{"id":"75ab89e9-80f6-4c20-8b85-9ad09c3e5bb2","arxiv_id":"2607.16046","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CHILI models hypergraph growth as label-biased copying of earlier edges and yields an analytic degree power law plus stochastic-EM and simulated-annealing inference.","lead":"This paper introduces CHILI, a growing-hypergraph model in which new edges copy old edges, with copying and node-addition biased by binary node labels and homophily. It derives the model's limiting label structure and degree power law, and fits the model to real co-sponsorship, school-contact, coauthorship, and email data. Why read it: it offers a likelihood-based framework for measuring homophily in polyadic interactions and a proof-of-concept that modeling edge dependence c","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transition matrix T factorizes label counts after marginalizing over the focal label; the CHILI process has a single shared focal label, so T is not the true transition law and the stationary distribution/power-law exponent rest on an incorrect matrix.","rationale":"I stressed the central analytic derivation rather than the algorithmic sections. The reader's weakest assumption — degree–label independence in the power-law derivation — is explicitly flagged by the paper and is a mean-field approximation that could be tested; but before that assumption is even reached, the transition matrix used to define q(k0,k1) is mis-specified. The two label counts are not conditionally independent given only (k0,k1): they share the focal node's label. The paper's T multiplies the two marginal distributions, which is equivalent to drawing the focal label twice independently. This is not a regime-dependent subtlety; it changes the support of the stationary distribution in the zero-parameter limit. Because the moments ⟨k0⟩, ⟨k1⟩, μ00, μ11, μ01 in Eq. (2) are all taken with respect to the claimed stationary distribution of this incorrect T, the power-law exponent is not established even if the mean-field assumption held. The paper's own simulations in Fig. 2 might still appear consistent if the product T is close to the true T for the tested parameters, but the analytic claim is stated as exact. The SEM and simulated-annealing contributions do not depend directly on this particular T, and the released code is a positive feature, but the central theoretical claim of the paper needs correction. A small numerical check would settle the issue definitively, so I do not regard this as an unfalsifiable objection. Because the claimed closed-form signatures are not merely unproved but demonstrably computed from the wrong transition law, the appropriate disposition is reject as stated, pending a corrected transition matrix and re-derived exponent.","tokens_in":19999,"tokens_out":7228,"duration_ms":70152,"concrete_test":"Simulate the exact CHILI process (e.g. with the authors' released code) at θ=(0,0,0,0,0,0) for 10^5 steps and record the empirical joint distribution q(k0,k1) of label counts in uniformly sampled edges. Separately compute the leading eigenvector of the product T defined in §II A 1, truncating at a sufficiently large edge-size bound. If the empirical distribution places all mass on (1,0) and (0,1), while the T eigenvector has positive mass on (1,1) or (0,0), the claimed stationarity relation fails. A second, less degenerate check: for a parameter set such as the Fig. 2 values (ρ+=0.8, ρ−=0.1, γ+=0.5, γ−=0.2, η+=0.4, η−=0.2), compare the stationary eigenvector of the exact joint transition above with that of the product T; any nonzero L1 difference invalidates Eq. (2).","verdict_should_be":"REJECT","load_bearing_attack":"In §II A 1 the paper defines t(k'_0|k0,k1) as the probability of observing k'_0 label-0 nodes conditional only on (k0,k1), i.e. after marginalizing over the focal node's label, and similarly t(k'_1|...). It then sets T(k'_0,k'_1|k0,k1) = t(k'_0|k0,k1) t(k'_1|k0,k1). But in the generative process, the focal node's label is a single latent variable that couples the two label counts. Given seed counts (k0,k1), the exact transition law is\n\nP(k'_0,k'_1|k0,k1) = (k0/(k0+k1)) A0(k'_0) B1(k'_1) + (k1/(k0+k1)) A1(k'_0) B0(k'_1),\n\nwhere A0 = 1 + Binom(k0−1, ρ+) + Pois(γ++η+), A1 = Binom(k0, ρ−) + Pois(γ−+η−), B0 = Binom(k1, ρ−) + Pois(γ−+η−), and B1 = 1 + Binom(k1−1, ρ+) + Pois(γ++η+). The product-of-marginals T includes cross terms (focal label 0 for k'_0 together with focal label 1 for k'_1) that the actual process never realizes. This is not a small approximation: when ρ=γ=η=0, every edge has size exactly 1, yet T assigns positive probability to states (1,1) and (0,0). Therefore the leading eigenvector of T is not the stationary joint distribution of label counts in edges, and the moments μij entering Eq. (2) are not the CHILI moments. The paper's central analytical signatures are unsupported; this error precedes the mean-field assumption flagged in Appendix A.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CHILI, a generative model for growing labeled hypergraphs in which each new edge is a noisy copy of a randomly sampled previous edge, with label-dependent probabilities for copying, extant-node addition, and novel-node addition. The authors derive two analytic signatures: the stationary joint distribution q(k0,k1) of label counts in a uniformly random edge is claimed to be the leading eigenvector of a transition matrix T (Section II A 1), and the asymptotic degree distribution is claimed to be a power law with exponent ζ = 1 + (⟨k0⟩+⟨k1⟩)/(1+ρ+(μ00+μ11−1)+2ρ−μ01) (Eq. 2, Appendix A). For inference, the paper proposes stochastic expectation-maximization for parameter estimation and a simulated-annealing algorithm for community detection, with experiments on synthetic and six empirical hypergraphs. The central difficulty is that the factorized transition matrix T is not the exact transition law of the generative process because the focal node's label couples the two count coordinates; the claimed stationary distribution and all moments entering the degree exponent are therefore computed from the wrong object.","tokens_in":20504,"tokens_out":3350,"duration_ms":30868,"significance":"If the analytic derivations were correct, the paper would provide a valuable bridge between edge-dependent generative hypergraph models and statistically principled inference, with a closed-form degree exponent and a tractable stationary distribution that could guide model interpretation. The SEM algorithm is clearly specified and the empirical case studies (e.g., the coauthorship gender interpretation, Enron core-fringe disassortativity) are plausible and interesting. The public code and reproducible experimental setup are strengths. However, the analytic centerpiece is not the model's true stationary distribution, which undermines both the theoretical claims and the interpretive use of the moments in Eq. (2). The simulation and inference contributions are partially independent of that derivation, but the paper's stated central claims rest on the incorrect transition matrix.","major_comments":[{"comment":"The transition matrix T is defined as a product of marginals after marginalizing over the focal node's label: T(k'_0,k'_1|k0,k1) = t(k'_0|k0,k1)t(k'_1|k0,k1). This is not the transition law of CHILI. Given seed counts (k0,k1), the focal node's label is a single latent variable; the exact kernel mixes the two labels before drawing the two count coordinates. The product T includes unphysical cross terms such as focal label 0 for k'_0 simultaneously with focal label 1 for k'_1. Example: with ρ=γ=η=0, every edge has size exactly 1, yet T gives positive probability to (1,1) and (0,0). Thus the Perron eigenvector of T is not q(k0,k1), and the moments μij used in Eq. (2) and Appendix A are not the CHILI moments. The central analytic claims are unsupported as stated.","section":"II A 1, Eqs. (3)-(5)"},{"comment":"The degree power-law derivation explicitly assumes independence of the degree d of a node and the label ratio k0/k in the sampled seed edge. This is acknowledged as a mean-field assumption but is neither tested numerically nor qualified with a parameter regime. Given that the degree exponent in Eq. (2) is sold as a closed-form signature, some validation (e.g., comparing the theoretical exponent to simulated degree distributions across a parameter grid, not just one curve in Fig. 7) is needed. Moreover, this issue is downstream of the incorrect stationary distribution; even if the mean-field assumption held, the moments μij would still be computed from the wrong T.","section":"Appendix A, Eq. (A8)"},{"comment":"No conditions are stated for the existence of a Perron eigenvector with eigenvalue 1 or for convergence of q_t to q. The matrix T is substochastic for some parameters because an edge copy can fail and the available-node caps can alter the Poisson rates; the paper does not discuss whether T is stochastic, primitive, or even nonnegative with spectral radius 1 across the parameter range. The spectral plots in Fig. 2 assume these properties. This is secondary to the incorrect factorization, but it is load-bearing for the claimed stationary distribution and for the use of λ2 as a convergence rate.","section":"II A 1, stationarity and Perron eigenvector"},{"comment":"The empirical community detection comparison against reported values from prior work is not apples-to-apples: the authors use a fixed parameter vector with zero novel-node addition and an approximate likelihood (Appendix D), and they concede that belief-propagation spectral clustering achieves higher ARI on senate-bills than CHILI. The claim that CHILI 'outperforms' competitors should be framed as proof-of-concept; this is consistent with the paper's own wording but the abstract and introduction overstate it.","section":"IV C, Fig. 6"}],"minor_comments":[{"comment":"The relation between the exponent ζ and the slope of the degree density or complementary CDF in Fig. 7 should be stated explicitly; the current phrasing invites a factor-of-one confusion.","section":"Eq. (2)"},{"comment":"The entrywise division s′ ← x ⊘ y is unclear when y contains zero entries; a brief statement on numerical handling (e.g., zero-belief edges) would improve reproducibility.","section":"Algorithm 1"},{"comment":"The intersection-based approximation is justified for large n, but several empirical data sets have m much larger than n (Table I). A small sensitivity analysis on one empirical data set would help gauge the approximation error.","section":"Appendix D"},{"comment":"The caption contains a likely typo: 'γ− = η−3' should probably read 'γ− = η− = 3'. Please clarify.","section":"Fig. 7 caption"},{"comment":"The abstract overstates the derivations by not flagging the mean-field assumption or the likelihood approximation used in community detection.","section":"Abstract / Introduction"}],"recommendation":"major_revision","confidential_remarks":"The central analytic claim is internally inconsistent: the stationary distribution used in the paper is for a factorized process that does not match CHILI, and the error is visible in a trivial limiting case. This is a load-bearing defect that cannot be fixed by presentation changes alone; the authors need to either derive the exact transition kernel and its stationary distribution, or reframe the paper around the SEM and empirical contributions, which are partially independent of the incorrect derivation. If the exact kernel is intractable, the authors should say so explicitly and provide numerical evidence that their factorized approximation is accurate in the regimes used. The manuscript is otherwise well organized and the inference experiments are a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a load-bearing flaw in its asymptotic analysis. In Section II A 1, the transition matrix T is defined as the product of two marginal distributions: t(k'_0|k0,k1) t(k'_1|k0,k1). But the generative process has a single focal node whose label couples the two counts. The exact transition is a mixture of two product distributions — focal label 0 with probability k0/(k0+k1), focal label 1 with probability k1/(k0+k1). The product form includes cross terms that the process never generates. The extreme case makes the problem obvious: when all six parameters are zero, every edge has size exactly 1, so the only possible transitions from (k0,k1) are to (1,0) and (0,1). Yet T assigns positive probability to (0,0) and (1,1). Therefore the leading eigenvector of T is not the stationary joint distribution of label counts, and the moments μij used in the power-law exponent (Eq. 2) are not the model's moments. The Appendix A mean-field assumption is a separate, secondary issue; this factorization error precedes it.\n\nWhat the paper does well: the CHILI model itself is clearly specified and is a natural extension of the hyperedge copy model with label-dependent parameters. The SEM-based parameter inference is a sensible and useful contribution, and the empirical results on legislative cosponsorship and other datasets are interesting, even if they are interpreted qualitatively. The community detection experiments are honest about computational cost and mixed performance. The paper is also transparent about its limitations, including the mean-field assumption and the likelihood approximation used for simulated annealing.\n\nThe soft spots beyond the transition matrix error: SEM consistency is claimed but not proved; there is no uncertainty quantification on the estimated parameters; and the stationarity conditions for the count process are not stated. These are addressable. The transition matrix error is not — it is a mathematical mistake that invalidates the paper's central analytical results as stated.\n\nThis paper deserves a serious referee but not acceptance in its current form. The model and inference are salvageable, and the empirical parts are worth preserving. The authors should either correct the transition matrix derivation and redo the asymptotic analysis, or reframe the spectral and power-law results as approximate statements verified by simulation. I would not cite the current analytic results, but I would be interested in reading a revision.\n\nRecommendation: send to peer review, with a referee who checks the transition matrix carefully. The paper is substantial enough to warrant the effort, but the current central claims should not stand.","headline":"The CHILI model and inference pipeline are worth engaging, but the central analytic claims rest on a transition matrix that is not the actual transition law of the model.","tokens_in":20933,"tokens_out":2688,"would_cite":false,"duration_ms":29123,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C65","05C82","60J20","62F10","91D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A label-aware copying model for growing hypergraphs predicts a power-law degree distribution with an exponent computable in closed form from six parameters.","keywords":["hypergraph growth","homophily","edge copying","power-law degree distribution","stationary label distribution","stochastic expectation maximization","community detection","generative model"],"falsifier":"Simulate the model with parameters that create strong degree-label correlation (e.g., high same-label copying probability and a large imbalance in novel-node addition rates), fit the tail of the empirical degree distribution over millions of edges, and compare the fitted exponent to the closed-form value from the paper's equation; a systematic mismatch would falsify the mean-field assumption. A second check: diagonalize the transition matrix over the full parameter grid to confirm that the leading eigenvalue is 1 and simple; if it is not in some regime, the Perron-eigenvector statement for the","tokens_in":19900,"feed_emoji":"📈","tokens_out":5020,"duration_ms":48301,"temperature":0.7,"pith_summary":"The paper introduces CHILI, a generative model in which each new hyperedge is a noisy copy of an earlier edge, with the copying and node-addition steps biased by binary node labels. It derives that, in the long-time limit, the distribution of label counts inside a uniformly random edge is the leading eigenvector of a transition matrix, and that the degree distribution is a power law whose exponent is a closed-form function of six parameters. On the inference side, it shows the model defines a likelihood over labeled hypergraphs, supports a stochastic expectation-maximization algorithm for parameter recovery, and uses simulated annealing on an approximate likelihood for community detection. The payoff is a framework that relaxes the edge-independence assumption in hypergraph models while keeping tractable asymptotics.","feed_headline":"Hypergraph growth with homophily has a closed-form degree exponent","feed_subtitle":"A copy-and-add model predicts the long-run label mix as an eigenvector and the degree tail as a power law — testable on real data.","key_machinery":"The central object is the transition matrix T(k0′,k1′|k0,k1), the probability that a copied edge with k0 label-0 and k1 label-1 nodes yields a new edge with k0′ and k1′ labels. Conditional on the focal node's label, the counts of the two labels evolve independently as a binomial (copy step) plus a Poisson (addition step), so each entry factors into two tractable convolutions. The stationary label-count distribution is the leading eigenvector of T, and its moments are fed into a mean-field balance equation that produces the degree power law.","core_discovery":"At the heart of the paper is the claim that a simple mechanistic rule — every new edge copies an old one, with label-aware probabilities for copying, adding extant nodes, and adding novel nodes — is enough to reproduce two signature properties of real-world hypergraphs: a heavy-tailed degree distribution and tunable assortative mixing. The stationary joint distribution q(k0,k1) of label counts in a random edge is shown to be the Perron eigenvector of a doubly-indexed matrix T whose entries are binomial-Poisson convolutions. From q, the paper computes the moments that enter the power-law exponent ζ = 1 + (⟨k0⟩+⟨k1⟩)/(1+ρ+(µ00+µ11−1)+2ρ−µ01), so both the label-mixing asymptotics and the degree","pith_inferences":["A testable diagnostic follows: measure the degree-tail exponent of a real hypergraph and invert the closed-form formula to infer an effective homophily strength, without fitting the full generative process.","The same eigenvector machinery plausibly extends to k-ary labels, turning the transition matrix into a higher-order tensor; the corresponding Perron eigenvector could support a spectral algorithm for multiway community detection.","Because the mean-field independence assumption is untested, a corrected joint degree-label balance equation would likely be needed for regimes where hubs concentrate in one label; this is a natural next step, not taken in the paper.","The transient regime dominated by the antisymmetric second eigenvector suggests the model can mimic systems with unequal group sizes even though the stationary state is symmetric — a feature worth exploiting in generative data augmentation."],"forward_implications":["If the eigenvector and power-law claims hold, the model gives a closed-form benchmark: simulators and empirical data sets can be checked against the predicted exponent with no free parameters.","The likelihood formulation enables parameter inference via stochastic expectation-maximization with per-step cost linear in the number of edges, making the six homophily parameters estimable on real hypergraphs.","Community detection by simulated annealing on the approximate CHILI likelihood can outperform edge-independent baselines (spectral clustering, modularity maximization) on data sets like legislative cosponsorship networks, showing that edge dependence is a recoverable signal.","The spectral decomposition of the transition matrix predicts slow transients along the second eigenvector when same-label copying is strong, offering a natural explanation for persistent label imbalance in finite observation windows."],"fun_headline_variants":["Hypergraph growth: homophily yields exact power law","Copy-and-add model gives closed-form degree exponent","Homophily in hypergraphs: exact degree tail exponent","Edge copying with homophily: new degree power law","Tunable assortative hypergraphs via copying model"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The power-law degree exponent rests on the mean-field assumption that a node's degree is independent of the ratio of label-0 to label-1 nodes in the edge that copied it (Appendix A, Eq. A8), an assumption the paper states but does not validate.","fun_headline_variants_meta":{"raw":{"variants":["Hypergraph growth: homophily yields exact power law","Copy-and-add model gives closed-form degree exponent","Homophily in hypergraphs: exact degree tail exponent","Edge copying with homophily: new degree power law","Tunable assortative hypergraphs via copying model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000568,"raw_usage":{"total_tokens":2556,"prompt_tokens":805,"completion_tokens":1751,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":1673}},"tokens_in":549,"tokens_out":1751,"duration_ms":14164,"temperature":1.0,"reasoning_tokens":1673,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:30:39.703958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the model with parameters that create strong degree-label correlation (e.g., high same-label copying probability and a large imbalance in novel-node addition rates), fit the tail of the empirical degree distribution over millions of edges, and compare the fitted exponent to the closed-form value from the paper's equation; a systematic mismatch would falsify the mean-field assumption. A second check: diagonalize the transition matrix over the full parameter grid to confirm that the leading eigenvalue is 1 and simple; if it is not in some regime, the Perron-eigenvector statement for the","supporting_citations":[],"review_version":1}