell with probability proportional to incoming mass, so lineages compete by converting each other's matter. The flow itself combines the affinity gradient with a concentration-gradient diffusion term weighted by $\\alpha(x)$, and reintegration tracking (Eq. 6) moves matter while conserving total mass. This machinery turns \"species\" into unique points in parameter space and gives the system a heritable, selectable substrate.","core_discovery":"The central claim is that Flow-Lenia's mass conservation does not merely stabilize Lenia but changes the kind of phenomena the system can exhibit. With total mass fixed and matter moving along the gradient of an affinity map via reintegration tracking, spatially localized patterns become the generic outcome rather than a rare search product, and the update rule itself can be embedded in the moving matter through a parameter map. The paper's parameter-mixing rule selects one incoming parameter vector with probability proportional to the mass arriving with it, which lets one creature convert another's matter and thus creates competition on heritable parameters. In simulations lasting 500,000 steps with mutation beams, the number of distinct parameters grows sublinearly with mutation rate, parameter trajectories project into branching tree-like structures, and count-based and non-neutral evolutionary activity follow power-law responses to mutation rate and differ systematically across the vanilla, dissipative, and food models. The authors read these observations as evidence of an intrinsic fitness landscape and of emergent evolutionary dynamics without any externally imposed fitness function.","pith_inferences":["A neutral control that decouples mutation from mass-weighted selection would sharpen the paper's central distinction; I would expect the branching trees to depend on the softmax competition, which the current simulations do not directly test.","The paper's provisional species definition, one species per unique parameter vector, likely undercounts biological species; redefining species as coherent clusters moving in parameter space, as the authors suggest, could change the evolutionary-activity numbers.","The inverse dissipative result hints that predictions about dissipation as a driver of life may be model-dependent; varying beam rates and input-zone layouts could test whether any dissipation regime raises activity.","A testable extension would be to let the embedded parameters include kernel or growth-function parameters beyond the weight vector, at the cost of losing fast Fourier convolution; the resulting phenotypes might show richer heritability."],"forward_implications":["Mass conservation alone acts as a regularizer: the same random parameter sets that mostly explode or vanish in Lenia produce spatially localized, sometimes complex patterns in Flow-Lenia.","Simple evolution strategies suffice to optimize Flow-Lenia creatures for directed motion, angular motion, obstacle navigation, and chemotaxis, where the same optimization recipe on Lenia is unstable and yields exploding patterns.","Embedding parameters in matter enables multispecies coexistence and competition; the sublinear growth of distinct parameter counts with mutation rate indicates that lineages compete rather than accumulate neutrally.","Evolutionary-activity measures respond systematically to conditions: activity decays as a power law with mutation rate, and the dissipative and food variants alter activity and diversity, with mass-corrected measures reversing the raw ordering.","If correct, Flow-Lenia offers a quantitative arena for testing theories about when dissipative dynamics or shared resources promote or hinder open-ended evolution."],"supporting_citations":[{"why":"Defines Lenia, the continuous cellular automaton that Flow-Lenia extends by adding mass conservation.","marker":"Chan, 2019, 2020"},{"why":"Supplies reintegration tracking, the semi-Lagrangian transport method that conserves total mass in the flow update.","marker":"Moroz, 2020"},{"why":"Introduces the evolutionary-activity framework used to measure whether evolution is taking place.","marker":"M. A. Bedau & Packard, 1996"},{"why":"Adds the non-neutral activity measure and the observation that high mutation rates reduce activity, used as a quantitative benchmark.","marker":"Droop & Hickinbotham, 2012"},{"why":"Provides prior evidence that conservation of matter increases evolutionary activity, motivating the design choice.","marker":"Hickinbotham & Stepney, 2015"},{"why":"Earlier Flow-Lenia report that established the parameter-embedding mechanism and mixing rules used here.","marker":"Plantec et al., 2023"},{"why":"The evolution-strategy optimizer used for the directed-motion, navigation, and chemotaxis tasks.","marker":"Salimans et al., 2017"},{"why":"Shows that comparable Lenia creatures require diversity search and curriculum methods, the baseline Flow-Lenia's optimization is contrasted with.","marker":"Hamon et al., 2024"},{"why":"A previous attempt at large-scale intrinsic-evolution simulation in Lenia that the multispecies experiments extend.","marker":"Chan, 2023"}],"fun_headline_variants":["Mass conservation unlocks evolutionary dynamics in Lenia","Flow-Lenia: evolution emerges from fixed total matter","Conserved mass in Lenia spawns branching creature lineages","Lenia with fixed mass evolves creatures autonomously","Flow-Lenia: emergent evolution from a single constraint"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the stochastic softmax parameter-mixing rule and the mutation beams create genuine selection on heritable parameter vectors, so the branching trees and evolutionary-activity scores measure adaptation rather than neutral mixing or an artifact of the softmax competition.","fun_headline_variants_meta":{"raw":{"variants":["Mass conservation unlocks evolutionary dynamics in Lenia","Flow-Lenia: evolution emerges from fixed total matter","Conserved mass in Lenia spawns branching creature lineages","Lenia with fixed mass evolves creatures autonomously","Flow-Lenia: emergent evolution from a single constraint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1471,"prompt_tokens":944,"completion_tokens":527,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":453}},"tokens_in":560,"tokens_out":527,"duration_ms":5655,"temperature":1.0,"reasoning_tokens":453,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:07:04.892983+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the multispecies simulations with the same mass dynamics and mutation beams but replace the mass-proportional softmax sampling of Eq. 8 with a uniform random choice among the parameter vectors arriving in each cell. If evolutionary-activity measures and branching trees persist, selection on parameters is not needed for the observed dynamics; if they collapse to linear parameter growth, the mass-weighted competition is the causal mechanism claimed.","supporting_citations":[],"review_version":1}