{"id":"d6ed5b87-4d46-4a50-a85b-525a2091adf4","arxiv_id":"2507.09132","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GPAWP prunes low-importance graph prompts using gradient-based importance masks and retunes the remaining prompts, improving few-shot node classification over HGPrompt on ACM, DBLP, and Freebase.","lead":"GPAWP is a graph prompt learning method that scores each prompt weight by gradient sensitivity, prunes low-scoring weights, and retrains the survivors, producing small gains in few-shot node classification on three heterogeneous graph benchmarks. The work matters as an efficiency fix for graph prompt tuning, though the pruning thresholds are calibrated by hand and the reported gains are modest.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pruning a semantic prompt via λ_i=0 does not remove the subgraph: Eq. 6's (1+p_i^s) term becomes 1, so 'eliminating negative prompt labels' misdescribes the implemented operation.","rationale":"The reader identified the absolute-gradient importance measure as failing to distinguish 'negative' from merely 'uninformative' prompts. That is a valid concern, but the most load-bearing issue is more concrete and more damaging to the stated mechanism: in the semantic-prompt branch, the pruning operation defined by Eqs. 8 and 6 does not remove the prompt token from the model. Because the aggregation always adds 1 to the learned scalar, setting λ_i=0 changes the multiplier from (1 + p_i^s) to 1, which is the neutral no-prompt baseline, not deletion of the subgraph. Therefore, the paper's central assertion that GPAWP 'eliminate[s] negative prompt labels' is not supported by the equations as written. The empirical comparisons against HGPrompt and random pruning may still be valid, and the parameter-count reduction is real, so I do not move the verdict to REJECT; however, the method description and the interpretation of the semantic-pruning ablation (Fig. 6) would need major revision. This is an internal inconsistency rather than a disagreement with consensus, and it can be settled by a straightforward forward-pass check. Because the reader's verdict was already CONDITIONAL and this concern reinforces that conditionality (albeit for a different reason), I leave the overall verdict unchanged.","tokens_in":22304,"tokens_out":12597,"duration_ms":149798,"concrete_test":"Run a forward pass on the ACM one-shot task with a trained GPAWP prompt. For a semantic token with importance below δ, compare three variants: (a) original p_i^s; (b) \\hat{p}_i^s = 0 as in Eq. 8 (λ_i=0); (c) the subgraph S_i removed from \\hat{G}(S) so the ReadOut sum excludes it. If (b) matches (a-with-p_i^s=0) rather than (c), the pruning operation does not eliminate the prompt's influence—it only resets the multiplier to 1. Also report the number of subgraphs actually used in inference before and after pruning; if unchanged, 'eliminating negative prompt labels' is not what the implementation does.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GPAWP 'eliminate[s] negative prompt labels' by pruning low-importance semantic tokens and feature blocks. For feature prompts this is true: Eq. 14 multiplies node embeddings by \\hat{P}_f, so zeroed blocks genuinely drop features. For semantic prompts, however, the pruning operation as specified does not eliminate the prompt token. Eq. 8 defines \\hat{P}_s = λ · P_s with λ_i ∈ {0,1}, so a pruned token becomes \\hat{p}_i^s = 0. But the aggregation in Eq. 6 (and Retuning Eq. 15) is ReadOut{ (1 + p_i^s) ⊙ ReadOut(S_i) | S_i ∈ \\hat{G}(S) }, summing over all subgraphs. Thus λ_i=0 gives (1 + 0) ⊙ ReadOut(S_i) = 1 · ReadOut(S_i); the subgraph remains in the computation with the default neutral weight, and the number of subgraphs used in inference is unchanged. Consequently, 'negative prompt labels' are not removed; they are reset to a no-prompt baseline. The reported reduction in trainable parameters (Table IV) is real, but the mechanism described in the abstract and Section V—hierarchically pruning away negative prompts—is not what the semantic-prompt branch implements. The importance computation in Eq. 9 is also presented inconsistently ('set λ_i to 0' vs. an absolute gradient at the mask), but the more decisive issue is that even a correctly computed importance score cannot override the fact that λ_i=0 leaves the subgraph in the aggregate. If the intended operation was to drop the subgraph, the term should be removed from the sum or multiplied by λ_i directly, not by (1 + λ_i p_i^s). This is an internal inconsistency, not a matter of contestable framing: as written, the method cannot 'eliminate' a semantic prompt, only neutralize it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GPAWP, a framework for heterogeneous graph prompt learning that first tunes feature prompts (Pf) and semantic prompts (Ps) on a downstream node-classification task, then evaluates the importance of each semantic prompt token and each feature prompt block via the expected absolute gradient of the downstream loss with respect to associated mask variables (Eqs. 9 and 11), prunes low-importance tokens/blocks using thresholds δ and β, and finally retunes the surviving prompts. The authors report that this procedure eliminates negative prompts, improves one-shot node classification over HGPrompt on ACM, DBLP, and Freebase, and reduces optimized prompt parameters from 67 to 12 on ACM. The paper includes an ablation study (Table III), a random-pruning comparison, and parameter/time-efficiency comparisons.","tokens_in":22702,"tokens_out":4203,"duration_ms":52198,"significance":"If the mechanism worked as claimed, the paper would provide a modest but useful efficiency lever: using importance scores to prune uninformative prompt components before retuning, with consistent gains across three heterogeneous graphs. The paper deserves credit for including a component ablation (Table III) showing that evaluation-and-pruning plus retuning beats tuning-only, for comparing against random pruning, and for reporting parameter savings in Table IV. However, the semantic-pruning operation is misdescribed, and the importance metric does not actually measure whether a prompt is positive or negative; these issues bear directly on the central claim. The contribution is incremental, but the empirical evidence suggests the underlying retuning-after-pruning idea has value if the mechanism is clarified and the claims are adjusted.","major_comments":[{"comment":"Semantic prompt pruning does not eliminate the corresponding subgraph. When λ_i=0, Eq. (8) sets \\hat{p}_i^s=0, but the aggregation in Eq. (6) and the retuning aggregation in Eq. (15) use (1+p_i^s)⊙ReadOut(S_i). A pruned token therefore becomes 1·ReadOut(S_i), so the subgraph S_i remains in the sum with a neutral weight; it is not removed from the computation. Consequently, the abstract's claim that GPAWP 'eliminate[s] negative prompt labels' and Section II's description of 'remov[ing] negative prompts' overstate what the semantic branch actually does. This can be fixed by changing the aggregation to multiply by λ_i directly (or by dropping the term from the ReadOut sum), or by revising the claims to say that low-importance semantic prompts are neutralized rather than eliminated.","section":"V-B, Eqs. (6), (8), (15)"},{"comment":"The importance score I_{p_i^s}=E|∂L_down/∂λ_i| is an absolute gradient magnitude. It measures the sensitivity of the loss to the mask, not the sign or direction of the prompt's effect, so it cannot by itself distinguish 'positive prompts' from 'negative prompts'. A prompt with a large gradient magnitude could be beneficial or harmful; a prompt with a small gradient magnitude is merely uninformative at the current operating point. In addition, the text says that to evaluate p_i^s one should 'set λ_i to 0 and all other λ values to 1,' but Eq. (9) computes the derivative of the loss with respect to λ_i, presumably at the current λ=1, which is not the same as the loss difference after setting λ_i=0. Please either use a sign-aware score (e.g., expected loss change when the mask is zeroed) or reframe the method as pruning uninformative prompt components rather than negative ones.","section":"V-B, Eqs. (9) and (11)"},{"comment":"The pruning thresholds δ=0.6 and β=0.4 are selected from the importance-score distributions and task-specific objectives on the same three datasets. The paper states that increasing β to 0.5 or decreasing it to 0.3 changes performance, and similarly for δ, but it does not report a validation procedure or a held-out criterion for choosing these values. Since the central 'adaptive' pruning claim depends on these thresholds, the authors should specify how the thresholds are set (e.g., on a validation split) and provide sensitivity results with standard deviations. Without this, the method is a dataset-calibrated heuristic rather than an adaptive algorithm whose thresholds transfer.","section":"VI-A and VI-C, Eqs. (12)-(13)"}],"minor_comments":[{"comment":"The equation references in the pseudocode are swapped: the aggregation using Pf should refer to Eq. (5), and the aggregation using Ps should refer to Eq. (6).","section":"Algorithm 1, lines 7 and 9"},{"comment":"The Freebase GAT cell contains a duplicated '20.28±0.08', and several GPAWP cells repeat the same value; these typographical artifacts should be cleaned up.","section":"Table II"},{"comment":"The sentence 'For Ps we set the threshold β=0.4 and for semantic prompt Ps we set the threshold δ=0.6' should read 'for feature prompt Pf' in the first clause; the names β and δ are also inconsistent with Eqs. (12) and (13).","section":"VI-A, hyperparameter paragraph"},{"comment":"The statement about threshold sensitivity ('Increasing β to 0.5 ... lowering it to 0.3 ... δ>0.7 ... δ<0.5') is presented without a table or figure; please add the supporting results or remove the quantitative claims.","section":"VI-C"},{"comment":"Eq. (9) is attributed to Michel et al. [30], but that work computes expected absolute gradients over input samples for attention heads; the adaptation to mask variables for graph prompts should be clarified so the relationship to [30] is precise.","section":"V-B, citation [30]"}],"recommendation":"major_revision","confidential_remarks":"The paper is a modest extension of HGPrompt, and the empirical evidence for pruning-plus-retuning is reasonably consistent. The main risk is that the semantic-pruning mechanism is not what the abstract claims, and the importance metric does not measure sign; these are fixable with reframing or a change to the aggregation. I would not reject, but the authors need to address the mismatch between the claimed 'elimination' of negative prompts and the actual neutralization operation, and they should be transparent about threshold calibration. No code release is mentioned, which weakens reproducibility for a paper whose contribution is largely empirical."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a workmanlike empirical paper: it takes sensitivity-based prompt pruning and lottery-ticket retraining, applies them to heterogeneous graph prompts, and shows consistent one-shot node-classification gains across ACM, DBLP, and Freebase, with real parameter savings (67 to 12 on ACM). The ablation in Table III is the most convincing part: pruning plus retuning beats tuning alone on all three datasets, and random pruning does not. Second thing: the paper's headline claim is not what the math does. The stress-test note is correct: for semantic prompts, Eq. 6 aggregates over all subgraphs with weight (1 + p_i^s), so setting p_i^s to zero does not drop the subgraph; it leaves it in with weight 1. That is neutralization, not elimination. The abstract and Section V say 'eliminate negative prompt labels,' which overstates the implemented operation. The feature-prompt branch (Eq. 14) genuinely drops zeroed blocks, so the parameter reduction is real, but the semantic branch does not prune away subgraphs.\n\nWhat the paper does well: the combination is new, the experiments are internally consistent, and the random-pruning control is a good touch. The authors are honest that thresholds beta and delta are calibrated to the score distributions and task objectives, which is a limitation they acknowledge.\n\nSoft spots, in order of weight: (1) The semantic-pruning mechanism is misdescribed as elimination; the term should be multiplied by lambda_i directly or dropped from the sum if that is the intent. (2) The importance score is an absolute gradient magnitude, which measures sensitivity, not whether a prompt is helpful or harmful; 'negative prompt' is not supported. (3) The text says z-score normalization is applied to importance scores, but Algorithm 2 compares raw scores to thresholds. That is a concrete inconsistency. (4) No code or data are released, which makes the threshold choices hard to audit. None of these are fatal to the empirical claim, but they need to be fixed before the paper is publishable.\n\nThe paper is for readers who work on graph prompt efficiency; they will get a plausible recipe and a useful baseline. It deserves a serious referee, but the verdict should be major revision, not acceptance as is.","headline":"Useful empirical recipe for pruning heterogeneous graph prompts, but the semantic-prompt branch neutralizes rather than eliminates, so the headline mechanism claim needs correction before this is publishable.","tokens_in":23239,"tokens_out":3050,"would_cite":true,"duration_ms":33403,"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":"GPAWP prunes low-importance semantic tokens and feature blocks from graph prompts, retunes the survivors, and beats full-prompt tuning on three heterogeneous benchmarks while shrinking prompt parameters from 67 to 12 on ACM.","keywords":["heterogeneous graph prompt learning","graph prompt pruning","importance evaluation","weight pruning","node classification","few-shot learning","parameter efficiency","graph neural networks"],"falsifier":"Take the prompts pruned by GPAWP and re-insert each one individually into the retuned model, then measure downstream validation loss: if re-adding a so-called negative semantic token or feature block improves the loss for many tasks, the importance score is not identifying harmful prompts. Alternatively, compare GPAWP against retuning the same number of randomly chosen prompt components over many random seeds; if the random subsets match or beat the importance-selected subsets, the gradient-based scoring is not the source of the gain.","tokens_in":22121,"feed_emoji":"✂️","tokens_out":8866,"duration_ms":96212,"temperature":0.7,"pith_summary":"Graph prompt learning adapts a pre-trained GNN to a new task by adding a small set of tunable prompts, here a semantic prompt that weights each node-type subgraph and a feature prompt that reweights node features. The paper argues that not all prompt tokens and prompt blocks are useful, and that some can even hurt. GPAWP evaluates each semantic prompt token and each feature prompt block with a gradient-based importance score, prunes the low-scoring parts, and then retunes the surviving prompts. On one-shot node classification, it reports higher Micro-F and Macro-F than the HGPrompt baseline on ACM, DBLP, and Freebase while cutting prompt parameters from 67 to 12 on ACM. If correct, this is a simple efficiency lever for prompt tuning on heterogeneous graphs, where trainable prompt size and tuning cost currently grow with graph complexity.","feed_headline":"Pruning graph prompts slashes parameters, lifts one-shot accuracy","feed_subtitle":"A gradient score drops low-value prompt tokens and blocks, then retunes the survivors with fewer parameters.","key_machinery":"The machinery has three layers. First, a heterogeneous graph template splits a graph into $|A|+1$ homogeneous subgraphs, one per node type plus the complete topology; the semantic prompt $P_s = [p_s^0, \\dots, p_s^{|A|}]$ weights each subgraph during ReadOut aggregation, while the feature prompt $P_f$ is multiplied elementwise into node embeddings before pooling. Second, importance scoring attaches a binary mask $\\lambda_i$ to each semantic token and computes $I_{p^i_s} = \\mathbb{E}_x |\\partial L_{\\mathrm{down}}(x)/\\partial \\lambda_i|$; the feature prompt is split into $t$ equal blocks, each with mask $\\eta_j$, scored as $I_{p^j_f} = \\mathbb{E}_x |\\partial L_{\\mathrm{down}}(x)/\\partial \\eta_j|$. Third, after z-score normalization, tokens with scores below $\\delta=0.6$ and blocks with scores below $\\beta=0.4$ are masked out, and the surviving prompts are re-tuned from their pruned initialization, following the lottery-ticket retraining idea that sparse subnetworks can match full networks when retrained.","core_discovery":"The paper's central claim is that hierarchical pruning of graph prompts is both feasible and beneficial: after training full semantic and feature prompts on the downstream task, one can measure each component's importance as the expected absolute gradient of the downstream loss with respect to a binary mask on that component, prune every component whose normalized score falls below a granularity-specific threshold, and then retrain the remaining prompts initialized from the pruned values. The authors report that this removes what they call negative prompt labels and yields a model that in one-shot node classification outperforms the previous heterogeneous graph prompt method HGPrompt on ACM (Micro-F 72.98 vs 71.60, Macro-F 71.04 vs 68.14), DBLP (82.57 vs 79.25, 81.31 vs 78.00), and Freebase (25.32 vs 23.67, 16.13 vs 14.60), while reducing downstream prompt parameters from 67 to 12 on ACM, 68 to 24 on DBLP, and 71 to 58 on Freebase. The ablation results show that both evaluation-and-pruning and the subsequent retuning stage contribute to the gain, with feature-prompt pruning contributing more than semantic-prompt pruning.","pith_inferences":["My inference: because the paper's importance score is an absolute gradient, the method is better interpreted as pruning uninformative components than as proving those components have a negative effect; a signed influence measure or a re-insertion test would be needed to identify true negatives.","My inference: the calibrated thresholds ($\\delta=0.6$, $\\beta=0.4$) are likely dataset- and score-distribution-dependent; applying GPAWP to a new graph would probably require re-calibrating them rather than carrying them over, since the paper ties threshold choice to the observed distribution.","My inference: the same masked-gradient scoring could be extended to continuous soft masks or per-relation prompt tokens, allowing pruning to become differentiable and done jointly with training, whereas the paper only prunes after full tuning.","My inference: because feature-prompt pruning contributed more than semantic-prompt pruning in the ablation, the largest redundancy on these benchmarks may lie in node feature dimensions rather than subgraph structure, suggesting future heterogeneous graph prompt designs should consider input-feature compression."],"forward_implications":["Prompt tuning on heterogeneous graphs can be made parameter-efficient without sacrificing accuracy: GPAWP reports better one-shot Micro-F and Macro-F than HGPrompt on ACM, DBLP, and Freebase while shrinking prompt parameters, for example from 67 to 12 on ACM.","Pruning alone is not enough: the retuning stage is necessary, since removing it lowers ACM Micro-F from 74.07 to 71.07 in the ablation study.","Prompt importance is distributed unevenly and differs by granularity: feature prompt blocks show a left-skewed importance distribution while semantic prompt tokens show a right-skewed one on ACM, so treating all prompts as equally important is a real miss.","The framework also saves tuning time: per-epoch training on DBLP drops from 0.0521 seconds for HGPrompt to 0.0460 seconds for GPAWP, with test time down from 0.0458 to 0.0339 seconds.","The efficiency gains are largest in low-shot regimes: GPAWP holds the best Micro-F and Macro-F from 1 to 5 shots on ACM, which fits the label-scarce setting that motivates graph prompt learning."],"supporting_citations":[{"why":"Gives the heterogeneous graph template and the semantic/feature prompt design that GPAWP tunes, evaluates, and prunes.","marker":"[19]"},{"why":"Supplies the block-wise soft-prompt pruning strategy that GPAWP adapts for feature prompt blocks.","marker":"[27]"},{"why":"Motivates the retuning stage, where pruned prompts are initialized from the surviving components and trained again.","marker":"[33]"},{"why":"Provides the task template that turns node classification into a similarity prediction task compatible with the link-prediction pre-training objective.","marker":"[17]"},{"why":"Sources the expected-sensitivity importance formulation used in Eqs. (9) and (11).","marker":"[30]"},{"why":"Provides the ACM dataset and the heterogeneous attention baseline used in the node classification experiments.","marker":"[49]"},{"why":"Supplies the DBLP dataset and the heterogeneous graph benchmark library used for baseline settings.","marker":"[50]"},{"why":"Supplies the Freebase dataset used in the experiments.","marker":"[51]"}],"fun_headline_variants":["Pruned graph prompts cut parameters 80%, boost one-shot accuracy","GPAWP: weight pruning makes graph prompts leaner and stronger","Hierarchical prompt pruning beats HGPrompt with far fewer parameters","Adaptive pruning slashes graph prompt parameters, raises accuracy","Fewer graph prompts, better node classification via pruning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a prompt part which barely changes the prediction loss when switched off is a harmful 'negative' prompt that can be removed, and that the two score thresholds (0.6 and 0.4) correctly separate negatives from positives; low sensitivity does not by itself prove harm.","fun_headline_variants_meta":{"raw":{"variants":["Pruned graph prompts cut parameters 80%, boost one-shot accuracy","GPAWP: weight pruning makes graph prompts leaner and stronger","Hierarchical prompt pruning beats HGPrompt with far fewer parameters","Adaptive pruning slashes graph prompt parameters, raises accuracy","Fewer graph prompts, better node classification via pruning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1549,"prompt_tokens":1003,"completion_tokens":546,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":461}},"tokens_in":619,"tokens_out":546,"duration_ms":5830,"temperature":1.0,"reasoning_tokens":461,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:03:18.983696+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the prompts pruned by GPAWP and re-insert each one individually into the retuned model, then measure downstream validation loss: if re-adding a so-called negative semantic token or feature block improves the loss for many tasks, the importance score is not identifying harmful prompts. Alternatively, compare GPAWP against retuning the same number of randomly chosen prompt components over many random seeds; if the random subsets match or beat the importance-selected subsets, the gradient-based scoring is not the source of the gain.","supporting_citations":[{"cited_title":"Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning,","cited_arxiv_id":null,"evidence_quote":"Gives the heterogeneous graph template and the semantic/feature prompt design that GPAWP tunes, evaluates, and prunes."},{"cited_title":"XPrompt: Exploring the extreme of prompt tuning,","cited_arxiv_id":null,"evidence_quote":"Supplies the block-wise soft-prompt pruning strategy that GPAWP adapts for feature prompt blocks."},{"cited_title":"The lottery ticket hypothesis: Finding sparse, trainable neural networks,","cited_arxiv_id":null,"evidence_quote":"Motivates the retuning stage, where pruned prompts are initialized from the surviving components and trained again."},{"cited_title":"Graphprompt: Unifying pre- training and downstream tasks for graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the task template that turns node classification into a similarity prediction task compatible with the link-prediction pre-training objective."},{"cited_title":"Are sixteen heads really better than one?","cited_arxiv_id":null,"evidence_quote":"Sources the expected-sensitivity importance formulation used in Eqs. (9) and (11)."},{"cited_title":"Heterogeneous graph attention network,","cited_arxiv_id":null,"evidence_quote":"Provides the ACM dataset and the heterogeneous attention baseline used in the node classification experiments."},{"cited_title":"Are we really making much progress? revisiting, benchmarking and refining heterogeneous graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the DBLP dataset and the heterogeneous graph benchmark library used for baseline settings."},{"cited_title":"Freebase: a collaboratively created graph database for structuring human knowledge,","cited_arxiv_id":null,"evidence_quote":"Supplies the Freebase dataset used in the experiments."}],"review_version":1}