{"id":"f05be46f-9881-42ef-bbe7-a48e83b9975f","arxiv_id":"2411.17676","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Instance-aware node prompts built from a compact PHM projector and vector quantization improve few-shot graph classification over static and attention-based prompt baselines.","lead":"The paper introduces a graph prompt learning method that generates a different prompt for each node in a graph, using vector quantization to keep the prompts stable and compact. It reports improved few-shot accuracy on molecule property prediction and citation tasks compared with prior graph prompting baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (7) defines Xp = X + P but P is produced in the hidden dimension of the frozen GNN, while molecular node features have only 9 dimensions; as written, the central prompt-injection step is not executable.","rationale":"The reader's conditional verdict rests on exactly the right soft spot: Eq. (7) is the point where the proposed method touches the frozen backbone, and it is dimensionally ambiguous. For the abstract's claim to hold, the prompted graph must be a valid input to the frozen GNN. A dimensional analysis of the paper's own definitions shows that the equation cannot be verified from the text: d is used for both the raw feature space (9 for molecular data) and the hidden space of H, and no projection to the input feature dimension is stated. The concern is not a matter of disagreement with the community; it is internal to the manuscript. I also checked whether the training objective could rescue the ambiguity. The consistency loss in Eq. (17) and the EMA updates in Eqs. (18)-(19) do provide a learning signal to the PHM generator, so that part is not the weak point. The overclaim about 'consistently surpassing' baselines (Tables 1, 2, and 6 contain cells where GPF or GPF-plus is higher) is real but secondary; the primary threat is that the method as written cannot be executed. I did not find evidence of intentional misrepresentation, and the issue is addressable with a clarified equation, an explicit output dimension for the up-projector, and released code. Therefore the conditional verdict stands, and no harsher adjustment is warranted.","tokens_in":19324,"tokens_out":9030,"duration_ms":83915,"concrete_test":"Obtain the authors' implementation (or ask for an erratum) and instrument a forward pass on BBBP 50-shot random split: print the shapes of X, H, and P immediately before Eq. (7). If P.shape[1] does not equal X.shape[1], the equation as written cannot be executed. If the code instead adds P to H or projects H to 9 before adding, rerun the reported Table 1 BBBP result (85.62) under the corrected, clearly specified injection rule and check whether the advantage over GPF/GPF-plus survives; either outcome settles whether the manuscript describes the method that produced the numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's central claim requires that a frozen GNN can consume a prompted graph Gp = (Xp, A). Section 4.1 states H = fθ(G) with H in R^{|V|×d}, then P = gΦ(H), Xp = X + P (Eq. 7). The same symbol d is used for the raw feature dimension and the GNN hidden dimension. Appendix A (Table 4) shows molecular datasets have 9 node features, while the 5-layer GIN backbone (Appendix C) operates in a larger hidden space. Therefore P in R^{|V|×d} cannot be added to X in R^{|V|×9} unless gΦ explicitly projects to 9 dimensions and the notation is corrected. The text says P is projected back to d and 'matching the same shape as X,' which is internally inconsistent. If P is hidden-dimensional, Eq. (7) is undefined; if P is input-dimensional, then Eq. (7) relies on an unstated projection and the claim that prompts are generated in the hidden space and added to raw features is unsupported. This is load-bearing because the whole mechanism — and the comparison to input-space prompt methods like GPF — depends on this injection step being well-defined. No code or commit hash is provided to resolve the ambiguity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"IA-GPL is a graph prompting method that replaces the static task-level prompts of prior work (GPF, GPF-plus, GPPT, All-in-One) with node-level prompts generated per input instance. The pipeline is: (i) a frozen pre-trained GNN encodes the input graph to node representations H; (ii) a parameter-efficient bottleneck of PHM layers maps H to intermediate prompts p_c; (iii) vector quantization over a learnable codebook with multinomial sampling and an EMA update yields quantized prompts p_q, combined with a shared static prompt p_s; (iv) the final node-level prompt P is added to the original node features (Eq. 7), and the prompted graph is fed into the frozen backbone for classification. The paper reports full-shot and few-shot experiments on eight MoleculeNet datasets under random and scaffold splits, three citation datasets, and PPI, with five pre-training strategies, comparing against supervised, fine-tuning, linear-probing, and prompt-learning baselines, plus ablations, efficiency analysis, codebook visualization, and hyperparameter-sensitivity studies.","tokens_in":19603,"tokens_out":15626,"duration_ms":125199,"significance":"The contribution is meaningful and the experimental effort is unusually extensive: 12 datasets, 5 pre-training strategies, two split regimes, few-shot and full-shot regimes, plus ablations and efficiency measurements. If the method description is made precise, the empirical claim is plausible: average gains over the strongest prompting baselines are consistent (Table 2: 61.59 vs. 60.09 for GPF under scaffold 50-shot; Table 5: 74.93 vs. 73.91 for GPF-plus under full-shot scaffold), the few-shot gains are in the regime where prompt methods are claimed to help, and the VQ clustering rationale is a concrete, falsifiable mechanism for the OOD gains. The method is genuinely parameter-efficient (Table 3: ~20K tunable parameters vs. 1.86M for fine-tuning) and the ablation isolates the PHM and VQ components. The main caveat is that the paper's own description is internally inconsistent at the central injection step and in the backbone specification, so the reported numbers cannot currently be reproduced or even unambiguously interpreted; this is fixable but must precede any final judgment.","major_comments":[{"comment":"Section 4.1, Eq. (7): the central prompt-injection step is not well-defined as written. The text defines H in R^{|V| x d} as the output of the frozen GNN, defines P = g_Phi(H), and then says P in R^{|V| x d} is 'matching the same shape as X so that they can be added back to the original node features,' with X_p = X + P. The same symbol d is used in Section 3 for the raw node-feature dimension, which is 9 for every MoleculeNet dataset in Table 4, while the frozen 5-layer GIN backbone described in Appendix C operates in a far larger hidden dimension. Therefore P cannot simultaneously live in the hidden dimension of H and have the same shape as X. The authors must either (a) explicitly project the prompt to the input feature dimension d_X, in which case the statements 'projected back to d dimensions,' the bottleneck asymmetry, the codebook dimension in Eqs. (13)-(15), and the consistency loss in Eq. (17) all need to be restated in the prompt dimension, or (b) state that prompts are injected into a hidden representation, which contradicts the text's repeated claim that prompts are added to the original node features and would change the basis of the comparison to input-space methods such as GPF. Because the prompted graph G_p = (X_p, A) is the input to the frozen backbone in every experiment, this ambiguity is load-bearing for the method description and must be resolved (including with code or a pseudocode block) before the contribution can be evaluated.","section":"Section 4.1, Eq. (7)"},{"comment":"Section 5.1 vs. Appendix C: the experimental backbone is described inconsistently. Section 5.1 states 'GCN is adopted as our backbone model,' while Appendix C states that for molecular datasets 'we adopt the widely used 5-layer GIN' and for citation networks 'we adopt 2-layer Graph Transformers'; Section 6 and the ablation discussion in Section 5.3 also refer to 'the huge pre-trained GIN model.' The pre-trained checkpoints and frozen backbone determine H, the space in which prompts are generated, as well as the validity of the comparison across baselines. A single consistent statement of which architecture(s) generated Tables 1, 2, 5-9 is required for reproducibility, and the authors should also clarify which pre-trained checkpoints (e.g., from Hu et al. 2020) were used.","section":"Section 5.1 vs. Appendix C"}],"minor_comments":[{"comment":"The text in Section 5.2 states that IA-GPL 'attains optimal results across these eight datasets' under the scaffold 50-shot setting, but in Table 2 GPF achieves a higher ROC-AUC on ClinTox (65.18 +/- 1.76 vs. 63.28 +/- 3.52); similarly, in Table 1, GPF-plus exceeds IA-GPL on MUV (59.93 +/- 0.83 vs. 59.32 +/- 1.13). Suggest rephrasing these claims to 'best average' or explicitly reporting per-dataset exceptions.","section":"Section 5.2, Tables 1-2"},{"comment":"The code repository is referenced twice ('You can also visit our code repository'), but no URL, repository name, or commit hash is provided; given the ambiguity in Eq. (7), providing the actual implementation of the injection step would materially resolve the main concern.","section":"Section 6 and Appendix C"},{"comment":"Eq. (14) draws from a 'Multinomial distribution over the logits,' but a Multinomial distribution requires probabilities; the authors should state explicitly that the negative-distance scores are normalized (e.g., by a softmax over -d_i^c/tau) before sampling, and clarify the role of tau in that normalization.","section":"Section 4.3, Eq. (14)"},{"comment":"In the EMA update, the count c_j updated in Eq. (18) is used as the denominator in Eq. (19) for the same batch; please clarify whether the count used in the mean is the updated running count or the per-batch count, since this affects the effective codebook learning dynamics.","section":"Section 4.4, Eqs. (18)-(19)"},{"comment":"The inference procedure requires two forward passes through the frozen GNN, one on G to produce H for prompt generation and one on G_p for prediction, but Table 3 reports only training time and memory; a statement of this inference overhead (and whether it is included in the reported training time) would improve the efficiency analysis.","section":"Section 5.3 and Table 3"},{"comment":"Several typos and inconsistencies should be corrected: 'IA-GPF' for 'IA-GPL' in Section 5.2; 'instance-agonist' for 'instance-agnostic' in Section 4.3 and Figure 3; 'virgina.edu' in the author block; and the first author's email domain (brandeis.edu) differs from the stated affiliation (University of Connecticut).","section":"Throughout"},{"comment":"No statistical significance tests are reported for the few-shot tables, where several comparisons fall within one standard deviation (e.g., Table 2, HIV: 66.52 +/- 2.10 vs. 65.59 +/- 2.31); paired significance tests across the five runs would strengthen the superiority claims.","section":"Section 5.2, Tables 1-2"}],"recommendation":"major_revision","confidential_remarks":"I regard the two internal contradictions (the Eq. (7) dimension mismatch and the GCN vs. GIN backbone statement) as fixable within the scope of a revision, and the underlying idea as sound and well-motivated, so reject is not warranted. I would urge the editor to require the authors to provide either the code or a precise pseudocode description of the prompt-injection step as a condition of the revision, since the main tables rest on a mechanism the paper does not currently specify unambiguously. The 'optimal across all eight datasets' claim in Section 5.2 should also be checked against Table 2 before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful few-shot graph prompting paper, but it needs a revision before I'd take it as written. The core idea—generate per-node prompts from GNN hidden states, then quantize them against a codebook with EMA updates—is sensible, and the experiments are about as thorough as you'd want: eight MoleculeNet datasets under both splits, five pre-training strategies, node-level tasks, ablations, efficiency, and hyperparameter sensitivity. The gains over GPF-plus are modest (roughly 1-1.5 AUC points) but consistent, and the codebook visualization actually supports the instance-awareness story.\n\nNow the soft spots. The Eq. (7) dimension issue is real and load-bearing. The paper defines H in R^{|V|xd} as GNN hidden states, then P = gPhi(H), and Xp = X + P. But for the molecular datasets X has only 9 features (Table 4). The text tries to have it both ways: it says P is projected back to d dimensions \"matching the same shape as X,\" but d is the hidden dimension, not the input dimension. As written, the injection step is not executable. This is not a cosmetic typo; the whole mechanism depends on where P lives. A clear fix would state that the up-projector maps to the input feature dimension, or introduce a separate dimension symbol.\n\nSecond, the novelty claim is overstated. The introduction and contributions say IA-GPL is \"the first graph prompting method capable of generating distinct prompts based on different instances.\" That is false: GPF-plus already generates per-node prompts with attention, and the paper acknowledges this in Section 2 before arguing advantages. The novelty is in the specific PHM + VQ + EMA pipeline, which is fine, but the claim needs rephrasing.\n\nThird, \"IA-GPL attains optimal results across these eight datasets\" in the scaffold-split discussion is wrong: on ClinTox, GPF is clearly better (65.18 vs 63.28). Small overstatement, but it should be corrected.\n\nFinally, no code or commit hash, no significance tests, and the gap over GPF-plus is sometimes within one standard deviation. The appendices are the strength, but they don't include code or error bars.\n\nWho is this for? Anyone working on prompt tuning for GNNs. It is a credible incremental contribution. Deserves peer review—a serious referee can check the fix for Eq. (7), the reworded novelty, and ideally a code link. I would recommend conditional acceptance after a minor-to-major revision.","headline":"A solid few-shot graph prompting extension with a thorough experimental appendix, but a dimension mismatch in the central injection step and an overstated novelty claim need fixing before it is citable as written.","tokens_in":20147,"tokens_out":3394,"would_cite":false,"duration_ms":30080,"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 generating a distinct prompt for every input graph, via a parameter-efficient PHM bottleneck and vector quantization, outperforms fixed task-level prompts on few-shot graph classification while keeping the pretrained…","keywords":["graph prompt learning","instance-aware prompts","graph neural networks","vector quantization","few-shot learning","parameter-efficient fine-tuning","molecular property prediction","out-of-distribution generalization"],"falsifier":"Run the released implementation on a molecular dataset with 9-dimensional raw features and a hidden dimension $d>9$: if the addition $X_p=X+P$ in Eq. (7) fails or requires an undocumented projection, then the published mechanism is not reproducible as written. Independently, replace the vector-quantization step with simple averaging of the PHM outputs while keeping all other hyperparameters fixed; if out-of-domain ROC-AUC on scaffold splits does not drop, the paper's central attribution of OOD gains to the codebook is falsified.","tokens_in":19086,"feed_emoji":"🧪","tokens_out":8344,"duration_ms":65961,"temperature":0.7,"pith_summary":"Existing graph prompt learning methods attach one fixed prompt to every graph in a task, even though graphs within a task differ widely; the paper argues this static prompt is the key limitation and that prompts should be generated per input instance. IA-GPL therefore produces a distinct prompt for each node by passing the frozen GNN's hidden representations through a parameter-efficient bottleneck built from parameterized hypercomplex multiplication layers, then quantizes those prompt vectors through a small trainable codebook updated by exponential moving average. The claimed result is that this frozen-backbone method beats fixed-prompt baselines (GPF, GPF-plus, All-in-One) and often approaches or exceeds full fine-tuning on eight molecular datasets and three citation networks under few-shot and full-shot settings, with the largest gains in out-of-domain scaffold splits. If the claim holds, prompt tuning on graphs becomes per-instance adaptation rather than task-level patching, which would matter for any application where a single prompt must cover structurally heterogeneous inputs.","feed_headline":"Per-instance prompts beat fixed graph prompts in few-shot tests","feed_subtitle":"A frozen GNN plus a hypercomplex bottleneck and codebook quantization tops molecular and citation benchmarks","key_machinery":"The load-bearing object is the prompt generator $g_\\Phi(H) = \\mathrm{UpProject}(\\mathrm{ReLU}(\\mathrm{DownProject}(H)))$, with both projectors implemented as PHM layers. A PHM layer replaces the full weight matrix $W$ with a sum of Kronecker products $M = \\sum_{i=1}^n A_i \\otimes S_i$, so the number of trainable parameters scales roughly as $1/n$ of a fully connected layer. The intermediate per-node prompts $P_c = g_\\Phi(H)$ are then mapped through vector quantization: for each prompt, $M$ codebook vectors are sampled from a multinomial distribution over negative squared Euclidean distances scaled by a temperature, and the quantized prompt $p_q$ is the average of those sampled vectors. The final prompt is $p_f = p_q + \\beta p_s$, where $p_s$ is a shared learnable prompt. Codebook vectors are updated by exponential moving average rather than backpropagation, which the paper argues prevents representation collapse and stabilizes training; the training objective adds a consistency loss $\\lambda \\sum_i \\|p_{q,i} - p_{c,i}\\|_2^2$ to the cross-entropy loss.","core_discovery":"On its own terms, the paper establishes that instance-aware prompts, generated from the input graph's own hidden representation rather than learned as a single task-level vector, improve graph prompt learning. The authors report that IA-GPL achieves the highest average ROC-AUC among graph prompting methods on eight MoleculeNet benchmarks under both random split (68.46 in 50-shot) and scaffold split (61.59 in 50-shot), and under full-shot scaffold split (74.93), along with the best accuracy on Cora, CiteSeer, and PubMed in the 100-shot node-level setting. They further report that IA-GPL obtains the best result in 27 of 32 configurations tested across five pretraining strategies. They attribute the improvement to two design choices acting together: the PHM bottleneck keeps per-instance generation cheap, and vector quantization with EMA-updated codebooks forces distinct, stable prompts whose clustering transfers better to out-of-distribution instances.","pith_inferences":["The same PHM-bottleneck-plus-codebook recipe could be dropped into other frozen encoders, such as text or vision transformers where fixed soft prompts are standard; the paper's mechanism does not depend on graph-specific message passing beyond the hidden representations it starts from.","The paper leaves the dimensionality of the injection $X_p = X+P$ unspecified: molecular node features are 9-dimensional while hidden prompts live in the GNN hidden dimension $d$. A concrete implementation would need an explicit projection or padding, and the stated equation is not self-consistent as written.","One can test directly whether the codebook vectors carry semantic meaning by probing whether clusters of quantized prompts align with chemical functional groups across molecules; the paper's t-SNE suggests clustering but does not quantify this.","An ablation fixing codebook size and varying only the sampling temperature would separate the contribution of stochastic sampling from the contribution of quantization itself."],"forward_implications":["IA-GPL reports the best average ROC-AUC among graph prompting methods on eight molecular datasets in 50-shot random split (68.46) and scaffold split (61.59), and in full-shot scaffold split (74.93).","On 100-shot node-level tasks over Cora, CiteSeer, and PubMed, IA-GPL reports the highest accuracy among prompting baselines, with an average of 72.06.","Across five pretraining objectives, IA-GPL is reported as the best prompting method in 27 of 32 configurations, so the gains are not tied to one particular pretrained backbone.","With roughly 20K trainable parameters (about 1.08% of fine-tuning's 1.86M), the method keeps the GNN frozen and uses comparable training time and slightly less GPU memory than fine-tuning.","The performance gap over static prompts is larger under scaffold split than under random split, which the paper takes as evidence that vector quantization's clustering property helps out-of-domain generalization."],"supporting_citations":[{"why":"Supplies the PHM layer construction used for the parameter-efficient prompt bottleneck.","marker":"Zhang et al. (2020)"},{"why":"Supplies the vector quantization formalism used to discretize intermediate prompts.","marker":"GRAY (1998)"},{"why":"Supplies the EMA codebook-update procedure that stabilizes vector quantization training.","marker":"Roy et al. (2018)"},{"why":"Supplies the EMA update in quantized transformer spaces, the direct template for the codebook update.","marker":"Angelidis et al. (2021)"},{"why":"Provides the GPF and GPF-plus fixed-prompt baselines that IA-GPL must beat.","marker":"Fang et al. (2023)"},{"why":"Provides the All-in-One multi-task prompt baseline and the subgraph-based task unification used in the method.","marker":"Sun et al. (2023)"},{"why":"Supplies the edge-prediction pretraining strategy used for the main molecular experiments.","marker":"Jin et al. (2020)"},{"why":"Supplies the eight MoleculeNet molecular datasets used for evaluation.","marker":"Wu et al. (2018)"},{"why":"Supplies the unification of node and graph tasks into a general graph-level prompt learning setting.","marker":"Liu et al. (2024)"}],"fun_headline_variants":["Per-instance graph prompts outperform fixed prompts in few-shot tests","Instance-aware prompting lifts graph learning across 32 settings","Graph prompts get a per-instance makeover with codebook quantization","Tailored graph prompts beat generic ones on MoleculeNet and citation","IA-GPL: per-instance prompts cut label costs, boost accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a prompt vector computed in the GNN's hidden space can be added directly to the raw input features ($X_p = X+P$) and the frozen backbone will still read the result, yet on the molecular datasets raw features are 9-dimensional while hidden prompts are $d$-dimensional, so the stated injection is not dimensionally well-defined.","fun_headline_variants_meta":{"raw":{"variants":["Per-instance graph prompts outperform fixed prompts in few-shot tests","Instance-aware prompting lifts graph learning across 32 settings","Graph prompts get a per-instance makeover with codebook quantization","Tailored graph prompts beat generic ones on MoleculeNet and citation","IA-GPL: per-instance prompts cut label costs, boost accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000535,"raw_usage":{"total_tokens":2559,"prompt_tokens":917,"completion_tokens":1642,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1554}},"tokens_in":533,"tokens_out":1642,"duration_ms":11964,"temperature":1.0,"reasoning_tokens":1554,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:50:57.544854+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released implementation on a molecular dataset with 9-dimensional raw features and a hidden dimension $d>9$: if the addition $X_p=X+P$ in Eq. (7) fails or requires an undocumented projection, then the published mechanism is not reproducible as written. Independently, replace the vector-quantization step with simple averaging of the PHM outputs while keeping all other hyperparameters fixed; if out-of-domain ROC-AUC on scaffold splits does not drop, the paper's central attribution of OOD gains to the codebook is falsified.","supporting_citations":[{"cited_title":"Quantization","cited_arxiv_id":null,"evidence_quote":"Supplies the vector quantization formalism used to discretize intermediate prompts."},{"cited_title":"Extractive opinion summarization in quantized transformer spaces","cited_arxiv_id":null,"evidence_quote":"Supplies the EMA update in quantized transformer spaces, the direct template for the codebook update."},{"cited_title":"One for all: Towards training one graph model for all classification tasks","cited_arxiv_id":null,"evidence_quote":"Supplies the unification of node and graph tasks into a general graph-level prompt learning setting."}],"review_version":1}