{"id":"7a8203c3-8333-4ea7-be02-7b4aa70d9246","arxiv_id":"2607.26533","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"AgentGFM treats every node as an agent that controls source reception, signal channel, and halting during graph propagation, and shows improved cross-domain transfer in 1-shot node and graph classification.","lead":"AgentGFM lets each node in a graph decide how it gathers and forwards information during message passing, using a shared policy trained across many graphs. It reports strong cross-domain 1-shot classification results on ten node and five graph benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not established: many reported margins are within noise and Tmax is selected on the test targets.","rationale":"The reader identified feature alignment via truncated SVD as the weakest assumption. That is a genuine reproducibility concern: the node datasets have incompatible feature dimensions (Table 7), and Section 4.1 does not specify whether SVD is applied per graph or jointly, so the 'shared space' may be only a common dimensionality rather than a common semantic coordinate system. However, even a flawed alignment could still allow transfer if the learned information-flow policy is largely topology-driven, and a corrected alignment might change the numbers without disproving the mechanism. The more load-bearing issue for the paper's central claim is the quality of the experimental evidence itself. The SOTA statement is an empirical claim, and the reported tables do not establish that the improvements over the strongest baselines are real rather than sampling noise. The lack of paired significance tests is especially important because all methods are evaluated on the same 100 support-set draws. The additional test-set selection of Tmax in Table 4 compounds the risk that the reported configuration is optimistic. These problems do not prove the method is wrong, but they do mean the central claim is not yet supported. The reader's conditional verdict already requires code and significance testing, so the verdict should remain CONDITIONAL rather than being moved to acceptance or rejection.","tokens_in":19287,"tokens_out":12658,"duration_ms":125789,"concrete_test":"Request the per-run 1-shot accuracies for AgentGFM and the strongest baseline on each of the ten node-classification datasets, then compute paired bootstrap 95% confidence intervals for the per-dataset difference. If fewer than six of ten intervals exclude zero, the claimed SOTA advantage is not supported by the reported experimental protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that AgentGFM achieves state-of-the-art cross-domain 1-shot transfer rests on Tables 1–3, but the paper reports only mean accuracy and standard deviation over 100 support-set samples with no significance testing. Several headline wins are very small: PubMed 52.96 vs 52.16 for GraphAny, Chameleon 30.54 vs 30.32 for GCOPE, and Squirrel 24.20 vs 23.77 for GCOPE. On CiteSeer, AgentGFM is 6.93 points worse than R-GFM (44.32 vs 51.25), despite the claim of being 'competitive'. Because all methods use the same 100 support-set draws, the comparisons are paired; overlapping standard deviations do not tell us whether the differences are significant, and a paired test could easily show that several apparent wins are noise. In addition, Section 5.3.4 selects Tmax = 5 based on test-set accuracy in Table 4, meaning the reported configuration is chosen on the evaluation targets. This can inflate the reported performance. Without raw per-run results, a validation-based hyperparameter choice, or released code, the empirical basis for the SOTA claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes AgentGFM, a graph foundation model that treats each node as an agent executing a shared policy for node-specific information-flow control. The model aligns node features via truncated SVD, then runs a recurrent predict–act–observe–correct process in which each node decides source reception, low-/high-frequency channel selection, and gain-aware halting. The encoder is pretrained on source graphs with masked reconstruction, prediction-observation alignment, and three regularization objectives, then evaluated under leave-one-dataset-out cross-domain 1-shot protocols on node, graph, and large-scale benchmarks. The paper claims state-of-the-art average ranking and consistent gains over existing GFMs, supported by ablations and a mechanism-level case study.","tokens_in":19640,"tokens_out":7053,"duration_ms":55841,"significance":"If the empirical claims hold, AgentGFM would be a meaningful step toward topology-adaptive, transferable graph representations: node-specific propagation decisions under a single shared policy are a plausible alternative to fixed message-passing schemes, and the formulation is clean and mechanism-level. The paper provides a detailed description of the policy components, comprehensive coverage of ten node datasets and five graph datasets, and ablations isolating each design choice, plus explicit hyperparameter settings and an honest discussion of computational overhead. However, the central SOTA claim rests on comparisons with small margins, without significance testing, and on a rollout horizon selected on the target benchmarks; these issues must be addressed before the claim is convincing.","major_comments":[{"comment":"The default maximum rollout horizon Tmax = 5 is selected by comparing average rank on the same ten target datasets used for evaluation. This is test-set hyperparameter tuning: the configuration that produced the reported results was chosen to maximize performance on the benchmarks it is then evaluated on. This can inflate the reported gains and weakens the transfer claim. Please either fix Tmax a priori based on validation sources, or present Table 4 as a separate sensitivity analysis and run the main comparison with a pre-specified value; if such a value is chosen from source-domain validation alone, details of that protocol should be reported.","section":"§5.3.4 (Table 4)"},{"comment":"The feature-alignment step described as 'aligns their node features into a shared space using truncated SVD' is never specified: the paper does not state what matrix is decomposed, how the 50-dimensional shared space is computed, or whether the SVD is fit on source graphs only or on a combination of source and target graphs. If target-graph features are included in the SVD fit, the leave-one-dataset-out protocol is violated; if not, the procedure must be described in full to be reproducible. Given that this alignment is the first step of the cross-domain pipeline and is load-bearing for transfer, please provide the exact algorithm and a sensitivity analysis of the shared dimension.","section":"§4.1 and Appendix B.2"},{"comment":"The evaluation reports only means and standard deviations over 100 support-set samples. Since all methods are evaluated under identical support-set draws, paired significance tests are the appropriate tool, and several headline wins are small relative to the reported variability (e.g., PubMed 52.96 vs 52.16 for GraphAny; Chameleon 30.54 vs 30.32 for GCOPE; Squirrel 24.20 vs 23.77 for GCOPE; MUTAG 59.52 vs 58.30 for R-GFM). Without such tests, the statements 'obtains the best performance on nine datasets' and 'ranks first on all five datasets' are not supported. At minimum, report paired bootstrap or permutation test results and effect sizes; also clarify how 'competitive' on CiteSeer (44.32 vs 51.25 for R-GFM) is assessed given that the margin is 6.93 points in favor of the baseline.","section":"§5.1.2, Tables 1–2"},{"comment":"The 'prediction–observation' feedback is an internal self-consistency loop: the predicted observation b_o_i and the actual observation o_i are both generated by the same network (the predicted one by f_pred and the actual one by the model's own rollout), and L_pred in Eq. (28) trains the former to match the latter. The discrepancy in Eq. (23) therefore does not measure error with respect to any external ground truth, so the claim that it 'reflects the mismatch between the node's current expectation and the information returned by the graph' is only true in a model-internal sense. This is not circular for the transfer claim, which is tested on held-out datasets, but it should be acknowledged and its effect on the mechanism should be analyzed, e.g., by checking whether the correction gate in Eq. (24) is sensitive to the discrepancy or simply learns a constant.","section":"§4.3, §4.6, Eq. (28)"}],"minor_comments":[{"comment":"In the variant name, 'Gain-A ware Halt' should be 'Gain-Aware Halting'.","section":"§5.3.1"},{"comment":"The large-scale evaluation uses a different pretraining corpus (eight fixed source graphs) than the leave-one-out protocol of §5.2.1; while each target is still unseen, the source set is not the same across the two tables, so the results are not directly comparable and this discrepancy should be noted.","section":"§5.2.3 (Table 3)"},{"comment":"The values of λ_high and θ_0 are reported in Table 10 but the paper does not describe how they were chosen; a brief explanation for at least these two hyperparameters would improve reproducibility.","section":"Appendix B.2 / Table 10"},{"comment":"The source-reception score γ^t_{u→i} is a sigmoid and is thus not normalized over the incoming neighbors; the paper should clarify whether this is intentional and how it interacts with the normalized weights α in Eq. (12).","section":"§4.4, Eq. (10)"},{"comment":"The claim that 82.42% of nodes halt by the second rollout step would be more informative if accompanied by the standard deviation across datasets, since the distribution is described as varying by dataset.","section":"§5.3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and the method is well motivated. The main concern is the empirical support for the SOTA claim: the combination of test-set hyperparameter selection and absence of significance testing is a serious issue at a venue requiring rigorous empirical validation. No code is released, and the SVD alignment is underspecified. I recommend major revision with a focus on these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: AgentGFM is a genuine new mechanism for cross-graph transfer—node-level control of information flow under a shared policy—but the empirical SOTA claim is not yet established, because several headline wins are within noise and the main hyperparameter is tuned on the test targets.\n\nWhat is new: the combination of source-reception gating, low/high-frequency channel selection, and gain-aware halting, all driven by a single shared policy trained with a predict-act-observe-correct loop. That is distinct from Policy-GNN and GDPNet, which adapt propagation but are task-specific, and from GraphAny/MoE-style GFMs, which adapt components but not propagation. The math is coherent; equations 10–26 hang together. Credit is due for the evaluation breadth: ten node-level datasets, five graph-level, three large-scale, plus ablations isolating each mechanism, and an honest efficiency table. The large-scale protocol differs from the leave-one-out scheme, but that is a minor inconsistency, not a fatal one.\n\nThe soft spots are empirical. No code is released, so nothing is independently checkable. Table 4 picks T_max=5 by best average rank on the same target benchmarks that are later reported; that is test-set tuning. Several SOTA margins are tiny: PubMed 52.96 vs 52.16 for GraphAny, Chameleon 30.54 vs 30.32, Squirrel 24.20 vs 23.77. With standard deviations of 4–9 points, those differences are plausibly noise; a paired test over the shared 100 support-set draws would settle it. The aggregate average rank of 1.10 is the strongest evidence, and it needs that significance support. Also, \"competitive\" on CiteSeer is a stretch when the gap to R-GFM is 7 points.\n\nThe central transfer claim is not circular—target graphs are held out during pretraining. The predict-observe-correct loop trains the predictor to match the model's own observations (Eq. 28), which is a self-consistency constraint rather than a leakage. The truncated-SVD feature alignment is a big assumption, but it appears to work across very different feature spaces.\n\nWho this is for: anyone working on graph foundation models or adaptive propagation mechanisms. The idea is worth engaging with, and the paper deserves a serious referee; it should not be rejected outright. I'd send it to review with a requirement for code, significance testing, and a validation-based hyperparameter choice.","headline":"Real conceptual step in graph foundation models, but the SOTA claim rests on tiny margins and test-set-tuned hyperparameters; worth serious refereeing, not acceptance as-is.","tokens_in":20105,"tokens_out":4212,"would_cite":true,"duration_ms":35745,"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":"AgentGFM's central claim is that treating each node as an agent that controls its own information flow—deciding which neighbors to hear, which signal channel to use, and when to stop—under one shared policy, transfers across unseen graphs…","keywords":["graph foundation models","cross-domain transfer","information-flow control","node agents","adaptive propagation","one-shot learning","heterophilous graphs","halting policy"],"falsifier":"Replace the truncated-SVD feature alignment in Section 4.1 with a fixed random projection of the same dimension, keep everything else identical, and rerun the leave-one-dataset-out one-shot protocol; if accuracy does not drop on heterophilous targets, the alignment is not carrying the transfer, and if it collapses, the premise is confirmed.","tokens_in":19107,"feed_emoji":"🕸️","tokens_out":3947,"duration_ms":33134,"temperature":0.7,"pith_summary":"The paper proposes that a graph foundation model can transfer across unseen graphs if each node is treated as an agent that decides, at each step, where to receive information, which frequency channel to use, and when to stop, all under a single shared trainable policy. It claims this node-level information-flow control outperforms fixed propagation schemes in cross-domain one-shot classification. If true, it means adaptive propagation decisions can generalize rather than being tied to a specific graph distribution, which matters for real graphs whose local structures vary widely.","feed_headline":"Nodes that steer their own information flow win cross-domain tests","feed_subtitle":"AgentGFM treats each node as an agent choosing sources, signal channels, and a stopping time under one shared policy.","key_machinery":"The central object is the shared information-flow policy that each node executes locally. It produces edge-wise source-reception strengths γ, edge-wise channel gates ρ interpolating low- and high-frequency messages, and a node-wise halting score κ whose discrete decision is trained with a straight-through estimator. A forwarding budget b and prediction–observation feedback cues ε condition these decisions, while a feedback-gated correction T merges the predicted and actual observations; the whole policy is trained with masked attribute reconstruction, prediction–observation alignment, and three regularizers that discourage indiscriminate reception, excessive transmission, and overlong rollouts.","core_discovery":"AgentGFM establishes that information-flow control can be realized as node-agent decision making under a shared end-to-end trainable policy. Each node runs a predict–act–observe–correct loop: it predicts the contextual information its current state should receive, acts through three decisions—source reception, signal-channel selection, and gain-aware node-wise halting—receives an observation from the graph, and corrects its state by the discrepancy between prediction and observation. The paper reports that this design achieves the best average rank of 1.10 across ten node-classification datasets, the best average rank of 1.00 across five graph-classification datasets, and strong results on large-scale graphs, with particularly clear gains on heterophilous graphs where fixed propagation rules absorb misleading neighbor signals.","pith_inferences":["Beyond the paper, a direct test of the weakest link would be to replace the truncated-SVD feature alignment with a learned per-domain encoder; if transfer accuracy holds, the alignment is not the limiting factor, and if it collapses, the SVD step is doing more work than the paper credits.","Since all node agents share one policy but never exchange explicit messages about their decisions, a natural extension is to add lightweight agent-to-agent coordination or a global context token that conditions the shared policy on graph-level statistics.","The information-flow decisions are trained only through reconstruction and alignment losses; adding a downstream-aware objective during pretraining might sharpen source selection and halting behavior on target graphs.","Because over 82% of nodes halt by the second rollout step, an adaptive per-graph horizon could cut the rollout cost further without hurting accuracy, especially on large graphs."],"forward_implications":["Cross-domain one-shot transfer improves on heterophilous graphs, where fixed message-passing schemes are most likely to absorb misleading neighbor signals.","Node-specific halting saves computation because most nodes stop early, and the method never constructs dense higher-order neighborhoods, keeping complexity at O(RT|E|d) for bounded rollout rounds and horizons.","The same shared policy transfers across node-level and graph-level tasks, so the mechanism is not tied to a particular task head.","The prediction–observation discrepancy provides a self-supervised training signal that lets the model correct node states without downstream labels.","The learned halting and reception distributions differ across datasets, confirming that the policy produces node-specific trajectories rather than a uniform propagation rule."],"supporting_citations":[{"why":"Policy-GNN learns node-specific propagation iterations and is the adaptive-propagation baseline AgentGFM extends to cross-graph transfer.","marker":"[10]"},{"why":"GDPNet learns adaptive neighborhood selection, providing prior evidence that propagation can be conditioned on graph signals.","marker":"[28]"},{"why":"GraphMAE supplies the masked autoencoder reconstruction objective that AgentGFM adapts for self-supervised pretraining.","marker":"[7]"},{"why":"BRIDGE is a feature-alignment GFM baseline that AgentGFM must beat in cross-domain one-shot comparisons.","marker":"[35]"},{"why":"GraphAny is a mixture-of-experts GFM baseline that adapts encoders to graph structures and is a strong competitor in the reported benchmarks.","marker":"[38]"},{"why":"GFT learns transferable propagation patterns, representing the structural-adaptation line of GFM research that AgentGFM distinguishes itself from.","marker":"[24]"}],"fun_headline_variants":["Each node decides its own flow: AgentGFM tops graph benchmarks","Node agents choose propagation paths, beating fixed rules","AgentGFM: nodes steer info flow, set graph AI record","Autonomous nodes pick sources and channels, win graph tasks","Graph model lets nodes control their own signal routes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single truncated-SVD linear projection maps very different node feature spaces into one shared 50-dimensional space that still preserves enough structural signal for the shared policy to act on; if that alignment loses transfer-relevant information, the whole cross-domain pipeline degrades.","fun_headline_variants_meta":{"raw":{"variants":["Each node decides its own flow: AgentGFM tops graph benchmarks","Node agents choose propagation paths, beating fixed rules","AgentGFM: nodes steer info flow, set graph AI record","Autonomous nodes pick sources and channels, win graph tasks","Graph model lets nodes control their own signal routes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00048,"raw_usage":{"total_tokens":2380,"prompt_tokens":956,"completion_tokens":1424,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":1343}},"tokens_in":572,"tokens_out":1424,"duration_ms":9683,"temperature":1.0,"reasoning_tokens":1343,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:23:28.956020+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the truncated-SVD feature alignment in Section 4.1 with a fixed random projection of the same dimension, keep everything else identical, and rerun the leave-one-dataset-out one-shot protocol; if accuracy does not drop on heterophilous targets, the alignment is not carrying the transfer, and if it collapses, the premise is confirmed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GDPNet learns adaptive neighborhood selection, providing prior evidence that propagation can be conditioned on graph signals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GraphMAE supplies the masked autoencoder reconstruction objective that AgentGFM adapts for self-supervised pretraining."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BRIDGE is a feature-alignment GFM baseline that AgentGFM must beat in cross-domain one-shot comparisons."},{"cited_title":"Bronstein, and Jian Tang","cited_arxiv_id":null,"evidence_quote":"GraphAny is a mixture-of-experts GFM baseline that adapts encoders to graph structures and is a strong competitor in the reported benchmarks."},{"cited_title":"Chawla, Chuxu Zhang, and Yanfang Ye","cited_arxiv_id":null,"evidence_quote":"GFT learns transferable propagation patterns, representing the structural-adaptation line of GFM research that AgentGFM distinguishes itself from."}],"review_version":1}