{"id":"78119831-70ae-4adb-aa05-1e3eb1ee4ca3","arxiv_id":"2412.12155","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SGPT adapts pre-trained unsigned GNNs to few-shot signed graph node classification and link sign prediction via balance-theory channel decomposition and lightweight prompts.","lead":"SGPT is a method that lets a graph neural network pre-trained on plain, unsigned networks be reused for few-shot tasks on signed networks, where links carry positive or negative meaning. It separates signed links into positive, negative, and topology channels with a balance-theory template, then tunes small prompts instead of the whole model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SGPT's core graph template (Eqs. 4-5) assumes unambiguous balance-theoretic path signs; when a node pair is connected by both balanced and unbalanced paths, the indicator binarizes the pair into both channels, breaking the claimed disentanglement.","rationale":"The reader's verdict is already CONDITIONAL, and the weakest assumption identified there matches the stress-test's most load-bearing concern: the balance-theory graph template relies on an unambiguous path-sign assignment that real signed graphs do not guarantee. The paper's own Eq. 4-5 and the indicator function make this explicit; the manuscript never defines what happens when A^k_P and A^k_N both contain the same pair, nor does it measure how often this occurs in the evaluated datasets. Because the graph template is the primary mechanism claimed to fix the graph-type divergence between pre-training and downstream, an empirical demonstration on low-balance or sign-shuffled graphs is necessary to attribute SGPT's gains to balance-theoretic disentanglement rather than to the additional hop-2 connections or the task template. The concern is not a logical contradiction in the paper; it is an untested empirical precondition. The proposed concrete test directly measures the ambiguity rate and ablates it, which would settle whether the central claim survives. I therefore keep the verdict at CONDITIONAL: the architecture may well work, but the stated reason for its core design choice is currently unsupported without balance-level diagnostics.","tokens_in":17166,"tokens_out":5824,"duration_ms":70184,"concrete_test":"For each of the seven datasets, compute the 2-hop ambiguity ratio r = |{(i,j) : A^2_P(i,j)=1 and A^2_N(i,j)=1}| / |{(i,j) : A^2_P(i,j)=1 or A^2_N(i,j)=1}|, and report it alongside the balance fraction (e.g., fraction of signed triangles with an even number of negative edges). Then construct synthetic signed graphs with the same degree sequence and positive/negative edge counts but controlled balance by randomly flipping a varying fraction of link signs (or by rewiring to force unbalanced triangles), and run SGPT under the same few-shot protocol. If SGPT's ROC-AUC degrades monotonically as r increases, or if its advantage over Variant-1 disappears when the two channels contain mostly overlapping edges, the disentanglement assumption of Eqs.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central novelty of SGPT is the graph template that disentangles mixed signed relationships into positive, negative, and topological channels, each asserted to contain a consistent link semantics compatible with a frozen unsigned GNN. This disentanglement is implemented by Eqs. 4-5, where A^k_P and A^k_N are obtained from products of the previous-hop matrices followed by an indicator I(·) that binarizes every nonzero entry to 1. The implicit assumption is that each node pair has a unique sign at each hop, i.e., that balanced and unbalanced paths do not both connect the same pair. In real signed networks, a pair (i,j) commonly has both an even-length path (making it positive by balance theory) and an odd-length path (making it negative). Then A^2_P(i,j) = 1 and A^2_N(i,j) = 1 simultaneously, so the same edge is inserted into both the positive and negative channels. The 'consistent type of internal link semantics' is therefore not achieved; the two channels become partially redundant and may carry nearly identical topological information. The problem compounds with hop count and with the density of mixed-sign neighborhoods. The paper asserts that balanced structures are prevalent in real-world networks and cites SGCL [28], but it reports no balance statistics (e.g., fraction of balanced triads, or fraction of node pairs with ambiguous path signs) for any of the seven datasets. It also provides no ablation on synthetic graphs with controlled balance levels. Consequently, the mechanism that supposedly justifies the graph template is unverified in exactly the regime where the method is claimed to excel. If the experimental gains are instead driven mainly by the extra 2-hop edges (which Variant-1 omits) or by the task template and prompts, the 'balance-theory disentanglement' contribution would not be validated. This is a load-bearing concern because the graph template is the paper's primary response to the pre-training/downstream graph-type divergence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SGPT (Signed Graph Prompt Tuning), a framework that adapts pre-trained unsigned GNNs to few-shot signed graph tasks (node classification and link sign prediction). SGPT introduces a graph template based on balance theory that separates multi-hop relationships into positive, negative, and topological channels; a task template that unifies downstream tasks into a link-prediction form; channel-specific feature prompts; and a semantic-prompt adapter that fuses channel embeddings. The authors evaluate on seven signed-network datasets against supervised SGNNs, pre-train/fine-tune methods, and graph prompt-tuning baselines, reporting ROC-AUC improvements across tasks and shots.","tokens_in":17493,"tokens_out":5687,"duration_ms":54849,"significance":"If the reported results are valid, SGPT would be a practically useful contribution to label-scarce signed graph learning, showing how abundant unsigned graph pre-training can be transferred to signed tasks without retraining the backbone. The paper is well structured, provides ablations that isolate the proposed components, includes sensitivity analyses for the hop number and prompt basis size, and offers complexity analysis. However, the main empirical claim is currently weakened by an evaluation protocol that gives signed-aware models extra sign information, and the core disentanglement property of the graph template is not validated under ambiguous multi-hop paths. These issues are addressable but must be fixed before the contribution can be assessed fairly.","major_comments":[{"comment":"The link sign prediction protocol is not fair across methods. The paper states that for the 30% message-passing links, signed graph models (including SGPT, SGCN, and SDGNN) know the link signs when generating embeddings or applying the graph template, while unsigned models treat these links as unsigned. This gives SGPT and other signed baselines access to 30% of the ground-truth signs in addition to the few-shot training labels, while unsigned prompt baselines (GPPT, GraphPrompt, GPF+, etc.) do not receive that information. The observed advantage over unsigned models may therefore reflect this extra supervision rather than the proposed templates or prompts. Please rerun experiments under a controlled protocol in which all methods see the same sign information (or none do), or add an ablation where unsigned models also receive the 30% signs as an explicit two-channel edge feature, so the contribution of the graph template and prompts can be isolated.","section":"§5.1 (Implementations)"},{"comment":"The graph template claims to disentangle mixed node relationships into channels with consistent link semantics, but the construction via the indicator function assumes that each node pair has a unique balance-theoretic sign at each hop. If a pair (i,j) is connected by both a balanced and an unbalanced k-hop path, then both A^k_P(i,j)=1 and A^k_N(i,j)=1 after applying I(·), so the same edge is inserted into both the positive and negative channels. The paper asserts that balanced structures are prevalent in real-world networks but reports no balance statistics (e.g., fraction of balanced triads or fraction of node pairs with mixed-sign paths) for any of the seven datasets, and provides no synthetic experiments with controlled balance levels. Please measure the prevalence of ambiguous path-sign pairs in the datasets or include a controlled synthetic study; if ambiguity is common, explain why the method remains effective or modify the template (e.g., using path counts or sign voting) to preserve disentanglement.","section":"§4.3, Eqs. (4)–(5)"},{"comment":"The paper claims that SGPT \"significantly outperforms\" existing methods, but no statistical significance testing is reported. Several differences in Table 2 are within one standard deviation (e.g., WikiEditor: SGPT 54.17±5.51 vs. GPF+ 52.61±5.28; WikiElec: GraphPrompt+ 51.23±9.11 vs. SGCN 51.06±1.52). Since the results are aggregated over 100 random tasks, paired significance tests (e.g., Wilcoxon signed-rank or paired t-test) should be reported for all main comparisons, and the wording should be adjusted to the actual statistical evidence.","section":"§5.2, Table 2 and Fig. 3"},{"comment":"Pre-training and prompt-tuning details are insufficient for reproducibility. The paper does not specify which unsigned graph(s) were used for pre-training, their size and domain, the node feature transformation, the number of pre-training epochs, the number of prompt-tuning epochs, batch size, or how the class prototypes in Eqs. (10)–(11) are initialized and updated. These details are necessary to reproduce the reported numbers and to allow the community to build on the method. Please add a reproducible experimental setup description or an appendix with these hyperparameters.","section":"§5.1 and §4.2"}],"minor_comments":[{"comment":"The GraphPrompt+ entry for Wikipedia-RfA (75.11±0.24) is identical to the Epinions entry (75.11±0.24) and appears to be a copy-paste error; please correct it.","section":"Table 2"},{"comment":"The dataset name \"Wikiedia-Editor\" in Table 2 should be \"Wikipedia-Editor\", and \"GraphPromp+\" in Section 5.2 should be \"GraphPrompt+\".","section":"Table 2 and §5.2"},{"comment":"\"preserving instinct graph structures\" should likely be \"preserving intrinsic graph structures\".","section":"§1"},{"comment":"References [9] and [10] are duplicates (both cite \"Signed graph attention networks\" by Huang et al.), and references [23] and [24] are also duplicates (both cite the same Mercado et al. paper); please consolidate.","section":"References"},{"comment":"The claim of being \"the first study\" should be supported by a more explicit search of the literature or softened; if prior signed-graph prompt-tuning work exists, it should be cited and discussed.","section":"§1"},{"comment":"It is unclear whether the class prototypes E=[e_P,e_N] and E=[e_{c1},...,e_{cm}] are learnable parameters, computed from the training set, or updated during prompt tuning; please clarify.","section":"§4.4, Eqs. (10)–(11)"}],"recommendation":"major_revision","confidential_remarks":"The evaluation protocol issue in Section 5.1 is the most serious problem: it makes the headline comparison against unsigned baselines unfair, and the fix (running a controlled protocol) may change the conclusions. The balance-theory ambiguity concern is also substantive because the graph template is the paper's central novelty. I would recommend inviting a revision and asking for (a) re-run experiments with a fair protocol, (b) balance/ambiguity statistics or synthetic tests, (c) significance testing, and (d) full reproducibility details. If the authors cannot provide these, the paper should not be accepted in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is the first paper to bring prompt tuning to signed graphs, and the overall idea is sound. But the core contribution -- the balance-theory graph template -- rests on an assumption the paper never checks, and the experimental protocol has a couple of unfair comparisons. I'd send it to a serious referee, but I wouldn't take the main claim at face value yet.\n\nWhat's genuinely new: SGPT is the first to adapt pre-trained unsigned GNNs to few-shot signed tasks. The graph template decomposes the signed graph into positive, negative, and topological channels using balance-theory path parity. The task template unifies node classification and link sign prediction into link prediction, which is a neat trick. Feature prompts and semantic prompts are lightweight and parameter-efficient. The experiments cover seven datasets and show consistent gains over supervised SGNNs, fine-tuning, and unsigned prompt-tuning baselines. The ablation study shows each component contributes, and the complexity analysis is honest.\n\nWhere it gets soft. The stress-test note is on target. Equations 4-5 use an indicator function that binarizes nonzero entries. If a node pair is connected by both balanced and unbalanced paths -- common in real signed networks -- that pair goes into both the positive and negative channels. The claimed 'consistent type of internal link semantics' doesn't hold. The paper cites SGCL for the prevalence of balanced structures but reports no balance statistics for any dataset and runs no synthetic experiments with controlled balance. So the mechanism that supposedly justifies the template is unverified where it matters most. The strong ablation of the graph template (Variant-1) shows it helps, but doesn't distinguish balance-theoretic disentanglement from simply adding 2-hop edges.\n\nThere's also a copy-paste error in Table 2: GraphPrompt+ has the same value (75.11±0.24) for Epinions and Wikipedia-RfA. The LSP protocol gives signed models the signs of 30% of links for message-passing, while unsigned baselines see only topology; that tilts the comparison. No significance testing, and the pre-training graph and prompt hyperparameters (basis count, bottleneck dimension) aren't specified for the main runs.\n\nBottom line: this is a genuine first step with a sensible design, but the central claim is under-supported. The paper deserves a serious referee, but the authors need to address the ambiguity issue, fix the protocol, and add statistical tests before the 'universal superiority' claim is credible.","headline":"First prompt-tuning framework for signed graphs with a sensible design and strong results, but the balance-theory disentanglement is under-validated and the evaluation has protocol issues; worth serious refereeing but needs revision.","tokens_in":18044,"tokens_out":3795,"would_cite":true,"duration_ms":35525,"reading_group":"yes","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 proposes SGPT, a prompt-tuning framework that transfers pre-trained unsigned GNNs to few-shot signed graph learning.","keywords":["signed graphs","prompt tuning","few-shot learning","graph neural networks","balance theory","link sign prediction","node classification","pre-training"],"falsifier":"Compute the structural balance level of the seven datasets, or split any signed graph into high-balance and low-balance subgraphs, and check whether SGPT's AUC drops as contradictory paths become more common; if performance stays flat even when balance is low, the channel decomposition itself is not what drives the reported gains.","tokens_in":17017,"feed_emoji":"🕸️","tokens_out":4690,"duration_ms":44954,"temperature":0.7,"pith_summary":"Pre-trained graph neural networks are usually built for unsigned graphs, where links only mean connection, while signed graph tasks need to know whether a link is trust or distrust. SGPT claims that a frozen unsigned GNN can be reused for both node classification and link sign prediction with very few labels, as long as the signed graph is first repackaged by templates that make it look like the pre-training data. If true, this would let practitioners skip training specialized signed GNNs and instead adapt abundant unsigned graph models with only a few thousand tunable prompt parameters. The paper reports consistent gains over supervised signed models and prior graph-prompt methods on seven datasets under few-shot settings.","feed_headline":"Pre-trained unsigned GNNs can handle signed graphs with few labels","feed_subtitle":"Balance-theory templates split trust and distrust into separate channels, so frozen encoders transfer without retraining.","key_machinery":"The load-bearing mechanism is the three-channel graph template built from balance theory. With $A^1_P = A^+$ and $A^1_N = A^-$, each further hop is computed as $A^k_P = I(A^{k-1}_P A^1_P + A^{k-1}_N A^1_N)$ and $A^k_N = I(A^{k-1}_P A^1_N + A^{k-1}_N A^1_P)$, where $I(\\cdot)$ binarizes non-zero entries; this encodes whether paths have an even or odd number of negative links. The positive, negative, and topological samples are encoded in parallel by the frozen pre-trained GNN, aligned by channel-specific feature prompts and fused by a semantic prompt, so the whole pipeline behaves like the pre-training link-prediction task.","core_discovery":"The central claim is that the structural gap between unsigned pre-training and signed downstream tasks can be closed by a balance-theory graph template that splits a signed graph into three channels: positive, negative, and purely topological. Multi-hop positive and negative adjacencies are built by recursive rules that encode the balance-theoretic sign of a path, so each channel contains links of one consistent semantics and can be fed into an unsigned GNN without violating its homophily assumption. A task template then rewrites both node classification and link sign prediction as prototype-based link prediction, matching the pre-training objective. Lightweight feature prompts adjust each channel's input space, and a bottleneck-adapter semantic prompt fuses the channels task-adaptively. The paper claims this is the first prompt-tuning framework to transfer pre-trained unsigned GNNs to few-shot signed tasks and that it outperforms existing supervised SGNNs and unsigned prompt-tuning baselines on seven benchmark datasets.","pith_inferences":["Editorial inference: the method's success should depend on how balanced the test graphs actually are; datasets with many contradictory paths between the same pair are where the binarizing indicator function in the recursive adjacency construction is least reliable.","Editorial inference: because the indicator function discards path multiplicity, a testable extension is to weight channels by counts or confidence of balanced versus unbalanced paths rather than binarizing them, which may improve robustness on noisy signed networks.","Editorial inference: the framework suggests a practical recipe beyond the paper's benchmarks: take any pre-trained unsigned GNN, add these templates, and probe it on directed or weighted signed graphs, provided the balance-theoretic sign rule still applies."],"forward_implications":["Signed graph node classification and link sign prediction can in principle be solved without training a dedicated signed GNN, reusing a frozen unsigned backbone instead.","Because only prompts are tuned, the downstream phase needs very few labels and far fewer tunable parameters than supervised fine-tuning, which helps in label-scarce industrial settings.","The same template and prompt design carries across different pre-trained backbones (GCN, GAT, GIN), so the method is not tied to one encoder architecture.","The task template makes two seemingly different tasks share one objective, so gains on one signed task can transfer to the other.","Balance-theoretic multi-hop relationships give the method a way to incorporate global structure beyond direct signed neighbors."],"supporting_citations":[{"why":"Supplies balance theory, the social-psychology principle that a triad is balanced with an even number of negative links, which motivates the graph template.","marker":"[6]"},{"why":"Provides the signed graph convolutional network that aggregates features based on balanced and unbalanced paths, the basis for decomposing relationships by path sign.","marker":"[3]"},{"why":"Defines the link-prediction self-supervised pre-training strategy used to train the unsigned GNN backbone.","marker":"[8]"},{"why":"Introduces universal feature prompts for GNNs, the design that SGPT adapts into channel-specific feature prompts.","marker":"[4]"},{"why":"A strong supervised signed GNN baseline that SGPT must beat and that motivates semantic-aware handling of signed relationships.","marker":"[11]"},{"why":"The GPPT method that reformulates node classification as link prediction, a precursor to the task template.","marker":"[29]"},{"why":"The ProG multi-task prompting framework that supplies the few-shot setting and prototype-based objective used in downstream evaluation.","marker":"[32]"},{"why":"Graph contrastive pre-training baseline that is compared against link-prediction pre-training for downstream signed tasks.","marker":"[44]"},{"why":"The HGPrompt framework whose few-shot node classification setting is followed and which unifies homogeneous and heterogeneous graphs.","marker":"[45]"}],"fun_headline_variants":["Prompt tuning adapts offline GNNs to signed graphs with only a few labels","SGPT: Balance-theory prompts bridge unsigned pretraining to signed tasks","Few-shot signed graph learning via transferable prompts on frozen GNNs","Signed graphs tamed by prompt tuning of pre-trained unsigned GNNs","Separating trust and distrust with prompts: SGPT for few-shot signed graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes balance theory holds on the test graphs, so the sign of any multi-hop relationship is determined by whether the number of negative links along the path is even, and contradictory paths can be safely binarized into positive or negative channels.","fun_headline_variants_meta":{"raw":{"variants":["Prompt tuning adapts offline GNNs to signed graphs with only a few labels","SGPT: Balance-theory prompts bridge unsigned pretraining to signed tasks","Few-shot signed graph learning via transferable prompts on frozen GNNs","Signed graphs tamed by prompt tuning of pre-trained unsigned GNNs","Separating trust and distrust with prompts: SGPT for few-shot signed graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1504,"prompt_tokens":977,"completion_tokens":527,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":428}},"tokens_in":593,"tokens_out":527,"duration_ms":5738,"temperature":1.0,"reasoning_tokens":428,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:03:50.995819+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the structural balance level of the seven datasets, or split any signed graph into high-balance and low-balance subgraphs, and check whether SGPT's AUC drops as contradictory paths become more common; if performance stays flat even when balance is low, the channel decomposition itself is not what drives the reported gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the link-prediction self-supervised pre-training strategy used to train the unsigned GNN backbone."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A strong supervised signed GNN baseline that SGPT must beat and that motivates semantic-aware handling of signed relationships."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Graph contrastive pre-training baseline that is compared against link-prediction pre-training for downstream signed tasks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The HGPrompt framework whose few-shot node classification setting is followed and which unifies homogeneous and heterogeneous graphs."}],"review_version":1}