{"id":"62ce56ba-9838-47b8-ac5b-b1269dfa569e","arxiv_id":"2501.15142","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DAGPrompT combines low-rank adaptation of a pre-trained GNN's weights and message passing with per-hop class prompts, and reports consistent accuracy gains over 14 baselines on few-shot node and graph classification, especially on heterophily graphs.","lead":"This paper proposes DAGPrompT, a method that adapts pre-trained graph neural networks to new tasks by lightly adjusting the network's internal weights and giving each network layer its own class-comparison prompt. On 10 graph benchmarks it reports accuracy gains of 1 to 6 points over 14 existing methods, largest on 'heterophily' graphs where connected nodes tend to carry different labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical SOTA claim may rest on hyperparameters selected against test labels; no validation split or runnable code is provided to rule this out.","rationale":"The reader's stated weakest assumption is that the rank-1 adjacency adaptation in Eq. 4 is too weak or collapses, making GLoRA's contribution vanish. The ablation study in Table 7 weakens that worry: removing GLoRA drops accuracy by only 1.9-2.5 points, and the remaining hop-specific prompting still beats every baseline on the largest-gain datasets. So the viability of the central claim does not hinge on the expressiveness of the rank-1 perturbation. The more fragile assumption is that the reported accuracies are honest estimates of expected performance. The paper never defines the data split protocol, never describes a validation set, and describes hyperparameter selection only as 'based on performance' through an 'end-to-end' grid search. In transductive node classification, with no held-out validation labels, the natural reading is that the grid was scored on the test nodes. That would invalidate the headline improvements, many of which are within one standard deviation of the runner-up. The theoretical section does not rescue this: Theorem 1 is a generic Rademacher bound with no argument tying it specifically to GLoRA's parameter count or to the actual optimization procedure. The correct disposition is still CONDITIONAL rather than REJECT, because the paper may have used a validation split that is simply underreported; the missing code and split details must be supplied before the empirical claim can be accepted. This is why I keep the reader's CONDITIONAL verdict unchanged while disagreeing on which assumption is the weakest link.","tokens_in":22571,"tokens_out":4618,"duration_ms":45591,"concrete_test":"Obtain the released code (the paper currently gives only 'GitHub' with no URL or commit) and rerun Table 3 with hyperparameters alpha in {0.1,0.3,0.5,0.7,0.9}, r in {8,16,32}, learning rate, and weight decay selected on a held-out validation split (e.g., a random 10% subset of the labeled training nodes) instead of the test/transductive labels, using the same protocol for every baseline. If DAGPrompT's advantage over GCOPE and H2GCN shrinks below the reported margins or reverses on several datasets, the SOTA claim is an artifact of test-set selection; if the margins hold, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Many structural risks can be set aside: even if the rank-1 adjacency adaptation in Eq. 4 collapses, Table 7 shows DAGPrompT without GLoRA still beats the best baseline on Texas (78.12 vs 75.85), so the hop-specific prompts alone would often carry the method. The load-bearing weakness is instead the experimental protocol. Section 5.2 states 'Hyperparameters are selected based on performance' via an 'end-to-end' grid search, and Section 5.1 describes no held-out validation split. If alpha, r, learning rate, and weight decay were chosen using the downstream test/transductive labels, every entry in Tables 3 and 4 is optimistically biased. That matters because the headline gains are often small relative to variance: Texas 5-shot is 80.64±3.75 vs GCOPE 75.85±2.36, Cornell 5-shot is 84.09±1.57 vs GPR-GNN 79.73±3.85, and several 10-shot gains are 1-2 points. Test-set selection could easily produce exactly this pattern. The paper also gives no working code URL or commit hash, so the reader cannot currently verify the search procedure. This concern attacks the central claim directly: if the numbers are selection artifacts, the 'new state-of-the-art' conclusion is unsupported, regardless of how expressive GLoRA is.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DAGPrompT, a graph prompt tuning method for few-shot node and graph classification, with two components: (i) GLoRA, a low-rank adaptation of the GNN projection matrices and the message-passing adjacency matrix (Eq. 4), and (ii) hop-specific layer-wise class prompts (Eqs. 5-8) that combine per-hop similarity scores with learned coefficients. The method is evaluated on 10 datasets against 14 baselines, reporting state-of-the-art accuracy with improvements up to 4.79% on node classification (Table 3) and up to 5.89% on graph classification (Table 4), along with ablations (Table 7), transfer experiments (Table 5), efficiency analysis (Table 6), and additional experiments in the appendix.","tokens_in":22752,"tokens_out":3563,"duration_ms":31584,"significance":"If the empirical results are taken at face value, the paper makes a useful contribution: it extends graph prompting to heterophily graphs, a setting where prior prompting methods freeze the encoder and underperform, and it does so with a small number of tunable parameters (about 6.4K on Chameleon) while maintaining competitive efficiency. The ablations give some evidence that both the low-rank adaptation and the hop-specific prompts contribute, and the transfer experiment (Table 5) shows a clear gain from cross-domain pre-training. The code is announced as available at GitHub, and the appendix provides extended experiments with a GAT backbone, full-shot settings, and parameter analyses. However, the central empirical claim currently rests on an incompletely described hyperparameter-selection procedure, and several reported gains are within one standard deviation of the best baselines, which makes the headline result not yet fully supported.","major_comments":[{"comment":"The paper states that 'Hyperparameters are selected based on performance' using an 'end-to-end' grid search, but it does not describe any held-out validation split or nested cross-validation. If the hyperparameters (alpha, rank r, learning rate, weight decay, and possibly temperature) were selected on the downstream test labels or transductive labels, every entry in Tables 3 and 4 would be optimistically biased. Since the paper's central claim is state-of-the-art accuracy, please specify the exact selection protocol, report the chosen hyperparameter values per dataset, and provide the code or a commit hash so that the protocol can be verified.","section":"Section 5.2"},{"comment":"Several headline improvements are within or close to one standard deviation of the best baseline, e.g., Texas 5-shot (80.64±3.75 vs GCOPE 75.85±2.36), Cornell 10-shot (85.13±2.08 vs GPR-GNN 82.87±4.02), and Wisconsin 5-shot (73.86±2.45 vs GCOPE 71.45±2.86). The paper does not report the number of seeds, significance tests, or confidence intervals. The claim of 'consistently outperforming' would be much more convincing with paired significance tests across the same splits and seeds, or at least a statement of how many runs were averaged.","section":"Table 3"},{"comment":"Appendix A states that for extremely large graphs P_A Q_A^T is reduced to a single edge-weight vector applied only to edges connected to nodes in the training set. This is a substantial restriction of the GLoRA message-passing adaptation, yet the paper does not analyze how such a local, scalar reweighting can reshape message passing for heterophily on large graphs such as Arxiv-year. Please discuss the expressiveness of this reduction, and clarify whether test nodes that are not adjacent to any training node receive any message-passing adaptation; if the reduction is used in the reported Arxiv-year experiments, its effect on the results should be evaluated explicitly.","section":"Appendix A"},{"comment":"Theorem 1 is a standard Rademacher generalization bound for a generic hypothesis class; it does not itself compare GLoRA with full-parameter fine-tuning. The assertion that 'low-rank adaptations like GLoRA reduce model complexity' is not instantiated: no bound on R_D(H) is derived for the specific hypothesis class of low-rank adapted GNNs, and the bound holds for any hypothesis class of bounded complexity. The claim that 'in few-shot settings, low-rank adaptation proves to be more effective' is therefore not proven by Theorem 1. Please either provide a specific complexity measure for the GLoRA parameterization or present the theoretical statement as a heuristic interpretation of the standard bound, with the experimental comparison in Section E.6 as the primary evidence.","section":"Section 4"}],"minor_comments":[{"comment":"The abstract reports an improvement 'up to 4.79%' while Section 1 claims 'up to 7.55%' and Table 4 shows a 5.89% improvement on Texas* for graph classification; please reconcile these numbers and state the setting for the maximum improvement.","section":"Abstract and Introduction"},{"comment":"In Eq. (8), the summation for tilde-S is over l=1 to L, while the layer collection in Eq. (5) and the loss in Eq. (9) are over l=0 to L. Please align the indexing or explain why the l=0 term is omitted in the final scoring.","section":"Eq. (8) and Eq. (9)"},{"comment":"In Algorithm 1, after computing H <- f(X, A; Theta), the line 'H <- matmul(H, A)' appears redundant or possibly erroneous, since message passing is already performed inside the GNN forward; please clarify or remove it.","section":"Algorithm 1"},{"comment":"Reference [6] lists only 'Diederik' as the author of the Adam paper; please use the full standard citation.","section":"References"},{"comment":"The appendix uses the name 'DAGPrompt' in Tables 8-10 and in Section E.2, while the main text uses 'DAGPrompT'; please standardize the naming throughout.","section":"Appendix E"},{"comment":"The paper states that 'Codes are available at GitHub' but does not provide a URL or commit hash throughout the manuscript; please include a working link.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the missing validation protocol for hyperparameter selection, which directly affects the credibility of the SOTA claim. The paper has a solid ablation and a sensible motivation, and the structural concerns about GLoRA are somewhat mitigated by the ablation showing that DAGPrompT without GLoRA still beats baselines on the three datasets tested. If the authors can provide a rigorous validation protocol and code release, the paper could be acceptable. I would not reject outright, because the protocol issue is fixable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know up front. DAGPrompT is a clean, sensible combination—low-rank adaptation of the GNN weights plus the adjacency, and hop-wise class prototypes with learned coefficients—and the empirical story is consistent: it beats the 14 baselines on all 10 datasets, with the largest margins on heterophily graphs. Second, the paper's central claim rests on a hyperparameter-selection protocol that is described too loosely to rule out test-set leakage. That is the load-bearing weakness, and it is real.\n\nWhat is actually new: GLoRA adapts both the projection matrices and the message-passing adjacency with low-rank factors (Eq. 4), and the hop-specific prompts (Eqs. 5–8) are a reasonable extension of prototype-based prompting. The ablations in Table 7 show each component earns its keep, and the GAT-backbone results in Appendix E.1 indicate the method is not backbone-specific. The ablation also shows that even without GLoRA, DAGPrompT beats the best baseline on Texas (78.12 vs 75.85), so the prompting alone carries much of the gain.\n\nThe soft spots are real but mostly concentrated in experimental reporting. Section 5.2 says hyperparameters are \"selected based on performance\" via an end-to-end grid search, and Section 5.1 describes no held-out validation split. If alpha, rank, learning rate, and weight decay were chosen on test/transductive labels, every entry in Tables 3 and 4 is optimistically biased. Several headline wins are within one standard deviation of the runner-up, which makes this concern material. The paper also gives no working code URL or commit hash, so the search procedure cannot be checked. Separately, the reported improvements are inconsistent (abstract says 4.79, introduction says 7.55, Section 5.3 says 2.43 average), and \"pioneering\" overstates things given ProNoG is cited as a heterophily-aware prompting method. The theoretical section is a generic Rademacher bound that does not actually analyze the GLoRA hypothesis class; it should be labeled intuition or replaced. There is also a duplicated sentence in Section 3.2.1.\n\nNone of this makes the method implausible. The combination is new relative to the cited literature, and the ablation evidence is internally consistent. But the SOTA claim is not yet verified. The paper deserves a serious referee—the question is interesting and the empirical pattern is promising—but the review must press for a clear validation protocol, code release, and significance tests. If those come back clean, this is a solid WWW-level contribution.","headline":"A plausible and well-ablated prompting method whose SOTA claim is undercut by an under-specified hyperparameter-selection protocol; deserves review, not blind trust.","tokens_in":23369,"tokens_out":2469,"would_cite":false,"duration_ms":21427,"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 graph prompting can be pushed to heterophily graphs by low-rank encoder adaptation plus hop-specific class prompts.","keywords":["graph neural networks","graph prompting","few-shot learning","heterophily graphs","low-rank adaptation","prompt tuning","node classification","graph classification"],"falsifier":"Train DAGPrompT on Texas with 5-shot labels, then freeze the hop-prompt module and randomly permute the learned adjacency-adjustment weights while keeping all other parameters fixed; if accuracy does not drop materially, the claim that GLoRA's message-passing adaptation drives the heterophily gain is not supported. A complementary check is to test whether the learned edge weights assign higher scores to same-label edges than the unmodified adjacency does.","tokens_in":22249,"feed_emoji":"📈","tokens_out":8695,"duration_ms":73087,"temperature":0.7,"pith_summary":"Existing graph prompting methods freeze the pre-trained GNN encoder and therefore fail on heterophily graphs, where connected nodes often have different labels and different hops carry different class distributions. The paper proposes DAGPrompT, which tunes the frozen encoder through a low-rank adaptation of both the projection matrices and the message-passing adjacency, and couples this with hop-specific class prompts that are combined by learned coefficients. The claim, supported by experiments on ten datasets and fourteen baselines, is that this distribution-aware design improves few-shot node and graph classification accuracy by up to 4.79 percent while keeping tunable parameters in the thousands. If the claim holds, graph prompting can be extended to complex, heterophily-rich graphs without full fine-tuning.","feed_headline":"Low-rank tuning extends graph prompting to heterophily graphs","feed_subtitle":"A few thousand tunable parameters lift few-shot accuracy by up to 4.79 percent on ten datasets.","key_machinery":"GLoRA (Graph Low-Rank Adaptation) is the module that carries the encoder-adaptation claim: it injects low-rank matrices into each GNN layer, $H^{(\\ell)} = (A + P_A^{(\\ell)} Q_A^{(\\ell)\\top}) H^{(\\ell-1)} (W_0^{(\\ell)} + P^{(\\ell)}Q^{(\\ell)\\top})$, so that message passing can up- or down-weight edges and the projection can shift to the downstream distribution while the original weights $W_0$ stay frozen. The second machinery is Hop-specific Graph Prompting: the model keeps the intermediate embeddings $H^{(0)},\\ldots,H^{(L)}$, creates layer-specific class tokens $P_c^{(\\ell)}$ as the mean of training embeddings plus a learnable prompt, computes cosine similarity at each hop, and combines the hop scores with learnable coefficients $\\gamma^{(\\ell)}$ initialized as $\\alpha(1-\\alpha)^\\ell$. Together these two mechanisms are what the paper claims makes prompting work on heterophily graphs: the first gives the encoder enough flexibility to separate different labels, and the second lets each hop contribute according to its own distribution.","core_discovery":"On its own terms, the paper's central discovery is that a pre-trained GNN can be prompted effectively on heterophily graphs if the encoder is lightly adapted rather than left fully frozen. The GLoRA module writes the adapted layer as $H^{(\\ell)} = (A + P_A^{(\\ell)} Q_A^{(\\ell)\\top}) H^{(\\ell-1)} (W_0^{(\\ell)} + P^{(\\ell)}Q^{(\\ell)\\top})$, with rank-1 factors on the adjacency and low-rank factors on the projection, so the model reweights which edges matter for message passing and adjusts feature projections while keeping the pre-trained weights intact. The hop-specific prompting module collects embeddings from every layer, builds a per-layer class prototype plus a learnable prompt for each class, scores each node against these prototypes at every hop by cosine similarity, and fuses the hop scores with learnable coefficients initially set to $\\alpha(1-\\alpha)^\\ell$. The paper reports that this combination reaches state-of-the-art accuracy on all ten datasets, with the largest gains on strongly heterophily graphs such as Texas (up to 4.79 percent) and Chameleon, and that removing either module degrades performance.","pith_inferences":["An editorial reading of the ablation table is that the hop-specific prompting module is the more consistently load-bearing component: removing GLoRA costs 1.9 to 2.5 points on the reported datasets, while removing the layer-specific prompts or the learned hop coefficients also costs between 0.8 and 3.5 points, so the exact division of credit is dataset-dependent.","A natural testable extension is to attach GLoRA to a heterophily-specific GNN backbone; if the gains there are smaller, the benefit of low-rank adjacency adaptation may be specific to GCN-style encoders whose message passing assumes homophily.","Because Appendix A reduces the adjacency update to a single edge-weight vector applied only to training-set edges, an implied conjecture is that most of GLoRA's message-passing benefit comes from reweighting edges incident to labeled nodes; checking performance when the edge weights are applied to all edges would test this directly."],"forward_implications":["Frozen-encoder prompting is not the only viable graph-prompting design; a lightly tuned encoder with low-rank updates extends prompting to heterophily graphs.","Hop-wise class tokens and learned hop weights let a prompting method exploit the fact that first-hop and later-hop neighborhoods can have opposite label patterns.","The method keeps the parameter-efficiency advantage of prompting: on Chameleon it uses roughly 6.4K tunable parameters at the downstream stage, compared with 331K for full fine-tuning.","Cross-domain pre-training transfers: pre-training on Texas and then prompting on Cornell, Wisconsin, or Chameleon outperforms training from scratch on those targets.","The full-shot evaluation shows the same qualitative pattern, with DAGPrompT improving accuracy by 3.79 percent on Chameleon even when labels are plentiful."],"supporting_citations":[{"why":"Supplies the low-rank adaptation idea that GLoRA transfers from language models to GNN layers.","marker":"[9]"},{"why":"Introduces GPPT, the first graph prompting method that reformulates node classification as link prediction, the baseline DAGPrompT extends and outperforms.","marker":"[25]"},{"why":"Introduces GraphPrompt, a unified graph prompting baseline that the hop-specific module builds on and compares against.","marker":"[16]"},{"why":"Defines the homophily ratio and the heterophily challenge that motivates the paper, and supplies a heterophily-aware baseline.","marker":"[42]"},{"why":"Provides the link-prediction pre-training approach used before prompting.","marker":"[17]"},{"why":"Contributes the GPF and GPF-Plus universal prompting baselines that freeze the encoder, which DAGPrompT argues is insufficient.","marker":"[7]"},{"why":"Provides All-In-One, the multi-task graph prompting baseline that reformulates downstream tasks into graph-level tasks, a reformulation DAGPrompT shares.","marker":"[26]"},{"why":"Supplies the Rademacher-complexity generalization bound used in Theorem 1 to argue low-rank adaptation suits few-shot settings.","marker":"[23]"},{"why":"Supplies the synthetic heterophily graph generation procedure and the observation that heterophily distributions vary across hops.","marker":"[18]"},{"why":"Contributes DGI, a pre-training baseline used in comparisons and as an alternative pre-training strategy.","marker":"[27]"}],"fun_headline_variants":["Graph prompting gets distribution-aware GLoRA and hop-specific prompts","Low-rank graph prompts adapt to heterophily, beating SOTA by 4.79%","DAGPrompT tunes graph prompts with hop-specific modules","Few thousand tunable parameters for heterophily graph prompting","GLoRA and hop prompts: new SOTA for graph prompt tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a one-dimensional low-rank adjustment to the graph's adjacency, reduced in Appendix A to a single learned edge-weight vector applied only to edges connected to training-set nodes, is expressive enough to reshape message passing so that nodes of different classes become separable.","fun_headline_variants_meta":{"raw":{"variants":["Graph prompting gets distribution-aware GLoRA and hop-specific prompts","Low-rank graph prompts adapt to heterophily, beating SOTA by 4.79%","DAGPrompT tunes graph prompts with hop-specific modules","Few thousand tunable parameters for heterophily graph prompting","GLoRA and hop prompts: new SOTA for graph prompt tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3507,"prompt_tokens":1035,"completion_tokens":2472,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":2379}},"tokens_in":651,"tokens_out":2472,"duration_ms":17174,"temperature":1.0,"reasoning_tokens":2379,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:35:53.856356+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DAGPrompT on Texas with 5-shot labels, then freeze the hop-prompt module and randomly permute the learned adjacency-adjustment weights while keeping all other parameters fixed; if accuracy does not drop materially, the claim that GLoRA's message-passing adaptation drives the heterophily gain is not supported. A complementary check is to test whether the learned edge weights assign higher scores to same-label edges than the unmodified adjacency does.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the link-prediction pre-training approach used before prompting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the synthetic heterophily graph generation procedure and the observation that heterophily distributions vary across hops."}],"review_version":1}