{"id":"ba16bf58-f2e5-4764-97ed-f6a03f626f0f","arxiv_id":"2605.28488","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Interprets MLVI in SBMs as srGW OT projection, proves asymptotic consistency of unregularized srGW, and demonstrates regularized version for simultaneous parameter recovery and model selection.","lead":"The paper recasts maximum likelihood variational inference for stochastic block models as a semi-relaxed Gromov-Wasserstein optimal transport problem and shows that unregularized versions recover parameters consistently in the limit while a regularized version performs joint inference and cluster number selection. A smart generalist might read it to understand how optimal transport ideas can simplify model selection in network clustering without separate grid searches.","discovery_kind":"unclear","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption correctly flags the finite-sample gap that the abstract itself flags. Because the abstract does not overclaim theoretical guarantees for the regularized estimator, the concern does not rise to a load-bearing flaw in the stated central claim. Full-text inspection would be needed only to check whether the empirical section overstates robustness; on the given material the argument is internally consistent.","tokens_in":1671,"tokens_out":269,"duration_ms":9369,"concrete_test":"Verify whether the full manuscript contains a formal consistency or selection-consistency theorem for the regularized srGW objective (beyond the unregularized case), or whether the model-selection claim rests solely on the reported experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract states a consistency proof for unregularized srGW in the asymptotic regime and an empirical demonstration that a regularized version performs joint parameter recovery and model selection. The reader's identified gap (asymptotic consistency not implying finite-sample reliability) is explicitly acknowledged in the abstract itself, which then pivots to the regularized formulation and its empirical behavior. No internal contradiction or unsupported leap is visible from the provided claims; the paper positions the regularization as an empirical fix rather than claiming a new theorem for it.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript interprets maximum likelihood variational inference for stochastic block models as an entropically regularized semi-relaxed Gromov-Wasserstein projection. It proves that unregularized srGW estimators are consistent for recovering the connectivity matrix and latent assignments in the asymptotic regime, while acknowledging that this consistency does not yield reliable finite-sample model selection. It then presents an empirical demonstration that an additional regularized formulation performs joint parameter recovery and selection of the number of clusters within a single optimization problem.","tokens_in":1779,"tokens_out":449,"duration_ms":20290,"significance":"If the consistency result is rigorously established, the work supplies a useful OT-based reinterpretation of MLVI for SBMs and a practical route to model selection that avoids separate grid search. The explicit recognition that asymptotic consistency alone is insufficient for finite-sample selection is a positive feature; the empirical evidence for the regularized joint estimator would need to be robust across multiple regimes to support the efficiency claim.","major_comments":[{"comment":"Abstract (final paragraph): the central claim that the regularized formulation 'yields estimators that simultaneously recover model parameters and select the number of clusters in a single optimization problem' rests entirely on empirical demonstration; because the manuscript does not supply a corresponding consistency or selection-consistency theorem for the regularized case, the load-bearing step from asymptotic unregularized recovery to finite-sample joint selection requires stronger empirical controls (e.g., explicit comparison against standard criteria such as BIC or ICL on the same simulated and real networks).","section":"Abstract"}],"minor_comments":[{"comment":"The notation for the semi-relaxed Gromov-Wasserstein objective and its relation to the SBM likelihood should be introduced with an explicit equation linking the two formulations (currently only described in prose).","section":null},{"comment":"Figure captions and experimental details should state the precise form of the sparsity-promoting regularization (e.g., the value of any additional penalty parameter) and the range of network sizes used to illustrate finite-sample behavior.","section":null}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the careful reading of the manuscript and the constructive feedback. We address the single major comment below.","responses":[{"response":"We agree that the central claim for the regularized formulation is supported by empirical evidence rather than a consistency theorem, as already noted in the manuscript. To strengthen the empirical validation, we will add explicit comparisons of the regularized srGW estimator against BIC and ICL on the same simulated and real networks in the revised version. This will provide more robust controls for the joint recovery and selection performance.","revision_made":"yes","referee_comment":"[Abstract] Abstract (final paragraph): the central claim that the regularized formulation 'yields estimators that simultaneously recover model parameters and select the number of clusters in a single optimization problem' rests entirely on empirical demonstration; because the manuscript does not supply a corresponding consistency or selection-consistency theorem for the regularized case, the load-bearing step from asymptotic unregularized recovery to finite-sample joint selection requires stronger empirical controls (e.g., explicit comparison against standard criteria such as BIC or ICL on the same simulated and real networks)."}],"tokens_in":1283,"tokens_out":253,"duration_ms":8569,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that maximum likelihood variational inference for stochastic block models is equivalent to an entropically regularized semi-relaxed Gromov-Wasserstein projection. Removing the entropy term yields estimators that recover both the connectivity parameters and the cluster assignments consistently in the large-sample limit. Adding a different regularizer then lets the same optimization pick the number of clusters without a separate grid search.\n\nThe equivalence itself is a clean observation that connects two literatures directly. The consistency proof for the unregularized case is the clearest new theoretical statement. The empirical claim that regularization solves the finite-sample selection problem is the practical payoff, and the abstract is explicit that the asymptotic result alone does not deliver reliable selection.\n\nThe soft spot is exactly the one the abstract flags: the gap between asymptotic recovery and finite-sample behavior. The selection result therefore rests on the experiments and on whatever sparsity-promoting term they introduce. Without seeing the full derivations or the experimental controls, it is difficult to judge how robust that fix is across different graph sizes and densities.\n\nThis is niche work aimed at people already working on SBM inference or on optimal transport for structured data. It is worth sending to a referee because the mapping and the consistency result are substantive enough to check, and the single-optimization selection idea addresses a concrete computational issue even if the empirical support needs closer examination.","headline":"The paper reinterprets SBM variational inference as semi-relaxed Gromov-Wasserstein, proves asymptotic consistency for the unregularized version, and shows empirically that regularization enables joint parameter recovery plus model selection.","tokens_in":2289,"tokens_out":358,"would_cite":false,"duration_ms":14645,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Unregularized semi-relaxed Gromov-Wasserstein estimators recover stochastic block model parameters and cluster assignments in the asymptotic regime.","keywords":["stochastic block model","optimal transport","gromov-wasserstein distance","model selection","variational inference","network clustering","asymptotic consistency"],"falsifier":"A finite-sample simulation in which the regularized srGW estimator selects a number of clusters different from the true value would challenge the claim of simultaneous reliable selection.","tokens_in":2580,"feed_emoji":"","tokens_out":557,"duration_ms":19515,"temperature":0.7,"pith_summary":"The paper reinterprets maximum likelihood variational inference for stochastic block models as a semi-relaxed Gromov-Wasserstein projection under entropic regularization. It proves that removing the regularization yields consistent recovery of both the connectivity matrix and the latent cluster assignments when the number of nodes tends to infinity. Because this consistency does not guarantee good model selection in finite samples, the authors add a regularization term that promotes sparsity in the cluster proportions, allowing the same optimization to recover parameters and select the number of clusters at once.","feed_headline":"Unregularized srGW recovers SBM clusters asymptotically","feed_subtitle":"A regularized version selects the number of clusters automatically in one optimization, avoiding grid searches.","key_machinery":"semi-relaxed Gromov-Wasserstein projection, which reformulates the variational inference objective and permits direct analysis of consistency for recovering SBM parameters.","core_discovery":"Maximum likelihood variational inference in stochastic block models is equivalent to a semi-relaxed Gromov-Wasserstein projection with entropic regularization. Unregularized srGW estimators consistently recover the SBM connectivity matrix and latent cluster assignments asymptotically. A regularized formulation performs both parameter estimation and selection of the number of clusters simultaneously in a single optimization problem.","pith_inferences":["The connection between MLVI and srGW may allow importing other optimal transport techniques to improve SBM inference.","Similar reformulations could apply to other network models beyond SBMs.","Finite-sample performance might improve further with adaptive regularization parameters."],"forward_implications":["Consistent asymptotic recovery of the connectivity matrix and assignments.","Joint inference and model selection avoids separate grid search procedures.","Regularization addresses the sparsity issue that prevents reliable finite-sample selection.","Provides an efficient alternative to traditional heuristic model selection in SBMs."],"fun_headline_variants":["srGW interprets MLVI as entropic projection in SBMs","Asymptotic SBM recovery via unregularized srGW","Regularized srGW automates SBM cluster selection","Joint SBM parameter and model selection with srGW"],"cache_read_input_tokens":64,"weakest_assumption_plain":"That sparsity-promoting regularization can be introduced to achieve reliable model selection in finite samples while preserving the recovery properties.","fun_headline_variants_meta":{"raw":{"variants":["srGW interprets MLVI as entropic projection in SBMs","Asymptotic SBM recovery via unregularized srGW","Regularized srGW automates SBM cluster selection","Joint SBM parameter and model selection with srGW"]},"model":"grok-4.3","cost_usd":0.005151,"raw_usage":{"total_tokens":2465,"prompt_tokens":594,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":51512000,"prompt_tokens_details":{"text_tokens":594,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1806,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":594,"tokens_out":65,"duration_ms":13440,"temperature":1.0,"reasoning_tokens":1806,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T09:43:56.033763+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A finite-sample simulation in which the regularized srGW estimator selects a number of clusters different from the true value would challenge the claim of simultaneous reliable selection.","supporting_citations":[],"review_version":1}