{"id":"3e473900-1af3-4e68-887e-8e223e7834e7","arxiv_id":"2411.09779","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Introduces a multilevel Wright-Fisher process built on the Kantorovich monad, derives a mixed Price equation, and fits multilevel genetic algorithms to synthetic data with variational and coalescent methods.","lead":"A new mathematical model describes evolution acting on cells, individuals, and groups at the same time, using a tool from category theory and probability. The authors use it to study when selection at different levels conflicts, and to fit genetic algorithms to simulated data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 3's proof is internally inconsistent: Eq. 19 has the κ term with the wrong sign, so the claimed conflict regime is unsupported; this is the load-bearing theoretical claim.","rationale":"The reader's weakest assumption points to the missing coalescent density in Eq. 32, which is a real and clear gap in the optimization section; I agree that the importance-sampling estimator is not computable as written and that the empirical coalescent results are unsupported. However, I judge the more load-bearing issue to be Prop. 3, because it is the main stated theoretical contribution: the paper claims to characterize parameter regimes where selection acts antagonistically across levels. The proof of Prop. 3 contains an algebraic sign inconsistency (Eq. 19) and ignores the κψ = κ contribution for uniform groups, so the central theoretical claim is not supported independently of the optimization concerns. If Prop. 3 cannot be repaired, the framework's core biological characterization is unproven; if it is repaired, the Eq. 32 issue still remains. Therefore my read does not change the reader's REJECT verdict.","tokens_in":14455,"tokens_out":16893,"duration_ms":171904,"concrete_test":"Independently re-derive Eqs. 18-20 from Eqs. 3-4 for the state constructed in Eq. 17. Compute f_{l',1}^* and f_{l',2}^* with the correct ψ for uniform and mixed groups, including the +κ contribution for uniform groups. Compare with the printed Eq. 19: if the correct coefficient of κ in f_{l',1}^* is negative (e.g., -κ d_l · 2(N-1)/N^2) rather than positive as printed, Prop. 3's proof fails as stated. Then recompute the inequality |Δ_l^{l'}| > Δ_l^l; if it still holds for large κ, the proposition may be salvageable, but the submitted proof is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Prop. 3 is the paper's central theoretical result, and its proof is the weakest link. In Eq. 19, the mixed level-l' group's fitness is written as f_{l',1}^* = N^{l'}μ_{l'} + κ(1 - ((N-1)^2+1)/N^2)(1 - d_l(...)). The coefficient is positive and d_l is a bounded metric, so 1 - d_l ≥ 0; increasing κ therefore increases f_{l',1}^*. But the proof immediately asserts that f_{l',2}^* - f_{l',1}^* can be made arbitrarily large by increasing κ. The printed equation has the opposite sign from what the argument needs. Separately, the proof states that uniform meta-populations at level l have cooperativity term 0, although Eq. 4 gives ψ = 1 for a uniform population, contributing +κ to f_l; Eqs. 18 and 20 do not track this constant. The constant may cancel in some comparisons, but the claim that |Δ_l^{l'}| > Δ_l^l is obtained without the corrected expressions. Thus the existence of conflict states for κ > κ* is not established. This matters because the abstract's characterization of antagonistic/cooperative regimes rests on Prop. 3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multilevel Wright-Fisher process on the Kantorovich monad, with fitness defined recursively across levels and a cooperativity term κ. It derives a mixed multilevel Price equation (Prop. 1), shows absence of conflict when κ=0 (Prop. 2), and claims that for sufficiently large κ there are population states with negative expected change in mean fitness at every level below the top (Prop. 3). The second half of the paper develops variational and stochastic optimization methods, including a multilevel coalescent sampler, and evaluates them on a small synthetic TSP-based problem.","tokens_in":14787,"tokens_out":14371,"duration_ms":146014,"significance":"If the theoretical results held, the paper would offer a useful, unified formalization of multilevel selection and a parameter-based criterion for antagonism between levels, together with a general inference framework. The category-theoretic construction is elegant, and the authors are honest in noting that Prop. 1 is a special case of Rice's equation. The synthetic experiments are a reasonable proof-of-concept. However, the proof of Prop. 3 contains algebraic sign errors, the coalescent-based estimator lacks a specified density, and the model definition in Eq. (6) appears to be missing a normalization. These are load-bearing issues, so the contribution is presently conditional on substantial repairs.","major_comments":[{"comment":"Proposition 3 is not established by the proof as printed. In Eq. (19), the mixed level-l' group is assigned fitness f^{l',*}_1 = N^{l'} μ_{l'} + κ(1 - ((N-1)^2+1)/N^2)(1 - d_l(...)), while the uniform group is assigned f^{l',*}_2 = N^{l'} f^*_2 with no κ term. Because f^*_1 > f^*_2, μ_{l'} lies above f^*_2, and the coefficient of κ is nonnegative, this makes f^{l',*}_1 - f^{l',*}_2 increase with κ; the proof immediately asserts the opposite. The statement in the same paragraph that uniform level-l meta-populations have 'cooperativity term 0' also contradicts Eq. (4), which gives ψ=1 for a uniform population and hence a contribution κ to f_l. The central conflict claim of the paper therefore rests on an equation with the wrong sign.","section":"§3, Eq. (19)"},{"comment":"Eq. (18)'s variance expression is not consistent with the state defined in Eq. (17). With h level-l meta-populations having fitness f^*_1 and N^{L-l}-h having fitness f^*_2, the variance of the level-l fitnesses contains the factor h(N^{L-l}-h), and the factor N/h appearing in Eq. (18) is unexplained; the displayed value also omits h entirely. Because Eq. (18) is used to bound Δ_l^l above, the bound Δ_l^l < μ/2 is unsupported.","section":"§3, Eq. (18)"},{"comment":"The importance-sampling estimate in Eq. (32) requires explicit densities for both the forward process P^{L,T}_{θ} and the multilevel coalescent process C(.|x^{T†}_L, θ). The paper specifies C only algorithmically, and the text states 'we assume that this can be done' without giving the density or its normalization. Without these densities, the ratio P/C is not computable, and the coalescent variants of VO, SPSA, and MC-EM, as well as the claim in Table 1 that coalescent simulations are beneficial, are not well-defined.","section":"§4.2, Eq. (32)"},{"comment":"The transition kernel in Eq. (5) uses \\tilde f, but Eq. (6) is printed as \\tilde f(x^{t,i}_l) = f(x^{t,i}_l) times a sum of parent fitnesses, rather than as f(parent) divided by the sum of fitnesses in the parent group. Taken literally, this does not define a Wright-Fisher reproduction step and invalidates the proof of Prop. 1, which requires \\tilde f to be a normalized reproduction probability. This needs to be corrected or the notation explained.","section":"§2, Eq. (6)"},{"comment":"The algebra in Example 1 conflates the normalized quantities \\tilde f^*_1, \\tilde f^*_2 with the raw fitness values f^*_1, f^*_2, and the simplifications leading to Eq. (23) do not follow from Eq. (22). Since the example is used to assert that f^{L-1} is non-decreasing for all states in a two-level model, that assertion is not supported as written.","section":"§3, Example 1, Eqs. (22)-(23)"}],"minor_comments":[{"comment":"'Kantorivich' is a typo for 'Kantorovich'.","section":"§2, first paragraph"},{"comment":"'Simultaneous Perturbation Stochastic Perturbation' should be 'Simultaneous Perturbation Stochastic Approximation' (SPSA).","section":"§4 heading"},{"comment":"'p_L is uniform' should read 'p_l is uniform over l=0,...,L-1', since level L is not updated.","section":"§3, Prop. 3 statement"},{"comment":"The permutation notation and Iverson brackets in Eq. (8) are not defined precisely; please spell out the indexing and the ranges of the permutations.","section":"§2, Eq. (8)"},{"comment":"The notation in the variance expression is garbled: the factor 'NP' has no clear meaning, and the formula should be rewritten with explicit definitions of the sums and variances.","section":"§3, Eq. (13)"},{"comment":"Several comparisons are within one standard deviation (for example, VO Wasserstein 0.25±0.02 vs. 0.23±0.03), so the statement that coalescent simulations are beneficial for all methods should be qualified or supported by significance testing.","section":"§5, Table 1"}],"recommendation":"reject","confidential_remarks":"The paper contains two largely independent contributions: a theoretical analysis of multilevel selection and a learning framework for multilevel genetic algorithms. Both have load-bearing gaps in the current version: the proof of Prop. 3 has a sign error in the central equation, and the coalescent estimator is not accompanied by a computable density. I would view a resubmission favorably if the authors can repair Prop. 3, provide a concrete density for the multilevel coalescent process, and clarify the normalization in Eq. (6)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely interesting framework, but the main theoretical result has a load-bearing proof error and the coalescent estimator is not actually computable as written. I would send it to review—the model is worth engaging—but it should not be accepted without major revision.\n\nWhat is actually new: the multilevel Wright-Fisher process on the Kantorovich monad, with recursive Wasserstein fitness, and the attempt to fit such processes with VO, SPSA, and MC-EM using both forward and coalescent simulation. The model definition is clear, and the categorical framing is structural rather than decorative: Wasserstein distance recurses naturally across levels. Prop. 1 is correctly labeled a special case of Rice's equation; that is an honest reduction, not overclaiming, though the abstract's wording overstates the novelty. Prop. 2 is straightforward and sound.\n\nThe soft spots are real. Prop. 3 is the centerpiece of the theoretical section and its proof, as printed, does not work. Eq. 19 gives the mixed-level group a positive κ term, so increasing κ makes f_1^{l',*} larger, not smaller; the text then claims f_2^{l',*} − f_1^{l',*} can be made arbitrarily large. That is the opposite sign from what the argument needs. The proof also drops the constant κ term for uniform groups without explanation. Since the abstract's antagonistic/cooperative characterization rests on this proposition, this is not a typo-level issue. The examples are plausible but do not rescue the general claim.\n\nSecond, the coalescent-based estimators in Sec. 4.2 are not fully specified. Eq. 32 requires the density of the coalescent sampler C(.|x_T^†, θ), but the paper only describes the algorithm and says \"we assume this can be done.\" Without a density, the importance weights in Eq. 32 and the MC-EM objective in Eq. 37 cannot be computed. The synthetic experiments are a toy TSP, five runs, ten epochs; they illustrate the pipeline but provide little evidence for the claim that coalescent training helps. No code or data are included.\n\nWho should read this: people working on multilevel selection or evolutionary computation who want a formal, recursive language for levels. The model formulation and Prop. 2 are useful even if Prop. 3 falls. A serious referee should get it, but the authors need to fix Prop. 3 or demote it to a conjecture, and they need to give a real coalescent density or drop the importance-sampling estimator.","headline":"A genuinely interesting multilevel framework with a load-bearing proof error in Prop. 3 and an underspecified coalescent estimator; worth reviewing, but needs major revision.","tokens_in":15290,"tokens_out":4292,"would_cite":false,"duration_ms":39729,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92D15","60B05","18B99"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a multilevel Wright-Fisher process on the Kantorovich monad supports a mixed multilevel Price equation, and that above a cooperativity threshold $\\kappa^*$ expected mean fitness can decrease at every level below the…","keywords":["multilevel selection","Wright-Fisher process","Kantorovich monad","Price equation","Wasserstein distance","coalescent analysis","variational optimization","genetic algorithms"],"falsifier":"Write down and normalize the multilevel coalescent density $C(\\cdot|x_T^\\dagger_L,\\theta)$ for a small case, then check the importance-sampling identity in Eq. 32 numerically; if the density is intractable or its normalizing constant is not 1, the forward-coalescent VO, SPSA, and MC-EM objectives are not well defined.","tokens_in":14240,"feed_emoji":"🧬","tokens_out":7626,"duration_ms":70999,"temperature":0.7,"pith_summary":"This paper proposes a unified mathematical model of multilevel selection: a Wright-Fisher process whose states live in the Kantorovich monad, so populations of populations can be nested to any depth. The authors aim to show that this single construction supports both theory and learning: it yields a mixed multilevel Price equation that splits expected trait change into a probabilistic mixture over levels, and it gives a parameter threshold $\\kappa^*$ at which, for any level below the top, high cooperativity $\\kappa$ makes expected mean fitness decrease for some population states. On the learning side, the paper aims to show that genetic algorithms over the same process can be fit to data by minimizing a recursive Wasserstein objective, and that combining forward simulation with a multilevel coalescent sampler improves the fit on a synthetic traveling-salesman problem. A sympathetic reader would care because this is a rare attempt to have one formal framework that both characterizes when levels of selection conflict and provides inference algorithms for fitting such multilevel processes.","feed_headline":"High cooperativity can make mean fitness fall at every level below the top","feed_subtitle":"A multilevel Wright-Fisher model plus Price equation gives a threshold where lower levels lose fitness.","key_machinery":"The central object is the Kantorovich monad on the category of 1-bounded compact metric spaces: the operation $B$ sends a metric space $X$ to the space of Borel probability measures on $X$, with the Wasserstein distance as metric, and the monad's unit and multiplication let a population be a measure over genotypes, a meta-population a measure over populations, and so on indefinitely. The multilevel Wright-Fisher transition uses a recursive fitness $f_l(x_l)=\\kappa\\psi(x_l)+\\sum_i f_{l-1}(x_{l-1}^{(i)})$, where $\\psi$ is the average pairwise similarity, one minus the Wasserstein distance, inside the meta-population. The fitness recursion is what makes the mixed Price equation and the conflict threshold computable; the multilevel coalescent process, which walks backward through the same level assignments, is what makes the importance-sampled training objectives possible. The paper defines a multilevel coalescent process $C(\\cdot|x_T^\\dagger_L,\\theta)$ that assigns parent states backwards and fills undefined states by forward sampling.","core_discovery":"The paper's central claim is that a multilevel Wright-Fisher process on the Kantorovich monad has a tractable selection decomposition and a sharp conflict regime. Proposition 1 derives a Mixed Multilevel Price Equation, $E[\\phi_{t+1}-\\phi_t]=\\sum_{l<L}p_l\\,\\mathrm{Cov}(\\phi_t,\\hat\\Omega^t_l)$, in which the expected change in a trait is a probability-weighted covariance sum over the level chosen to reproduce. Proposition 2 shows that with no cooperativity ($\\kappa=0$) expected mean fitness increases at every level. Proposition 3 shows that once $\\kappa$ exceeds a level-dependent threshold $\\kappa^*$, for every level $l<L-1$ there are population states in which expected mean fitness decreases. Two worked examples delimit the top level: for $L=2$, $N=2$ mean fitness at the top level never decreases, while for $N\\ge 5$ the authors find states where it does. The optimization sections claim that the same Wasserstein-based hierarchy supports variational and gradient-based fitting, and that a forward-plus-coalescent estimator improves learning of the fitness landscape in a small TSP test problem.","pith_inferences":["Beyond the paper, a reader could turn the conflict threshold into a model-selection test: fit the model with and without a cooperativity term and see whether the inferred $\\kappa$ crosses $\\kappa^*$, using fitness trajectories from real multilevel systems to estimate the number of active levels.","The same Wasserstein recursion could define a distance-to-data objective for tumor phylogenies, where clones at multiple scales play the role of the TSP genotypes; the paper notes this direction but does not test it.","The importance-sampling estimate in Eq. 32 would become a fully operational estimator only when the coalescent density $C(\\cdot|x_T^\\dagger_L,\\theta)$ is written down explicitly; supplying that density is the natural next step for making the combined objectives practical.","If the mixed estimator is validated, the monotone improvement guarantee of smoothing-based optimization suggests the forward-coalescent VO estimator should inherit a descent property in expectation, though the paper does not prove that for the mixed version."],"forward_implications":["At $\\kappa=0$ the model reproduces the classical guarantee that expected mean fitness cannot decrease, at every level simultaneously.","For $\\kappa>\\kappa^*$, selection at any level below the top can be antagonistic to mean fitness at that level, giving a concrete parameter-based definition of when levels of selection conflict.","Conflict at the top level is not generic: it depends on the population size and on the genotype fitness and distance parameters, so a fitted model must estimate these before predicting top-level behavior.","If the combined forward-coalescent objectives are valid, the paper's method offers a way to fit hierarchical genetic algorithms and multilevel evolutionary models from a single observed final population.","The framework treats the number of levels as a design choice, so the same sampler and objective can be used to compare models with different numbers of levels."],"supporting_citations":[{"why":"Supplies the Kantorovich monad and the recursive Wasserstein metric structure that define the state space.","marker":"[9]"},{"why":"Provides the stochastic Price equation of which the mixed multilevel Price equation is a specialization.","marker":"[16]"},{"why":"Supplies the recursive expansion of the stochastic Price equation used to situate Proposition 1.","marker":"[17]"},{"why":"Provides the smoothing-based optimization and SMO updates used by the variational optimizer.","marker":"[15]"},{"why":"Provides simultaneous perturbation stochastic approximation used for gradient-free parameter learning.","marker":"[18]"},{"why":"Provides the Monte-Carlo expectation-maximization scheme used as the third fitting method.","marker":"[19]"},{"why":"Motivates the coalescent analysis that the multilevel coalescent sampler adapts to backward simulation.","marker":"[11]"}],"fun_headline_variants":["Cooperativity threshold flips fitness gains at lower levels","Why strong cooperation can hurt lower-level fitness","Multilevel selection: a sharp threshold for fitness decline","Price equation meets Kantorovich: selection's cooperative flip"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The combined forward-coalescent training objectives rest on the assumption that the probability of a coalescent trajectory can be computed exactly under both the forward and coalescent processes; the paper states \"we assume that this can be done\" without giving the coalescent density $C(\\cdot|x_T^\\dagger_L,\\theta)$.","fun_headline_variants_meta":{"raw":{"variants":["Cooperativity threshold flips fitness gains at lower levels","Why strong cooperation can hurt lower-level fitness","Multilevel selection: a sharp threshold for fitness decline","Price equation meets Kantorovich: selection's cooperative flip"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000298,"raw_usage":{"total_tokens":1729,"prompt_tokens":953,"completion_tokens":776,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":714}},"tokens_in":569,"tokens_out":776,"duration_ms":6769,"temperature":1.0,"reasoning_tokens":714,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:19:46.655234+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Write down and normalize the multilevel coalescent density $C(\\cdot|x_T^\\dagger_L,\\theta)$ for a small case, then check the importance-sampling identity in Eq. 32 numerically; if the density is intractable or its normalizing constant is not 1, the forward-coalescent VO, SPSA, and MC-EM objectives are not well defined.","supporting_citations":[{"cited_title":"The metric monad for probabilistic nondeterminism","cited_arxiv_id":null,"evidence_quote":"Supplies the Kantorovich monad and the recursive Wasserstein metric structure that define the state space."},{"cited_title":"Evolutionary Theory: Mathematical and conceptual foundations","cited_arxiv_id":null,"evidence_quote":"Provides the stochastic Price equation of which the mixed multilevel Price equation is a specialization."},{"cited_title":"How Many Levels Are There? How Insights from Evolutionary Transitions in Individuality Help Measure the Hierarchical Complexity of Life","cited_arxiv_id":null,"evidence_quote":"Supplies the recursive expansion of the stochastic Price equation used to situate Proposition 1."},{"cited_title":"Smoothing-based optimization","cited_arxiv_id":null,"evidence_quote":"Provides the smoothing-based optimization and SMO updates used by the variational optimizer."},{"cited_title":"Stochastic optimization, stochastic approximation and simulated annealing","cited_arxiv_id":null,"evidence_quote":"Provides simultaneous perturbation stochastic approximation used for gradient-free parameter learning."},{"cited_title":"A review of Monte Carlo-based versions of the EM algorithm","cited_arxiv_id":"2401.00945","evidence_quote":"Provides the Monte-Carlo expectation-maximization scheme used as the third fitting method."},{"cited_title":"Inferring phylogenies","cited_arxiv_id":null,"evidence_quote":"Motivates the coalescent analysis that the multilevel coalescent sampler adapts to backward simulation."}],"review_version":1}