{"id":"791d6fcf-22c2-4a55-aa10-7d0f56f3ccc7","arxiv_id":"2607.23448","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A GP-guided neural parametric model learns threshold-to-solution maps for expensive constrained problems, enabling fast prediction and one-step refinement for arbitrary unseen thresholds.","lead":"CBA-BO trains one neural map from constraint thresholds to near-optimal designs during Bayesian optimization, then answers new threshold queries with a cheap prediction plus optional one local BO step. That matters for engineers who keep retuning feasibility limits on expensive simulations and cannot afford a full restart each time.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline win-rate-vs-N trend rests on a protocol that forbids baselines from pooling evaluations across thresholds — even though f and c_j are identical for every threshold, so one shared GP could serve all queries with no MLP at all.","rationale":"The reader's stated weakest assumption (GP-gradient fidelity and smoothness/single-valuedness of x*(θ)) is a real concern but is (a) acknowledged by the authors in the Conclusions, (b) partially mitigated by the one-step refinement and by the honest reporting in Table 1 and the Figure 3 visualizations showing where the mapping breaks down, and (c) less load-bearing than the evaluation protocol, because even a perfectly smooth x*(θ) would not establish the claimed advantage if a trivially pooled baseline matches it. The reader's rationale did flag that the large-N advantage 'partly follows from giving baselines a split budget by construction' — my analysis sharpens this from a partial inflation to the single most load-bearing issue: the split budget plus forced independence removes the one piece of information sharing (pooled surrogate data on identical f, c_j) that requires none of the paper's machinery. I keep the verdict at CONDITIONAL (UNCHANGED relative to the reader), because the contribution is coherent and the fix is well-defined: the condition list should specifically require the pooled-GP LogcEI baseline (and ideally a pooled multi-output/multi-task GP) at matched per-threshold budgets, alongside the reader's requested code release. If the pooled baseline is competitive, the paper's contribution reduces to the Direct Prediction use case (zero marginal-cost queries), which Table 1 shows is currently unreliable on 5+ of 16 problems without refinement — a substantially narrower claim than 'one model for all thresholds'. Credit where due: the ablation against random-initialization refinement (Table 6) is a good-faith attempt to isolate the PCM's contribution, and the N=1000 feasibility table reports failures openly; but Table 6's baseline still uses only 200 random samples rather than the optimization-focused pooled data a real BO run would produce, so it does not settle the issue.","tokens_in":30687,"tokens_out":2657,"duration_ms":79318,"concrete_test":"Implement a pooled-data LogcEI baseline: maintain one ModelListGP over all accumulated expensive evaluations (shared 30 LHS + the same 170-evaluation acquisition budget, spent however); then, for each queried threshold θ, refit nothing new — just optimize LogcEI(x; θ) under the pooled GP with the same 2 per-threshold evaluations CBA-BO uses for prediction+refinement. Run it on the five problems where CBA-BO's N-scaling advantage is largest (CEC2006-g01, CEC2010-c13, PressureVessel, ThreeTruss, Car) at N ∈ {5, 20, 50}, 10 runs, same win-rate rule. If pooled LogcEI achieves ≈50% win rate against CBA-BO refinement, the amortization advantage comes from data sharing rather than the learned PCM, and the central claim must be narrowed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that CBA-BO's win rate over constrained-BO baselines grows as the number of queried thresholds N increases (Figure 2), demonstrating amortization. But the protocol (Experimental Settings) gives each baseline only (170/N)+2 evaluations per threshold, optimized fully independently. This handicaps baselines in a way that is not intrinsic to the problem: the objective f and constraints c_j are the *same* black-box functions for every threshold — only the bound θ_j changes. An evaluation (x, f(x), c(x)) is therefore valid training data for every ECOP in the family. Nothing about cEI/LogcEI/SCBO prevents fitting a single ModelListGP on all pooled evaluations and simply re-optimizing the acquisition function per threshold, which is exactly what CBA-BO itself does (its GPs are fit on the pooled database D_t; the threshold only enters the PCM loss, Eq. 12). So the 'independence' of baselines is an artificial constraint of the experimental design, not of the methods. A pooled-GP LogcEI baseline with 2 per-threshold acquisition evaluations would match CBA-BO's evaluation budget and data access while requiring no learned threshold→solution map. If that baseline is competitive, the demonstrated N-scaling advantage is attributable to trivial data sharing, and the PCM/MLP machinery (the paper's claimed contribution) is load-bearing only in the Direct Prediction setting — where Table 1 shows feasibility as low as 39.5% (g01), 68–72% (ThreeTruss, Car), and refinement does not help KeaneBump at all (92.5→92.5). The strongest claim ('one model for all thresholds') thus currently stands on a comparison against baselines denied the most obvious reuse of their own data. The paper's dismissal of multi-task/transfer BO ('pre-defined finite tasks') does not address this, because here there is literally one shared function, not a task distribution.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper proposes CBA-BO, a framework for expensive constrained optimization in which the constraint thresholds θ vary over a domain Θ. A parametric constraint model (PCM), an MLP h_φ(θ), is trained concurrently with Bayesian optimization using gradients of GP-based LCB/UCB surrogate objectives (Eqs. 9–13), plus a monotonicity regularizer (Eq. 14) encoding that relaxing thresholds cannot increase the optimal objective. After a single 200-evaluation training run, the PCM predicts solutions for arbitrary unseen thresholds; a one-step local LogcEI refinement (Eq. 17) corrects prediction errors. An intent-guided threshold recommendation utility (Eqs. 18–22) is demonstrated on the car side-impact problem. Experiments on 16 problems compare against CMA-ES, cEI, ALBO, SCBO, and LogcEI under a protocol where N thresholds share a fixed total budget; win rates grow with N (Figure 2, Tables 12–13), and feasibility at N=1000 is reported with and without refinement (Table 1).","tokens_in":31079,"tokens_out":3894,"duration_ms":66844,"significance":"If the amortization claim holds under fair controls, this is a practically useful contribution: engineers genuinely re-solve ECOPs under varying thresholds, and a single 200-evaluation training run yielding a reusable threshold→solution map with millisecond inference would be valuable. The empirical work is substantial: 16 problems, 10 runs, win-rate protocol against five baselines, N up to 1000, ablations of PCM guidance and monotonic regularization, a PCM-vs-random-initialization refinement control (Table 6), pairwise monotonicity tests (Table 7), hyperparameter sensitivity (Tables 9–11), and honest reporting that direct prediction is unreliable on problems with active or switching constraints (e.g., 39.5% feasibility on g01, Table 1). The acknowledgment that the method degrades for high-D or irregular feasible regions is appropriately candid.","major_comments":[{"comment":"The headline N-scaling claim rests on a protocol that forbids baselines from pooling data across thresholds, even though f and c_j are the *same* functions for every threshold — only the bound θ_j changes, so every evaluation (x, f(x), c(x)) is valid training data for all N ECOPs. CBA-BO itself exploits exactly this (its GPs are fit on the pooled database D_t). A pooled-GP LogcEI baseline — one ModelListGP fit on all evaluations, with the acquisition re-optimized per threshold at (170/N)+2 evaluations each — would match CBA-BO's budget and data access with no learned PCM. If that baseline is competitive, the Figure 2 trend reflects trivial data sharing rather than the PCM, which is the claimed contribution. This control is load-bearing and must be added.","section":"§Experimental Settings, Figure 2, Tables 12–13"},{"comment":"Related but distinct: at N=50 each baseline gets (170/50)+2 ≈ 5.4 evaluations per threshold after the 30 shared LHS points — below what any GP-based method needs to fit even a single-output surrogate meaningfully. The monotone increase of win rate with N therefore partly reflects driving baselines into a degenerate regime rather than an intrinsic amortization advantage. The paper should either (a) report absolute objective values as a function of N for all methods (Tables 14–18 show only one θ per N), or (b) add a fixed per-threshold budget comparison with total-cost accounting, so the efficiency claim is separated from the budget-splitting artifact.","section":"§Experimental Settings"},{"comment":"The abstract's claim that CBA-BO 'directly predicts solutions for arbitrary unseen threshold configurations without additional optimization' is in tension with Table 1: direct prediction feasibility is 39.5% (g01), 68.2% (ThreeTruss), 69.2% (g07), 72.1% (Car), and KeaneBump is not improved at all by refinement (92.5±15.8% both rows). The practical method is PCM + per-threshold local LogcEI refinement, which is itself a small per-threshold BO. The framing 'one model for all thresholds' should be qualified, and the conditions under which direct prediction suffices (inactive constraints, smooth solution manifolds, per the Appendix visualization) should be stated in the main text, not only the Appendix.","section":"Abstract, Table 1, §Local Refinement"},{"comment":"The control meant to isolate the PCM's contribution trains GPs on 200 *randomly sampled* points, whereas CBA-BO trains on the BO-collected database D_T. This conflates two things: PCM-guided initialization and PCM-guided data acquisition (the batch strategy of Eq. 16 selects candidates that shape D_T). The clean control is LogcEI refinement initialized from a random point but with GPs fit on CBA-BO's own D_T (or on data collected by threshold-agnostic BO). Two rows (g01: 47.5%, GKXWC2: 45.5%) already show the random-init baseline winning; the authors' explanation (simple landscapes) is plausible but the confound should be removed before concluding the PCM learning strategy is what helps.","section":"§Analysis of PCM-based Initialization, Table 6"}],"minor_comments":[{"comment":"Presenting win rates as a dense grid of numbers (16 problems × 5 N values × 5 baselines) is hard to parse; a line plot of win rate vs. N per problem (small multiples) with baseline panels would communicate the central trend far better.","section":"Figure 2"},{"comment":"The main text underplays the N=1 column of Table 13: against LogcEI, CBA-BO scores 0% on g01, GKXWC1, KeaneBump, PressureVessel and 10% on g07. Since N=1 is the only setting where the baseline comparison is not affected by budget splitting, these losses deserve explicit discussion rather than the single sentence about LogcEI being 'highly competitive'.","section":"§Experimental Results and Analysis"},{"comment":"The surrogate objective switches discontinuously between LCB_f and total violation at the feasibility boundary; the text says gradients come from 'GP surrogate gradients' but does not discuss how the max(0,·) kinks in Eq. (10) and the switch in Eq. (11) are handled in the chain rule of Eq. (13). A sentence on subgradient choices would help reproducibility.","section":"Eq. (11)"},{"comment":"The monotonicity regularizer is applied to the LCB surrogate f̂, not the true f; since the monotonicity property holds for f* but only approximately for LCB_f (which includes a −βσ term), the regularizer can penalize legitimate uncertainty-driven variation. Please comment on this choice, and on the g07 regression in Table 7 (81.9% → 72.5%).","section":"Eq. (14)"},{"comment":"The threshold domains (e.g., [210×3, 110×6] for g01) are given without justification. How were θ_U values chosen relative to the constraint values at the unconstrained optimum? The difficulty of the learned mapping depends strongly on this choice.","section":"Table 4"},{"comment":"Code/data availability is not stated; given the number of tuned components (Table 5 lists ~20 settings), release of implementation would substantially strengthen the paper. Typos: 'A extra analysis' (§Ablation Study); inconsistent spacing artifacts throughout ('optimizationproblems', 'constraint-bounds in practice are hard to be determined').","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The requested pooled-GP baseline is cheap to implement (the authors already have the BoTorch infrastructure and the pooled database), so the revision burden is modest relative to its importance. The outcome genuinely could go either way: if CBA-BO still beats pooled-GP LogcEI, the paper becomes considerably stronger; if not, the honest contribution is the PCM's direct-prediction capability plus amortized inference cost, which is still publishable but narrower than the current framing. The N=1000 feasibility study and the monotonicity analysis are nice touches that suggest a careful group."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: CBA-BO is a real engineering-facing idea—learn one MLP map from constraint bounds θ to solutions while you run BO, then query arbitrary new bounds with optional one-step LogcEI polish—and the experiments are broader than most constrained-BO papers. The soft spot is that Figure 2’s “wins grow with N” story is inflated by a protocol that forces every baseline to restart with a split budget, even though f and c_j are identical across thresholds.\n\nWhat is actually new is the online PCM: train h_φ(θ) with GP LCB/UCB gradients (plus a soft monotonic regularizer), batch-sample candidates from the map, and keep a pooled database. That is a sensible synthesis of parametric/Pareto-set learning with expensive constrained BO, not just multi-task BO with a finite task list. The intent-guided recommendation case on Car is practical and clearly specified. Sixteen problems, ablations (All Rand / No Mono), N=1000 feasibility, PCM-vs-random init for refinement, and hyperparameter checks are honest empirical work. Direct prediction plus one refine is a clean two-stage story; when the map is good, amortization is real.\n\nThe stress-test lands. Nothing in cEI/LogcEI/SCBO prevents one shared ModelListGP on all evaluations and re-optimizing the acquisition per θ. CBA-BO itself pools D_t that way. So the headline N-trend partly measures “we allowed data sharing and they didn’t,” not only the value of the MLP. Direct prediction is where the map must carry the load—and Table 1 shows weak feasibility on g01, ThreeTruss, Car; KeaneBump does not improve under refine. GP-gradient fidelity and smoothness of x*(θ) remain the structural bet; the authors admit high-D / irregular regions. No code. Multi-task/transfer baselines are waved off a bit quickly.\n\nWho it is for: people who actually retune feasibility bounds on simulators and want a reusable θ→x prior, not a theory rewrite. Worth a serious referee if they demand a pooled-GP multi-query baseline, fixed total-budget curves, and code. I would engage, not dismiss.","headline":"Useful amortization idea for multi-threshold ECOPs, but the N-scaling win rates partly rest on baselines that are forbidden from pooling the shared (f,c) data.","tokens_in":31865,"tokens_out":559,"would_cite":false,"duration_ms":20091,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"One learned model maps any constraint threshold to a near-optimal design, so changing feasibility limits no longer means restarting expensive optimization.","keywords":["constrained Bayesian optimization","parametric optimization","threshold-to-solution mapping","expensive black-box constraints","Gaussian process surrogates","intent-guided recommendation","engineering design"],"falsifier":"On a problem with a highly irregular or multi-modal feasible region, train CBA-BO once with the stated budget, then query one thousand random unseen thresholds; if direct prediction plus one-step refinement loses to per-threshold LogcEI given the same total evaluations, the amortization claim fails.","tokens_in":31504,"feed_emoji":"⚙️","tokens_out":838,"duration_ms":20456,"temperature":0.7,"pith_summary":"Engineers often cannot fix constraint thresholds in advance and must try many feasibility settings on the same expensive black-box design problem. Standard constrained Bayesian optimization treats every new threshold vector as a fresh problem, wasting evaluations and ignoring that the underlying objective and constraints are shared. This paper claims that a neural network can be trained, inside a single Bayesian optimization run, to map any threshold vector directly to a high-quality solution. After roughly two hundred shared evaluations the model answers arbitrary new threshold queries instantly; one local Bayesian refinement step then corrects residual errors. The practical payoff is amortization: the more thresholds an engineer later queries, the larger the advantage over methods that re-solve each setting from scratch, plus a simple intent-guided recommender that suggests better thresholds matching stated tighten/lock/loosen preferences.","feed_headline":"One model answers every constraint threshold","feed_subtitle":"Train once on shared expensive evaluations; predict and refine designs for any new feasibility limit.","key_machinery":"The parametric constraint model (PCM): an MLP h_φ(θ) that outputs a candidate design for any threshold vector θ. It is trained by gradient descent on a feasibility-aware surrogate objective built from GP lower/upper confidence bounds, optionally regularized for monotonicity under threshold relaxation, then refined by one local constrained expected-improvement step.","core_discovery":"After a single shared Bayesian-optimization budget, a parametric constraint model learns a reusable map from continuous constraint-threshold vectors to near-optimal designs. Direct prediction plus one local LogcEI refinement then yields better feasible solutions than independent constrained BO or evolutionary runs on the same total evaluation budget, and the advantage grows as the number of queried thresholds increases.","pith_inferences":["The same threshold-to-solution map could serve as a warm-start generator for multi-fidelity or multi-objective constrained design loops that also vary resource or preference parameters.","If the PCM’s monotonicity regularizer is strengthened, the model may become a differentiable surrogate for sensitivity analysis of active constraints across the threshold domain.","Failure modes on narrow or disconnected feasible sets suggest a natural hybrid: fall back to local trust-region BO automatically when PCM prediction variance or violation exceeds a threshold."],"forward_implications":["Engineers can explore many feasibility–performance trade-offs after one optimization campaign instead of restarting for every new threshold.","Win-rate gains over independent constrained BO widen as the number of queried thresholds grows, because learning cost is amortized.","Direct PCM prediction already recovers high feasibility on many problems; one local refinement step lifts feasibility above 90–99 % on the harder cases tested.","An intent-guided recommender can propose new thresholds that improve the objective while obeying user tighten/lock/loosen preferences without manual trial-and-error."],"fun_headline_variants":["One model maps all constraint thresholds to designs","Learn once, predict designs for any threshold","Single parametric model covers every constraint bound","Threshold-agnostic BO reuses one shared model","Predict then refine: solutions for unseen thresholds"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Independent Gaussian-process surrogates supply gradients accurate enough, from only about two hundred expensive evaluations, to train one global neural map over the whole continuous family of thresholds.","fun_headline_variants_meta":{"raw":{"variants":["One model maps all constraint thresholds to designs","Learn once, predict designs for any threshold","Single parametric model covers every constraint bound","Threshold-agnostic BO reuses one shared model","Predict then refine: solutions for unseen thresholds"]},"model":"grok-4.5","effort":"low","cost_usd":0.003981,"raw_usage":{"total_tokens":1208,"prompt_tokens":708,"num_sources_used":0,"completion_tokens":50,"cost_in_usd_ticks":39808000,"prompt_tokens_details":{"text_tokens":708,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":450,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":708,"tokens_out":50,"duration_ms":6659,"temperature":1.0,"reasoning_tokens":450,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T21:53:28.147601+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a problem with a highly irregular or multi-modal feasible region, train CBA-BO once with the stated budget, then query one thousand random unseen thresholds; if direct prediction plus one-step refinement loses to per-threshold LogcEI given the same total evaluations, the amortization claim fails.","supporting_citations":[],"review_version":1}