{"id":"0ed26643-f503-4a87-bad3-c21a4688b8b5","arxiv_id":"2505.04121","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VGP uses low-rank virtual node, edge, and node prompts to adapt frozen Vision GNNs, matching or exceeding full fine-tuning on ten vision and nine graph classification benchmarks.","lead":"Vision Graph Prompting freezes a Vision GNN backbone and trains low-rank graph, edge, and node prompts, reaching accuracy close to or above full fine-tuning while updating about 5 percent of the parameters. The paper's core insight, that semantic information in vision graphs is low-rank, is supported only by qualitative PCA visualizations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reproducibility gap: Eq. 9 and Eq. 12 specify different edge-prompt updates, so the reported full-fine-tuning-comparable accuracy cannot be verified from the paper as written.","rationale":"The reader's weakest-assumption analysis identifies the KNN/PCA circularity, which is a real limitation of the paper's claimed semantic insight but would not, by itself, overturn the central PEFT claim: the method could still be an effective parameter-efficient adapter even if the low-rank observation is a tautology. The more load-bearing concern is the unstated gap between Eq. 9 and Eq. 12, which the reader listed only as a secondary reservation. These equations describe different operations—one is a normalized convex replacement of the center feature, the other is an unnormalized additive term appended after the aggregation update—and no derivation shows how one follows from the other. Because the central claim is empirical and the paper's ablations depend on which formula was actually used, this ambiguity blocks verification of the headline result. The released code makes the check straightforward, and the issue is fixable by a correction and clarification, so conditional acceptance remains the appropriate verdict rather than rejection. The reader's verdict is therefore unchanged, but the stress-test should make the Eq. 9/Eq. 12 discrepancy the primary item to resolve.","tokens_in":935,"tokens_out":2177,"duration_ms":81292,"concrete_test":"Download the released code and instrument the ViG block on CUB. Log the exact edge-prompt computation: is the center node first replaced by the normalized convex combination of Eq. 9 and then passed through the frozen convolution, or is the unnormalized term sum of beta MLP_s(x_j) P_e added after the aggregation update as in Eq. 12? Then re-run with beta = 0 and with Eq. 9's normalization removed to see which variant reproduces the reported CUB accuracy of 87.4 and the ablation trajectory in Table 3. If neither variant reproduces the reported numbers, the method as specified is not reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's concern about the low-rank insight being a KNN artifact is plausible but not the most load-bearing issue: even if the semantic story is trivial, VGP could still work as a frozen-backbone adapter. The more serious problem is the unstated mathematical gap between the two formal specifications of the edge prompt. Eq. 9 defines a normalized convex replacement of the center node: x_c ← (beta / ||N_s(x_c)||) sum over s_n in N_s(x_c) of s_n P_e + (1-beta) x_c. Eq. 12 instead writes f_hat(x_i, P_e) = (1-beta) x_i + g_hat(x_i, P_n) W_update + sum over x_j in N_hat(x_i) of beta MLP_s(x_j) P_e. These are not algebraically equivalent: Eq. 12 drops the normalization, applies the blend to the whole pre-update residual rather than to the convolution input, and adds the edge term as an unnormalized additive contribution after the aggregation update. If beta = 0, Eq. 12 still contains the node prompt inside g_hat but does not reduce to Eq. 9 applied before a standard update. If Eq. 9 were applied literally, the input to the graph convolution would change, which is not what Eq. 12 describes. The paper provides no derivation connecting Eq. 9 to Eq. 12, and the ablation in Table 3 does not state which formulation was implemented. Since the central claim is that the full VGP pipeline reaches full fine-tuning accuracy, the exact form of the edge prompt is load-bearing: a reader cannot tell what algorithm produced Tables 1 and 3 without inspecting code. The availability of released code makes this concern checkable rather than fatal, but it must be resolved before the central empirical claim can be fully accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Vision Graph Prompting (VGP), a parameter-efficient fine-tuning method for Vision GNNs. It introduces three prompt components—SeLo-Graph, SeLo-Edge, and SeLo-Node—that are designed around the claimed observation that semantically connected nodes in vision graphs share low-rank PCA structure. The method is evaluated on ten image classification datasets and nine chemistry/biology graph benchmarks, reporting accuracy comparable to or better than full fine-tuning while updating roughly 5% of the parameters. The paper also includes ablations of the three components, rank dimension, and blending factors, plus an efficiency analysis.","tokens_in":14590,"tokens_out":4459,"duration_ms":41794,"significance":"If the claims hold, VGP would be a useful contribution to parameter-efficient fine-tuning for graph-based vision backbones, an area with few existing prompting methods. The paper's strengths are the breadth of evaluation (external benchmarks for both vision and graph tasks), the component-level ablation, the efficiency table, and the public code release. However, the central 'low-rank semantic' insight is not established as a property of semantics rather than a byproduct of the KNN graph construction, and the headline comparison with full fine-tuning lacks statistical support. The method may still work as an adapter, but the paper as written overstates the interpretability and the quantitative certainty of its core claims.","major_comments":[{"comment":"The two formal specifications of the edge prompt are not algebraically equivalent. Equation (9) applies a normalized convex combination to the center node before convolution, while Eq. (12) applies the edge term as an unnormalized additive residual after the update and drops the normalization. The manuscript provides no derivation linking the two, and the ablation in Table 3 does not state which formulation was implemented. Because Tables 1 and 3 depend on the exact edge-prompt update, a reader cannot reproduce the reported numbers from the text; please specify the implemented form, justify it, and reconcile Eq. (9) with Eq. (12).","section":"Section 4.3, Eq. (9) and Section 4.5, Eq. (12)"},{"comment":"No standard deviations, number of seeds, or significance tests are reported. The headline claim 'comparable to full fine-tuning' rests on average accuracy gaps (e.g., 89.6 vs. 89.0 in Table 1) that are smaller than typical run-to-run variation on fine-grained benchmarks, and several per-dataset entries (CUB, NABirds, Flowers, CIFAR10) are actually lower than full fine-tuning. Please report multiple seeds with error bars and a statistical test for the key comparisons, especially the comparison against full fine-tuning and against the strongest baseline InsVP.","section":"Section 5.3, Tables 1-3"},{"comment":"The observation that connected nodes share dominant PCA components is confounded by the KNN rule in Eq. (1). Since ViG connects each node to its nearest neighbors in the same feature space, connected nodes are close by construction and will share principal components even without semantic structure. The visualizations in Figure 2 and Appendix A.3 do not include a control condition (e.g., random node pairs matched for feature distance, or KNN graphs on non-semantic features), so the paper's stated core insight that 'semantic information within vision graphs predominantly resides in the low-rank components' is not currently supported. Please add such a control or temper the claim accordingly.","section":"Section 3 and Appendix A.3"},{"comment":"The extension to chemistry/biology tasks is justified by the hypothesis that low-rank patterns exist in molecular and protein graphs, but no analysis of that property is provided for these datasets. Since the entire prompt design is motivated by the low-rank semantic observation, the graph-task results do not demonstrate that the low-rank mechanism transfers; they only show that the specific set of prompts improves GIN prompting baselines. Please provide a low-rank analysis on graph datasets or reframe the contribution of these experiments.","section":"Section 5.3, Table 2"}],"minor_comments":[{"comment":"Equation (6) writes cosine similarity as a raw dot product, but cosine similarity normally requires normalization; please clarify whether the top-K selection uses normalized features.","section":"Section 4.2, Eq. (6)"},{"comment":"Equation (9) uses \\|N_s(x_c)\\| to denote set cardinality, which is unconventional, and the parenthesis after N_s(x_c) is unclosed. Please fix the notation.","section":"Section 4.3, Eq. (9)"},{"comment":"Figure 4 has no axis labels and the red star is not described in the caption, so the dependence of accuracy on alpha and beta cannot be read reliably. Please add axes labels, a legend, and error bars if multiple seeds were used.","section":"Section 5.4, Figure 4"},{"comment":"Appendix A.3 estimates an observed rank of 50-60 for CUB and Flowers, but the experiments use r=32; the link between the observed rank and the chosen hyperparameter is not explained. Please clarify how the chosen rank follows from the analysis.","section":"Appendix A.3"},{"comment":"The statement that all methods were trained for 100 epochs and then '10 additional epochs for probing the optimal result' is ambiguous; please clarify whether this is an evaluation protocol, early stopping, or a model selection procedure.","section":"Section 5.2"},{"comment":"Table 5 reports FLOPs only for full fine-tuning and VGP; without the same metric for the other PEFT baselines, the conclusion of 'little additional overhead' is not fully contextualized. Please include FLOPs for the compared methods or qualify the claim.","section":"Appendix A.1, Table 5"}],"recommendation":"major_revision","confidential_remarks":"The discrepancy between Eq. (9) and Eq. (12) is the most serious issue; I would ask the authors to confirm the implemented update with the released code. The KNN confound is real but fixable with a control experiment. I do not see grounds for rejection if the implementation is clarified and the quantitative claims are supported by multiple seeds and appropriate statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked about arXiv:2505.04121 (VGP). Short version: it's a solid, well-tested PEFT method for Vision GNNs, with the first prompting framework specifically for ViG, but the paper's central \"low-rank semantics\" claim is weakly supported and there's a formal gap between two key equations that needs fixing before I'd fully trust the reproducibility.\n\nWhat's new: VGP combines VPT-style virtual nodes, low-rank adapters, and graph-structure prompts into a single method for frozen ViG backbones. The three components—SeLo-Graph, SeLo-Edge, SeLo-Node—are sensible, and the ablation in Table 3 shows each adds a clear gain. Experiments cover ten vision benchmarks and nine graph benchmarks with standard splits, and the code is released. That's more than many PEFT papers ship.\n\nWhere it stumbles: first, the motivating observation that semantically connected nodes share PCA components is likely an artifact of ViG's KNN graph construction—nodes are connected precisely because they're close in feature space, so sharing dominant PCs is almost forced. The authors even acknowledge the KNN construction in Eq. 1, but don't address this trivial explanation. Second, no error bars or seeds are reported anywhere, which makes the \"comparable to full fine-tuning\" claim hard to judge; the gaps on individual datasets often look within run-to-run noise. Third, and more load-bearing: Eq. 9 defines the edge prompt as a normalized convex replacement of the center node, while Eq. 12 in the analysis section defines an additive, unnormalized update that does not reduce to Eq. 9 even at beta=0. The paper never connects these. Since the ablation doesn't say which version was implemented, a reader cannot reproduce the reported accuracies from the text alone. The code may resolve this, but the paper should state the exact algorithm.\n\nNone of these are fatal—the method appears to work, and the low-rank component design is a legitimate engineering contribution even if the \"semantic\" motivation is overclaimed. The formal inconsistency is fixable with a short derivation or a clarification that Eq. 9 is a conceptual simplification.\n\nWho should read this: anyone working on parameter-efficient adaptation for graph-structured vision backbones, or on combining prompting with GNNs. It's a niche architecture, so not a must-read for the broader PEFT community, but it's a genuine addition.\n\nRecommendation: worth a serious referee. I'd ask for a major revision that adds seeds/error bars, resolves Eq. 9 versus Eq. 12, and either de-emphasizes or re-frames the PCA observation.","headline":"A useful PEFT method for Vision GNNs, but the low-rank semantic story is likely a KNN artifact and the Eq. 9/12 mismatch needs fixing before reproduction.","tokens_in":15103,"tokens_out":2606,"would_cite":true,"duration_ms":24210,"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":"Semantically connected vision-graph nodes share dominant PCA components, so a frozen ViG with three low-rank prompts matches full fine-tuning.","keywords":["vision graph prompting","parameter-efficient fine-tuning","low-rank decomposition","vision GNN","visual prompt tuning","graph neural networks","semantic low-rank","transfer learning"],"falsifier":"Take a ViG and compute the PCA spectrum of node features for three graph constructions — KNN, spatial adjacency, and random neighborhoods; if the eigenvalue decay is just as sharp for non-semantic neighborhoods, the low-rank 'semantic' claim is an artifact, and VGP should be re-tested with random low-rank prompt projections as a control.","tokens_in":14045,"feed_emoji":"🕸️","tokens_out":6817,"duration_ms":63023,"temperature":0.7,"pith_summary":"Vision GNNs represent an image as a graph of patches, and this paper argues that the semantic information carried by such a graph is concentrated in low-rank components of the feature space. Building on that observation, it introduces Vision Graph Prompting, a parameter-efficient fine-tuning method that freezes the ViG backbone and trains three low-rank prompt modules at the graph, edge, and node levels. The method is reported to match or exceed full fine-tuning across ten vision classification datasets while using about 94.6% fewer trainable parameters, and to transfer to chemistry and biology graph benchmarks. A curious reader would care because this reframes visual prompting for graph backbones as a low-rank semantic operation rather than a set of attention-style tokens.","feed_headline":"Low-rank prompts let frozen vision graphs match full fine-tuning","feed_subtitle":"Three tunable prompt layers reach 89.6% average accuracy on 10 datasets with 94.6% fewer trainable parameters.","key_machinery":"The load-bearing mechanism is the semantic low-rank decomposition: a shared MLP $MLP_s: \\mathbb{R}^d \\to \\mathbb{R}^r$ maps each $d$-dimensional node feature to a compact semantic vector with $r \\ll d$, and three low-rank prompt matrices $P_g, P_e, P_n \\in \\mathbb{R}^{r \\times d}$ fold that vector back into the full feature space at the graph, edge, and node levels. SeLo-Graph uses virtual nodes to alter topology, while SeLo-Edge and SeLo-Node modify the aggregation and update steps of graph convolution respectively. This carries the argument by ensuring the trainable prompt capacity is spent on the presumed semantic subspace, filtering high-frequency local detail.","core_discovery":"The paper's central claim is that semantically connected nodes in a ViG graph share dominant PCA components, so semantic information lies in a low-dimensional subspace of the latent feature space. To exploit this, VGP projects each node feature into a low-rank semantic space with a shared MLP and composes it with three prompt matrices: SeLo-Graph appends trainable virtual nodes that dynamically link to nearby nodes, SeLo-Edge propagates the aggregated low-rank neighbor semantics into each node update, and SeLo-Node blends the low-rank semantic component back into the original node feature. The appendix estimates the semantic rank at roughly 50 to 60 against a feature dimension of 768. With the backbone frozen, only the prompts and a downstream head are trained; experiments report an average accuracy of 89.6% versus 89.0% for full fine-tuning on the ten vision datasets, and gains on seven of nine chemistry and biology graph datasets.","pith_inferences":["An extension the paper does not test: because ViG builds edges by K-nearest-neighbor search in the very feature space used for the PCA analysis, the low-rank 'semantic' property may be partly a by-product of that construction rule; re-running the prompts on spatially built graphs would separate semantic low-rankness from KNN-induced similarity.","If the low-rank semantics is generic across graph backbones, the same three prompts could be applied to other graph-based vision models or point-cloud GNNs by retraining only the low-rank MLP and prompt matrices, without changing the recipe.","The reported sensitivity table suggests rank could be set per block from the PCA spectrum of that block's features, rather than fixed globally at 32, potentially improving parameter efficiency on deeper models.","The strong graph-domain results imply a testable hypothesis the paper leaves open: molecular and protein graphs may inherit low-rank structure from their pre-training features, so VGP-style prompts might work even without task-specific graph pre-training."],"forward_implications":["A frozen ViG can be adapted to downstream classification at accuracy comparable to full fine-tuning, cutting trainable parameters by about 94.6% on average and adding only about 3.1% extra FLOPs.","On the ten vision benchmarks, VGP beats the strongest applied visual-prompting baseline by an average of 5.0 accuracy points, including on fine-grained and small-sample datasets.","The same three prompt modules transfer to chemistry and biology graph classification, outperforming full fine-tuning on seven of nine datasets and previous graph-prompting adapters.","The method is robust to its main hyperparameters: rank $r=32$ and blending factors $\\alpha=\\beta=0.2$ work across datasets with a broad plateau."],"supporting_citations":[{"why":"Defines the ViG backbone and K-nearest-neighbor graph construction that VGP freezes and prompts.","marker":"Han et al., 2022"},{"why":"Introduces visual prompt tuning in feature space, the main baseline VGP extends from Transformer attention to graph topology.","marker":"Jia et al., 2022"},{"why":"Provides the input-space visual prompting baseline included in the comparison.","marker":"Bahng et al., 2022"},{"why":"Supplies the DAM-VP cluster-level prompting baseline and the training/evaluation protocol used for vision datasets.","marker":"Huang et al., 2023"},{"why":"Supplies the InsVP instance-level baseline and the dataset protocol the vision experiments follow.","marker":"Liu et al., 2024"},{"why":"Supplies GPF-Plus, the universal graph prompt method that defines the graph-task benchmarks and GIN pre-training setup.","marker":"Fang et al., 2023"}],"fun_headline_variants":["Low-rank semantics unlock efficient prompting for vision graphs","Frozen vision GNNs match full fine-tuning with low-rank prompts","Semantic low-rank prompts beat full fine-tuning on vision graphs","Prompting vision graphs via low-rank semantic decomposition","Low-rank prompts: vision graph adaptation with 94% fewer params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that semantically connected graph nodes share dominant PCA components because visual semantics is genuinely low-rank; since edges are drawn by K-nearest-neighbor search in the same feature space, this concentration could be an artifact of the graph-construction rule rather than a fact about semantics.","fun_headline_variants_meta":{"raw":{"variants":["Low-rank semantics unlock efficient prompting for vision graphs","Frozen vision GNNs match full fine-tuning with low-rank prompts","Semantic low-rank prompts beat full fine-tuning on vision graphs","Prompting vision graphs via low-rank semantic decomposition","Low-rank prompts: vision graph adaptation with 94% fewer params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1288,"prompt_tokens":931,"completion_tokens":357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":274}},"tokens_in":547,"tokens_out":357,"duration_ms":3606,"temperature":1.0,"reasoning_tokens":274,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:36:33.937030+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a ViG and compute the PCA spectrum of node features for three graph constructions — KNN, spatial adjacency, and random neighborhoods; if the eigenvalue decay is just as sharp for non-semantic neighborhoods, the low-rank 'semantic' claim is an artifact, and VGP should be re-tested with random low-rank prompt projections as a control.","supporting_citations":[],"review_version":1}