{"id":"b01abf02-ff16-4c64-b14a-8a01d12f221e","arxiv_id":"2505.09391","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural network that produces inexact ADMM subproblem solutions for convex quadratic programs is shown to converge under learned residual conditions, and runs faster than Gurobi, SCS, and OSQP on synthetic benchmarks, though the conditions are violated on some instances.","lead":"This paper trains an LSTM network to generate cheap approximate answers to the hard subproblem inside ADMM, an iterative solver for quadratic programs, and proves the approximation still converges if it meets certain error conditions. It reports speedups of up to 7x over Gurobi, 28x over SCS, and 22x over OSQP on its test problems, at the price of small optimality gaps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's hypotheses fail on the paper's own benchmark: Assumption 1 is impossible for Random QP (A is 2000×1000), and Section 5.3.2 admits conditions (14),(16) are significantly violated there, so the convergence claim does not cover the reported method.","rationale":"The paper's central contribution is a convergence guarantee for a learned inexact ADMM. That guarantee is only meaningful if the theorem's hypotheses are satisfied by the actual algorithm. Reading the full text, two hypotheses fail on the evidence presented. First, Assumption 1 is dimensionally impossible for Random QP because m>n; the authors never note this. Second, the conditions (13)-(16) are not part of the training loss (23) or the forward pass of Algorithm 2, and Section 5.3.2 explicitly reports violations for Random QP and only approximate adherence elsewhere. Since these conditions are the bridge between the neural solver and the inexact ADMM theory, the theorem does not cover the method as evaluated on a headline benchmark. The abstract's 'near-perfect compliance... ensuring algorithm convergence' is therefore stronger than what the body supports; the reader's CONDITIONAL verdict correctly captures this. I do not see an internal algebraic error in Proposition 1 or Theorem 1 under their stated assumptions; the issue is applicability rather than derivation. The concrete test would settle whether enforcement or rescoping is needed. The paper is honest about the gap in Section 5.3.2 and the code is public, so a conditional acceptance with a request to enforce or rescope the claim is the right call.","tokens_in":22701,"tokens_out":10445,"duration_ms":96980,"concrete_test":"Run the released code on the Random QP test set and compute rank(A); if rank(A)<m=2000, Assumption 1 fails. Then instrument Algorithm 2 to evaluate, at every iteration, the inequalities (13)-(16) using the paper's constants (c_x=c_z=1, delta=0.9, tau=0.1) and beta_x,beta_z from (39), plus (17) with the learned alpha^k; report per-instance violation counts and magnitudes. If the violation count is nonzero on Random QP (or on any reported benchmark), the claim that Theorem 1 applies is refuted; one would need to enforce the conditions (e.g., accept the LSTM step only when (13)-(16) hold, otherwise fall back to a proximal/exact subproblem step) and re-measure speedups, or rephrase the guarantee as conditional.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The convergence guarantee in Theorem 1 is conditional on Assumption 1 (Range(A)=R^m) and on Conditions (13)-(16) holding at every iteration. The implemented method does not establish either hypothesis. For the Random QP benchmark (Appendix D, Table 3), n=1000 and m=2000, so A is a 2000×1000 matrix with rank at most 1000; Range(A) cannot equal R^2000, and Assumption 1 is identically false on this benchmark. Second, the self-supervised loss (23) minimizes primal/dual residuals only; nothing in Algorithm 2 enforces (13)-(16) or the line-search condition (17). Section 5.3.2 states that \"significant violations occur for Random QP instances\" and concludes there is \"a subtle yet critical gap between inexact computation and provable convergence.\" Consequently, the abstract's statement that compliance with theoretical assumptions leads to \"ensuring algorithm convergence\" is not supported for the reported experiments. The central claim that I-ADMM-LSTM maintains provable primal-dual convergence would require either enforcing the conditions or restricting the claim to instances satisfying them.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes I-ADMM-LSTM, a self-supervised LSTM-based inexact ADMM for convex quadratic programs. The method replaces the exact ADMM subproblem solves with LSTM-generated approximate solutions, trains the network by minimizing primal and dual residuals over a window of iterations, and optionally adds a feasibility-restoration stage. The theoretical contribution is a convergence theorem (Theorem 1) showing that, if the learned iterates satisfy the sufficient-decrease and subgradient-boundedness conditions (13)-(16) and the line-search condition (17), together with an energy-boundedness assumption and certain matrix inequalities, then the ADMM iterates converge sublinearly to an optimal solution. The experimental section reports large speedups over Gurobi, SCS, and OSQP on several QP benchmarks, with an empirical validation section assessing whether the theoretical conditions hold.","tokens_in":22932,"tokens_out":24160,"duration_ms":199112,"significance":"If the conditional convergence theorem is sound, it would provide a useful template for learning-based inexact ADMM with provable guarantees under verifiable conditions. The LSTM architecture is parameter-efficient, the self-supervised loss avoids labeled data, and the code is publicly available. The empirical results on several benchmarks are encouraging. However, the paper's central advertised claim, that the implemented method maintains provable primal-dual convergence, is not established: the training loss and Algorithm 2 do not enforce the theorem's hypotheses, the paper's own Section 5.3.2 reports significant violations of those hypotheses on the Random QP benchmark, and some of the theory's assumptions are impossible on that same benchmark. The conditional theorem itself appears to be a genuine sufficient-condition result rather than a tautology, but it does not cover the method as implemented.","major_comments":[{"comment":"The convergence guarantee in Theorem 1 is conditional on Conditions (13)-(16) and the line-search condition (17) holding at every iteration, but the self-supervised loss (23) and Algorithm 2 do not enforce these conditions, and Section 5.3.2 states that 'significant violations occur for Random QP instances' and that there is 'a subtle yet critical gap between inexact computation and provable convergence.' Therefore the abstract's claim that 'near-perfect compliance with the theoretical assumptions, consequently ensuring algorithm convergence' is not supported by the paper's own experiments.","section":"§5.3.2, §3.1, Algorithm 2"},{"comment":"Assumption 1 requires Range(A)=R^m, i.e., A must have full row rank. For the Random QP benchmark, Table 3 reports n=1000 and m=2000, so A is a 2000×1000 matrix and its range cannot equal R^2000. Thus Theorem 1 cannot cover the Random QP results reported in Table 2, and no verification of full row rank is provided for any of the other datasets.","section":"§4, Assumption 1; Table 3"},{"comment":"Proposition 1 and Theorem 1 are proved for a constant scalar penalty parameter ρ appearing in the energy definitions (28)-(30), in the conditions (13)-(16), and in the dual update y^{k+1}=y^k+ρ(A\\tilde{x}^{k+1}-z^{k+1}) used in the proof of Lemma 1. Algorithm 2, however, uses an iteration-dependent diagonal penalty matrix ρ^{k+1} (Eq. (25)) and updates y using \\tilde{z}^{k+1} reconstructed from the approximate solution of the linear system (19), which need not equal A\\tilde{x}^{k+1} unless that linear system is solved exactly. No condition among (13)-(17) controls the mismatch A\\tilde{x}^{k+1}-\\tilde{z}^{k+1}, and no argument is given for the time-varying ρ; as stated, Theorem 1 therefore does not apply to Algorithm 2.","section":"§4 vs §3.2, Eqs. (25), (28)-(30)"}],"minor_comments":[{"comment":"The proof contains two visible typos: 'by (17) and , yields' in C.2 is missing a condition or reference, and 'The monotonicity and boundedness of {\\tilde{E}^k} is a ensure' in C.3 is grammatically corrupted. Please correct these.","section":"Appendix C.2, C.3"},{"comment":"The |IL| and |IU| columns are reported as zero for every instance, including problems that clearly have finite lower and upper bounds (e.g., Random QP with l≤Ax≤u and SVM). This is inconsistent with the dataset descriptions in Appendix D and should be clarified.","section":"Table 3"},{"comment":"The abstract's phrase 'near-perfect compliance with the theoretical assumptions' is contradicted by Section 5.3.2's report of 'significant violations' on Random QP. The abstract should be reworded to state the conditional nature of the guarantee and the observed violations.","section":"Abstract and §5.3.2"},{"comment":"Algorithm 1 uses \\tilde{z}^{k+1} in the dual update, but \\tilde{z}^{k+1} is never defined in Section 2.3. It should be defined as A\\tilde{x}^{k+1} (or explicitly related to Section 3.1's reconstruction) so that the algorithm matches the theoretical dual update used in the proof.","section":"Algorithm 1, line 4"},{"comment":"Section 5.3.2 describes condition-compliance behavior qualitatively (e.g., 'rare violations in early iterations' and 'significant violations' for Random QP) but does not report quantitative counts or aggregate statistics over the 50 test instances. Adding fractions of violating iterations would strengthen the empirical validation.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about the theory-practice gap in Section 5.3.2, which is to its credit, but the abstract and several high-level statements overclaim. The main issue is not the conditional theorem itself but the fact that the implemented algorithm does not satisfy the theorem's hypotheses, including on a benchmark where Assumption 1 is impossible. A revision that either enforces the conditions at runtime or carefully restricts the convergence claims to cases where they are verified would be needed. The adaptive-ρ and \\tilde{z}-reconstruction mismatches are additional structural gaps that must be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea is worth knowing, but the paper sells a convergence guarantee it doesn't actually deliver for its own flagship benchmark. The theory is a conditional inexact-ADMM result; the LSTM implementation is not shown to satisfy the conditions, and Section 5.3.2 admits significant violations on Random QP. That benchmark also violates Assumption 1 by construction: a 2000×1000 matrix A cannot have Range(A)=R^2000. The abstract's \"near-perfect compliance\" and \"ensuring algorithm convergence\" overstate what is demonstrated.\n\nWhat's genuinely new: a parameter-efficient coordinate-wise LSTM that produces inexact solutions to the OSQP subproblem, wrapped in an inexact-ADMM framework with a two-stage feasibility restoration. That is a legitimate L2O contribution, and the convergence analysis is a fair adaptation of Bai et al./Xie to learned subproblem solvers. The experiments cover several QP families, and the method is often faster than Gurobi/SCS/OSQP while beating other L2O baselines on feasibility. Section 5.3.2 is refreshingly honest about the gap.\n\nSoft spots, in order of severity. First, the load-bearing gap: Theorem 1 requires Assumption 1 and conditions (13)–(16) at every iteration. The loss (23) does not enforce them, and the paper reports they fail for Random QP. Since Random QP has m=2000, n=1000, Assumption 1 is impossible. So the main advertised result—provable convergence of the implemented method—does not cover at least one of the paper's own test sets. Second, the speedups are GPU inference versus CPU Gurobi/OSQP, which is an unfair comparison; Appendix E partly acknowledges this, but the abstract and highlights do not. Third, there are no error bars or variance across the 50 test instances. Fourth, the theory constants beta_x and beta_z are chosen after the fact to make (31)–(33) hold, which is fine mathematically but decouples the theory from what the LSTM actually learns.\n\nThis is fixable. The referee should ask for an honest abstract, either a verifier that enforces (13)–(16) or a restricted claim, same-hardware timing, and error bars. The central idea is solid and the proof structure is plausible, so it deserves serious refereeing, not desk rejection.","headline":"Interesting L2O-ADMM idea with a real theory-implementation gap: the convergence theorem's hypotheses fail on the paper's own Random QP benchmark, and the GPU-vs-CPU speedups need a caveat.","tokens_in":23539,"tokens_out":2335,"would_cite":true,"duration_ms":22450,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C20","90C25","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural-approximated ADMM provably converges to optimal solutions of convex quadratic programs when learned iterates satisfy specified inexactness conditions.","keywords":["quadratic programs","learning to optimize","inexact alternating direction method of multipliers","long short-term memory networks","self-supervised learning","convergence guarantees","ADMM"],"falsifier":"Choose a Random QP instance from the paper's benchmark, run Algorithm 2, and evaluate conditions (14) and (16) after the first 50 iterations; if either inequality is violated at a late iteration where the algorithm still reports convergence, then Theorem 1's hypotheses fail for that run and any observed convergence is not explained by the theorem.","tokens_in":22426,"feed_emoji":"🧮","tokens_out":6615,"duration_ms":66502,"temperature":0.7,"pith_summary":"This paper tries to establish that a learning-augmented alternating direction method of multipliers (ADMM) can replace the exact subproblem solves with approximations produced by a recurrent neural network and still converge to an optimal solution of a convex quadratic program. The authors cast the learned solver as an inexact ADMM, prove an energy-descent convergence theorem under five explicit conditions on the neural iterates, and add a short exact-ADMM refinement stage to restore feasibility. If the central claim is right, learning-based acceleration of ADMM does not have to give up convergence guarantees: approximate subproblems, no per-iteration factorization, and a provable path to optimality can coexist.","feed_headline":"Neural ADMM keeps convergence despite approximate subproblems","feed_subtitle":"A new proof guarantees optimality of a learned ADMM that never solves subproblems exactly, with benchmark speedups up to 28x.","key_machinery":"The load-bearing machinery is the inexact ADMM framework for the two-block reformulation min (1/2)x^T Q x + p^T x + I_{l≤z≤u}(z) subject to Ax=z, where the x- and z-subproblems are solved only approximately. The five conditions are sufficient decrease of the augmented Lagrangian, gradient and subgradient bounds proportional to iterate increments, and a line-search decrease on the final update; condition (17) controls the relaxation step. The convergence proof works through an energy function $E^k = L_\\rho(x^k,z^k,y^k) + \\Gamma^k$, where $\\Gamma^k$ is a weighted sum of squared iterate increments, and the matrix inequalities (31)-(33) make the energy decrease monotonically. From that descent, the paper derives $O(1/K)$ decay of the squared residuals and $O(1/\\sqrt{K})$ decay of the subdifferential distance, giving sublinear convergence to an optimal solution.","core_discovery":"The central claim is Theorem 1: under Assumption 1 (Range(A)=R^m) and conditions (13)-(16), with the line-search condition (17), if the energy sequence is bounded below and matrix inequalities (31)-(33) hold, then Algorithm 2 generates iterates whose Lagrangian, augmented Lagrangian, and energy sequences converge to a common limit F*, whose distance to the Lagrangian subdifferential goes to zero, and whose limit points are optimal solutions of the original QP. This is stated as the learned method maintaining primal-dual convergence while satisfying residual thresholds. The theorem is the bridge that turns the neural solver into a provably convergent optimization algorithm, provided the neural outputs obey the inexact-ADMM conditions at every iteration.","pith_inferences":["Editorial inference: if the sufficient-decrease and gradient conditions were enforced as constraints or penalty terms during training, the deployed network would satisfy the theorem's hypotheses by construction, closing the theory-practice gap the paper reports for Random QP instances.","Editorial inference: the proof is largely agnostic to the specific neural architecture, so any learned oracle that meets conditions (13)-(17) inherits the convergence claim; the result can be read as a certification template for other learned ADMM subproblem solvers.","Editorial inference: Assumption 1 excludes rank-deficient constraint matrices, so a natural stress test is to apply the method to QPs with redundant constraints, where the dual sequence may drift; constraint preprocessing or dual normalization would be a testable fix.","Editorial inference: the theorem's convergence guarantee concerns the idealized iterates satisfying all five conditions, not the trained model as such; users should monitor the violation of (14) and (16) online and switch to the refinement stage when violations persist."],"forward_implications":["If the trained network satisfies the hypotheses at every iteration, the learned solver is guaranteed to drive primal and dual residuals to zero, so it can be used as a convergent approximate QP solver without per-iteration factorization.","Because only approximate subproblem solves are required, each iteration avoids the matrix factorization that dominates classical solvers; the reported wall-clock speedups come from the resulting cheaper iterations.","The sublinear rates ($O(1/K)$ on residuals, $O(1/\\sqrt{K})$ on subdifferential distance) make the guarantee practical for moderate-accuracy regimes rather than for high-precision asymptotics.","The two-stage refinement, which runs a short exact-ADMM phase with a single factorization, restores exact feasibility and brings objective values in line with established solvers, so the neural phase only needs to land in a neighborhood of the solution.","The stopping criterion based on the composite residual $R^k$ is justified by the theorem: as $R^k$ goes to zero, the iterates approach a stationary point, which is globally optimal by convexity."],"supporting_citations":[{"why":"Supplies the inexact ADMM framework and the sufficient-decrease and gradient conditions that the learned method must satisfy.","marker":"(Bai et al., 2025)"},{"why":"Establishes relative-error inexact ADMM criteria that justify controlled subproblem approximation within ADMM.","marker":"(Xie, 2018)"},{"why":"Provides the QP reformulation, the condensed linear systems, and the preconditioning approach that the algorithm builds on.","marker":"(Stellato et al., 2020)"},{"why":"Shows learned approximations can respect inexact-method tolerances and is the direct precursor for LSTM-based approximate solves.","marker":"(Gao et al., 2024)"},{"why":"Contributes the DC3 baseline and the Convex QP (RHS) task protocol used in the experimental evaluation.","marker":"(Donti et al., 2021)"}],"fun_headline_variants":["Neural ADMM: approximate subproblems, proven convergence","28x speedup: LSTM-based QP solver with convergence guarantee","Learned inexact ADMM: fast QP solving with rigorous convergence","Provably convergent neural ADMM beats OSQP, SCS, Gurobi"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the neural network outputs satisfy conditions (13)-(16) and the line-search condition (17) at every iteration; the paper's own Section 5.3.2 reports significant violations of these conditions on Random QP instances.","fun_headline_variants_meta":{"raw":{"variants":["Neural ADMM: approximate subproblems, proven convergence","28x speedup: LSTM-based QP solver with convergence guarantee","Learned inexact ADMM: fast QP solving with rigorous convergence","Provably convergent neural ADMM beats OSQP, SCS, Gurobi"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1509,"prompt_tokens":900,"completion_tokens":609,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":531}},"tokens_in":516,"tokens_out":609,"duration_ms":5865,"temperature":1.0,"reasoning_tokens":531,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:33:38.016731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a Random QP instance from the paper's benchmark, run Algorithm 2, and evaluate conditions (14) and (16) after the first 50 iterations; if either inequality is violated at a late iteration where the algorithm still reports convergence, then Theorem 1's hypotheses fail for that run and any observed convergence is not explained by the theorem.","supporting_citations":[{"cited_title":"An inexact admm for separable nonconvex and nonsmooth optimization","cited_arxiv_id":null,"evidence_quote":"Supplies the inexact ADMM framework and the sufficient-decrease and gradient conditions that the learned method must satisfy."},{"cited_title":"On inexact admms with relative error criteria","cited_arxiv_id":null,"evidence_quote":"Establishes relative-error inexact ADMM criteria that justify controlled subproblem approximation within ADMM."},{"cited_title":"Ipm-lstm: A learning-based interior point method for solving nonlinear programs","cited_arxiv_id":null,"evidence_quote":"Shows learned approximations can respect inexact-method tolerances and is the direct precursor for LSTM-based approximate solves."},{"cited_title":"Dc3: A learning method for optimization with hard constraints","cited_arxiv_id":null,"evidence_quote":"Contributes the DC3 baseline and the Convex QP (RHS) task protocol used in the experimental evaluation."}],"review_version":1}