{"id":"d6e32942-c1b7-4428-9c1b-9edde8e8613f","arxiv_id":"2501.05845","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A hybrid pipeline that uses annealing-machine solutions on Louvain-compressed graphs to initialize a graph neural network improves solution quality on large MaxCut, MIS, and graph partitioning instances beyond the annealer's native size limit.","lead":"An annealing machine, a specialized device for solving small binary optimization problems, is paired with graph neural networks to handle much larger graphs. The authors compress a large graph, let the annealing machine solve the compressed pieces, and use those solutions to initialize a neural solver for the original graph.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AM-specific contribution is untested: without a control replacing AM labels with random or heuristic labels on compressed graphs, Table 4 gains could come from any multiresolution supervision rather than from AM knowledge.","rationale":"Reader's weakest assumption (Louvain alignment) is real, but I think the sharper load-bearing problem is causal attribution. The paper's Section 3 asserts alignment without evidence; Table 4 compares AM-guided mrGNN to unguided mrGNN, which cannot separate 'AM label quality' from 'any external labels improve local GNNs and thus main solver initialization.' The right experiment is a label ablation: random labels and greedy labels on the same compressed graphs. If random labels already produce the 25-35% MIS gains, then the central 'AM beyond its limits' claim is unsupported, regardless of whether the alignment assumption holds. The Hamiltonian-loss proxy is also fragile, especially with violations reported in Table 2, so the control should report MIS size and violations, not just loss. This keeps the verdict CONDITIONAL — the idea is plausible and the positive comparisons are suggestive — but the AM-specific transfer remains unestablished.","tokens_in":9916,"tokens_out":5457,"duration_ms":58426,"concrete_test":"Re-run the pipeline on the n=100k/150k MIS graphs (d=4,5) with x_AM_i in Algorithm 1 replaced by (a) random binary vectors and (b) greedy-solver solutions on each Louvain-compressed graph, keeping all other settings identical. Then compute final Hamiltonian, binarized MIS size, and violations for 50 samples and report means/standard deviations. If either control attains relative-loss differences close to Table 4, the AM's knowledge is not the active ingredient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3's Multiresolution Guidance II treats AM solutions as supervision for local GNNs, but the only comparison that isolates AM is mrGNN+AM vs mrGNN (Table 4). Because mrGNN's local GNNs are unsupervised, the comparison conflates the value of AM labels with the value of having any external signal, and with the Louvain hierarchy itself. The paper explicitly assumes alignment between compressed and original QUBOs ('there should be certain alignment'), yet nowhere validates that assumption. A negative control is missing: if random binary vectors or cheap heuristic solutions (e.g., greedy) on the same Gi produce similar relative-loss differences, then the central 'AM beyond limits' claim is not established — the pipeline would work with any labels. The absence of variance/error bars in Table 4 compounds this: the 25-35% MIS figures could be within run-to-run noise. The single most load-bearing concern is therefore causal attribution: the paper has not shown that the AM's specific solutions, rather than the multiresolution scaffolding or the act of supervised guidance, drive the reported gains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a pipeline that combines annealing-machine (AM) solvers with graph neural networks (GNNs) for large combinatorial optimization. Given an original graph G_P, the method applies Louvain compression to obtain a hierarchy of smaller graphs, solves each compressed graph on a CMOS annealing machine, uses the resulting binary solutions as supervised targets for local GNNs, aggregates the learned node representations into initial features for a global GNN solver that minimizes the QUBO Hamiltonian on the original graph. Experiments on synthetic d-regular graphs (n=50k, 100k, 150k; d=3,4,5) for MaxCut, MIS, and Graph Partition compare three variants (rGNN, mrGNN, mrGNN+AM), reporting Hamiltonian loss, violation counts, and wall-clock time, plus comparisons with greedy, HB, and a neural baseline on MIS. The main reported findings are relative loss improvements of roughly 25-35% for MIS at 100k and 150k nodes, smaller gains for MaxCut and GP, and mixed timing results.","tokens_in":10093,"tokens_out":6172,"duration_ms":57857,"significance":"If the improvements are attributable to the AM's guidance, the paper would demonstrate a practical scheme for extending AM-style solution quality to graphs beyond hardware limits, with a credible scale of 150k nodes and multiple problem formulations. The paper's strengths include evaluation on large synthetic graphs, reporting of constraint-violation counts alongside Hamiltonian loss, explicit wall-clock measurements, and comparisons with classical baselines. However, because the experimental design does not isolate the AM-specific contribution and lacks statistical error bars, the central causal claim is not yet established; the correct scope of the contribution depends on the additional controls recommended below.","major_comments":[{"comment":"The central claim that AM knowledge causes the gains is not tested. The comparison mrGNN vs mrGNN+AM conflates three factors: the Louvain hierarchy, the presence of any external supervision signal for the local GNNs, and the specific AM solutions. The paper explicitly assumes alignment between compressed QUBOs and the original problem (\"there should be certain alignment,\" Section 3), but no experiment validates that assumption. A negative control is required: run the identical pipeline with random binary labels, and with cheap heuristic labels (e.g., greedy solutions on each compressed graph), in place of xAM_i. If these controls yield similar relative loss differences on the same test graphs, then the AM-specific \"beyond limits\" claim is not supported; if they do not, the control would establish causal attribution. This is the load-bearing issue for the paper's title and abstract claim.","section":"Section 3 (Multiresolution Guidance I and II); Table 4"},{"comment":"All relative loss differences are reported as single point estimates. The text says solutions were sampled 50 times per target graph, yet Table 4 gives no standard deviation, confidence interval, or significance test for Δrel. For the headline MIS numbers (25-35% at n=100k and 150k), the spread across degrees within a fixed n (e.g., 25.06%, 28.11%, 33.25% at 100k) is comparable to possible run-to-run noise. Reporting mean ± std over the 50 samples, and ideally paired differences between mrGNN and mrGNN+AM on identical graph instances, is necessary to establish that the improvements are not noise.","section":"Section 5 (Solution Quality and Convergence); Table 4"},{"comment":"The abstract and introduction state that the approach reduces \"the overall converge time\" / \"overall convergence time.\" Table 4's ΔT is total execution time with ΔT>0 defined as mrGNN being faster. Under that definition, mrGNN+AM is slower by 21-89% for MIS at n=100k and 150k, and is meaningfully faster mostly at n=50k. For MaxCut and GP, the timing differences are mixed and often small. Thus the general speed claim is contradicted by the reported data; the paper should either restrict the claim to the cases where it holds or report a separate convergence-time metric (e.g., epochs to convergence) if that is what is meant.","section":"Abstract, Section 1, and Table 4"},{"comment":"The primary quality metric is the Hamiltonian loss evaluated on binarized solutions, which is the same objective minimized by all three GNN variants. This creates a self-referential evaluation: a method that more aggressively minimizes the QUBO Hamiltonian will look better by construction. The violation counts partly mitigate this for constraints, and the MIS comparison against classical heuristics provides an external check, but for MaxCut and GP there is no ground-truth or independent objective. The paper should acknowledge this limitation explicitly and, if possible, validate on small graphs with known optima or compare solution sizes and partition balances against additional external baselines.","section":"Section 5 (Solution Quality and Convergence)"}],"minor_comments":[{"comment":"The map ψω is typed as R → {0,1}, but xGNN is a vector of node assignments; the codomain should be {0,1}^{|V|} (or the map should be described as applied elementwise).","section":"Equation (3)"},{"comment":"The line \"¯ri_v ← deg(v)P_u∈Vni deg(u) ¯fni\" is missing the division operator; it should read deg(v) / Σ_{u∈Vni} deg(u).","section":"Algorithm 1, line 7"},{"comment":"The number of Louvain levels s and the sizes of each compressed graph are not reported; without these, it is hard to assess the AM load and the claim that all Gi are below the 100k-variable limit.","section":"Section 4 (Data Generation)"},{"comment":"The caption of Figure 2 contains a typo (\"for from a sample\"), and the figure's y-axis label is missing.","section":"Section 5 (Node Decision Assignment Uncertainty)"},{"comment":"The text \"We omitrGNN\" is missing a space, and Table 2's \"AM / GNN\" notation is not defined in the caption; it should be stated explicitly which side of the slash corresponds to which solver.","section":"Section 5 (Graph Solver Variants)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: This is a workshop-style empirical contribution with a promising scale of experiments, but the missing isolation control is a standard and essential requirement for the causal claim being made. The issues are fixable within a revision: adding random-label and heuristic-label controls, reporting variance from the 50 samples, and correcting the convergence-time claim. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper offers a genuinely new combination: Louvain compression, AM-solved subproblems as supervision for local GNNs, and feature pooling to initialize a global GNN. Each part is known, but I haven't seen this integration before. The empirical work is solid in scale: 50k-150k node random d-regular graphs, three problems, and the MIS gains at 100k and 150k nodes (25-35% relative loss improvement over the mrGNN baseline) are large enough to take seriously.\n\nThe most serious problem is attribution. The only comparison isolating AM supervision is mrGNN vs mrGNN+AM. But mrGNN's local GNNs are purely unsupervised; mrGNN+AM adds a supervised MSE term. So the improvement could come from any external labels, not necessarily from the AM's solution quality. A negative control with random or greedy labels on the same compressed graphs is missing. That control is essential before accepting the 'AM beyond its limits' claim. The paper explicitly assumes alignment between compressed and original QUBOs but never validates it; this control would do exactly that.\n\nSecond, the abstract claims reduced convergence time, but Table 4 shows for MIS at 100k and 150k the AM version is 20-90% slower. The text even notes mrGNN is 'consistently faster,' which conflicts with the 50k MIS rows. That needs a correction.\n\nThird, no variance is reported. They sample 50 times, but the reported loss is the minimum; without standard deviations, the 25-35% figures could be within noise. Fourth, the Louvain alignment assumption is stated but untested; a quick validation comparing AM solutions on compressed graphs to random would help.\n\nDespite these, the paper is well-written and honest about missing ground truth. The engineering contribution is real and the results are suggestive. I'd send it to a serious referee, but with the expectation that the attribution experiment is added. For a workshop paper it's fine; for a journal it needs the control and error bars.","headline":"A plausible AM-GNN transfer pipeline, but the AM's specific contribution is untested because the only control misses the value of any external supervision.","tokens_in":10724,"tokens_out":3286,"would_cite":true,"duration_ms":30822,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that an annealing machine can solve combinatorial optimization problems larger than its native variable limit by supervising local graph neural networks on Louvain-compressed subproblems and using the learned features to…","keywords":["combinatorial optimization","annealing machine","graph neural network","QUBO","Louvain decomposition","multiresolution guidance","maximum independent set","knowledge transfer"],"falsifier":"Take graphs small enough that exact or high-quality solutions are available (e.g., random d-regular graphs up to a few thousand nodes), run the full pipeline, and check two things: whether the AM's solution on the compressed graph, lifted back through the Louvain mapping, correlates with the true optimum on the original graph, and whether mrGNN+AM's lower Hamiltonian loss corresponds to actually better objective values (e.g., larger independent sets or more balanced cuts). If the lifted AM solutions are uncorrelated with optimality while the pipeline still shows lower Hamiltonian loss, the central claim of knowledge transfer fails and the apparent gain is an artifact of measuring quality with the same Hamiltonian the solver minimizes.","tokens_in":9659,"feed_emoji":"🧩","tokens_out":9540,"duration_ms":84266,"temperature":0.7,"pith_summary":"Annealing machines are precise solvers for QUBO problems, but each machine has a hard cap on the number of variables it can handle. This paper proposes to break that cap by compressing the problem graph with Louvain community detection, letting the annealing machine solve each compressed graph, and using those solutions as labels for small local graph neural networks. The node representations those networks learn are mapped back to the original graph and averaged to initialize a larger GNN, which then solves the original problem unsupervised. The authors report that on maximum independent set instances with 100k and 150k nodes, this AM-guided pipeline lowers the final Hamiltonian loss by roughly 25–35% compared with the same GNN architecture without AM guidance, with fewer constraint violations. If the transfer works as claimed, it offers a practical way to apply a fixed annealing machine to problems far beyond its hardware limit.","feed_headline":"Annealing machine solves graphs beyond its hardware limit","feed_subtitle":"Compressed QUBOs let a 100k-variable machine guide a GNN on 150k-node instances, cutting loss by up to 35%.","key_machinery":"The load-bearing mechanism is the AM-guided multiresolution GNN pipeline. It consists of three coupled pieces: (i) Louvain decomposition produces a sequence of compressed graphs $\\{G_i\\}_{i=1}^s$ with a mapping from each artificial node back to a community of original nodes; (ii) a Guiding block trains a local GNN on each $G_i$ with a loss that combines the QUBO Hamiltonian $H_{QUBO}$ and the squared-error alignment $\\ell(x^{GNN}_{i,t}, x^{AM}_i) = \\|x^{GNN}_{i,t} - x^{AM}_i\\|^2$ against the annealing machine's binary solution, so the AM acts as a teaching signal; (iii) a Mapping module projects each learned node vector $\\bar{f}_{n_i}$ onto the original nodes in that community, weighted by each node's degree divided by the community's total degree, then averages the $s$ projected feature matrices $\\bar{R}_i$ to form the initial features $R$ for the main GNN solver on the original graph $G_P$, which is trained by minimizing $H_{QUBO}$ alone. The key design choice is that the AM never sees the original large graph; its knowledge enters only through the features that initialize the global solver.","core_discovery":"The central discovery is that the annealing machine never needs to see the original large graph: its solutions on Louvain-compressed graphs act as supervision for local GNNs, and the feature vectors learned by those local GNNs, lifted to original nodes through an inverse-degree-weighted Louvain mapping and averaged over resolutions, form an initialization for an unsupervised global GNN solver that is better than random initialization. On maximum independent set instances with 100k and 150k nodes, this AM-guided solver achieves 25–35% lower final Hamiltonian loss than the multiresolution GNN without AM guidance, with fewer violations; on MaxCut and Graph Partition the improvements are smaller and mostly appear at 150k nodes. The paper also reports that the AM-guided solver reduces late-training assignment shifts, and that classical heuristics time out on the larger graphs while the proposed pipeline finishes in about twelve minutes on average. The authors read these results as evidence that GNN flexibility can carry the annealing machine's accuracy to problem sizes beyond its native variable limit.","pith_inferences":["The same distillation could apply to any expensive solver—quantum annealers, coherent Ising machines, or specialized CMOS annealers—so solver improvements become portable across hardware generations.","Because local GNNs are roughly 9× faster than the AM, one could run the AM once on a training set of compressed graphs, then deploy a purely GNN pipeline at inference time, trading a small quality loss for large speedups.","The inverse-degree weighting in the mapping module is a free design choice; ablating it (e.g., uniform averaging within communities) would reveal whether the reported gains come from the weighting or simply from community-aware feature pooling.","Testing on graphs with known optima would separate the paper's Hamiltonian-loss measure from true objective quality, since a lower Hamiltonian could in principle come from penalty-term exploitation rather than better combinatorial solutions."],"forward_implications":["An annealing machine with a fixed variable limit can be applied to arbitrarily larger instances by letting a GNN carry its solution knowledge across resolutions, so hardware scaling is no longer the only route to larger problems.","The final solver runs on the original graph, so the output size is unbounded by the AM's limit; the AM only needs to handle the largest compressed subproblem.","On maximum independent set instances of 100k–150k nodes, the AM-guided solver lowers the final Hamiltonian loss by roughly 25–35% relative to the unsupervised multiresolution GNN, with fewer constraint violations.","AM guidance stabilizes training: the main solver exhibits fewer late-stage assignment shifts, which makes early stopping more reliable.","The pipeline remains practical in wall-clock terms: it finishes 150k-node MIS instances in about 12 minutes on average, whereas the tested classical heuristics time out after two hours."],"supporting_citations":[{"why":"provides the Louvain community-detection method used to generate the hierarchy of compressed graphs and the reverse mapping to original nodes.","marker":"[2]"},{"why":"describes momentum annealing, the specific annealing mechanism used to solve the compressed QUBO problems and produce supervision labels.","marker":"[23]"},{"why":"supplies the base physics-inspired GNN solver for QUBO problems that this paper extends with AM guidance.","marker":"[25]"},{"why":"gives the QUBO formulation framework that defines the Hamiltonian cost functions and target problems.","marker":"[10]"},{"why":"provides the classical MIS heuristic baseline whose failure to scale motivates the large-graph regime.","marker":"[3]"},{"why":"supplies the implementation of Louvain decomposition used to generate compressed graphs in the experiments.","marker":"[13]"},{"why":"provides a neural baseline whose MIS results are compared against the proposed model.","marker":"[27]"},{"why":"supports the choice of Louvain over other hierarchical community-detection methods by noting its homogeneity.","marker":"[9]"}],"fun_headline_variants":["Annealing machines guide GNNs past hardware limits","AM-infused GNN solves beyond native size limits","GNN learns from annealing machine on compressed graphs","Compressed QUBOs let AM steer GNN beyond its limit","AM-guided GNN cuts loss 35% on huge graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the QUBO matrices built from Louvain-compressed graphs are aligned enough with the original problem's Hamiltonian that the annealing machine's solutions on those small graphs teach the local GNNs something true about the large graph; if that alignment fails, the transferred features are no better than random and the reported quality gains—measured by Hamiltonian loss without ground truth—would be an artifact.","fun_headline_variants_meta":{"raw":{"variants":["Annealing machines guide GNNs past hardware limits","AM-infused GNN solves beyond native size limits","GNN learns from annealing machine on compressed graphs","Compressed QUBOs let AM steer GNN beyond its limit","AM-guided GNN cuts loss 35% on huge graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1302,"prompt_tokens":928,"completion_tokens":374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":294}},"tokens_in":544,"tokens_out":374,"duration_ms":3895,"temperature":1.0,"reasoning_tokens":294,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:06:30.882429+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take graphs small enough that exact or high-quality solutions are available (e.g., random d-regular graphs up to a few thousand nodes), run the full pipeline, and check two things: whether the AM's solution on the compressed graph, lifted back through the Louvain mapping, correlates with the true optimum on the original graph, and whether mrGNN+AM's lower Hamiltonian loss corresponds to actually better objective values (e.g., larger independent sets or more balanced cuts). If the lifted AM solutions are uncorrelated with optimality while the pipeline still shows lower Hamiltonian loss, the central claim of knowledge transfer fails and the apparent gain is an artifact of measuring quality with the same Hamiltonian the solver minimizes.","supporting_citations":[{"cited_title":"Binary optimization by momentum annealing","cited_arxiv_id":null,"evidence_quote":"describes momentum annealing, the specific annealing mechanism used to solve the compressed QUBO problems and produce supervision labels."},{"cited_title":"Quantum bridge analytics i: a tutorial on formulating and using qubo models.Annals of Operations Research, 314(1):141–183, 2022","cited_arxiv_id":null,"evidence_quote":"gives the QUBO formulation framework that defines the Hamiltonian cost functions and target problems."},{"cited_title":"Graph neural networks for maximum constraint satisfaction","cited_arxiv_id":null,"evidence_quote":"provides a neural baseline whose MIS results are compared against the proposed model."},{"cited_title":"Graph communities in neo4j: Four algorithms at work","cited_arxiv_id":null,"evidence_quote":"supports the choice of Louvain over other hierarchical community-detection methods by noting its homogeneity."}],"review_version":1}