{"id":"157b6df6-bbfd-4ff7-b9a9-4a711deedd2f","arxiv_id":"2501.19113","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Genetic AI uses evolutionary game dynamics over data rows and columns to rank discrete solutions without training data or predefined weights.","lead":"This paper proposes an optimization method that turns data columns into genes and data rows into organisms, then runs an evolutionary game to rank which solutions and features matter most. The authors claim the method needs no training data or predefined weights, and demonstrate it on two small flight-choice examples.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The convergence of γ* to a meaningful optimum is never established, because the ESE is a fixed point of user-chosen update rules, no objective links it to optimization, and the unproved Δ-bound (Eq. 18) plus division by r_i (Eq. 31) make the dynamics ill-defined on the claimed domain.","rationale":"The reader's REJECT verdict is appropriate. The concrete strength of the paper is a fully specified update scheme and a self-consistent mixing idea, but these do not bridge the gap to 'optimization'. The abstract claims universality and absence of external parameters; however, φ_j, γ(0), strategy choice, α, and stopping iteration are external and influence the output. The formal dynamics are under-specified: Eq. (18) is an unproved convenience assumption, Eq. (31) contains a division by organism fitness that can be zero, there is no convergence proof, and the two examples are toy-sized with post hoc interpretation. The reader's Eq. (18) concern is real but narrower; the more load-bearing issue is the missing objective and the dependence of the result on permissible user choices. A strategy-dependence check on the simple example would settle whether any unique 'optimal solution' exists; the expected outcome based on the paper's own figures is that the ranking changes, which would directly refute the strongest abstract claims.","tokens_in":15705,"tokens_out":12598,"duration_ms":142110,"concrete_test":"Use the paper's simple flight data (Table I) and reproduce Algorithm 1 for the two pairs GS-Dominant+OS-Balanced and GS-Altruistic+OS-Selfish, both with the paper's uniform γ(0) and φ_inv. Compute the converged organism fitness r_i from Eq. (12) at k=30 in both cases. If the highest-r organism differs between the two strategy pairs, then the 'optimal solution' depends on an external choice and the central claim is not supported. As a second gate, record max_j|Δ_j| over all iterations: if it exceeds 1, the unproven Eq. (18) premise fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires the converged gene fitness γ* and the resulting organism ranking to be an optimization result obtained without external choices. That condition is not met. The paper never defines the objective being optimized or proves a correspondence between an ESE of Eqs. (19)-(20) and an optimum of the data; the examples are interpreted post hoc. The 'without external parameters' premise is directly contradicted by the setup: the gene-variant fitness functions φ_j (Sec. III B, Eqs. 6/36/37), the initial gene fitness γ(0) (Sec. IV A), the strategy pairing, and the mixing coefficients α (Sec. V E, Eqs. 33-35) are user choices that change the ranking, and Sec. V E 2-3 explicitly allows preference- and training-determined choices. For the simple flight example, the two strategy pairs in Fig. 2 produce qualitatively different gene-fitness vectors, so the 'optimal' outcome is strategy-dependent. Formally, Eq. (18) is assumed, not proven; and OS-Selfish in Eq. (31) divides by r_i, which is zero for any all-zero row, so the claimed applicability to 'all problems that can be formulated in matrix form' is not even well-defined. The reported ESEs are therefore properties of the chosen equations and data, not a demonstrated property of the data itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces 'Genetic AI', an evolutionary-game procedure for discrete matrix-form decision problems. Rows and columns of the input matrix are interpreted as organisms and genes, respectively; a normalized gene-fitness vector γ is updated by replicator equations using four hand-defined strategies (Dominant, Altruistic, Balanced, Selfish), and the organism fitness is the linear score r_i = ω_i·γ. The authors claim that the resulting evolutionary stable equilibrium identifies optimal solutions and feature relevance 'without external parameters', and they illustrate the method on a toy flight-selection problem and a ten-flight example, including a self-consistent scheme that adapts strategy-mixing weights. The paper contains concrete update equations and pseudocode, but the central optimality and parameter-free claims are not established.","tokens_in":16082,"tokens_out":6936,"duration_ms":67259,"significance":"If the central claim were true, this would be a notable contribution: a parameter-free, data-only method that simultaneously ranks discrete solutions and identifies feature relevance would be useful across many applied domains. The paper has some merits: the formal machinery is explicit, the four update rules are stated in enough detail to reproduce the two examples, and the self-consistent mixing extension is a genuine attempt to remove one layer of user choice. However, the manuscript provides no formal definition of optimality, no convergence theorem, no comparison against known optima, and no reproducibility artifact; the reported 'optimal' rankings change with strategy choices. The significance is therefore prospective rather than demonstrated.","major_comments":[{"comment":"The central claim requires that the ESE of Eqs. (19)–(20) is an optimization result, but no objective function is ever defined and no theorem states that a fixed point of the replicator dynamics corresponds to a minimum, maximum, or Pareto-optimal point of any data-derived criterion. The organism fitness in Eq. (12) is a linear weighted score, yet γ is updated by arbitrary rules, so the final γ⋆ is a fixed point of those rules, not a solution of an optimization problem. The paper's own Sec. VII leaves the question 'why does it work?' open; this is a load-bearing gap, not a cosmetic one.","section":"Secs. I, IV A, VII"},{"comment":"Eq. (18) assumes −1 ≤ Δ_j ≤ 1, but this bound is not proved for any of the four strategies or for the linear combinations in Eqs. (33)–(34). If Δ_j < −1, the update in Eq. (19) can produce non-positive gene fitness values before normalization; if Δ_j > 1, the update can overshoot and oscillate. The convergence plots in Sec. VI therefore rest on an unverified premise. A proof of the bound, or a numerical check of the bound for all rows and columns and all iterations, is required before the reported ESEs can be interpreted.","section":"Sec. IV A, Eq. (18)"},{"comment":"The claim of 'no external parameters' is contradicted by the design choices required by the method. The gene-variant fitness functions φ_j (Eqs. (6), (36), (37)), the initial gene fitness γ(0), the choice of the strategy pair (GS, OS), and the mixing coefficients in Eqs. (33)–(34) all affect the outcome. The paper itself states in Sec. III B that the choice of φ_j influences the simulation, and Sec. V E 2–3 explicitly reintroduces preferences and training. Even in the 'ab initio' route of Sec. VI C, the strategy sets S_g, S_ω and α(0)=0.5 are user inputs. The simple example in Sec. VI A gives different converged gene-fitness rankings for DomBal and AltSel (Fig. 2), so the claimed 'optimal' result is strategy-dependent.","section":"Secs. III B, V E, VI A"},{"comment":"Eq. (31) divides by r_i, and Eq. (23) also divides by r_i via µ_ij. For any row of Φ(Xp) that is entirely zero, r_i = 0 by Eq. (12), so OS-Selfish and OS-Balanced are undefined for that organism. The paper claims applicability to all matrix-form problems, but this domain excludes matrices with zero rows unless a special-case rule is supplied. The manuscript does not provide such a rule.","section":"Sec. V D, Eqs. (23) and (31)"},{"comment":"The convergence claims are heuristic. Sec. VI D asserts that 'it seems plausible' OS-Balanced prevents complete gene-fitness accumulation, and Sec. VII leaves the question 'why does it work?' open. No convergence theorem is given, and the two examples have no known ground-truth optimum against which the reported rankings are checked. As a result, the claimed universality of the method is not supported by the evidence within the manuscript.","section":"Secs. VI D and VII"}],"minor_comments":[{"comment":"The maximum in Eqs. (36) and (37) is written with the bound 1 ≤ ℓ ≤ 1; it should presumably be 1 ≤ ℓ ≤ n. As written, the formula is not a column-wise normalization.","section":"Sec. VI, Eqs. (36) and (37)"},{"comment":"The text says that 'time' and 'stops' are correlated with κg12 = 0.83, but the displayed matrix gives κg12 = 0.67 and κg23 = 0.83. The index is inconsistent with the matrix.","section":"Sec. VI A, Eq. (52)"},{"comment":"There are several typographical or typesetting issues: Eq. (12) contains a doubled comma, Eq. (72) is missing a closing bracket, and the sign tuples in Eqs. (51) and (55) are formatted inconsistently, for example '(−−, −)'.","section":"Throughout"},{"comment":"The footnote suggests ρ = max_r_t as an alternative when all initial organism fitness values are equal, but if all r_t are zero then ρ = 0 and the term (r_i − r_t)/ρ in Eq. (30) is still undefined.","section":"Sec. V D, footnote 20"},{"comment":"Reference [1] is incomplete: it lacks an author and a title, reading only 'Optimization algorithms (1970) pp. 13–31'.","section":"Bibliography, reference [1]"}],"recommendation":"reject","confidential_remarks":"The paper is a methods proposal without validation, and its central claims are much stronger than the evidence. The lack of a formal optimality notion, the unproved bound in Eq. (18), and the acknowledged dependence on user choices together mean that the method, as presented, does not deliver 'ab initio' optimization. A completely rewritten version that defines the optimization objective, provides convergence guarantees, and benchmarks against known optima could be considered in the future, but that is beyond a revision of the current manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: the paper is a new heuristic with a novel set of strategy equations and a self-consistent mixing loop, but the load-bearing claim—that it does optimization without external parameters—does not survive contact with the manuscript.\n\nWhat's actually new: Equations (21)-(31) define four strategies (Dominant, Altruistic, Balanced, Selfish) that update feature weights (gene fitness) via replicator dynamics, and Eqs. (58)-(61) let the mixing coefficients alpha themselves evolve by the same replicator rule. I don't recall that exact combination in the EGT/EMO literature. The paper is also honest in places: it admits the gene variant fitness function phi_j and initial gamma(0) are user choices, and that the strategy pairing changes the outcome (see Fig. 2, where DomBal and AltSel give different rankings for the flight example).\n\nThe soft spots are real and central. There is no formal definition of what \"optimal\" means for this method, and no theorem connecting an ESE of Eqs. (19)-(20) to any optimum of the input matrix. Eq. (18) is simply assumed (\"let us assume\") and never proven; if Delta_j leaves [-1,1] the replicator update can misbehave. OS-Selfish in Eq. (31) divides by r_i, which is zero for any all-zero row, so the claimed domain \"all problems that can be formulated in matrix form\" is not even well-defined. And the examples are toy-sized, with post hoc explanations of why a particular gene \"wins.\" The conclusion that converged gamma* encodes feature relevance is therefore an interpretation of hand-crafted equations, not a demonstrated property of the data.\n\nThat said, the paper is not a mess. It is readable, the analogies are clear, and the authors flag many limitations themselves (Sections V E 2-3 discuss predefined choices and training). The problem is the abstract and intro overpromise \"ab initio,\" \"universal,\" and \"without external parameters,\" which the body contradicts.\n\nWho is this for? Someone working on evolutionary heuristics or feature weighting might find the strategy zoo worth a look, but should not cite it as a proven method. It deserves peer review in the sense that a serious referee could help the authors either add a precise objective and convergence analysis, or reposition the paper as a heuristic proposal with benchmarks against existing adaptive-weight MOO methods. As is, I would not accept it as a finished contribution.\n\nRecommendation: send it to review only if the venue wants speculative new ideas; otherwise desk reject with a clear message about the missing formal foundation.","headline":"A clearly written new evolutionary heuristic whose central claim—that it solves multi-objective problems without external choices—is not supported by the paper's own formalism or examples.","tokens_in":16561,"tokens_out":2187,"would_cite":false,"duration_ms":22126,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A22","90C29"],"pacs":[],"model":"deepseek-v4-flash","headline":"Genetic AI can solve any discrete multi-objective problem that can be written as a matrix by simulating gene competition, with no preset weights or training data.","keywords":["Genetic AI","multi-objective optimization","evolutionary game theory","replicator equations","evolutionarily stable equilibrium","data-less training","feature relevance","decision problems"],"falsifier":"Run Genetic AI on a large random matrix whose entries are drawn independently from a fixed distribution and inspect the converged gene fitness $\\gamma$. An unstructured matrix should yield roughly uniform $\\gamma$ at the equilibrium; if $\\gamma$ is consistently skewed in a fixed direction across many seeds, the skew comes from the update equations rather than from the data. A second check is to test whether Eq. (18) is violated for any strategy pair on random inputs; if it is, the normalization step is masking an instability in the update that defines the method.","tokens_in":15491,"feed_emoji":"🧬","tokens_out":7640,"duration_ms":69565,"temperature":0.7,"pith_summary":"The paper introduces Genetic AI, a method that treats each column of an input matrix as a gene, each row as an organism, and runs an evolutionary game in which gene fitness values are updated by replicator equations until the system reaches an evolutionarily stable equilibrium. The central claim is that the converged gene fitness vector $\\gamma$ ranks the relevance of the data features, and the organism with the highest fitness at that equilibrium is the optimal solution. The intended payoff is multi-objective optimization without training data, predefined weights, or externally chosen rules. The paper argues this applies to any discrete problem that can be formulated in matrix form and demonstrates the claim on two flight-choice decision problems. If the claim is correct, the data's own internal structure, rather than a user's preferences, determines the objective weights and the winning solution.","feed_headline":"Evolutionary game ranks options with no preset weights","feed_subtitle":"A data matrix becomes a gene pool; the fittest organism is the optimal solution.","key_machinery":"The machinery is a replicator-equation update on the gene fitness vector $\\gamma$. Each iteration accumulates per-gene changes from a gene strategy and an organism strategy, $\\Delta_j = \\sum_i (\\Delta^g_{ij} + \\Delta^\\omega_{ij})$, and updates $\\tilde\\gamma_j^{(k+1)} = \\gamma_j^{(k)}(1+\\Delta_j)$, followed by normalization. The four strategies are the central objects: GS-Dominant rewards gene variants above 50 percent, OS-Balanced penalizes genes that dominate an organism's fitness, GS-Altruistic transfers fitness between kin genes based on similarity, and OS-Selfish shifts fitness between related organisms based on fitness gaps. A second replicator loop (Eqs. 58–61) adjusts the mixing coefficients $\\alpha$ from measured strategy effects, making the dynamics fully self-consistent. The paper's argument rests on the claim that this iterative process reaches an evolutionarily stable equilibrium in which $\\gamma$ reflects true feature relevance.","core_discovery":"The paper's central discovery is that a fixed data matrix can be turned into an evolutionary system whose equilibrium solves the optimization problem. In that system, data features are genes, data sets are organisms, and four strategies—Dominant, Altruistic, Balanced, Selfish—govern how fitness flows between them. The simulation iteratively updates the gene fitness vector $\\gamma$ and the organism fitness values until a nontrivial evolutionarily stable equilibrium is reached; at that point the converged $\\gamma$ is read as feature relevance and the fittest organism as the optimal solution. The paper also shows a self-consistent variant in which the mixing coefficients $\\alpha$ of the strategies are themselves updated by replicator equations, so the simulation chooses its own balance of strategies without external tuning.","pith_inferences":["Editorial inference: If $\\gamma$ truly encodes feature relevance, the method offers a new feature-selection heuristic for tabular data, one that can be benchmarked against established importance scores.","Editorial inference: The unproven bound in Eq. (18) suggests a stress test: run the simulation on random matrices and check whether the equilibrium stays nontrivial and whether $\\gamma$ stays uniform; if not, part of the ranking may be an artifact of the update rule.","Editorial inference: The four strategies can be treated as a basis set and the $\\alpha$-replicator loop run over a larger strategy library; the paper sketches this direction but does not test it.","Editorial inference: Because organisms are never mutated or replaced, the method is closer to a ranking algorithm than to a search-based optimizer; its evolution is the evolution of weights, not of candidate solutions. This may limit its scope to selection among given candidates."],"forward_implications":["Any discrete optimization problem representable as a matrix becomes solvable without labeled training data or user-chosen weights; the simulation itself sets the objective weights.","Because each data package $X_p$ is simulated independently, the optimal solution and feature rankings update dynamically when the input changes, with no retraining step.","The strategy pair GS-Dominant + OS-Balanced is claimed to expose symmetries in the data, while GS-Altruistic + OS-Selfish exposes correlations; a self-consistent mix yields a single ranking in one run.","The converged gene fitness vector $\\gamma$ provides an interpretable output: an ordering of feature relevance that explains why the winning solution wins.","The same machinery can, in principle, train an AI model without training data, since the input matrix supplies both the population and the fitness landscape, according to the paper's universality claim."],"supporting_citations":[{"why":"Supplies the replicator-equation and evolutionarily stable equilibrium apparatus the gene-fitness update is built on.","marker":"[8, 9]"},{"why":"Provides the gene's-eye view of organisms as replicator machines, which motivates mapping data features to genes and data sets to organisms.","marker":"[16]"},{"why":"Defines evolutionary multi-objective optimization as the baseline methodology Genetic AI contrasts with (discrete fixed population vs continuous solution generation).","marker":"[3]"},{"why":"Weighted-preference EMO variant used as the comparison point for Genetic AI's claim of dynamically adapting weights.","marker":"[4]"},{"why":"Supplies the self-consistency analogy (Kohn-Sham iteration) that motivates the alpha-mixing replicator loop.","marker":"[21]"},{"why":"Grounds the 'ab initio' claim by defining first-principles computation as requiring no external parameters.","marker":"[23, 24]"}],"fun_headline_variants":["Evolutionary game finds optimal solution with no preset weights","Self-consistent evolutionary game optimizes without tuning","Data matrix evolves: fittest organism is the optimal answer","Four strategies, one game: evolution picks the best option","ab initio evolution optimizes without prior rules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the accumulated update $\\Delta_j$ always satisfies $-1 \\le \\Delta_j \\le 1$ (Eq. 18); the paper introduces this with \"let us assume\" and does not prove it, so if the bound fails the gene fitness can leave its valid range and the reported equilibrium could be an artifact of the equation design rather than a property of the data.","fun_headline_variants_meta":{"raw":{"variants":["Evolutionary game finds optimal solution with no preset weights","Self-consistent evolutionary game optimizes without tuning","Data matrix evolves: fittest organism is the optimal answer","Four strategies, one game: evolution picks the best option","ab initio evolution optimizes without prior rules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00112,"raw_usage":{"total_tokens":4603,"prompt_tokens":827,"completion_tokens":3776,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":443,"completion_tokens_details":{"reasoning_tokens":3700}},"tokens_in":443,"tokens_out":3776,"duration_ms":22896,"temperature":1.0,"reasoning_tokens":3700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:15:47.615104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Genetic AI on a large random matrix whose entries are drawn independently from a fixed distribution and inspect the converged gene fitness $\\gamma$. An unstructured matrix should yield roughly uniform $\\gamma$ at the equilibrium; if $\\gamma$ is consistently skewed in a fixed direction across many seeds, the skew comes from the update equations rather than from the data. A second check is to test whether Eq. (18) is violated for any strategy pair on random inputs; if it is, the normalization step is masking an instability in the update that defines the method.","supporting_citations":[{"cited_title":"In this case it pays to use the training data to determine the optimal combination of evolution- ary strategies, but customize γ(0) according to e.g","cited_arxiv_id":null,"evidence_quote":"Defines evolutionary multi-objective optimization as the baseline methodology Genetic AI contrasts with (discrete fixed population vs continuous solution generation)."},{"cited_title":"Yang, Optimization algorithms (1970) pp","cited_arxiv_id":null,"evidence_quote":"Weighted-preference EMO variant used as the comparison point for Genetic AI's claim of dynamically adapting weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the self-consistency analogy (Kohn-Sham iteration) that motivates the alpha-mixing replicator loop."}],"review_version":1}