{"id":"c0b24073-aa57-4391-b064-560a036ca62b","arxiv_id":"2608.13079","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CGD is a graph diffusion model that projects intermediate samples toward the feasible set during reverse denoising, then hands the fixed binary decisions to a continuous solver, achieving reported speedups of up to 425x on two MINLP benchmarks.","lead":"This paper adds a feasibility projection step into a graph-based diffusion model that generates the binary decisions of mixed-integer programs, then solves the remaining continuous problem with a standard optimizer. On transmission switching and portfolio problems it reports large speedups over a commercial MINLP solver while keeping objective gaps near 1 to 7 percent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central mechanism and speedup claim rely on a per-step feasibility projection, but the AC implementation in Appendix C is a 597-binary MILP whose cost and use at every reverse step are unmeasured and inconsistent with the reported 0.11s sampling time; the theory only covers convex projections.","rationale":"The reader's weakest_assumption identifies exactly the gap I find most load-bearing: the per-step projection operator is the mechanism that distinguishes CGD from prior diffusion solvers, and its cost and implementation are not isolated. My reading of Appendix C strengthens this concern. The projection described there is a MILP with binary variables, not a convex program, so the theoretical results in Section 5.1 do not cover it. The reported runtime data (Table 3) make it implausible that this MILP is solved 100 times per 500-bus instance, yet the paper does not say how the projection is approximated or when it is triggered. This matters because the central speedup claim (up to 425x) depends on the projection being cheap enough to be negligible within the 0.11s sampling time, and the central methodological claim depends on the projection actually steering every reverse step. The paper is honest about several limitations: it states that projection provides locally improved transitions rather than a feasible trajectory guarantee, that connectivity is necessary but not sufficient for AC feasibility, and that warm-starting the MINLP solver does not reduce certification time. Those disclosures lower the risk of overclaiming, but they do not resolve the projection-implementation gap. The requested code release and separate projection-cost reporting would settle it. I therefore agree with the reader's CONDITIONAL verdict and see no reason to move it.","tokens_in":21910,"tokens_out":4749,"duration_ms":52314,"concrete_test":"Instrument Algorithm 1 (line 4) with per-call timing around Π_Cξ for IEEE 500-bus with T=100, and report the number of MILP invocations, the total projection time, and the average time per call. Then re-run Appendix C's MILP on a typical relaxed prediction: if one solve takes materially longer than 1ms, the reported 0.11s sampling time cannot accommodate 100 MILPs; if the projection is invoked only on violation or replaced by a heuristic, report that explicitly and check whether the same empirical results are obtained when the projection is removed entirely.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core assertion is that projecting every reverse step onto the feasible set yields fast, feasible discrete decisions (Algorithm 1, step 4; Eq. 9). The only concrete projection supplied for AC transmission switching (Appendix C) is a mixed-integer linear program over binary line variables, with big-M flow constraints and a Hamming objective. This is not the convex relaxation Cξ of Eq. (8), and the MILP is neither differentiable nor cheap: at 500-bus scale it has 597 binary variables, and if solved at each of T=100 reverse steps, the reported 0.11s average sampling time would require roughly 1ms per MILP, which is not credible for a general-purpose MILP solver. The paper never states whether this MILP is solved at every t, only on violation, or replaced by a heuristic; nor does it report projection time separately from 'CGD sampling' in Table 3. Section 5.1 allows 'an application-specific correction with the same interface' when exact projection is impractical, but that does not bridge the gap: Propositions 1-2 and Corollary 3 all assume exact projection onto a closed convex set, so none of the paper's guarantees apply to the operator actually used in the headline AC experiments. The speedup claim therefore rests on an unmeasured and theoretically mismatched component: if the MILP is actually run per step, the 425x speedup is implausible; if it is not, the 'constraint-aware diffusion' mechanism is not what was evaluated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Constrained Graph Diffusion (CGD), a graph-based conditional diffusion model that generates the binary decision block of a mixed-integer optimization problem and then solves the remaining continuous problem with a numerical optimizer. Feasibility is enforced by inserting a projection operator into every reverse diffusion step, projecting a relaxed decision onto a continuous relaxation of the combinatorial feasible set before the next denoising step. The method is evaluated on AC optimal power flow with transmission switching (IEEE 9-, 197-, and 500-bus systems) and on mixed-integer portfolio optimization. The paper reports lower Hamming distance, higher exact reconstruction, zero observed downstream infeasibility for most CGD settings, and wall-clock speedups up to 425x over a joint MINLP solver.","tokens_in":22171,"tokens_out":3958,"duration_ms":42404,"significance":"The paper's decomposition strategy is sensible: amortizing the combinatorial search with a generative model and relegating the continuous variables to a standard solver is a natural way to attack structured MINLPs. The geometric core is a genuine strength: Proposition 1, Proposition 2, and Corollary 3 are stated cleanly and their proofs are correct, and the paper is honest about what the theory does not imply, explicitly noting that repeated projection gives locally improved transitions rather than a certificate of sampling from the constrained distribution. The experimental design is also comparatively thorough, with several learning-based baselines, separate reporting of discrete quality, downstream feasibility, objective gap, and runtime, and a useful negative result on warm-starting. The significance of the framework, however, is conditional on the projection operator living up to the role assigned to it in Algorithm 1.","major_comments":[{"comment":"The theoretical guarantees (Proposition 1, Proposition 2, Corollary 3) are established only for exact projection onto a nonempty, closed, convex set C_xi, as defined in Eq. (8). The AC transmission-switching projection in Appendix C, however, is a mixed-integer linear program over binary line-status variables z_e with additional flow and supply variables, minimizing a Hamming objective. This is a nonconvex discrete optimization problem, not the convex relaxation C_xi of Eq. (8). Consequently, none of the paper's projection guarantees applies to the operator actually used in the headline AC experiments. The sentence in Section 5.1 allowing 'an application-specific correction with the same interface' when exact projection is impractical does not bridge this gap, because no analogue of Propositions 1-2 or Corollary 3 is provided for the discrete MILP operator. This is load-bearing: the entire conceptual justification for constraint-aware generation rests on the geometry of exact convex projection.","section":"Section 5.1, Eq. (8)-(9); Section 6.1; Appendix C"},{"comment":"There is a direct inconsistency between Algorithm 1 and the described AC implementation. Algorithm 1 applies the projection at every reverse step, unconditionally, for t = T down to 1. Appendix C, in contrast, states that 'the projection is applied during diffusion sampling whenever the predicted topology violates the load-relevant connectivity condition,' i.e., conditionally on violation. These two descriptions cannot both describe the evaluated method. Moreover, Table 3 reports only 'CGD sampling' time and does not report the number of projection solves, the time spent inside the projection MILP, or whether the MILP was invoked at every step, only on violation, or replaced by a heuristic. At the 500-bus scale the projection has 597 binary variables; a per-step invocation at T=100 steps would imply roughly 1 ms per MILP solve, which is not credible, while a violation-only invocation would mean the experiments did not test the per-step projection mechanism that the paper claims. The paper must state the actual invocation schedule and report projection cost separately.","section":"Algorithm 1, step 4; Appendix C, final paragraph; Table 3"},{"comment":"The headline speedup of 425.2x on the 9-bus system compares CGD's end-to-end time (0.33 s) with Gurobi's full termination time (140.32 s) at a MIP gap tolerance of 1e-4. This is not an equal-footing comparison: CGD returns a feasible solution without an optimality certificate, whereas Gurobi's reported time includes proving the certificate. The paper's own Figure 4 shows that Gurobi reaches the same 0.01% gap level as CGD after 30.02 s, giving a 91.0x ratio rather than 425.2x. For the 197- and 500-bus systems, no time-to-same-gap comparison is reported, so the 22.0x and 40.0x speedups in Table 3 similarly mix certified and uncertified solution times. To support the speedup claims, the paper should report a time-to-target comparison, or clearly state and justify that the appropriate comparison is against full solver termination.","section":"Section 6.1, Table 3 and Figures 4-6"}],"minor_comments":[{"comment":"There is a typo in the text: 'solving the resulting nonconvex mixed-integer AC-OPF with GUROBIGurobi' should be 'with Gurobi.'","section":"Section 6.1"},{"comment":"The sentence 'These predictions are averaged, as in Eq. equation 24' has a duplicated word and an awkward reference; it should read 'as in Eq. (24)'.","section":"Section 6, Baselines"},{"comment":"The description of the portfolio reference solver as 'ECOS BBSOLVER' is unclear; please specify the exact solver and version used for the MIQP reference solutions.","section":"Appendix D"},{"comment":"The load-bus threshold epsilon is defined but its numerical value is never reported. Since the projection's behavior depends on which buses are classified as loads, the value of epsilon should be given for each benchmark.","section":"Appendix C"},{"comment":"There is a typo: 'AC-OPF with tranmission switching instances' should be 'transmission switching.'","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The central idea is attractive and the geometric propositions are solid, but the AC implementation currently does not match the method as described and the speedup claim rests on an unmeasured, potentially very expensive component. I would ask for a revised version that either implements an actual convex (or otherwise cheap and differentiable) projection for AC transmission switching, or re-scopes the theoretical claims and the experimental reporting to match the discrete MILP projection actually used. The paper would also benefit from a time-to-same-gap comparison in Table 3 so that the speedup claims are not overstated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe one thing to know: this is a serious, well-engineered paper that deserves a shot at peer review, but the headline AC speedup rests on a projection whose actual implementation the theory does not cover and whose runtime is never isolated. The core novelty—projecting relaxed discrete predictions back into the feasible set at every reverse-diffusion step, then rounding and solving the continuous completion—is real and produces the best feasibility/objective trade-offs across the board. Propositions 1, 2, and Corollary 3 are correct and clean, and the authors are honest about warm-starting not helping, connectivity being necessary but not sufficient, and no global optimality certificate.\n\nThe soft spot is the gap between the convex-projection theory and the AC experiments. Stage (2) of Algorithm 1 applies Π_Cξ at every t, and the theory assumes an exact projection onto a closed convex set. But the AC projection described in Appendix C is a MILP over 597 binary line variables, minimizing Hamming distance to the predicted topology under flow constraints. That is a discrete projection, not the convex relaxation of Eq. (8). The paper never states whether this MILP is actually solved at every reverse step, only on violation, or approximated; nor does it report projection time separately from 'CGD sampling' in Table 3. The appendix's phrase 'whenever the predicted topology violates' hints at a conditional use, but the main-text algorithm does not say that. If the MILP runs at every one of T=100 steps on 597 binaries, the 0.11s average sampling time is not credible. If it runs only on violation, then the evaluated mechanism is not the one described. Either way, the authors must clarify.\n\nSecond, the speedup is real but the headline is optimistic. Table 3 compares CGD's end-to-end solution (with ~0.01–2% objective gap) against Gurobi's full convergence to 10^-4 MIP gap. Figure 4 shows a fairer comparison at equal gap: 91x on 9-bus, not 425x. The disclosure in Figures 4–6 softens this, but the abstract's 'up to 425x' overstates the practical gain.\n\nThe portfolio experiments are cleaner, with a convex projection and an exact certificate for threshold recovery. Training on Gurobi's reference decisions is a standard choice; I don't see circularity in the geometric claims. Missing code/data is the main reproducibility blocker.\n\nWho's it for: ML-for-optimization, power systems, anyone using learning to generate discrete decisions for structured MINLPs. It will be cited if the code appears. I would bring it to reading group and would accept for peer review with major revision: separate the projection time, reconcile Algorithm 1 with Appendix C, and tone down the speedup claim.","headline":"A well-engineered diffusion-plus-projection pipeline for discrete decisions in MINLPs, with sound convex-projection theory and promising results, but the AC speedup claim is undercut by an unmeasured MILP projection that the stated theory does not cover.","tokens_in":22783,"tokens_out":4093,"would_cite":true,"duration_ms":41671,"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":"Injecting a feasibility projection into every reverse diffusion step yields feasible binary decisions for mixed-integer programs and up to 425x faster solves.","keywords":["diffusion models","mixed-integer nonlinear programming","constraint-aware generation","feasibility projection","optimal transmission switching","portfolio optimization","learning to optimize","graph-based generative models"],"falsifier":"Measure the wall-clock time of the projection call $\\Pi_{C_\\xi}$ alone on the IEEE 500-bus benchmark across all 100 reverse steps of a single sample. If solving the load-relevant connectivity MILP over 597 binary variables at each step takes longer than the reported 0.11 s total sampling time, the claim that the exact projection is applied at every step cannot be sustained. A second check is to verify on every test instance that the thresholded terminal point $T(u_0)$ belongs to $Z_\\xi$ and induces a feasible AC-OPF, since the paper states the projection does not guarantee either condition.","tokens_in":21651,"feed_emoji":"⚡","tokens_out":13172,"duration_ms":113350,"temperature":0.7,"pith_summary":"Mixed-integer nonlinear programs are slow because the discrete variables force an exponential search. The paper proposes Constrained Graph Diffusion (CGD), a graph-based diffusion model that learns the distribution of the binary decisions and projects every intermediate denoising estimate onto a convex relaxation of the combinatorial feasible set before the next step. Once a discrete decision is generated, the remaining continuous variables are recovered by a standard numerical solver, so the combinatorial search is amortized across related instances. On AC optimal power flow with transmission switching and on discrete portfolio optimization, CGD reports zero observed downstream infeasibility across the tested settings, objective gaps of 0.01–1.76% on the power benchmarks, and end-to-end speedups up to 425x relative to the tested joint solver.","feed_headline":"Constraint-aware diffusion speeds up mixed-integer solves 425x","feed_subtitle":"It projects every denoising step onto the feasible set, then solves the continuous part with a standard solver.","key_machinery":"The load-bearing object is the feasibility projection $\\Pi_{C_\\xi}$ placed inside every reverse transition. At reverse step $t$, the denoiser's noise prediction is converted into a clean embedded estimate, the tanh map bounds it to $[-1,1]^m$, and the map $e^{-1}$ converts it to a relaxed decision $\\hat{z}_t = (\\hat{y}_{0,t}+1)/2 \\in [0,1]^m$; the projection $u_t = \\Pi_{C_\\xi}(\\hat{z}_t)$ then solves the convex program minimizing $\\|u - \\hat{z}_t\\|_2^2$ over the relaxation $C_\\xi$, and the corrected point is reinserted via the consistent noise estimate $\\varepsilon_t^C = (y_t - \\sqrt{\\bar{\\alpha}_t} e(u_t))/\\sqrt{1-\\bar{\\alpha}_t}$, which replaces the model's raw prediction in the standard DDPM update. The antipodal embedding $y_0 = 2z^* - 1$ preserves Hamming geometry. Proposition 1 shows the projection residual is the gradient of the squared distance and that projection cannot increase the distance to any feasible reference; Proposition 2 shows the projected reverse kernel is the 2-Wasserstein-closest feasible kernel and that the correction reduces distance to every relaxed-feasible endpoint. The training objective adds a feasibility regularizer that projects the Monte Carlo mean of $K$ perturbed denoiser predictions and backpropagates the residual through the relaxed predictions with a stop-gradient on the projected target.","core_discovery":"The paper's central claim is that incorporating a feasibility projection operator inside every reverse diffusion step—rather than applying it only at the end—steers generated discrete decisions toward the feasible set without destroying the learned data distribution. The projection is applied to the relaxed decision in $[0,1]^m$, mapping it onto the convex relaxation $C_\\xi$ of the combinatorial feasible set, and the corrected point is reinserted into the reverse update through a consistent noise estimate. After the final step, a thresholding map recovers a binary decision and the original mixed-integer problem reduces to a continuous program. The paper reports that this repeated-projection protocol improves Hamming distance, exact reconstruction, downstream feasibility, and objective gap relative to no-projection and final-step-projection variants on both benchmark families. The paper is explicit about the scope of the guarantee: the projection enforces only relaxed feasibility, and discrete feasibility is certified only when the recovery map satisfies a slack condition, so downstream feasibility and optimality are attributed to the continuous solver rather than to the diffusion model itself.","pith_inferences":["Inference: the projection cost, not reported separately, is the main hidden variable in the speedup claim. If the connectivity projection is solved as a MILP over 597 binary variables at each of 100 reverse steps on the 500-bus case, the reported 0.11 s sampling time would imply the projection is approximated, and the combinatorial difficulty would have moved into the sampling loop rather than dis","Inference: Proposition 2's Wasserstein-minimal-intervention result frames the projection as a constrained posterior-mean correction, which suggests the same reinsertion mechanism could be transferred to flow-matching or consistency-model samplers that share the posterior-mean parameterization.","Inference: the appendix's warm-start experiment indicates that for these power instances the solver bottleneck is optimality certification rather than primal solution quality; combining CGD with learned branching or cut-selection guidance is the natural next step, since feasible starts alone do not speed up the solver.","Inference: the slack certificates (Propositions 5 and 6) define a monitorable condition during generation: if the relaxed slack at the terminal point is too small to absorb the rounding displacement, thresholding can break combinatorial feasibility. Tracking this slack could serve as an early-warning signal for retraining or for switching to a repair map."],"forward_implications":["On structured mixed-integer nonlinear programs with a convex relaxation of the discrete constraints, the online combinatorial search can be replaced by one diffusion pass plus one continuous solve, amortizing the hardest part of the computation across related instances.","Applying the projection at every reverse step improves discrete accuracy and downstream objective gap relative to applying it only at the end, even in settings where the final-step projection already achieves zero infeasibility.","The generative model represents a distribution over discrete assignments, so it can provide multiple near-optimal feasible decisions for a single instance, something a deterministic predictor cannot do.","The reported end-to-end runtime is dominated by the continuous completion stage on the two larger power networks, so the speedup is largest when the discrete block is the main bottleneck of the joint problem."],"supporting_citations":[{"why":"Supplies the DDPM posterior-mean coefficients and reverse update that CGD's projected transition modifies.","marker":"(Ho et al., 2020)"},{"why":"DIFUSCO, the graph-based diffusion baseline that CGD extends and compares against.","marker":"(Sun & Yang, 2023)"},{"why":"Prior evidence that diffusion models lack native constraint enforcement, motivating the projection operator.","marker":"(Christopher et al., 2024)"},{"why":"LTO-MIP baseline, a learning-to-optimize method for MINLPs with feasibility guarantees that CGD evaluates against.","marker":"(Tang et al., 2025)"},{"why":"Provides the IEEE AC-OPF test systems and benchmark formulation for the transmission-switching experiments.","marker":"(Babaeinejadsarookolaee et al., 2019)"},{"why":"Defines the mixed-integer portfolio benchmark and the MIQP reference generation procedure.","marker":"(Sambharya et al., 2023)"},{"why":"The numerical solver that produces reference solutions and solves the downstream continuous problems.","marker":"(Gurobi Optimization, LLC, 2026)"}],"fun_headline_variants":["Feasibility-projected diffusion hits 425x speedup","Projecting each denoising step solves MINLPs 425x faster","CGD: diffusion with stepwise projection, 425x speedup","Constraint-aware diffusion: 425x faster mixed-integer solves","In-loop feasibility projection yields 425x MINLP acceleration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the projection operator called at every reverse step is cheap enough and close enough to the true feasible set that it does not dominate runtime and does not steer samples away from good optima; for the 500-bus power benchmark the paper describes the projection as a MILP over 597 binary line variables, and it does not report the projection cost separately from the overall sampling time.","fun_headline_variants_meta":{"raw":{"variants":["Feasibility-projected diffusion hits 425x speedup","Projecting each denoising step solves MINLPs 425x faster","CGD: diffusion with stepwise projection, 425x speedup","Constraint-aware diffusion: 425x faster mixed-integer solves","In-loop feasibility projection yields 425x MINLP acceleration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1312,"prompt_tokens":934,"completion_tokens":378,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":288}},"tokens_in":550,"tokens_out":378,"duration_ms":4144,"temperature":1.0,"reasoning_tokens":288,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:12:42.489899+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the wall-clock time of the projection call $\\Pi_{C_\\xi}$ alone on the IEEE 500-bus benchmark across all 100 reverse steps of a single sample. If solving the load-relevant connectivity MILP over 597 binary variables at each step takes longer than the reported 0.11 s total sampling time, the claim that the exact projection is applied at every step cannot be sustained. A second check is to verify on every test instance that the thresholded terminal point $T(u_0)$ belongs to $Z_\\xi$ and induces a feasible AC-OPF, since the paper states the projection does not guarantee either condition.","supporting_citations":[],"review_version":1}