{"id":"2be169bd-b31f-4630-8626-c111f761f182","arxiv_id":"2505.12276","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A hypergraph Ricci flow is defined, proven to have a unique solution for all time, and used in a community detection method that scores well on several benchmarks.","lead":"Hypergraphs are networks where one connection can link more than two nodes, and this paper adds a geometric process that changes how strongly each connection counts. The authors use this process to find groups of nodes and report good results on test networks, but the method uses the true group labels when choosing its final output.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 chooses its cutoff by maximizing NMI against ground truth, so the reported empirical claims are oracle-fitted; this is the load-bearing weakness, not the ODE proof.","rationale":"The reader's weakest assumption identifies the oracle cutoff and the unproven community-separation mechanism; my reading agrees. The theorem proof has fixable constant errors (the exponent in (13) should use the maximum hyperedge size, and the nm factor in (12) is not the sharp constant), but the long-time existence conclusion survives once those are corrected, so the theory is not the main threat to the paper's central claim. The empirical claim, however, is not supported as reported because Algorithm 1 selects its cutoff by evaluating accuracy against ground truth, which is unavailable in unsupervised community detection. The paper also imports baseline scores from prior papers, so the comparisons are not controlled. These issues justify the reader's CONDITIONAL verdict: the applied contribution needs re-evaluation with a label-free cutoff and rerun baselines. I do not move the verdict because the theoretical contribution appears salvageable and the empirical gap is fixable with additional experiments.","tokens_in":15161,"tokens_out":8326,"duration_ms":87731,"concrete_test":"Use the released code (https://github.com/mjc191812/...) to re-run Algorithm 1 with the cutoff selected without labels: for each k, compute the candidate partitions for all cutoffs w'_max..w'_min, then choose the partition maximizing hypergraph modularity instead of NMI; report NMI for the D1–D3 series and for Table III. Also rerun Louvain, Girvan–Newman, LPA, Infomap, and Walktrap on clique-expanded inputs rather than importing scores. If the modularity-selected NMI falls materially below the reported oracle NMI (e.g., >0.1 on most datasets) or does not beat the rerun baselines, the robustness claim is not established for unsupervised use.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—'remarkable enhanced robustness to topological variations and competitive performance'—rests on Table III and Figures 1–3, all produced by Algorithm 1. In Algorithm 1, after each flow step the loop `for cutoff = w'_max, ..., w'_min` removes hyperedges with `w' > cutoff`, then `Calculate[s] the accuracy of community detection`; the output is the partition with the best accuracy (NMI) against ground truth. That is an oracle sweep over the threshold, so the reported NMI measures the best achievable value of a one-parameter family, not the quality of an unsupervised cutoff rule. The paper supplies no evidence that inter-community hyperedges end up with larger weights than intra-community hyperedges after the flow; the Section IV intuition ('edges with negative curvature are stretched... remove edges with large weights') is asserted, not proven. Without a label-free cutoff rule, the reported results cannot establish that Ricci flow separates hypergraph communities. In addition, baseline scores are imported from Lee and Shin [57] and Hacquard [46] rather than rerun on the same code paths, so the comparisons are not controlled. The theorem itself is less at risk: the long-time existence argument is standard once the constant in (13) is corrected to use the maximum hyperedge cardinality and the factor nm is replaced by the appropriate finite constant.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines an Ollivier-type Ricci curvature for weighted hypergraphs and proposes a Ricci flow that evolves hyperedge weights over time, proving long-time existence of solutions to the ODE system (Theorem III.2). It then introduces HyperRCD, a community detection algorithm that discretizes the flow and removes hyperedges with large weights after evolution. Experiments on synthetic DCSBM hypergraphs and seven real-world datasets claim that HyperRCD is robust to topological variations and competitive with or better than graph- and hypergraph-based baselines.","tokens_in":15418,"tokens_out":5982,"duration_ms":63061,"significance":"If the theoretical and empirical claims hold, the paper would provide a geometrically motivated hypergraph community detection method that avoids lossy clique or star expansion, with a rigorous existence theorem as a foundation. The flow construction is clearly defined, and the ODE existence argument is a standard finite-dimensional Lipschitz/ODE argument once the identified exponent issue is corrected. The availability of source code is a positive feature. However, the empirical evaluation is currently not sufficient to support the central performance claims because Algorithm 1 selects its cutoff using ground-truth labels, and baseline scores are imported from prior papers rather than rerun under controlled conditions. The community-separation property that underlies the algorithm is asserted rather than demonstrated. With a label-free cutoff rule and corrected experiments, the contribution could be valuable, but the manuscript as written overstates its empirical support.","major_comments":[{"comment":"Algorithm 1's inner loop selects the cutoff by maximizing community detection accuracy against ground-truth labels and then returns the best partition, so every reported NMI is the maximum over a one-parameter family whose parameter is chosen using the target labels. This is an oracle evaluation, not a test of an unsupervised community-detection rule. The paper should either define and evaluate a label-free cutoff rule (e.g., a fixed weight percentile or a data-driven threshold not involving the labels) or explicitly report the full cutoff-NMI curve and state that the selected value is an upper bound, not a prediction. As written, the central empirical claim in the Abstract and Conclusion is unsupported.","section":"§IV, Algorithm 1; §V.C, Table III and Figures 1–3"},{"comment":"The community-separation mechanism is asserted rather than established. The flow guarantees long-time existence, but nothing in Theorem III.2 implies that inter-community hyperedges end with larger weights than intra-community hyperedges, which is the premise of the 'remove edges with large weights' rule. The paper should add either a theorem or an empirical diagnostic on synthetic data showing that the weight distributions separate by ground-truth community membership after the flow. Without such evidence, the design principle of the algorithm is a conjecture.","section":"§IV, paragraph before Algorithm 1; §VI"},{"comment":"The lower bound in Eq. (10) uses exponent s(s−1)/2 for the specific hyperedge's cardinality, but to bound ψ(t) = min_l w_{h_l}(t) uniformly the exponent must be taken as max_l |h_l|(|h_l|−1)/2. The parenthetical statement that s here refers to the vertex count of the minimally weighted hyperedge does not resolve the non-uniformity, since the bound for each h_l is valid only with that hyperedge's own s. Additionally, the upper bound in Eqs. (11)–(12) uses an underspecified constant nm; it should be stated explicitly, for instance as E = Σ_l C(|h_l|, 2). These are local fixes that preserve the theorem's conclusion.","section":"§III, Theorem III.2 proof, Eqs. (10)–(13)"},{"comment":"All baseline scores are taken from Lee and Shin [57] and Hacquard [46] rather than rerun in the authors' environment. Different hardware, software, hyperparameter settings, and preprocessing can materially affect NMI values, so the comparisons in Table III and Figures 1–3 are not controlled. The authors should rerun the baseline methods on the same clique-expanded hypergraphs or hypergraph representations and report variance, or clearly present the imported scores as approximate and justify why the comparison is still valid.","section":"§V.B, paragraph 'Implementation Details'"}],"minor_comments":[{"comment":"The loop 'for cutoff = w'_max, ..., w'_min' should specify whether the sweep is over all distinct weight values, a fixed grid, or a continuous interval; otherwise the stopping criterion and computational cost are ambiguous.","section":"§IV, Algorithm 1"},{"comment":"The phrase 'If k = 2' should be 'If |h| = 2', since k is not otherwise defined in that remark and the condition concerns the cardinality of a hyperedge.","section":"§II, Remark II.2(i)"},{"comment":"The hyperedge is written h_l = {x_l1, x_l2, ..., x_lk}, which reuses k from the iteration counter; use a separate symbol such as s or r for the hyperedge cardinality.","section":"§IV, Eq. (14)"},{"comment":"Algorithm 1 says 'Calculate the accuracy of community detection' while the evaluation section defines NMI as the metric; replace 'accuracy' with 'NMI' or explicitly state that the two are used interchangeably.","section":"Algorithm 1 and §V.B"},{"comment":"Please describe how the DCSBM is used to generate hyperedges, since 'intra-community connectivity strength' alone does not specify the higher-order sampling process; this is important for reproducibility beyond the code link.","section":"§V.A"},{"comment":"There are minor typographical issues such as 'intra-comminity' in §V.C.2 and inconsistent spacing in some equations; a careful proofreading pass would improve readability.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the empirical protocol: the oracle cutoff sweep in Algorithm 1 invalidates the reported NMI numbers as unsupervised performance. The theorem proof is fixable with a uniform hyperedge-cardinality exponent, and the baseline-provenance issue can be addressed by rerunning baselines. I recommend requiring the authors to remove the oracle sweep or disclose it explicitly, add a label-free cutoff rule, and provide either a separation proof or a diagnostic showing that the flow actually separates inter- and intra-community hyperedge weights. If those changes are made, the paper may become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The theoretical core is genuinely new: the flow (7) acts directly on hyperedge weights rather than on a clique-expanded graph, and Theorem III.2 proves long-time existence for positive initial weights. The proof is a fairly standard ODE Lipschitz argument, and aside from a notational slip it essentially works. The experimental section, though, does not yet support the headline claims. Algorithm 1 sweeps the cutoff and selects the partition with the best NMI against ground-truth labels, so the reported numbers are the best of a one-parameter family, not the output of a label-free rule. That makes the robustness claims oracle-fitted.\n\nWhat the paper does well: it gives a clean curvature definition for hyperedges, a natural evolution equation, and an explicit complexity analysis. It honestly reports that Mushroom exceeds the time limit. The idea of avoiding clique expansion is worthwhile, and the authors are careful about the local Lipschitz estimates that underpin the existence theorem.\n\nThe weak spot is load-bearing. There is no argument, and no evidence beyond assertion, that after the flow inter-community hyperedges systematically carry larger weights than intra-community ones, which is what the removal rule needs. Section IV simply states that negative curvature stretches inter-community edges. Without a proof or a systematic empirical check, the algorithm is incomplete as an unsupervised method. Also, baseline scores are imported from Lee and Shin and from Hacquard rather than rerun under the same code paths, so the comparisons are not controlled.\n\nOn the proof itself: inequality (10) has each hyperedge with its own cardinality s, and moving to (13) uses a single s for the lower bound. That is fixable by replacing s with the maximum hyperedge cardinality. The same paragraph contains a confusing self-referential note about what s means; a referee should ask for a clean rewrite. The paper also imports two technical lemmas from the authors' own prior papers [47,48]. That is not damning, but those results should be checked, especially Remark II.2(ii), since the local Lipschitz continuity of the probability measures is doing real work.\n\nWho is this for? People working on discrete curvature flows for hypergraphs will find the theory section useful, and the community detection angle is interesting despite the current problems. I would not cite the empirical numbers as evidence. A serious editor should send this to peer review: the theoretical contribution is enough to warrant referee time, and the empirical issues are identifiable and fixable. The revision needs a label-free cutoff rule, rerun baselines, and either a separation theorem or a careful experiment showing that the flow actually separates communities.\n\nMy recommendation: engage with it, but treat the experimental claims as unvalidated until the oracle sweep is removed.","headline":"Genuinely new hyperedge-level Ricci flow with a fixable proof issue, but the community detection results are oracle-fitted because Algorithm 1 tunes its cutoff against ground truth; the theory deserves referee time, the experiments need rework.","tokens_in":15974,"tokens_out":2238,"would_cite":true,"duration_ms":25586,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C65","53E20","68R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that a Ricci flow on hyperedge weights has a unique solution for all time, and that its discretization, HyperRCD, detects communities in hypergraphs.","keywords":["hypergraph","Ricci flow","Ollivier-Ricci curvature","community detection","higher-order interactions","Wasserstein distance","long-time existence","hyperedge weights"],"falsifier":"Run HyperRCD on a degree-corrected stochastic block model hypergraph with two planted 50-node communities; if after the flow the heaviest hyperedges mostly lie inside communities rather than between them, the weight-separation premise fails.","tokens_in":14939,"feed_emoji":"🕸️","tokens_out":16449,"duration_ms":140702,"temperature":0.7,"pith_summary":"The paper tries to establish that a Ricci-flow equation defined directly on hyperedges—not on a graph approximation of the hypergraph—is mathematically well-posed and practically useful for community detection. It proves that the hyperedge weights solving the flow exist, are unique, and remain bounded away from zero and infinity for all $t \\geq 0$, and it packages a discrete version of the flow into an algorithm named HyperRCD. If the claim is right, community structure is revealed by the flow's own dynamics: hyperedges with negative curvature are stretched, and deleting the stretched, large-weight hyperedges separates communities. The experiments on synthetic and real hypergraphs are offered as evidence that this geometric prescription detects higher-order community structure without first flattening hyperedges into pairwise edges.","feed_headline":"Hypergraph Ricci flow is well-defined for all time","feed_subtitle":"The flow stretches negative-curvature hyperedges, so cutting the heaviest edges after evolution finds communities.","key_machinery":"The central object is the Ollivier–Ricci curvature of a hyperedge, $\\kappa_\\alpha(h_l) = 1 - W_{h_l}/d(h_l)$, with $W_{h_l}$ the sum of the $1$-Wasserstein distances between the probability measures $\\mu^\\alpha_{x_{l_i}}$ carried by the vertices of $h_l$ and $d(h_l)$ the hyperedge length. The flow $w'_{h_l} = -d(h_l)\\,\\kappa_\\alpha(h_l)$ turns this curvature into a weight update: negative curvature stretches a hyperedge, positive curvature shrinks it. The existence proof is carried by a local Lipschitz estimate for the map $w \\mapsto W_{h_l}(w) - d(h_l)(w)$, obtained through the optimal-transport duality of Proposition II.3, together with exponential bounds that prevent any weight from collapsing to zero or blowing up. These ingredients make the flow a well-defined dynamical system on the space of positive hyperedge weights.","core_discovery":"The paper's central claim is Theorem III.2: for any connected weighted hypergraph with initial weights $w_0$, the flow $$\\frac{d}{dt} w_{h_l}(t) = -d(h_l)\\,\\kappa_\\$\\alpha$(h_l) = W_{h_l}(w(t)) - d(h_l)(w(t))$$ has a unique solution for all $t\\in[0,\\infty)$, where $W_{h_l}$ is the sum of the $1$-Wasserstein distances between the vertex measures on $h_l$ and $d(h_l)$ is the hyperedge length. The proof bounds each weight below by exponential decay and bounds the total weight above by exponential growth, so the solution cannot hit zero or diverge in finite time. The authors then discretize the flow with step size $\\eta = 0.1$ and use it to define HyperRCD, which removes hyperedges whose evolved weights exceed a cutoff and outputs the resulting partition. On degree-corrected stochastic block model hypergraphs the method maintains high normalized mutual information as network size, degree, and community ambiguity vary, and on seven real-world datasets it beats graph-embedding baselines, reaching the best NMI on Zoo and NTU2012.","pith_inferences":["Editorial inference: An unsupervised variant could choose the deletion cutoff at the largest gap in the evolved weight distribution, removing the current reliance on ground-truth labels.","Editorial inference: The same ODE argument would extend to directed hypergraphs if the distance and measure definitions are made directed and remain Lipschitz in the weights; the paper does not prove this.","Editorial inference: The observed insensitivity to average degree suggests the flow regularizes hyperedge weights geometrically; comparing HyperRCD with spectral or modularity-based edge reweighting on the same synthetic series would test that mechanism directly."],"forward_implications":["The flow can be run for arbitrarily many discrete steps without degenerating: the continuous uniqueness and global-existence result rules out weight collapse or blow-up.","The discrete update with step size $\\eta = 0.1$ gives a concrete clustering rule: evolve weights, delete hyperedges heavier than a cutoff, and read off connected components.","On the synthetic series, the method's NMI stays above 0.6 up to 1000 nodes while several compared methods fall below 0.2, and it keeps high NMI when intra-community connectivity is weak.","On seven real-world benchmarks the method beats graph-embedding baselines on every dataset and reaches the best NMI on Zoo (0.981) and NTU2012 (0.836).","The $O(ED^3)$ cost, dominated by solving linear programs for Wasserstein distances, means the method does not finish within 48 hours on the large-cardinality Mushroom hypergraph."],"supporting_citations":[{"why":"Supplies the curvature and Wasserstein transport foundation that the hyperedge curvature generalizes.","marker":"[4]"},{"why":"Supplies the graph Ricci-flow community detection method that motivates HyperRCD and the alpha = 0.5 parameter choice.","marker":"[34]"},{"why":"Supplies the curvature-based baseline and the edge-transport perspective on hypergraph clustering, plus the experimental protocol.","marker":"[46]"},{"why":"Supplies the lemma that distances between vertices are locally Lipschitz in the hyperedge weights.","marker":"[47]"},{"why":"Supplies the local Lipschitz continuity of the vertex probability measures in the weights, used in the proof of Lemma III.1.","marker":"[48]"},{"why":"Supplies the optimal-transport duality used in the proof to bound differences of Wasserstein distances.","marker":"[49, 50]"},{"why":"Supplies the degree-corrected stochastic block model used to generate synthetic hypergraphs with planted community structure.","marker":"[51]"},{"why":"Supplies the dataset and evaluation protocol whose baseline scores are reused for comparison.","marker":"[57]"}],"fun_headline_variants":["Hypergraph Ricci flow: existence for all time enables community detection","HyperRCD: Ricci flow for robust hypergraph community detection","Long-time Ricci flow on hypergraphs improves community detection","Proven: hypergraph Ricci flow exists forever, finds communities","Curvature-driven hyperedge flow detects hypergraph communities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's load-bearing premise is that the curvature-driven flow pushes hyperedges crossing between communities to larger weights than hyperedges inside communities, so deleting the heaviest hyperedges—with a cutoff selected using ground-truth labels—recovers the true clusters.","fun_headline_variants_meta":{"raw":{"variants":["Hypergraph Ricci flow: existence for all time enables community detection","HyperRCD: Ricci flow for robust hypergraph community detection","Long-time Ricci flow on hypergraphs improves community detection","Proven: hypergraph Ricci flow exists forever, finds communities","Curvature-driven hyperedge flow detects hypergraph communities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2904,"prompt_tokens":882,"completion_tokens":2022,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":1939}},"tokens_in":498,"tokens_out":2022,"duration_ms":15137,"temperature":1.0,"reasoning_tokens":1939,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:37:53.577745+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HyperRCD on a degree-corrected stochastic block model hypergraph with two planted 50-node communities; if after the flow the heaviest hyperedges mostly lie inside communities rather than between them, the weight-separation premise fails.","supporting_citations":[{"cited_title":"Ricci curvature of markov chains on metr ic spaces,","cited_arxiv_id":null,"evidence_quote":"Supplies the curvature and Wasserstein transport foundation that the hyperedge curvature generalizes."},{"cited_title":"Commun ity detection on networks with ricci flow,","cited_arxiv_id":null,"evidence_quote":"Supplies the graph Ricci-flow community detection method that motivates HyperRCD and the alpha = 0.5 parameter choice."},{"cited_title":"Hypergraph clustering using ricci curvature: an edge transport perspective,","cited_arxiv_id":null,"evidence_quote":"Supplies the curvature-based baseline and the edge-transport perspective on hypergraph clustering, plus the experimental protocol."},{"cited_title":"Efficient monte carlo and greedy heur istic for the inference of stochastic block models,","cited_arxiv_id":null,"evidence_quote":"Supplies the degree-corrected stochastic block model used to generate synthetic hypergraphs with planted community structure."},{"cited_title":"I’m me, we’re us, and i’m us : Tri-directional contrastive learning on hypergraphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the dataset and evaluation protocol whose baseline scores are reused for comparison."}],"review_version":1}