{"id":"40e47c6e-efe2-434a-bf32-7ba7b022254b","arxiv_id":"2608.08406","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A lightweight GNN surrogate trained on seed-outcome pairs, combined with batched multi-swap simulated annealing, outperforms deeper learning-based influence maximization frameworks on tested benchmarks.","lead":"SIMBA replaces complex neural architectures and continuous optimization in learning-based influence maximization with a lightweight GNN surrogate and direct discrete search. It reports higher influence spread and lower time-to-solution than deeper frameworks, though its extreme gains over classical baselines on the Digg graph need an independent check.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Digg IC comparison with IMM is internally inconsistent under IMM's approximation guarantee; the paper does not rule out a baseline evaluation mismatch.","rationale":"The reader's weakest assumption identifies exactly the most load-bearing concern: baseline evaluation parity. The Digg IC result is not merely surprising but mathematically inconsistent with IMM's approximation guarantee under the stated diffusion model, making it the sharpest test of the paper's central claim. I considered other possible weaknesses, such as the surrogate possibly overfitting during discrete search or the uniform anchor making unseen nodes indistinguishable; however, these are either mitigated by the paper's final-simulation evaluation or are secondary to the comparative claim. The proposed concrete test directly settles whether the Digg comparison is fair, which determines whether the headline result survives. Since the reader already reached a CONDITIONAL verdict and my analysis supports that same conclusion without adding a new objection, no verdict adjustment is needed.","tokens_in":34692,"tokens_out":4230,"duration_ms":46452,"concrete_test":"Run the provided code to re-execute IMM and OPIM on Digg using the same IC parameters as SIMBA: p_{u,v}=1/d_in(v), 100 diffusion steps, seed budget 1% (2,796 nodes), and evaluate final spread with 5x100 Monte-Carlo simulations. Also inspect the baseline scripts to confirm they use these parameters rather than library defaults. If IMM's spread becomes at least 38%, the Table 1 comparison is invalid and the central claim needs re-benchmarking; if it remains near 7.4%, the discrepancy suggests SIMBA's evaluation protocol differs from IMM's and should be reconciled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparative claim (Section 5.1, Table 1) is that SIMBA outperforms all baselines in 62/72 settings. The sharpest evidence is Digg IC at 1% seeds: SIMBA (C) achieves 60.3% spread while IMM achieves 7.4%. This is incompatible with IMM's (1-1/e-epsilon) approximation guarantee (Tang et al., 2015) under the same IC model: if SIMBA's 60.3% is the true spread of a feasible seed set, then OPT >= 60.3%, so IMM should return at least (1-1/e-epsilon)*OPT, approximately 38% for small epsilon. A reported 7.4% therefore implies either different diffusion parameters, different simulation length, different seed budget, or an insufficient number of reverse-reachable sets for IMM. Appendix B states diffusion configurations are 'consistent with prior works,' but the paper does not provide per-baseline verification, and no baseline error bars are reported (Appendix E.2 gives standard deviations only for SIMBA). Because the headline comparison rests on these numbers, the evaluation mismatch is load-bearing. If the Digg comparison is unfair, the claim that a simple surrogate beats classical near-optimal algorithms on large graphs is unsupported; this could also affect other settings where baselines are copied from prior papers rather than rerun under identical conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper challenges the assumption that learning-based influence maximization (IM) requires complex neural architectures and continuous optimization over seed representations. SIMBA combines uniformly anchored node embeddings (all node embeddings initialized from a shared vector, with seed sets encoded by masking), a two-layer GraphSAGE surrogate trained to predict final infection states, and batched multi-swap simulated annealing that searches directly in discrete seed space. The framework is evaluated on six graphs under IC, LT, and SIS models at 1%, 5%, 10%, and 20% seed budgets, with reported influence spread computed from actual diffusion simulations rather than from the surrogate. The paper claims that SIMBA achieves the highest influence spread in 62 of 72 settings, substantially reduces time-to-solution, and retains performance with as little as 1% of the training data, and it offers ablations and a geometric analysis of the learned embeddings.","tokens_in":35001,"tokens_out":6311,"duration_ms":67382,"significance":"If the empirical claims hold, the paper makes a useful conceptual contribution: it suggests that the main bottleneck in prior neural-surrogate IM work is the continuous optimization/search strategy rather than surrogate capacity, and it offers a lightweight, data-efficient alternative. The manuscript has clear strengths: the method is described precisely, the code is publicly linked, the headline evaluation uses actual simulations so the reported spread is not circularly defined by the surrogate, and the ablation study (initialization, embedding dimension, binary-input variant) is relatively thorough. The main caveats concern the fairness and statistical grounding of the baseline comparisons: the anomalous Digg results are difficult to reconcile with the approximation guarantees of classical near-optimal baselines, baseline error bars are absent, and a baseline included in an appendix contradicts the statement that SIMBA is consistently the strongest. These issues are load-bearing and need to be resolved before the central claims can be accepted.","major_comments":[{"comment":"The reported Digg IC result at 1% seeds (SIMBA (C) 60.3% versus IMM 7.4%) is hard to reconcile with IMM's (1-1/e-epsilon) approximation guarantee under the same diffusion model. Since Appendix B states the IC setting is p_u,v = 1/d_in(v) with 100 simulation steps, and Appendix C says all methods are evaluated by the same 5x100 simulation rounds, a feasible seed set with true spread 60.3% would imply OPT >= 60.3%, so IMM should return at least roughly 38% (for small epsilon). The paper does not rule out an evaluation mismatch, for example baseline numbers being taken from prior papers run under different diffusion parameters, simulation lengths, or seed budgets. This is load-bearing for the central comparative claim, and the authors should either rerun all classical baselines under the exact evaluation protocol used for SIMBA or provide per-baseline verification that the same IC parameters, simulation length, and seed budget were used.","section":"Section 5.1, Table 1 (Digg IC) with Appendix B and C"},{"comment":"Only SIMBA's standard deviations are reported; all baselines are point estimates. Many of the claimed wins are small in magnitude (e.g., Power Grid SIS at 1% seeds is 1.9 for essentially all methods, and many Synthetic SIS differences are 0.2-0.4 percentage points), and the 62/72 count treats every setting equally regardless of effect size. Without error bars or significance tests for the baselines, the claim that SIMBA 'consistently achieves the strongest performance' is statistically unsupported. The authors should report variance over repeated runs for all baselines, or at least for the near-tie settings and the settings used to support the headline count.","section":"Section 5.1, Table 1, and Appendix E.2"},{"comment":"The headline claim that SIMBA achieves the highest influence spread in 62 of 72 settings counts only the baselines in Table 1. The appendix's comparison with DynaFLUX shows that DynaFLUX and DynaFLUX+ outperform SIMBA by large margins in many settings (e.g., Cora-ML IC at 1% seeds: 35.3/38.4 versus 14.2-17.2 for SIMBA; Jazz IC at 1% seeds: 26.0/26.3 versus 10.4-12.3). If DynaFLUX is a legitimate baseline, then the 'consistently strongest' claim is false as stated. If it is excluded because it requires access to the underlying diffusion model, the paper should say so explicitly and restrict the positioning to model-agnostic learning-based methods.","section":"Section 5.1 versus Appendix E.3, Table 8"},{"comment":"The claim that SIMBA's performance 'saturates rapidly' with a small fraction of training data is not clearly supported by the numbers. For example, on Jazz LT at 10% seeds, SIMBA (U) achieves 58.7% with 1% training data but only 48.0% with 100% training data, and there are other non-monotonic swings of several percentage points (e.g., Jazz SIS at 20% seeds: 28.9% with 20% data versus 30.8% with 100% data). Since no standard deviations are reported for these runs, the reader cannot distinguish signal from noise. The data-efficiency claim needs error bars or a paired analysis over repeated training-data fractions.","section":"Section 5.3, Table 3"}],"minor_comments":[{"comment":"With T0 = 1e-6 and alpha = 1, the acceptance rule never accepts a non-improving move, so the method is effectively a batched multi-swap hill-climbing search rather than simulated annealing; the authors should rename it or explicitly describe this configuration as a greedy variant.","section":"Algorithm 1 and Section 5 implementation details"},{"comment":"Several numeric entries are concatenated in the rendering (e.g., '41.549.9', '36.961.0', '8.526.9', and similar), making columns hard to parse; the tables should be reformatted so every number is clearly separated.","section":"Table 1 and other large tables"},{"comment":"The statement that diffusion configurations are 'consistent with prior works' is too vague for reproduction; please provide the exact parameters used for each baseline and, where baseline numbers are taken from prior papers, cite the source table or figure for each dataset.","section":"Appendix B"},{"comment":"The claim that both training and optimization 'scale linearly' is asserted informally; the search cost also depends on the number of steps I, batch size B, and swap size r, so a precise cost model would make the scalability discussion more rigorous.","section":"Section 4.4 and Section 5.2"},{"comment":"The geometric interpretation (e.g., 'linear scalability' and 'prevents uncalibrated local maxima') is not directly established by the embedding plots; these statements should be qualified as observations or supported by a quantitative analysis.","section":"Section 5.4"}],"recommendation":"major_revision","confidential_remarks":"The Digg anomaly in Section 5.1 is the single most important issue: it suggests a possible evaluation mismatch between SIMBA and the classical baselines, and if baseline numbers were copied from prior papers rather than rerun under identical conditions, the headline comparison could be invalid. I would ask the authors to provide raw logs or a rerun of IMM/OPIM on Digg under the exact evaluation protocol before the paper can be accepted. I also note that the 62/72 count excludes DynaFLUX, which is otherwise included in the appendix; the paper's self-positioning should be aligned with the full set of reported comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: SIMBA is a clean, simple recipe—two-layer GraphSAGE surrogate, uniformly anchored node embeddings, batched multi-swap simulated annealing—and the paper shows across many configurations that this lightweight combo beats heavier learning-based IM frameworks that rely on continuous latent optimization. That is a genuinely useful result, and the paper earns credit for testing it thoroughly: data-efficiency fractions, embedding ablation, 2D embedding visualization, binary-input variant, multiple diffusion models, five real graphs and one synthetic. The embedding-geometry explanation is speculative but presented as such.\n\nThe soft spot is the evaluation. The Digg IC table is internally inconsistent. SIMBA reports 60.3% spread at 1% seeds while IMM/OPIM/SubSIM report 7.4–7.6%. IMM has a (1-1/e-ε) approximation guarantee under the same IC parameters. If SIMBA's seed set truly reaches 60.3%, the optimum is at least that high, so IMM should return roughly 38% or better. It doesn't. That means the baselines were almost certainly run under different diffusion parameters, a different simulation length, a different seed budget, or a different number of RR sets. Appendix B says the configurations are 'consistent with prior works,' which resolves nothing—it just says the numbers might be copied. This is load-bearing: the paper's central claim is that a simple surrogate beats near-optimal classical algorithms on large graphs, and Digg is the largest graph.\n\nAlso, error bars exist only for SIMBA (Appendix E.2), and some of those are wide—Jazz LT has 48.0±15.3—so the headline spread differences aren't as sharp as the point estimates suggest. The data-efficiency experiments are noisy at 1% training data but the saturation pattern holds, so that part is fine.\n\nOn the positive side, the circularity burden is low: reported spread comes from actual diffusion simulations, not from the surrogate, and the code is linked. Self-citation isn't an issue. The authors should be asked to rerun IMM, OPIM, and SubSIM under identical diffusion configurations on Digg, report baseline error bars, and explain the Digg gap. If the gap survives, the result is important; if not, the paper's contribution is still the architecture lesson, but with more modest comparative claims.\n\nI'd send it to a serious referee with a request for a controlled rerun. It's a solid candidate after revision; right now the empirical headline is conditional on the Digg recheck.","headline":"SIMBA's simple surrogate-plus-discrete-search recipe is genuinely promising and thoroughly tested, but the Digg IC comparison is internally inconsistent with IMM's approximation guarantee, so the headline empirical claim needs a controlled baseline rerun before it can be trusted.","tokens_in":35468,"tokens_out":3587,"would_cite":true,"duration_ms":37138,"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":"This paper claims that a lightweight two-layer graph neural network surrogate paired with batched multi-swap simulated annealing can outperform complex learning-based influence maximization frameworks that rely on continuous optimization.","keywords":["influence maximization","neural surrogate","simulated annealing","graph neural network","discrete optimization","diffusion-model-agnostic","seed selection","data efficiency"],"falsifier":"On the largest graph's independent cascade benchmark, re-run SIMBA and the classical near-optimal baselines with identical propagation probabilities, 100 simulation steps, and the same seed budget; if the classical method's measured spread rises from the reported single digits to SIMBA's ~60% at 1% seeds, the central superiority claim would be an artifact of mismatched evaluation rather than a property of the method.","tokens_in":34527,"feed_emoji":"🕸️","tokens_out":10963,"duration_ms":102754,"temperature":0.7,"pith_summary":"This paper argues that effective learning-based influence maximization does not require expressive neural architectures or continuous optimization over seed representations. It introduces SIMBA, which combines a two-layer graph neural network surrogate, uniformly anchored node embeddings, and batched multi-swap simulated annealing to search the discrete seed space directly. On five real-world graphs and one synthetic graph, under three diffusion models, SIMBA reports the highest influence spread in 62 of 72 settings and substantially lower time-to-solution than existing frameworks. The authors interpret this as evidence that the optimization strategy, not surrogate expressiveness, is the bottleneck, and that a small fraction of diffusion observations is enough to train the surrogate.","feed_headline":"Two-layer neural net plus discrete search beats complex IM baselines","feed_subtitle":"The SIMBA framework gets the best spread in 62 of 72 settings while cutting time-to-solution.","key_machinery":"The load-bearing object is the uniformly anchored node embedding: every node's learnable representation is initialized at the same random vector $c \\in \\mathbb{R}^d$, so that message passing learns purely from graph topology and diffusion outcomes, and a seed set is encoded by zeros for non-seed nodes. This trick makes unseen nodes share the anchor as a special token, and it appears to make the learned surrogate smooth over novel seed configurations. The second half of the machinery is batched multi-swap simulated annealing, which samples $r$ nodes to remove and $r$ to add per candidate, evaluates all $B$ candidates in a single batched forward pass, and accepts swaps by the standard annealing rule; this keeps every candidate a valid size-$k$ set without gradients or continuous relaxations.","core_discovery":"SIMBA's central claim is that a deliberately simple pipeline can outperform sophisticated learning-based influence maximization. Every node embedding starts from the same random anchor, the seed configuration is expressed by masking non-seed embeddings to zero, and a two-layer graph neural network predicts the final infection state of every node. A batched multi-swap simulated annealer then generates candidate size-$k$ seed sets, evaluates all of them in one forward pass through the frozen surrogate, and keeps the best. The paper reports this combination reaches the highest influence spread in 86.1% of its 72 evaluation settings, and that the surrogate's predictions are smooth enough that expanding the search to nodes never seen during training improves rather than degrades results. The authors attribute this to the embedding geometry: anchored embeddings tend to form a low-dimensional manifold, which keeps surrogate-guided search away from uncalibrated local optima.","pith_inferences":["Editorial extension: the largest reported gap (a classical baseline at single-digit spread vs SIMBA around 60% at 1% seeds) is consistent with an evaluation-protocol mismatch; a controlled replication with identical simulation length and seed budget would determine whether the advantage is real.","Editorial extension: if the manifold explanation is right, then regularizing embeddings to stay near a low-dimensional subspace could further stabilize surrogate-guided search, which the paper does not test.","Editorial extension: the same template - cheap neural evaluator plus batched multi-swap search - may transfer to other combinatorial problems with expensive-to-simulate objective functions, such as sensor placement or experimental design.","Editorial extension: because SIMBA is diffusion-model-agnostic, a natural stress test is time-varying or competitive diffusion, where the surrogate would have to adapt without architectural changes; the paper only tests three static models."],"forward_implications":["If the results hold, learning-based influence maximization can drop autoencoders, latent seed vectors, and sheaf layers; a two-layer graph network plus swap-based annealing suffices.","Time-to-solution drops because training is quick and each search step is one batched forward pass, making the approach practical for graphs with hundreds of thousands of nodes.","Because the surrogate trains on a small fraction of seed-outcome pairs, diffusion-model-agnostic IM becomes usable where simulation data is scarce.","Expanding the search to nodes never seen during training improves spread, which means the surrogate extrapolates well beyond its training configurations."],"supporting_citations":[{"why":"Defines the influence maximization problem and the cascade/threshold diffusion models that the surrogate learns to predict.","marker":"(Kempe et al., 2003)"},{"why":"Provides the classical near-optimal baseline whose measured spread SIMBA must surpass; its approximation guarantee makes the comparison protocol decisive.","marker":"(Tang et al., 2015)"},{"why":"Supplies the continuous latent-space optimization framework and the synthetic random-graph benchmark that SIMBA challenges.","marker":"(Ling et al., 2023)"},{"why":"The sheaf-neural-network surrogate representing the complex architecture camp; SIMBA's lightweight surrogate is compared against it.","marker":"(Hevapathige et al., 2025)"},{"why":"Defines the two-layer neighbor-aggregation architecture used as SIMBA's lightweight surrogate.","marker":"(Hamilton et al., 2017)"},{"why":"Source of the largest benchmark graph where the most extreme spread differences appear.","marker":"(Lerman et al., 2012)"}],"fun_headline_variants":["SIMBA: two-layer GNN and annealing beat complex IM baselines","Simple GNN plus discrete search beats heavy IM frameworks","Two-layer net and annealing outperform complex IM baselines","SIMBA: light GNN and native search win influence spread","Surrogate-guided annealing beats deep IM baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes every baseline was run under exactly the same diffusion parameters, simulation lengths, and seed budgets as SIMBA on every dataset; if a classical near-optimal baseline was evaluated under a different protocol, the reported spread gaps could reflect evaluation conditions rather than method quality.","fun_headline_variants_meta":{"raw":{"variants":["SIMBA: two-layer GNN and annealing beat complex IM baselines","Simple GNN plus discrete search beats heavy IM frameworks","Two-layer net and annealing outperform complex IM baselines","SIMBA: light GNN and native search win influence spread","Surrogate-guided annealing beats deep IM baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000624,"raw_usage":{"total_tokens":2845,"prompt_tokens":853,"completion_tokens":1992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":1911}},"tokens_in":469,"tokens_out":1992,"duration_ms":15267,"temperature":1.0,"reasoning_tokens":1911,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:36:37.090639+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the largest graph's independent cascade benchmark, re-run SIMBA and the classical near-optimal baselines with identical propagation probabilities, 100 simulation steps, and the same seed budget; if the classical method's measured spread rises from the reported single digits to SIMBA's ~60% at 1% seeds, the central superiority claim would be an artifact of mismatched evaluation rather than a property of the method.","supporting_citations":[{"cited_title":"Influence maximization in near-linear time: A martingale approach","cited_arxiv_id":null,"evidence_quote":"Provides the classical near-optimal baseline whose measured spread SIMBA must surpass; its approximation guarantee makes the comparison protocol decisive."}],"review_version":1}