{"id":"2d892118-7920-4a7c-923c-9dce3f17b2a4","arxiv_id":"2608.08567","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A graph neural network framework that maps graphs with arbitrary feature spaces to a shared fixed-dimension space via a structural interaction graph, enabling zero-shot transfer across graphs and recovering knowledge-graph foundation models as a special case.","lead":"This paper introduces SIGIL, a method that turns any graph with arbitrary node features into a fixed-size representation by building a new graph over the feature dimensions and running message passing on it. A model trained on one graph can then be applied to completely new graphs without retraining, which is the goal of graph foundation models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Gram-matrix summarization in Eq. (4) cannot distinguish feature coordinates with identical alignment profiles, so the fixed-dimension transfer claim is only as strong as the preserved alignment statistics; the paper's own node-classification gap is the visible symptom.","rationale":"The reader's weakest assumption correctly targets the Gram-matrix summarization of Eq. (4): the Gram tensors are the only channel through which feature identities enter Θ(T). Because Proposition 1 makes Θ(T) feature-equivariant, two features with identical interaction profiles are not merely similar—they become identical inputs to the downstream model. The paper's own §5.3 limitation is an admission of this, and the node-classification gap is the empirical footprint. I do not think this makes the framework invalid; the construction is deterministic and parameter-free, and the equivariance proofs are coherent. But it does bound the scope of \"any downstream graph\": zero-shot transfer can only be expected when the task's label function is compatible with the alignment summarization. The link-prediction headline does not contradict this, because no analysis shows which part of the Hits@K gain comes from alignment statistics versus per-coordinate semantics. The synthetic two-feature test would turn the philosophical objection into a measured quantity. If it fails, the correct verdict remains CONDITIONAL with a re-scoped claim, as the reader argued; if it passes, the conditional can later be lifted. Either way, the current CONDITIONAL verdict is unchanged.","tokens_in":25798,"tokens_out":12798,"duration_ms":150564,"concrete_test":"Run a two-feature synthetic control: generate a family of attributed graphs where features f1 and f2 have identical row and column profiles in every Ā^(k) slice (for example, make f2 an i.i.d. copy of f1 after a node permutation), let source labels depend only on f1, and let target labels depend only on f2. Pretrain Sigil-lp on the source and evaluate zero-shot on the target. If performance collapses to chance, then Eq. (4)–Eq. (7) demonstrably conflate the two coordinates and the \"any graph / any downstream task\" claim is bounded by Gram sufficiency; if performance does not collapse, the representation retains enough per-coordinate information through the original X, and the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that H = XΘ(T), with Θ(T) depending only on the Gram tensors Ā^(k) of Eq. (4), is a representation that supports zero-shot transfer to any downstream graph. Because Θ(T) is feature-equivariant (Proposition 1), any two features f, f′ with identical rows in every Ā^(k) slice receive identical embeddings; the resulting H contains only their summed contribution and cannot separate them. The paper concedes this in §5.3: \"compressing arbitrary feature spaces into a fixed d_h leaves the original coordinates individually inaccessible to the decoder,\" and the 66.10 vs. 71.78 node-classification gap (Table 3) is direct evidence that the compression discards task-relevant signal. The link-prediction wins do not resolve this: Hits@K on the tested benchmarks may be driven by aggregate alignment statistics rather than by per-coordinate semantics, so the flagship result is consistent with the limitation rather than contradictory to it. The empirical headline is further weakened by Table 1 including Cora itself as a target and by the absence of any attributed-graph result from TFMLinker, but the deeper load-bearing issue is the unproven sufficiency of Gram-alignment features for arbitrary downstream tasks. The paper never states the condition under which tasks are compatible with this summarization, nor proves that the tested tasks satisfy it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SIGIL, a framework that lifts any attributed graph G=(V,E,X) to a structural interaction graph (SIG) whose nodes are the d feature dimensions and whose typed, weighted edges are Gram matrices of propagated feature alignments across orders k=0..K. A relational message-passing network embeds each feature dimension into a fixed d_h-dimensional space, and the original node features are transformed by H=XΘ(T) to a fixed-width representation. The paper proves feature-dimension and node-permutation equivariance (Lemma 1, Proposition 1) and shows that for one-hot relation features the SIG recovers ULTRA's relation graph (Theorem 1). It evaluates the resulting Sigil-lp on 11 link-prediction benchmarks, on ULTRA's KG reasoning suite, and Sigil-nc on 26 node-classification datasets.","tokens_in":26017,"tokens_out":8006,"duration_ms":80443,"significance":"The SIG construction is transparent, deterministic, and parameter-free at the graph-lifting stage, and Theorem 1's exact recovery of ULTRA's relation graph is a clean, testable equivalence rather than a fitted approximation. The equivariance results (Lemma 1, Proposition 1) are coherent and correctly stated. If the empirical claims are confirmed after fixing the evaluation issues, the framework would meaningfully unify relation-graph KGFMs with continuous-feature GFM designs. However, the advertised 'any downstream graph' guarantee is not supported as stated, and the empirical evidence is mixed, so the paper needs revision rather than acceptance in its current form.","major_comments":[{"comment":"Table 1 lists Cora as an attributed target, but the dagger marks Sigil-lp as pretrained on Cora. Because Cora is the pretraining graph, the result on Cora is not a zero-shot transfer result, and the sentence in §5.1 that Sigil-lp is 'the strongest fully-inductive method on every dataset, including the six it has never seen' and the abstract's 'strongest fully-inductive link predictor on every attributed benchmark' are overstated. Please exclude Cora from the zero-shot target set (or label it a sanity check) and re-state the claim as holding for the six unseen attributed benchmarks. In addition, TFMLinker is not reported on the attributed block, so the attributed comparison is effectively against UniLP alone; this should be acknowledged in the claim.","section":"§5.1, Table 1"},{"comment":"The SIG is defined entirely by the Gram matrices A-bar^(k). If two feature dimensions have identical rows in every slice, Eq. (7) produces identical embeddings for them, so H = XΘ(T) contains only their summed contribution; the model cannot separate such features. The paper explicitly concedes in §5.3 that 'compressing arbitrary feature spaces into a fixed d_h leaves the original coordinates individually inaccessible to the decoder,' and the 66.10 vs 71.78 node-classification gap in Table 3 is consistent with this loss. The abstract claims transferability to 'any downstream graph,' but no condition is stated under which a task is compatible with Gram-alignment features, nor is it proved that the tested tasks satisfy such a condition. Either a formal compatibility statement should be added, or the transfer claim must be weakened. This is the main theoretical gap.","section":"§3.1, Eq. (4); §5.3"},{"comment":"The text says Sigil-nc 'is competitive with node-classification specific GFMs' and 'ranks best or second-best on 10 of 26 datasets,' but the aggregate is 66.10% versus 71.78% for RGVT and 71.57% for TAG, and RGVT reaches top-2 on 22 of 26 datasets. Since RGVT is pretrained on ogbn-arxiv while Sigil-nc is pretrained on Cora (footnote to Table 3), the comparison is also not controlled for pretraining data. Please add per-dataset or paired statistical tests and either weaken the 'competitive' claim or include a Sigil-nc variant pretrained on the same source graph as the strongest baseline.","section":"§5.3, Table 3"}],"minor_comments":[{"comment":"The sentence 'it is the tightest of the three on Sigil-lp wins on C.ele, is second on USAir and very competitive on NS' is grammatically broken; please rephrase.","section":"§5.1, non-attributed results"},{"comment":"The method name is written 'SIGIL' in the title and 'Sigil' throughout the abstract and body; please standardize the capitalization.","section":"Title and abstract"},{"comment":"The claim that 'Sigil(0) reproduces ULTRA' is stated after numbers that differ by up to 2.1 MRR in the aggregate, and Appendix E shows larger per-dataset gaps (e.g., WN v1: 18.9 vs 64.8 MRR). Please qualify the reproduction claim and discuss the per-dataset variance.","section":"Table 2 and Appendix E"},{"comment":"The phrase 'which is an expected results and a key limitation for the GFM settings in advance' contains grammar errors; please revise.","section":"Appendix F, Figure 4a"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know. First, the core construction is real: SIGIL lifts any attributed graph to a graph over feature dimensions, with edges typed by propagation order and weighted by Gram alignments of aggregated edge features. I checked the formal parts — Theorem 1 (the relation graph of ULTRA is the discrete order-zero special case), Lemma 1, and Proposition 1 — and they are coherent. The symmetry arguments are standard but correctly executed. This is a new formalization; I don't know of prior work that makes the KGFM relation graph explicit as a Gram construction over one-hot relation indicators.\n\nSecond, the empirical headline is slightly ahead of the evidence. The single Cora-pretrained model does beat UniLP on the six attributed graphs it hasn't seen, which is a real result. But Table 1 also includes Cora itself as a target, TFMLinker is absent from the attributed block for lack of code, and the main link-prediction configuration uses order K=0, so the multi-order mechanism that differentiates SIGIL from a discrete relation graph is not what drives the flagship numbers. That's worth fixing before publication: drop Cora from the target set or mark it clearly, add a breakdown by order, and if possible release code and the TFMLinker comparison.\n\nThe stress-test concern about Gram summarization is fair but not fatal, and the paper is honest about it. Equation (4) collapses per-coordinate signal into pairwise alignment statistics; two features with identical alignment profiles are indistinguishable, and the node-classification gap (66.10 vs 71.78 for RGVT) is consistent with that loss. But the link-prediction results show that aggregate alignment can be enough for some tasks, and the paper never claims the compression is lossless. What's missing is a statement of which tasks are compatible with this summarization. That's a limitation, not an incoherence.\n\nThe KG experiment is the most convincing empirical part: Sigil(0) reproduces ULTRA within about 2 MRR, and the higher-order variants beat it on the entity-inductive group. The node-classification results are honestly reported as weaker.\n\nWho is this for? Anyone working on graph foundation models or knowledge graph reasoning. The unification claim is worth a serious look, and the paper deserves peer review with requests for code, a cleaned benchmark table, and a more careful statement of the summarization's limits.\n\nMy recommendation: send it out, but expect heavy revision.","headline":"A genuinely new lifting of attributed graphs to feature-interaction graphs, with a clean KGFM equivalence; the empirical claims are slightly ahead of the evidence, but this deserves serious referee time.","tokens_in":26584,"tokens_out":3152,"would_cite":true,"duration_ms":33370,"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":"The paper claims that any attributed graph can be lifted to a structural interaction graph over its feature dimensions, and that a relational message-passing network trained once on one graph yields fixed-dimension node representations…","keywords":["graph foundation models","fully-inductive link prediction","structural interaction graph","feature-space transfer","relational message passing","knowledge graph reasoning","node classification","permutation equivariance"],"falsifier":"Construct two graphs that have identical SIG tensors but different raw feature values with different task-relevant meanings, then check whether a Sigil model pretrained on one performs on the other at chance; if it does not transfer, the Gram-based summary has lost information the downstream task needs. A more specific test is to apply a feature-wise transformation on the target graph that leaves all pairwise alignment magnitudes unchanged and see whether link-prediction rankings change, since the SIG is meant to be the complete transferable description of the feature space.","tokens_in":25536,"feed_emoji":"🔗","tokens_out":9027,"duration_ms":85040,"temperature":0.7,"pith_summary":"Sigil addresses the obstacle that blocks graph foundation models: graphs arrive with different numbers of features, different feature meanings, and different label sets, so a model trained on one graph cannot usually even read another. The paper's proposal is to lift every graph to a structural interaction graph (SIG), whose nodes are the input feature dimensions and whose weighted, typed edges record how pairs of features align as signals propagate over the graph's connectivity. A relational message-passing network, with parameters acting only on a fixed hidden width, embeds those feature dimensions into one shared space and rewrites the original node features into a fixed dimension. The claim is that one model, pretrained on a single graph, then applies zero-shot to arbitrary unseen graphs for both link prediction and node classification, unifying regimes that the literature has treated separately.","feed_headline":"One Cora-trained model predicts links on every unseen attributed graph","feed_subtitle":"Sigil turns a graph's feature dimensions into a shared embedding space, then scores links without retraining.","key_machinery":"The structural interaction graph (SIG) is a tensor of Gram matrices $\\bar A^{(k)} = (\\bar X^{(k)})^\\top \\bar X^{(k)}$ over feature-dimension nodes, where $\\bar X^{(k)}$ is the per-node mean of pairwise feature interactions on edges after $k$ propagation steps; the directed variant produces four channel-pair slices per order. A relational message-passing update $\\Theta^{(t+1)} = \\Theta^{(t)} W_{\\mathrm{self}} + \\sum_k \\bar A^{(k)} \\Theta^{(t)} W_k$, with all parameters acting on the fixed hidden width $d_h$, produces feature embeddings, and $H = X\\Theta^{(T)}$ is the fixed-dimension node representation. The Gram matrices encode feature alignment, the stacking over orders encodes multi-hop connectivity, and node permutations cancel in the Gram product, giving the equivariance that makes the object transferable.","core_discovery":"The central claim is that pairwise Gram matrices of mean-aggregated edge features, computed after $k$-hop propagations, form a sufficient transferable description of a graph's feature space, and that relational message passing on the resulting tensor gives node representations whose dimension does not depend on the input graph. When the input features are one-hot relation indicators, the construction recovers exactly the relation graph of knowledge graph foundation models, making those models a discrete special case and adding multi-hop interaction orders they do not encode. A single model pretrained on Cora is reported to be the strongest fully-inductive link predictor on every attributed benchmark tested, to reproduce knowledge-graph reasoning results at order zero, and to yield a node encoder that ranks best or second-best on 10 of 26 node-classification benchmarks.","pith_inferences":["Beyond the paper: the SIG construction could be applied to edge features or temporal snapshots directly, which would give zero-shot transfer to graphs that have no node features at all, a regime the paper approaches only by synthesizing DeepWalk features.","Beyond the paper: the fixed-width compression is a deliberate bottleneck, so tasks that hinge on raw feature identity are the predictable failure mode; adding a compact per-feature descriptor alongside the alignment matrices is a testable repair.","Beyond the paper: the order-$K$ sweep behaves like a tunable receptive field with no per-graph training, so the optimal order could itself be predicted from local graph statistics, turning a design choice into a learnable property."],"forward_implications":["A single Sigil model pretrained on one attributed graph predicts links on any other attributed graph without retraining; the paper reports it as the strongest fully-inductive link predictor on all seven attributed benchmarks and competitive on four non-attributed ones.","Knowledge graph foundation models become a special case of the construction: with one-hot relation features at order zero the SIG equals the relation graph, and higher orders add multi-hop relation interactions those models lack.","Node classification on graphs with arbitrary label sets reduces to fitting a small decoder on fixed-width Sigil embeddings; this is competitive but not state of the art, with an average accuracy of 66.10 against 71.78 for the strongest specialized model.","Because the encoder parameters never depend on the input feature count, the same checkpoint can be deployed on any graph, making zero-shot deployment a matter of computing the SIG and running message passing."],"supporting_citations":[{"why":"Supplies the relation-graph construction and inductive benchmark that Theorem 1 reproduces and evaluates against.","marker":"Galkin et al. 2024"},{"why":"Defines fully-inductive node classification on arbitrary graphs and the analytical decoder Sigil pairs with.","marker":"Zhao et al. 2025"},{"why":"Supplies the equivariance recipe whose node, feature, and label symmetries Sigil is claimed to satisfy.","marker":"Finkelshtein et al. 2026"},{"why":"Supplies the relational message-passing update used as the SIG encoder.","marker":"Schlichtkrull et al. 2018"},{"why":"Supplies the NBFNet backbone used as the link-prediction decoder on the SIG embeddings.","marker":"Zhu et al. 2021"},{"why":"Supplies UniLP, the fully-inductive link-prediction baseline Sigil-lp must outperform on attributed graphs.","marker":"Dong et al. 2025"},{"why":"Formalizes fully-inductive node representation learning and supplies the RGVT baseline for node classification.","marker":"Lee et al. 2026"}],"fun_headline_variants":["SIGIL: one model predicts links on any unseen graph","From Cora to everywhere: a single GNN for all graphs","Feature dims become nodes: SIGIL unifies graph learning","Pretrain once on Cora, predict links on anything"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pairwise alignment statistics of propagated feature values, summarized in the Gram matrices, preserve enough task-relevant information for downstream prediction, because once the original features are mapped through the SIG the individual coordinates are no longer accessible.","fun_headline_variants_meta":{"raw":{"variants":["SIGIL: one model predicts links on any unseen graph","From Cora to everywhere: a single GNN for all graphs","Feature dims become nodes: SIGIL unifies graph learning","Pretrain once on Cora, predict links on anything"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1208,"prompt_tokens":921,"completion_tokens":287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":216}},"tokens_in":537,"tokens_out":287,"duration_ms":3782,"temperature":1.0,"reasoning_tokens":216,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:32:09.581986+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct two graphs that have identical SIG tensors but different raw feature values with different task-relevant meanings, then check whether a Sigil model pretrained on one performs on the other at chance; if it does not transfer, the Gram-based summary has lost information the downstream task needs. A more specific test is to apply a feature-wise transformation on the target graph that leaves all pairwise alignment magnitudes unchanged and see whether link-prediction rankings change, since the SIG is meant to be the complete transferable description of the feature space.","supporting_citations":[],"review_version":1}