{"id":"eca14dc8-51f8-4c2b-b05a-03fc9f41a839","arxiv_id":"2607.28980","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ProGFM transfers graph knowledge across domains by learning a prototype bank of per-edge, per-dimension propagation strengths and using them to modulate message passing on unseen graphs.","lead":"This paper proposes ProGFM, a graph model that treats the way different edges carry different feature dimensions as a transferable 'knowledge unit' across domains. It learns a bank of propagation prototypes on source graphs and applies them to new graphs without fine-tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-domain SVD leaves feature dimensions unaligned, so the shared linear transform and prototype bank may not actually transfer across domains.","rationale":"The strongest claim is that propagation relationships between edges and feature dimensions are transferable knowledge units that remain valid across domains with different feature semantics. This requires that the feature dimension index k is a meaningful, comparable slot across domains. The paper's SVD dimensionality alignment (Eq. 3) produces per-domain coordinate systems with no cross-domain correspondence; SVD gives no semantic alignment and even the sign of each component is arbitrary. Therefore, the scalar s_ij,k computed from these coordinates is not a domain-invariant unit, and the shared weight matrix W in Eq. 10 is applied to inputs with incompatible coordinate systems. This is the most load-bearing weakness because it attacks the foundation of the proposed transfer mechanism, not just an experimental detail. The concrete sign-flip test would settle whether this theoretical concern manifests empirically: if the model is invariant to SVD sign ambiguity, then the propagation knowledge is coordinate-robust; if not, the claimed semantics-free transfer is undermined. The reader's weakest_assumption identified essentially the same issue, and the conditional verdict is appropriate until this is tested. I do not raise the missing ProGFM row in Table 4 as the primary concern because, while a serious reporting flaw, it is fixable and does not directly invalidate the conceptual premise; it would be secondary evidence if the main concern is resolved.","tokens_in":17653,"tokens_out":9092,"duration_ms":490283,"concrete_test":"After pre-training, take a target graph and multiply its SVD-projected feature matrix \\tilde X (and hence H^(0)) by a random diagonal matrix with ±1 entries (random sign flips per dimension), repeated over 20 seeds. Since s_ij,k uses absolute differences, the propagation relationships are unchanged; only the input to the shared W changes. If target accuracy varies by more than the reported std, the model depends on the arbitrary SVD sign convention, confirming the alignment concern. If accuracy is stable, the concern is refuted.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The transfer mechanism hinges on the claim that propagation relationships s_ij,k are domain-semantics-free and can be clustered into a universal prototype bank. However, Section 3.2 uses SVD independently per graph to project features into a common d-dimensional space. SVD bases are dataset-specific and only defined up to sign (and arbitrary rotation in degenerate cases); dimension k in one graph is not the same latent direction as dimension k in another, and no alignment is applied. Thus Eq. 4's s_ij,k values from different domains are not comparable scalars despite both lying in [0,1]. Moreover, the shared transformation matrix W^(l) in Eq. 10 is applied to inputs living in different coordinate systems, so the same learned weights cannot correctly process source and target representations unless the bases are aligned. The paper explicitly declines to enforce semantic alignment (Sec. 3.1), but then treats the coordinate index k as a universal slot. If the SVD basis is arbitrary, both the prototype matching (Eq. 14) and the shared W are ill-founded, and the 'transferable propagation knowledge' may reduce to a fixed scalar quantization of a domain-dependent coordinate system.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ProGFM, a graph foundation model that treats the relationship between each edge and each feature dimension as a transferable knowledge unit. It first projects node features of each graph into a common dimensionality via SVD, then computes a scalar propagation relationship s_ij,k = 1 - r_ij,k from the normalized feature difference across an edge. K-means clustering of these scalars across source graphs forms a prototype bank B = {(p_c, alpha_c)}. During message passing, each edge receives a propagation-strength vector by quantizing its current-layer propagation relationships to the nearest prototype; the message is z_ij \\odot W h_j. ProGFM is pre-trained with the SGRL objective on multiple source graphs and then transferred to unseen target graphs with all parameters frozen (zero-tuning), with predictions made by prototype-based classification using labeled target nodes. Experiments cover one-shot/few-shot node classification, subgraph classification, and graph classification.","tokens_in":17959,"tokens_out":9507,"duration_ms":83590,"significance":"If the proposed transfer mechanism is sound, treating propagation relationships (quantified by relative feature differences) as knowledge units is a genuinely different approach from feature/structure alignment and could be a useful building block for graph foundation models. The empirical study is broad: leave-one-graph-out node classification and subgraph classification with 500 few-shot tasks per target, plus graph classification across social and protein domains, with a range of GNN, self-supervised, and GFM baselines. The consistent improvements on most node/subgraph settings are encouraging. However, the paper does not yet establish the central claim because (i) per-graph SVD leaves the coordinate systems unaligned, so the 'domain-agnostic' quantities are not comparable across domains; (ii) the graph-classification table omits the proposed method; and (iii) the default K is selected on target data, which conflicts with the zero-tuning claim. These issues are fixable but are not merely editorial.","major_comments":[{"comment":"The transfer mechanism is built on a coordinate system that is not shared across domains. Eq. (3) aligns feature dimensionality with a per-graph SVD. SVD bases are dataset-specific and only defined up to sign (and arbitrary rotation in degenerate subspaces), so coordinate k in one graph is not the same latent direction as coordinate k in another. Nevertheless, Eq. (10) applies a shared W^(l) to these representations and Eq. (14) matches the scalar s^(l)_ij,k to a global prototype p_c. Thus both the prototype matching and the shared linear transformation presuppose that the coordinate index k is a universal slot, which the paper explicitly declines to ensure (Sec. 3.1). If the per-graph projections are not aligned, the values s_ij,k from different domains are not comparable scalars and the 'transferable propagation knowledge' reduces to a scalar quantization of domain-dependent coordinate","section":"§3.2–§3.4, Eqs. (3), (10), (14)"},{"comment":"The graph-classification experiment is not reported as claimed. Table 4 contains no row labelled ProGFM; the row labelled 'FLT' is not defined in §4.2 or anywhere else, and §4.6 states that 'ProGFM achieves the best performance on all evaluated graph classification datasets.' As written, the reader cannot verify the central graph-classification claim. The table needs a correctly labelled ProGFM row (or the names must be fixed) and the accompanying text must refer to the row actually present.","section":"Table 4 / §4.6"},{"comment":"The zero-tuning claim is weakened by the selection of K. The paper says 'Based on the overall performance across different datasets, we set K=100 as the default configuration' (Sec. 4.8), where the 'different datasets' appear to be the target datasets used in the sensitivity analysis. Under the zero-tuning protocol of Eq. (21), target-domain data should not be used to choose model configuration. Please state how K is selected using only source-domain information, or explicitly report K as a fixed architectural choice with sensitivity analysis as a post-hoc diagnostic; otherwise the comparison is not a fair zero-tuning evaluation.","section":"§4.8 / Fig. 3"}],"minor_comments":[{"comment":"The ablation and sensitivity figures do not include error bars or significance statements. Given that Tables 1–2 report standard deviations over 500 tasks, the figures should include them as well.","section":"Fig. 2 / Fig. 3"},{"comment":"Notation: Eq. (8)/(9) denote learnable strengths as alpha_c, while Eq. (15) writes B_c=(p_c,a_c). Use a consistent symbol.","section":"Eq. (15)"},{"comment":"Typo: 'where where z^(l)_ij denotes' should be 'where z^(l)_ij denotes'.","section":"Eq. (17)"},{"comment":"The TIG row has a formatting issue: '45.50±8.7651.79±9.98' lacks a separator between the CiteSeer and PubMed entries.","section":"Table 1"},{"comment":"The 'w/o Learnable Strength' variant initializes fixed propagation strengths from N(1,1); since this distribution puts probability mass on negative values, please clarify whether negative propagation strengths are intended or should be restricted to positive values.","section":"§4.7"}],"recommendation":"major_revision","confidential_remarks":"The undefined 'FLT' row in Table 4 and the missing ProGFM row may indicate a formatting or naming error rather than a substantive omission, but it must be corrected before the graph-classification claims can be evaluated. The SVD-alignment issue is the main technical risk; if the authors cannot fix it, the central contribution is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: ProGFM is worth a serious look. The central idea—treat per-edge, per-dimension propagation strengths as transferable knowledge units, learned as a prototype bank—is genuinely new to the GFM literature and clearly stated. The mechanism is simple enough to reproduce. On node and subgraph classification it beats the baselines on most datasets, and the gains look consistent. I credit that.\n\nThe soft spots, in proportion. First, the SVD dimensionality alignment (Sec. 3.2) is likely the load-bearing weakness. Each graph is projected by its own SVD, so coordinate k in one domain is not the same latent direction as coordinate k in another. The paper explicitly declines semantic alignment, then treats dimension k as a universal slot in s_ij,k, the prototype matching (Eq. 14), and the shared weight matrix W^(l) (Eq. 10). A single linear transform applied across differently-aligned coordinate systems isn't well-founded. If the SVD bases are arbitrary, the 'transferable knowledge' may be an artifact of the shared scalar quantization rather than a genuine cross-domain invariant. This deserves a direct response: either align the bases, or argue why per-dimension scalars remain comparable without alignment. I suspect the empirical results partly rescue the method in practice, but the theory as written doesn't close this gap.\n\nSecond, the graph-classification evidence is not in the paper. Table 4 lists GCN, SCR, TFSGFM, TIG, and an undefined 'FLT'; ProGFM is absent. Section 4.6 claims ProGFM achieves best results on all datasets, but the table shows no such row. That's either a typo (FLT might be ProGFM) or a missing experiment. Either way, it must be fixed before the claim is credible. Third, the ablation and sensitivity analyses (Figures 2 and 3) have no error bars or numeric values; the K=100 choice appears to be made after looking at target datasets. Minor, but worth tightening. The prototype count being tuned on the target is a mild leak.\n\nThe citation pattern is okay; the relative-difference signal is known (DMP, cited as [42]), and the vocabulary idea appears in GFT/RiemannGFM, but the specific combination is new. The paper is honest about its own limitations; it doesn't overclaim much beyond the typo.\n\nWho this is for: anyone working on graph foundation models or cross-domain transfer. It deserves a serious referee—send it out, but expect major revision. I would not cite it in its current form, but I'd revisit after the alignment issue is handled.","headline":"A clean idea with solid node-level results, but the transfer story has an unaddressed SVD alignment hole and the graph-classification table doesn't include ProGFM.","tokens_in":18416,"tokens_out":2634,"would_cite":false,"duration_ms":23182,"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":"Propagation relationships between edges and feature dimensions are transferable knowledge units that let a graph model generalize without fine-tuning.","keywords":["graph foundation models","knowledge transfer","propagation relationships","prototype bank","message passing","cross-domain generalization","self-supervised pre-training","relative feature difference"],"falsifier":"Synthesize two graphs with identical edge-wise relative feature differences but opposite label-relationship patterns: in one, connected nodes with similar feature values share labels; in the other, connected nodes with dissimilar values share labels. Pre-train ProGFM on the first, freeze it, and evaluate on the second. If the model transfers without catastrophic loss, the relative-difference assumption survives; if accuracy collapses to chance, the propagation-relationship claim is refuted.","tokens_in":17558,"feed_emoji":"🔁","tokens_out":8118,"duration_ms":60427,"temperature":0.7,"pith_summary":"This paper argues that the missing transferable unit in graph foundation models is neither a token nor a patch but the propagation relationship between an edge and a feature dimension: how similar or dissimilar the two connected nodes are along that dimension. The authors propose ProGFM, which encodes each edge-dimension pair as a scalar derived from the relative feature difference, clusters these scalars from source graphs into a prototype bank, and attaches a learnable propagation strength to each prototype. A propagation-aware message passing rule then modulates each feature dimension of each neighbor's message by the matched strength. The paper's claim is that this bank of propagation patterns, once learned on source domains, transfers directly to unseen graph domains even when feature semantics and structures differ, because the patterns depend on relative differences rather than absolute meanings.","feed_headline":"Propagation patterns, not features, transfer across graph domains","feed_subtitle":"A frozen prototype bank of edge-dimension propagation strengths generalizes to unseen graphs with no fine-tuning.","key_machinery":"The central object is the propagation relationship scalar s_{ij,k}=1−r_{ij,k}, where r_{ij,k} is the relative feature difference between two connected nodes on feature dimension k. This scalar carries whether an edge tends to preserve or suppress information along that dimension: values near 1 mean connected nodes are similar on that dimension (homophily-like), values near 0 mean they are far apart. The prototype bank, built by K-means clustering of these scalars across source graphs, is a set of pairs (p_c, alpha_c) where p_c is a representative propagation relationship and alpha_c a learned propagation strength. The propagation-aware message passing in Eq. 18 does the main work: for each e","core_discovery":"The central claim is that propagation relationships — defined as s_{ij,k} = 1 − r_{ij,k}, where r is the normalized absolute feature difference between connected nodes on dimension k — can serve as domain-agnostic transferable knowledge units. ProGFM first projects all graphs to a shared feature dimensionality via SVD, then collects s-values across all edges and dimensions from pre-training graphs, clusters them with K-means into a propagation relationship prototype bank, and associates each prototype with a learnable scalar propagation strength. At each layer, the model recomputes the s-values in the current message space, matches each to its nearest prototype, retrieves the corresponding s","pith_inferences":["A natural extension would be to replace the hard nearest-prototype lookup with a continuous learned function from s-values to strengths, eliminating potential discontinuities at prototype boundaries.","The learned strengths could be interpreted as a data-driven notion of per-dimension homophily across domains; inspecting them might reveal which feature dimensions are consistently smoothing vs. sharpening across graphs.","The same logic suggests other relative measures — rank differences, normalized dot products, or angle-based similarities — could serve as alternative transferable units; comparing them would map the boundary of the claim.","The paper's zero-tuning protocol is strict; a lighter-weight consequence is that even partially updating only the prototype strengths could close most of the gap to full fine-tuning, which the ProGFM-tuning variant already hints at."],"forward_implications":["A frozen pre-trained model can be applied to an unseen domain with only labeled class prototypes, no gradient updates.","Propagation knowledge transfers across task granularities, from node classification pre-training to graph classification.","Each edge can multiply each feature dimension of its neighbor's message by a learned strength, enabling fine-grained adaptive propagation that fixed rules cannot express.","Only dimensionality alignment, not semantic alignment, is needed across domains, since propagation relationships are computed from relative differences."],"fun_headline_variants":["Propagation prototypes unlock cross-domain graph learning","Graph foundation models via transferable propagation knowledge","Edge-feature pairs as transferable units for graphs","Beyond alignment: propagation patterns generalize","ProGFM: Adaptive propagation for unseen graph domains"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim stands or falls on the assumption that equal relative feature differences between connected nodes always entail equal propagation behavior, so a prototype bank learned from source-domain differences stays valid on target domains after SVD alignment and across completely different feature semantics.","fun_headline_variants_meta":{"raw":{"variants":["Propagation prototypes unlock cross-domain graph learning","Graph foundation models via transferable propagation knowledge","Edge-feature pairs as transferable units for graphs","Beyond alignment: propagation patterns generalize","ProGFM: Adaptive propagation for unseen graph domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000341,"raw_usage":{"total_tokens":1715,"prompt_tokens":741,"completion_tokens":974,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":906}},"tokens_in":485,"tokens_out":974,"duration_ms":8628,"temperature":1.0,"reasoning_tokens":906,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T16:00:32.282990+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize two graphs with identical edge-wise relative feature differences but opposite label-relationship patterns: in one, connected nodes with similar feature values share labels; in the other, connected nodes with dissimilar values share labels. Pre-train ProGFM on the first, freeze it, and evaluate on the second. If the model transfers without catastrophic loss, the relative-difference assumption survives; if accuracy collapses to chance, the propagation-relationship claim is refuted.","supporting_citations":[],"review_version":1}