{"id":"bb713dcf-3dbe-414d-8687-9017ac88887c","arxiv_id":"2506.02842","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A directed sheaf Laplacian and the DSNN architecture bring edge-orientation information into sheaf-based graph learning with modest-but-consistent benchmark gains.","lead":"This paper adds edge direction to sheaf neural networks by encoding each directed edge as a complex phase in the sheaf's restriction maps. The resulting Directed Sheaf Neural Network posts the best or second-best accuracy on seven of nine node-classification benchmarks, though most gains are within one standard deviation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The directional inductive bias is real but partial: the DSL is a Hermitian magnetic-style operator whose symmetrized diagonal cannot distinguish in-degree from out-degree, so the 'faithfully integrates direction' claim is stronger than what is shown.","rationale":"Reading the paper in good faith, the construction is internally sound: L = δ*δ is Hermitian and PSD, the normalized operator is well-defined, and the synthetic experiments show a consistent improvement over undirected sheaf baselines. The reader's concern about the factor-of-2 discrepancy in Theorem 4 is legitimate but not the most load-bearing point: for all-directed graphs, L^F = 2L^(q) and the normalization satisfies D^F = 2D_s, so the normalized operators cancel to exact equality; for mixed directed/undirected graphs the theorem needs qualification, but this does not undermine the architecture. The deeper assumption is that a Hermitian phase encoding with a symmetrized degree diagonal captures directionality. That assumption is plausible and empirically useful, but it is not equivalent to capturing in/out asymmetry or non-normal flow. The paper's own language claims more than this, and the absence of a reversal or non-normal control leaves the boundary of the claim untested. This is a scope/accuracy issue that can be fixed without changing the architecture, so the reader's CONDITIONAL verdict remains appropriate.","tokens_in":18554,"tokens_out":14230,"duration_ms":174137,"concrete_test":"On the DSBM generator, create paired datasets with identical undirected topology, node features, and labels but with all edge orientations reversed, plus a third set with per-edge randomized orientation. Train Diag-DSNN and O(d)-DSNN alongside DirGNN under the same tuning budget, and compare test accuracy and the learned DSL eigenspaces. If DSNN's accuracy and learned operator are unchanged under full reversal (up to complex conjugation), the directional bias is only magnetic-phase-like; if performance changes with orientation, the current encoding does capture more than the symmetrized magnetic paradigm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 1 and Eq. (3) define the DSL through a Hermitian operator L = δ*δ whose degree normalization is D_uu = Σ_{e in Γ(u)} F*_{u⊴e}F_{u⊴e}, so in- and out-incidences contribute identically. In the trivial case this is exactly the symmetrized degree diagonal of the Magnetic Laplacian, and reversing every edge orientation sends T^(q) to its conjugate and L to its conjugate, leaving the spectrum unchanged. The directional signal is therefore confined to the phase T^(q); information carried by in/out degree asymmetry or by non-normal flow (e.g., PageRank-like propagation) is invisible to this operator. The experiments do not settle this: the synthetic DSBM study (Figure 2) compares DSNN only against NSD and NSD-Compl, not against a non-normal directed model or a reversed-orientation control. The real-data comparison against DirGNN and MagNet is supportive, but it does not isolate whether the gain comes from the Hermitian magnetic phase or from something the encoding intentionally omits. This is not a fatal contradiction, but the manuscript's framing ('directional awakening', 'faithfully integrates') overreaches relative to what a symmetrized-degree, real-spectrum operator can represent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Directed Cellular Sheaves, in which edge orientation is encoded into complex-valued restriction maps through a Hermitian phase matrix T^(q)=exp(i2πq(A-A^T)), and defines the Directed Sheaf Laplacian L^F=δ*δ together with its normalized version. On this operator the authors build the Directed Sheaf Neural Network (DSNN), a sheaf-diffusion architecture with a directional inductive bias, and report node-classification experiments on nine real-world benchmark datasets plus a synthetic DSBM study. The paper's central claims are that L^F captures both topology and direction, that it generalizes the Magnetic and Sign-Magnetic Laplacians, and that DSNN consistently outperforms baseline GNNs and SNNs.","tokens_in":18815,"tokens_out":8492,"duration_ms":84217,"significance":"The construction is attractive and largely sound: L^F is Hermitian and PSD by construction, its normalized version has spectrum in [0,2], and the architecture follows naturally from the directed sheaf diffusion equation. The synthetic DSBM experiment cleanly separates DSNN from undirected sheaf baselines, and the real-world results are strong enough to be of interest to the graph-learning community. If the claims are corrected, the paper would provide a principled way to inject directional information into sheaf neural networks and a useful incidence-matrix decomposition of magnetic-style Laplacians. However, the exact-equality theorems are overstated as written, and the empirical isolation of the directional mechanism is incomplete, so the advertised 'directional awakening' is currently stronger than what the manuscript demonstrates.","major_comments":[{"comment":"Theorem 4 states that L^F 'coincides' with the Magnetic Laplacian L^(q), but the proof in Part 2b derives L^F = 2 L^(q) for directed graphs without digons, and a similar factor of 2 arises when digons are present. The theorem as stated is therefore false; the correct unnormalized statement is L^F = 2 L^(q), while the normalized operators coincide for digon-free graphs. Since the DSNN uses the normalized Laplacian, the architectural relationship is salvageable, but every statement in the paper that relies on exact equality, including the abstract's and conclusion's claims that DSNN generalizes MagNet and SigMaNet, must be rephrased as equality up to the scaling factor.","section":"Section 3.3 (Theorem 4)"},{"comment":"Theorem 5 asserts that L^(q) = B B* for a trivial directed cellular sheaf. The proof, however, first establishes L^F = B B* and then invokes Theorem 4; because Theorem 4's exact equality is false, the claimed incidence-matrix decomposition of L^(q) holds only up to the same factor of 2 (or, equivalently, for the normalized version). The theorem should be restated and re-proved with the scaling made explicit, since this decomposition is one of the paper's advertised theoretical contributions.","section":"Section 3.3 (Theorem 5)"},{"comment":"The directional information carried by L^F is confined to the off-diagonal phase T^(q): the degree diagonal D_uu is a sum over all incident edges of F*_{u⊴e}F_{u⊴e}, so in-degree and out-degree contribute identically, and reversing every edge orientation conjugates L^F and leaves its spectrum unchanged. The manuscript's phrasing that L^F 'faithfully integrates the sign and directionality' (Section 1) and 'captures both topology and directional information' is therefore stronger than what the operator actually represents. This limitation should be stated explicitly, and the claim tempered unless a control experiment demonstrates that the phase information, rather than the learned sheaf parameterization, drives the empirical gains.","section":"Definition 1 and Eq. (3)"},{"comment":"The synthetic DSBM experiment compares DSNN only against NSD and NSD-Compl, which either ignore direction or use complex restriction maps without a magnetic phase. It does not include MagNet, SigMaNet, or DirGNN, the baselines that already carry a directional inductive bias. As a result, the conclusion that the large observed gains 'underscore DSNN's ability to leverage edge directionality' is not fully isolated: a simpler explanation is that any complex/magnetic encoding helps on this synthetic task, or that the sheaf parameterization is independently beneficial. Adding at least one directed spectral baseline (or a reversed-orientation control) would make the directional-bias claim supported by the controlled experiment.","section":"Section 5.2 (Figure 2)"}],"minor_comments":[{"comment":"The sentence 'mirroring a similar property that haolds for the standard graph Laplacian' contains a typo: 'haolds' should be 'holds'.","section":"Section 2.1"},{"comment":"At q=1/4, the definition T^(q)_{uv}=exp(i2πq(Auv-Avu)) with Auv=1 and Avu=0 gives T^(q)_{uv}=exp(iπ/2)=+i, but the text writes cos(-π/2)+i sin(-π/2)=-i. The indexing convention should be made consistent in this explanatory passage.","section":"Section 3.1 (Definition 1 rationale)"},{"comment":"The proof states twice that 'the last equation holds since T^(q)_{uu}=0 for any q'. In fact T^(q)_{uu}=exp(0)=1; the term vanishes because A_s,uu=0, not because T^(q)_{uu}=0. This typographical error should be corrected.","section":"Theorem 4 proof (Parts 2a and 2b)"},{"comment":"The phrase 'for a directed edge e=(u,v) or e=(v,e)' contains a typo: the second case should be e=(v,u), not e=(v,e).","section":"Section 3.1, text after Eq. (2)"},{"comment":"The paper promises that datasets and code are available on GitHub, but Appendix A does not actually provide the repository URL. Including the link would materially improve reproducibility.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is potentially publishable after revision, but the theoretical claims need to be corrected rather than merely softened. The false Theorem 4/5 equalities are the main barrier; the empirical section is strong but should include a directed baseline or orientation-reversal control to support the causal story about directionality. No concerns about author conduct or citation practice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for you. The DSNN is a legitimate extension of sheaf neural networks to directed graphs, and the core operator—the Directed Sheaf Laplacian defined as δ*δ—is sound. The genuinely new content is the learned complex restriction maps that place a magnetic phase on the head node's map, and the incidence factorization of the Magnetic Laplacian (Theorem 5). The synthetic DSBM experiment is clean and shows a large, reproducible advantage over NSD and NSD-Compl, so the directional inductive bias is doing real work.\n\nThe soft spots are in the claims, not the architecture. Theorems 4 and 5 state that the trivial DSL coincides with the Magnetic Laplacian, but the proofs show L^F = 2 L^(q) for digon-free directed graphs. The scaling factor is immaterial when the Laplacian is multiplied by learned weights, but the statements as written are false and should be corrected. Second, the abstract's 'consistently outperforms baseline methods' is too strong: on Cora, Diag-DSNN (87.36) is below GCNII (88.37), FAGCN (88.17), and GGCN (87.95), and many error bars overlap. The qualified claims in Section 5.1 are fine; the abstract should match them. Third, the paper says code is on GitHub, but Appendix A only lists baseline repositories—there is no link to the DSNN implementation. That is a minor but real reproducibility gap.\n\nThe stress-test note on symmetrized degree is on point. The DSL diagonal sums F*_{u⊴e}F_{u⊴e} over all incident edges, so in- and out-degree contribute identically; directional information lives only in the off-diagonal phases. The operator therefore cannot represent in/out asymmetry or non-normal flow. The paper's phrase 'faithfully integrates direction' overreaches. This is not fatal—the magnetic Laplacian paradigm is proven in practice—but the authors should temper the framing and ideally add a reversed-orientation control or a non-normal baseline.\n\nWho should read it: anyone working on directed GNNs or sheaf methods. It is a useful architectural contribution, not a paradigm shift. The theory statements need fixing and the empirical claims need calibration, but the central idea holds up. I would send it to peer review with a request for major revision rather than desk reject. The authors should correct the factor-of-2 statements, add the code link, and align the abstract with the actual results.","headline":"A genuinely useful directed sheaf architecture with a sound core, though the theorem statements overstate the match to the Magnetic Laplacian and the empirical claims need calibration.","tokens_in":19400,"tokens_out":4240,"would_cite":true,"duration_ms":39435,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C50","55N30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a Directed Sheaf Laplacian carries edge direction through a complex phase and that DSNN, the first sheaf network with this directional bias, consistently outperforms GNN and SNN baselines on nine node-classification…","keywords":["directed cellular sheaf","directed sheaf Laplacian","sheaf neural networks","graph neural networks","edge directionality","heterophilic graphs","magnetic Laplacian","sign-magnetic Laplacian"],"falsifier":"Take any directed graph, randomly permute the orientations of its edges so $A-A^\\top$ changes while the undirected skeleton and degrees stay fixed, and compare DSNN's accuracy with $q=1/4$ against the same model with $q=0$: the directional-bias claim predicts a clear accuracy drop, and if accuracy is unchanged, the phase encoding is not carrying the directional signal.","tokens_in":18313,"feed_emoji":"🧭","tokens_out":14733,"duration_ms":126851,"temperature":0.7,"pith_summary":"Sheaf neural networks move features between nodes through linear restriction maps, but until now those maps have ignored which way an edge points. This paper introduces the Directed Cellular Sheaf, in which the restriction map at the tail of a directed edge is multiplied by the unit-modulus phase $T^{(q)}=\\exp(i2\\pi q(A-A^\\top))$, so the edge's direction is written into the algebra of the sheaf rather than into a separate aggregation rule. From this sheaf the authors define the Directed Sheaf Laplacian $L^{\\tilde F}$ and the Directed Sheaf Neural Network (DSNN), which they describe as the first sheaf architecture with an explicit directional inductive bias. Their stated result is that DSNN consistently outperforms a broad set of GNN and SNN baselines on nine real-world node-classification benchmarks, with the largest gains on heterophilic graphs, and that on synthetic directed stochastic block models the non-directional sheaf baselines collapse while DSNN retains high accuracy. A sympathetic reader should care because the construction shows directionality can be added to sheaf models with a single algebraic change, at a constant computational overhead, while recovering existing magnetic-Laplacian models as special cases.","feed_headline":"A directed sheaf Laplacian beats GNN baselines on 9 node tasks","feed_subtitle":"The first sheaf network with a built-in directional bias generalizes MagNet and SigMaNet.","key_machinery":"The load-bearing object is the Directed Sheaf Laplacian $L^{\\tilde F}=\\tilde\\delta^*\\tilde\\delta$, built from a Directed Cellular Sheaf whose edge restriction maps are $F_{u\\unlhd e}\\in\\mathbb{R}^{d\\times d}$ at the head and $F_{v\\unlhd e}=F_{v\\unlhd e}^{0}\\,T^{(q)}_{uv}$ at the tail. The directional memory is the Hermitian matrix $T^{(q)}=\\exp(i2\\pi q(A-A^\\top))$, which equals $1$ on undirected edges and $-i$ on directed edges at $q=1/4$, making the off-diagonal blocks of the Laplacian complex and orientation-sensitive while the spectrum remains real and nonnegative. DSNN's update is the discretized diffusion $X^{(t+1)}=\\mathrm{diag}(1+\\varepsilon)X^{(t)} - \\sigma(\\tilde L^{\\tilde F}_N (I_n\\otimes W_1)X^{(t)}W_2)$, with restriction maps learned from concatenated node features and a complex ReLU activation. The theoretical bridge to existing models is Theorem 4: a trivial directed sheaf makes $L^{\\tilde F}$ equal to the Magnetic Laplacian, and at $q=1/4$ to the Sign-Magnetic Laplacian.","core_discovery":"The paper's central claim is that direction is not an extra ingredient to bolt onto a sheaf; it can be placed inside the restriction maps themselves. A Directed Cellular Sheaf assigns complex vector spaces to nodes and edges, keeps the head node's restriction map real, and sets the tail node's map to a real map followed by the Hermitian phase $T^{(q)}=\\exp(i2\\pi q(A-A^\\top))$. At $q=1/4$, a directed edge $(u,v)$ contributes a factor $-i$, an undirected edge contributes $1$, and the resulting Directed Sheaf Laplacian $L^{\\tilde F}=\\tilde\\delta^*\\tilde\\delta$ is Hermitian positive semidefinite with normalized spectrum in $[0,2]$. The authors prove that for a trivial one-dimensional sheaf this operator coincides with the Magnetic Laplacian $L^{(q)}$ and, at $q=1/4$, with the Sign-Magnetic Laplacian, and they give both of those matrices a new node-to-edge incidence factorization $L^{(q)}=\\hat B\\hat B^*$. On top of this operator, DSNN learns its restriction maps from node features and diffuses the node signal through the normalized Directed Sheaf Laplacian, which is the mechanism behind the reported gains on the nine benchmarks.","pith_inferences":["Because the phase encoding lives in the sheaf rather than in a model-specific layer, the same trick could be dropped into any spectral pipeline that currently symmetrizes a directed graph: replace the symmetrized adjacency with $A_s\\odot T^{(q)}$ and inherit a tunable directional bias.","Theorem 4's factor-of-two gap between $L^{\\tilde F}$ and $L^{(q)}$ on digon-free graphs suggests the choice of degree normalization is not neutral; ablating in-degree vs out-degree vs symmetrized degree could separate true directional signal from degree asymmetry.","A control the paper does not run would be to randomly permute edge orientations while keeping the graph skeleton fixed; DSNN's stated mechanism predicts accuracy should drop, and measuring that drop would isolate the directional component.","The construction is stated for complex vector spaces and ordinary graphs, but the same Hermitian phase can be attached to restriction maps on hyperedges or higher-dimensional cells, a direction the authors list as future work but do not test."],"forward_implications":["DSNN reduces to NSD on undirected graphs and to MagNet and SigMaNet in the trivial-sheaf limit, so those architectures become special cases of one operator.","The incidence factorization $L^{(q)}=\\hat B\\hat B^*$ gives the Magnetic and Sign-Magnetic Laplacians a Hermitian coboundary decomposition, which yields simpler positive-semidefiniteness proofs than the original constructions.","On the nine real benchmarks, DSNN is reported to beat the second-best baseline on seven of nine and every directional GNN baseline on all nine, with the largest single gain over directional baselines of 15.59 points on Cornell.","On synthetic DSBM graphs, DSNN variants hold 86-96% accuracy while all non-directional sheaf baselines except one fall to chance-level 20%, supporting the claim that the directional bias, not complex parameters alone, drives the improvement.","Adding the directional phase costs only a constant multiplicative factor (about 4) in the forward pass, leaving DSNN's asymptotic complexity equal to NSD's."],"supporting_citations":[{"why":"Supplies the cellular-sheaf Laplacian construction that the Directed Sheaf Laplacian extends to directed graphs.","marker":"[1]"},{"why":"Supplies the Neural Sheaf Diffusion baseline and the diffusion equation that DSNN generalizes.","marker":"[2]"},{"why":"Provides the MagNet directional GNN baseline and the magnetic-Laplacian spectral paradigm DSNN is compared against.","marker":"[3]"},{"why":"Introduces the Magnetic Laplacian that Theorem 4 recovers as the trivial directed-sheaf case of the new operator.","marker":"[20]"},{"why":"Introduces the Sign-Magnetic Laplacian that DSNN recovers at $q=1/4$.","marker":"[21]"},{"why":"Defines sheaf neural networks and the diffusion process underlying DSNN's update rule.","marker":"[23]"},{"why":"Provides Dir-GNN, a directional message-passing baseline that DSNN outperforms on all nine real datasets.","marker":"[16]"},{"why":"Source of the Texas, Wisconsin, Cornell, and Film benchmarks used in the real-world evaluation.","marker":"[25]"},{"why":"Source of the Citeseer, PubMed, and Cora benchmarks and of heterophilic-graph baselines.","marker":"[26]"},{"why":"Source of the Squirrel and Chameleon benchmarks and the heterophily evaluation protocol.","marker":"[27]"}],"fun_headline_variants":["First directed sheaf neural network outperforms GNNs","New directed sheaf Laplacian beats baselines on 9 tasks","DSNN: embedding direction into sheaves for better learning","A directional sheaf that generalizes magnetic Laplacians","Beats GNNs with a directed sheaf Laplacian"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a directed edge's meaning is fully captured by multiplying the tail node's restriction map by the complex phase $e^{i2\\pi q(A-A^\\top)}$, so any directional signal that only a non-Hermitian, non-normal operator could see is invisible to this Laplacian.","fun_headline_variants_meta":{"raw":{"variants":["First directed sheaf neural network outperforms GNNs","New directed sheaf Laplacian beats baselines on 9 tasks","DSNN: embedding direction into sheaves for better learning","A directional sheaf that generalizes magnetic Laplacians","Beats GNNs with a directed sheaf Laplacian"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000306,"raw_usage":{"total_tokens":1760,"prompt_tokens":960,"completion_tokens":800,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":714}},"tokens_in":576,"tokens_out":800,"duration_ms":8072,"temperature":1.0,"reasoning_tokens":714,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:15:32.359586+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any directed graph, randomly permute the orientations of its edges so $A-A^\\top$ changes while the undirected skeleton and degrees stay fixed, and compare DSNN's accuracy with $q=1/4$ against the same model with $q=0$: the directional-bias claim predicts a clear accuracy drop, and if accuracy is unchanged, the phase encoding is not carrying the directional signal.","supporting_citations":[{"cited_title":"Neural sheaf diffusion: A topological perspective on heterophily and oversmoothing in gnns.Advances in Neural Information Processing Systems, 35:18527–18541, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the Neural Sheaf Diffusion baseline and the diffusion equation that DSNN generalizes."},{"cited_title":"Magnet: A neural network for directed graphs, 2021","cited_arxiv_id":null,"evidence_quote":"Provides the MagNet directional GNN baseline and the magnetic-Laplacian spectral paradigm DSNN is compared against."},{"cited_title":"Fluxes, Laplacians, and Kasteleyn’s theorem","cited_arxiv_id":null,"evidence_quote":"Introduces the Magnetic Laplacian that Theorem 4 recovers as the trivial directed-sheaf case of the new operator."},{"cited_title":"Sigmanet: One laplacian to rule them all","cited_arxiv_id":null,"evidence_quote":"Introduces the Sign-Magnetic Laplacian that DSNN recovers at $q=1/4$."},{"cited_title":"Edge directionality improves learning on heterophilic graphs","cited_arxiv_id":null,"evidence_quote":"Provides Dir-GNN, a directional message-passing baseline that DSNN outperforms on all nine real datasets."},{"cited_title":"Two sides of the same coin: Heterophily and oversmoothing in graph convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Source of the Citeseer, PubMed, and Cora benchmarks and of heterophilic-graph baselines."},{"cited_title":"A critical look at the evaluation of gnns under heterophily: Are we really making progress? InThe Eleventh International Conference on Learning Representations, 2023","cited_arxiv_id":null,"evidence_quote":"Source of the Squirrel and Chameleon benchmarks and the heterophily evaluation protocol."}],"review_version":1}