{"id":"80a9c007-6fd1-4ad0-8438-6388aa74733c","arxiv_id":"2505.18505","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A particle-dynamics-based hypergraph message passing framework (HAMP) is proposed, with first- and second-order variants, an anti-over-smoothing guarantee, and strong empirical results on heterophilic node classification.","lead":"This paper reframes hypergraph message passing as interacting particle systems, where hyperedges act as fields that attract, repel, and stabilize node features through Allen-Cahn forces. It introduces first-order and second-order variants and claims a theoretical guarantee against over-smoothing, with competitive node classification results especially on heterophilic hypergraphs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prop. 5.4 does not follow from the paper's separation lemmas: the L2 separation bound in Prop. 5.2 is an upper bound on a ratio, so both mean separation and variance can decay to zero while the bound holds; the two-sentence 'proof' of Prop. 5.4 asserts the needed positive bound without deriving it.","rationale":"The paper's empirical contribution is real: HAMP-I and HAMP-II perform competitively, and the ablations show that repulsion and the Allen-Cahn term help. But the headline theoretical claim—a proven positive lower bound on hypergraph Dirichlet energy—is what would distinguish the method from diffusion-based HNNs. Looking for the place where the proof has to do the work, I found that Prop. 5.4's proof is not a proof. The only quantitative statement actually established is an upper bound on the ratio of within-group variance to squared center distance (Prop. 5.2). Such an upper bound is compatible with convergence to consensus: if every node position is multiplied by a decaying amplitude, the ratio is unchanged and the Dirichlet energy decays. Connectedness of the hypergraph does not by itself prevent that, since the graph Laplacian's spectral gap only controls the rate of decay in the purely diffusive case and says nothing about the forced dynamics. The proof also rests on Eq. (46), an assumed lower bound on the repulsive interaction term that is never established from the model or data, and on Lemma 4.1 of [27], which is not reproduced. The second-order variant is explicitly deferred to a remark. I therefore agree with the reader's REJECT, though my emphasis is slightly different: the central gap is not only the structural assumption, but that even granting it, the separation theorem does not imply a lower bound on Dirichlet energy. A revised paper would need either to prove a genuine lower bound with explicit constants, or to soften the claim to 'empirically mitigates over-smoothing'.","tokens_in":24808,"tokens_out":8968,"duration_ms":73578,"concrete_test":"Analytical check: attempt to derive a uniform lower bound on E(x(t)) from Lemmas B.1, B.3, B.4 and Prop. B.5, with explicit constants. The decisive sub-check is whether those inequalities can rule out the scaling family x_i(t) = a(t) y_i, where y is a fixed two-cluster configuration on a connected hypergraph and a(t) -> 0. This family keeps lambda(t) = cM2(t)/||bar x^(1)-bar x^(2)||^2 constant, so the Prop. 5.2 conclusion holds, while E(x(t)) = a(t)^2 E(y) -> 0. If the inequalities cannot contradict this family—or if a numerical RK4 integration of Eq. 21 on a small connected two-group hypergraph satisfying (46) and (70) drives E below any fixed epsilon > 0 at large T—then Prop. 5.4 is unsupported and the theoretical anti-over-smoothing claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the anti-over-smoothing guarantee: a positive lower bound on hypergraph Dirichlet energy (Prop. 5.4). It is not established. Prop. 5.2/B.5 proves only that lambda(t) = cM2(t)/||bar x^(1) - bar x^(2)||^2 satisfies an upper bound lambda(t) <= lambda_- + (lambda(0)-lambda_-)e^{-mu t} (Eq. 9, Appendix B.5). This constrains the ratio, not either factor. In particular cM2 and the squared center distance may both shrink as a(t)^2 with a(t)->0 while the ratio stays bounded; then the hypergraph Dirichlet energy E(x) = sum H_i,e H_j,e ||x_i - x_j||^2 (which scales with the squared fluctuations) decays to zero, i.e., exactly the over-smoothing regime of Def. 5.1. The proof of Prop. 5.4 in Appendix B.7 is a two-sentence assertion—'Since the hypergraph is connected, there is a positive bound between different clusters'—with no lower bound on the center distance or on E(x). Lemma B.3's inequality (34) has an inhomogeneous term -C cM2 and is not used to prove any fixed epsilon > 0. Additionally, the route to Prop. 5.2 depends on the unquantified structural assumption (46) (existence of c_m with P_e^s >= c_m (bar x^(1)-bar x^(2))) and on Lemma 4.1 of [27], which is not stated; even granting these, the lower-bound conclusion does not follow. For Eq. 5, Prop. 5.3 is quoted from [19] and Prop. 5.4's extension is deferred to Remark B.8 ('can be proved in a similar way'), so no independent proof appears. The implemented algorithms (Appendix A) also use MLPs, activation sigma, initial-state residual beta X(0), and noise, none covered by the ODE analysis; but the decisive gap is that Prop. 5.2 is not a no-over-smoothing statement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes HAMP, a hypergraph message-passing framework inspired by interacting particle systems, with first-order (HAMP-I) and second-order (HAMP-II) variants and a stochastic extension. Hyperedges are treated as fields that exert attraction, repulsion, and Allen-Cahn forces on node features. The central theoretical claim is that these dynamics maintain a strictly positive lower bound on the hypergraph Dirichlet energy, thereby preventing over-smoothing (Prop. 5.4). The paper also reports node-classification experiments on nine real-world hypergraph benchmarks, ablations, depth-accuracy experiments, and t-SNE visualizations.","tokens_in":25367,"tokens_out":5495,"duration_ms":45187,"significance":"If the anti-over-smoothing theorem were established, the paper would provide a useful bridge between collective-dynamics theory and hypergraph neural networks, and the empirical results on heterophilic benchmarks are competitive. The particle-system analogy is clearly explained, and the ablation studies (Tables 2, 5, 6) usefully isolate the contributions of repulsion, Allen-Cahn, and noise terms. However, the theoretical guarantee is the main advertised contribution, and it is not supported by the supplied proofs. The empirical results alone, while solid, do not compensate for the unsupported central claim.","major_comments":[{"comment":"The claimed positive lower bound on the Dirichlet energy does not follow from Prop. 5.2/B.5. Inequality (9) is an upper bound on the ratio λ(t) = cM2(t)/‖x̄(1)-x̄(2)‖²; it does not lower-bound either factor. Both cM2(t) and the squared center distance may decay at the same rate, keeping λ(t) bounded while E(x) = Σ_{i,j} Σ_e H_{i,e}H_{j,e}‖x_i-x_j‖² tends to zero, which is exactly the over-smoothing regime of Def. 5.1. The proof of Prop. B.7 asserts 'there is a positive bound between different clusters' without deriving any quantitative lower bound on the center distance or on E(x). This is a load-bearing gap in the main advertised guarantee.","section":"Section 5, Prop. 5.4 / Appendix B, proof of Prop. B.7"},{"comment":"The key lower-bound inequality (34) relies on the unquantified assumption (46), which postulates a constant c_m such that P_e^s(mean vectors) ≥ c_m (x̄(1)-x̄(2)). No construction or verification of such c_m is given for the interaction coefficients f_β used in the models or for general connected hypergraphs. This assumption already presumes the two-group separability that the theorem is supposed to establish, so the hypotheses of Prop. 5.2 are not checkable or guaranteed by the model design.","section":"Appendix B, Eq. (46) and Lemma B.3"},{"comment":"The second-order separation result is quoted from [19], and its extension to the Dirichlet-energy lower bound is deferred with 'can be proved in a similar way' (Remark B.8). The stochastic version is asserted in Remark B.9 without proof. Consequently, the paper supplies no independent proof for HAMP-II or for the SDE formulation, even though both are part of the advertised framework.","section":"Section 5, Props. 5.3-5.4 and Appendix B, Remarks B.8-B.9"},{"comment":"The implemented algorithms use a nonlinearity σ, an MLP classifier, explicit noise ϵB(t), and a residual connection βX(0), whereas Propositions 5.2-5.4 are stated for the deterministic ODEs (4)-(5) with no activation, no noise, and no such residual term. No argument is provided that the discrete stochastic updates inherit the ODE separation properties, so the theoretical anti-over-smoothing guarantee does not directly cover the evaluated models.","section":"Appendix A, Algorithms 1-2 vs. Section 5"},{"comment":"The proof of the L2 separation invokes 'Lemma 4.1 in [27]' without stating the lemma or verifying its hypotheses. Since [27] is a prior work by the same group and the conditions needed here (including the existence of c_m in Eq. (46)) are not established, the proof is incomplete as written.","section":"Appendix B, proof of Prop. B.5"}],"minor_comments":[{"comment":"In the bullet list, 'we should that HAMP achieves competitive results' should read 'we show that HAMP achieves competitive results'.","section":"Section 1, Contributions"},{"comment":"The notation cM2 is used inconsistently: Eq. (16) defines it via centered moments, while Prop. 5.2 defines cM2(t) := M2(x^(1)) + M2(x^(2)). Please reconcile the definitions.","section":"Section 5 and Appendix B"},{"comment":"Prop. 5.4 refers to 'the conditions of Theorem 5.2' and 'Theorem 5.1 in [19]', but the manuscript's own results are numbered as Propositions; the cross-reference should be aligned with the actual numbering.","section":"Section 5, Prop. 5.4"},{"comment":"Remark B.9 states that the separability of the stochastic system 'also holds' without proof; this should at minimum be labeled as a conjecture or supported by a proof.","section":"Appendix B, Remark B.9"},{"comment":"The noise ablation is shown only on the Senate dataset in the main text; since the full ablation tables appear in Appendix C.2, a brief statement in the main text directing readers to those tables would improve clarity.","section":"Section 6.3, Fig. 3"}],"recommendation":"reject","confidential_remarks":"The theoretical results rely heavily on prior works with overlapping authorship ([19], [27], [43]), and the novel proof content is concentrated in Prop. 5.2 and its appendix, where the main gaps occur. The editor may also wish to consider whether the theoretical framework is sufficiently connected to the evaluated algorithms, given the mismatch between the ODE analysis and the implemented discrete stochastic updates."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central theoretical claim — a positive lower bound on hypergraph Dirichlet energy — is not proven. Prop 5.2 gives an upper bound on the ratio cM2 / ||mean1 - mean2||^2. That constrains the ratio, not either factor. Both the variance and the center distance can decay to zero together while the ratio stays bounded, which is exactly the over-smoothing regime. The proof of Prop 5.4 is two sentences and just asserts that connectedness gives a positive bound between clusters; no lower bound on the center distance or on the energy is derived. So the advertised guarantee doesn't follow from the math in the paper.\n\nWhat is new: the mapping of hypergraph message passing to interacting particle systems with attraction, repulsion, and Allen-Cahn forcing is a fresh angle, and the second-order variant is a reasonable extension. The empirical results are decent: on heterophilic benchmarks, HAMP-I and HAMP-II beat the listed baselines by 1–3%, and the depth experiments show they degrade more gracefully than the baselines. The ablation studies are thorough and the reported gains are consistent.\n\nSoft spots: beyond the missing proof, the structural assumption (46) — existence of c_m with P_e^s >= c_m (mean difference) — does much of the work and is not justified. Prop 5.2 relies on Lemma 4.1 from [27] which is not stated. The implemented algorithms use MLPs, activations, residual connections, and noise, none of which appear in the ODE analysis; the theorem is about a continuous system, not the discrete model that is actually run. The dependence on self-cited earlier results (Prop 5.3 from [19], Lemma from [27]) is not by itself a flaw, but here it means the genuinely new part — the lower bound — is exactly the part that isn't proved.\n\nThis paper is for readers working on hypergraph neural networks and over-smoothing; the particle-system formulation is worth knowing, and the empirical comparison is useful. But the paper overstates what is established. A serious referee would send it back for a major rewrite of Section 5 and the appendix, or for the authors to either prove the lower bound under stated assumptions or soften the claim to what actually follows (e.g., the ratio bound). I'd recommend engaging with it as a work-in-progress, not as a finished theoretical contribution.","headline":"The paper's advertised anti-over-smoothing guarantee is not proven: the central lower-bound claim on Dirichlet energy does not follow from the ratio estimates they prove, though the particle-system framing and the empirical results are worth a look.","tokens_in":25883,"tokens_out":1995,"would_cite":false,"duration_ms":16501,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that adding repulsion and Allen-Cahn damping to hypergraph message passing keeps the Dirichlet energy positive, so deep networks do not over-smooth.","keywords":["hypergraph neural networks","message passing","particle systems","over-smoothing","Dirichlet energy","Allen-Cahn force","heterophily","stochastic differential equations"],"falsifier":"Measure the layer-wise hypergraph Dirichlet energy of a trained HAMP-I on a connected, two-class hypergraph whose interaction coefficients satisfy the paper's sign assumptions: if the energy decays to zero across integration steps while classification accuracy stays above chance, the claimed positive lower bound is false. On a multi-class benchmark such as Cora or Walmart, the same measurement tests whether the assumed two-group structure is ever realized in practice.","tokens_in":24622,"feed_emoji":"⚛️","tokens_out":9915,"duration_ms":69355,"temperature":0.7,"pith_summary":"This paper argues that hypergraph message passing should be viewed as an interacting particle system, with each hyperedge acting as a field that exerts forces on the nodes it contains. The proposed framework, Hypergraph Atomic Message Passing (HAMP), adds repulsive forces that push different classes apart and an Allen-Cahn damping force that balances attraction against repulsion, with an optional Brownian-noise term for interaction uncertainty. The central theoretical claim is that this dynamics keeps a positive lower bound on the hypergraph Dirichlet energy—a measure of how distinct node representations remain—so node states cannot collapse into a single indistinguishable configuration even in deep networks. If true, this removes the main obstacle to stacking many message-passing layers and also helps on heterophilic hypergraphs, where different classes are connected to each other and purely attractive diffusion fails.","feed_headline":"Hyperedge fields keep hypergraph networks from over-smoothing","feed_subtitle":"Treating hyperedges as force fields with attraction, repulsion, and damping keeps classes separate as depth grows.","key_machinery":"The central object is the hypergraph Dirichlet energy $\\mathcal{E}(x)=\\operatorname{tr}(x^\\top Lx)=\\sum_{i,j}\\sum_{e\\in E}H_{i,e}H_{j,e}\\|x_i-x_j\\|^2$, where $L$ is the hypergraph Laplacian; over-smoothing is defined as exponential decay of this energy with depth. The mechanism carrying the argument is the composite field $F_i=\\sum_{e:i\\in e}F_i^e+F^d$, whose gradient yields the first-order dynamics $\\dot x_i=\\sum_{e:i\\in e}\\sum_{j\\in e}f_\\beta(x_i,x_j,e)(x_j-x_i)+f_d(x_i)$, with sign-changing $f_\\beta$ controlling attraction versus repulsion and $f_d$ the Allen-Cahn force $\\delta(1-x_i^2)x_i$ acting as damping. The proof couples a lower bound on the growth of the inter-group center distance $\\|\\bar x^{(1)}-\\bar x^{(2)}\\|$ with an upper bound on the within-group second moment, so their ratio obeys a Lyapunov-type separation inequality; because a connected hypergraph has at least one cross-group edge, this separation forces the Dirichlet energy to stay above a positive floor.","core_discovery":"The paper's central claim is that adding repulsion and Allen-Cahn forcing to hypergraph message passing turns a purely diffusive propagation—which is known to drive the Dirichlet energy to zero exponentially—into a dynamics whose Dirichlet energy stays bounded below by a positive constant. In HAMP, each hyperedge acts as a field that superposes interaction forces on its nodes, with pair coefficients $f_\\beta$ that are positive inside a group (attraction) and negative between groups (repulsion). Under a two-group sign assumption on these coefficients, the paper proves an $L^2$ separation bound for the first-order system and invokes a bi-flocking result for the second-order Cucker–Smale-type system, yielding Proposition 5.4: for a connected hypergraph, the Dirichlet energy has a strictly positive lower bound during propagation. The two instantiations, HAMP-I and HAMP-II, then achieve competitive node-classification accuracy on nine benchmarks, with the clearest gains on heterophilic datasets and with HAMP-II remaining stable as depth increases.","pith_inferences":["The paper leaves implicit that the same two-group separation argument should extend to slowly evolving temporal hypergraphs; testing HAMP on dynamic incidence structures would show whether the Dirichlet-energy floor survives topology drift.","The theory is proved for the continuous ODE limit, while the implemented networks use MLPs, activations, and noise; measuring layer-wise Dirichlet energy in the discrete model and comparing it with the predicted positive floor would close that gap.","Because the Allen-Cahn force comes from a double-well potential, HAMP can be read as a soft clustering regularizer; one could test whether the learned equilibria align with class proportions even when labels are scarce.","The two-group sign pattern assumed in the proof is strong for datasets with more than two classes; an adaptive or learned sign pattern might be needed in general, and the energy lower bound would have to be re-proved for that setting."],"forward_implications":["Deep HAMP networks can keep discriminative node representations: the Dirichlet energy stays above a positive floor, so accuracy does not degrade sharply as layers are added.","The second-order variant HAMP-II supports deeper message passing than first-order and diffusion-based baselines, as shown by depth-accuracy experiments on Cora, Citeseer, and Congress.","The framework extends to heterophilic hypergraphs, where repulsion keeps different classes from merging; HAMP reports 1–3% accuracy gains over prior baselines on Congress, Senate, Walmart, and House.","Adding Brownian noise to the deterministic dynamics improves both accuracy and stability, indicating that stochastic message passing can model interaction uncertainty rather than hurting learning.","The particle interpretation provides a physically interpretable account of hypergraph neural networks, connecting node classification to equilibria of attractive-repulsive systems and suggesting a principled way to design new propagation rules."],"supporting_citations":[{"why":"supplies the Allen-Cahn double-well potential whose gradient gives the damping force $f_d$ used in both HAMP variants.","marker":"[1]"},{"why":"provides the exponential decay bound for graph diffusion energy that HAMP aims to overcome.","marker":"[9]"},{"why":"supplies the attractive-repulsive Cucker-Smale model with Rayleigh friction whose bi-flocking theorem underlies Proposition 5.3 and the second-order analysis.","marker":"[19]"},{"why":"defines the hypergraph propagation matrix and the simplified HGNN that HAMP generalizes and compares against.","marker":"[20]"},{"why":"provides the two-group stochastic opposition dynamics and the lemma used in the $L^2$ separation proof of Proposition 5.2.","marker":"[27]"},{"why":"motivates the first-order opinion-dynamics form of Eq. 4 and the connection to heterophilious dynamics.","marker":"[33]"},{"why":"supplies the benchmark dataset suite, baseline numbers, and training recipe used in the experimental section.","marker":"[42]"},{"why":"provides the Allen-Cahn message passing with attractive and repulsive forces for graphs that HAMP extends to hypergraphs.","marker":"[43]"},{"why":"is the control-diffusion ODE hypergraph baseline whose official results are reproduced for comparison.","marker":"[47]"}],"fun_headline_variants":["Repulsive hyperedges beat over-smoothing in deep networks","Particle physics for hypergraphs: separation without smoothing out","Force-field hyperedges keep classes apart as layers deepen","Attraction plus repulsion: hypergraph message passing that goes deep"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that interactions split cleanly into two groups with positive intra-group and non-positive inter-group coefficients, and that a fixed constant bounds the repulsion-induced separation of the two group centers from below; if real data violates that split, or if the discretized MLP dynamics leave the regime of the analyzed ODEs, the claimed lower bound on the Dirichlet energy does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Repulsive hyperedges beat over-smoothing in deep networks","Particle physics for hypergraphs: separation without smoothing out","Force-field hyperedges keep classes apart as layers deepen","Attraction plus repulsion: hypergraph message passing that goes deep"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1257,"prompt_tokens":933,"completion_tokens":324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":255}},"tokens_in":549,"tokens_out":324,"duration_ms":3362,"temperature":1.0,"reasoning_tokens":255,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:29:49.303775+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the layer-wise hypergraph Dirichlet energy of a trained HAMP-I on a connected, two-class hypergraph whose interaction coefficients satisfy the paper's sign assumptions: if the energy decays to zero across integration steps while classification accuracy stays above chance, the claimed positive lower bound is false. On a multi-class benchmark such as Cora or Walmart, the same measurement tests whether the assumed two-group structure is ever realized in practice.","supporting_citations":[{"cited_title":"A microscopic theory for antiphase boundary motion and its application to antiphase domain coarsening.Acta metallurgica, 27(6):1085–1095, 1979","cited_arxiv_id":null,"evidence_quote":"supplies the Allen-Cahn double-well potential whose gradient gives the damping force $f_d$ used in both HAMP variants."},{"cited_title":"Gorinova, Stefan D Webb, Emanuele Rossi, and Michael M","cited_arxiv_id":null,"evidence_quote":"provides the exponential decay bound for graph diffusion energy that HAMP aims to overcome."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the attractive-repulsive Cucker-Smale model with Rayleigh friction whose bi-flocking theorem underlies Proposition 5.3 and the second-order analysis."},{"cited_title":"Hypergraph neural networks","cited_arxiv_id":null,"evidence_quote":"defines the hypergraph propagation matrix and the simplified HGNN that HAMP generalizes and compares against."},{"cited_title":"Collective dynamics of opposing groups with stochastic communica- tion.Vietnam Journal of Mathematics, 49:619–636, 2021","cited_arxiv_id":null,"evidence_quote":"provides the two-group stochastic opposition dynamics and the lemma used in the $L^2$ separation proof of Proposition 5.2."},{"cited_title":"Heterophilious dynamics enhances consensus.SIAM Review, 56(4):577–621, 2014","cited_arxiv_id":null,"evidence_quote":"motivates the first-order opinion-dynamics form of Eq. 4 and the connection to heterophilious dynamics."},{"cited_title":"Equivariant hyper- graph diffusion neural operators","cited_arxiv_id":null,"evidence_quote":"supplies the benchmark dataset suite, baseline numbers, and training recipe used in the experimental section."},{"cited_title":"ACMP: Allen-cahn message passing with attractive and repulsive forces for graph neural networks","cited_arxiv_id":null,"evidence_quote":"provides the Allen-Cahn message passing with attractive and repulsive forces for graphs that HAMP extends to hypergraphs."},{"cited_title":"Hypergraph dynamic system","cited_arxiv_id":null,"evidence_quote":"is the control-diffusion ODE hypergraph baseline whose official results are reproduced for comparison."}],"review_version":1}