{"id":"937954d5-7740-4236-8ca2-7ee7fe05a5a3","arxiv_id":"2606.02777","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A novel covariance function integrating additive GPs with Vecchia approximation enables scalable emulation of large-scale computer experiments with mixed quantitative and qualitative inputs.","lead":"The paper proposes a new covariance function for additive Gaussian processes combined with Vecchia approximation to build scalable emulators for computer experiments that have both quantitative and qualitative inputs. Smart generalists might read it to understand how to handle large-scale mixed-input simulations more efficiently in engineering and scientific modeling.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No explicit check that the mixed-input additive covariance preserves Vecchia conditioning accuracy","rationale":"The reader's weakest assumption directly identifies the same unverified pairing between the new covariance and Vecchia. Because the full text is stated to be available yet the abstract supplies no derivation or error bound, the load-bearing gap remains exactly where the reader located it; no stronger internal inconsistency is visible from the given material.","tokens_in":1565,"tokens_out":309,"duration_ms":12793,"concrete_test":"Generate a 500-point design with 4 quantitative + 2 qualitative inputs; compute exact GP log-likelihood under the proposed covariance and under Vecchia (m=30 nearest neighbors) using the same ordering; if the absolute difference exceeds 2% of the exact value on more than one replicate, the accuracy-preservation claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the new covariance (additive GP over quantitative + qualitative inputs) can be paired with Vecchia without material loss of accuracy. Vecchia relies on a sparse approximation to the precision matrix via ordered conditionals; an additive construction that treats qualitative factors via separate kernels can produce dense cross-factor blocks that violate the local-neighbor assumption underlying Vecchia ordering. Without a proof that the resulting covariance still admits a sufficiently sparse Cholesky factor or an empirical bound on the Kullback-Leibler divergence between exact and Vecchia likelihoods, the scalability claim rests on an unverified structural assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes a novel scalable framework for emulating large-scale computer experiments with mixed quantitative and qualitative inputs. The core contribution is a new covariance function that integrates additive Gaussian processes to accommodate the mixed inputs, paired with the Vecchia approximation to achieve computational scalability. The authors claim that this modeling framework enables effective extension of existing large-scale emulation methods to the mixed-input setting.","tokens_in":1688,"tokens_out":321,"duration_ms":14474,"significance":"If the new additive covariance construction can be shown to preserve sufficient accuracy under Vecchia approximation, the work would address a practical challenge in computer experiment emulation and provide a reusable template for extending other scalable GP methods. The proposal of an additive GP covariance for mixed inputs combined with Vecchia is a direct response to a common modeling need, but its load-bearing assumption requires explicit support.","major_comments":[{"comment":"The central scalability claim rests on the unverified assumption that the proposed additive covariance for mixed inputs can be paired with Vecchia approximation without material loss of accuracy. Vecchia relies on ordered conditional approximations that assume local dependence; an additive construction separating quantitative and qualitative kernels can induce dense cross-factor blocks. No proof that the resulting precision matrix admits a sufficiently sparse Cholesky factor, nor any empirical bound on the Kullback-Leibler divergence between exact and Vecchia likelihoods, is supplied to substantiate the claim that existing large-scale methods extend effectively.","section":"Abstract / Proposed modeling framework and demonstration"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive comments on our manuscript. The primary concern is the need for stronger verification that the proposed additive covariance preserves accuracy under the Vecchia approximation. We address this point below and indicate the revisions we will make.","responses":[{"response":"We acknowledge that the current manuscript does not supply a formal proof of sparsity in the precision matrix or explicit empirical bounds on the KL divergence. The empirical demonstrations in the paper show that the overall framework performs well on large mixed-input problems, but they do not isolate the approximation error in the manner suggested. In the revision we will add a dedicated subsection that (i) derives the sparsity pattern induced by the additive covariance under standard Vecchia orderings and (ii) reports numerical KL-divergence comparisons across a range of quantitative/qualitative dimensions and sample sizes. These additions will directly address the concern that the additive construction may produce dense cross-factor blocks incompatible with Vecchia.","revision_made":"yes","referee_comment":"[Abstract / Proposed modeling framework and demonstration] The central scalability claim rests on the unverified assumption that the proposed additive covariance for mixed inputs can be paired with Vecchia approximation without material loss of accuracy. Vecchia relies on ordered conditional approximations that assume local dependence; an additive construction separating quantitative and qualitative kernels can induce dense cross-factor blocks. No proof that the resulting precision matrix admits a sufficiently sparse Cholesky factor, nor any empirical bound on the Kullback-Leibler divergence between exact and Vecchia likelihoods, is supplied to substantiate the claim that existing large-scale methods extend effectively."}],"tokens_in":1172,"tokens_out":341,"duration_ms":10610,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main things to know are that the authors build a covariance by adding separate GP kernels for the quantitative and qualitative inputs, then apply Vecchia to make the whole thing scale to large computer experiments. They show that this lets existing large-scale emulation techniques carry over to the mixed-input setting.\n\nThe covariance construction itself is the concrete new piece. Handling the two input types additively is a straightforward modeling move that avoids forcing a single kernel across everything, and the claim that prior Vecchia-based methods extend under this framework is the part that could matter for people who actually run these emulators.\n\nThe soft spot is the one flagged in the stress-test note. Vecchia gets its speed from sparse conditionals on local neighbors, but an additive covariance can produce cross-factor dependence that does not decay with distance in the ordering. If the qualitative kernels create denser blocks than the approximation assumes, the error could be larger than expected. The abstract says they demonstrate effective extension, yet without seeing explicit checks—such as KL divergence between the exact and approximated likelihoods or empirical sparsity of the resulting Cholesky—the scalability benefit stays partly assumptive. This is a moderate rather than minor concern because it sits right on the central claim.\n\nThe work engages the standard GP and Vecchia literature without circularity or obvious internal contradictions. It is aimed at the niche doing statistical emulation for computer experiments with categorical factors. A reader already working on scalable GPs would find the construction and the extension claim worth examining.\n\nIt deserves peer review. The problem is practical and the proposal is specific enough that referees can check the approximation details directly.","headline":"The paper gives a new additive covariance for mixed quantitative-qualitative inputs in GPs and pairs it with Vecchia, but the structural compatibility with Vecchia's local conditioning is not obviously guaranteed.","tokens_in":2183,"tokens_out":407,"would_cite":false,"duration_ms":23285,"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":"A new covariance function for mixed inputs paired with Vecchia approximation extends scalable emulation methods to computer experiments with both quantitative and qualitative factors.","keywords":["computer experiments","Gaussian processes","mixed inputs","Vecchia approximation","emulation","covariance function","scalability"],"falsifier":"Direct numerical comparison on a large mixed-input dataset in which the new emulator shows either substantially higher prediction error or no computational gain relative to standard large-scale methods applied only to the quantitative inputs.","tokens_in":2469,"feed_emoji":"","tokens_out":488,"duration_ms":18255,"temperature":0.7,"pith_summary":"The paper establishes a scalable framework for emulating computer experiments that take both numerical and categorical inputs. It introduces a covariance function built from additive Gaussian processes to manage the mixed inputs and applies the Vecchia approximation to keep computations feasible at large scales. The authors show that this pairing lets existing large-scale emulation techniques work on the mixed case. A reader would care because many simulation studies now include qualitative factors yet face severe computational limits when data grow large.","feed_headline":"Covariance function scales emulation for mixed-input computer experiments","feed_subtitle":"Additive GPs combined with Vecchia approximation let large-scale methods handle qualitative factors too.","key_machinery":"The new covariance function that integrates additive Gaussian processes with Vecchia approximation; it encodes the mixed-input structure while supplying the sparsity needed for large-scale computation.","core_discovery":"The central claim is that a covariance function integrating additive Gaussian processes to handle mixed quantitative and qualitative inputs, when used with the Vecchia approximation, forms a modeling framework under which methods already developed for large-scale computer experiments can be effectively extended while preserving the required accuracy and scalability.","pith_inferences":["The framework might be tested by replacing the Vecchia step with other sparse approximations to check whether accuracy holds under different sparsity patterns.","Applications in design optimization could treat qualitative factors as first-class inputs without separate categorical encoding tricks.","The additive structure may allow straightforward incorporation of additional input types if they can be expressed through similar additive covariance terms."],"forward_implications":["Existing large-scale emulation algorithms developed for quantitative inputs become directly usable on problems that also contain qualitative inputs.","Prediction and uncertainty quantification remain feasible at scales previously limited to purely quantitative experiments.","The same modeling structure supports extension of multiple different large-scale techniques rather than requiring entirely new algorithms for the mixed case."],"fun_headline_variants":["Additive GPs with Vecchia scale mixed-input emulation","Covariance framework extends emulation to qualitative factors","Vecchia approx supports large-scale mixed-input GPs","Scalable emulation for mixed quantitative and qualitative inputs"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That the proposed covariance function for mixed inputs remains accurate and computationally tractable once the Vecchia approximation is applied.","fun_headline_variants_meta":{"raw":{"variants":["Additive GPs with Vecchia scale mixed-input emulation","Covariance framework extends emulation to qualitative factors","Vecchia approx supports large-scale mixed-input GPs","Scalable emulation for mixed quantitative and qualitative inputs"]},"model":"grok-4.3","cost_usd":0.004244,"raw_usage":{"total_tokens":1973,"prompt_tokens":497,"num_sources_used":0,"completion_tokens":57,"cost_in_usd_ticks":42440500,"prompt_tokens_details":{"text_tokens":497,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1419,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":497,"tokens_out":57,"duration_ms":10326,"temperature":1.0,"reasoning_tokens":1419,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T11:39:53.324102+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Direct numerical comparison on a large mixed-input dataset in which the new emulator shows either substantially higher prediction error or no computational gain relative to standard large-scale methods applied only to the quantitative inputs.","supporting_citations":[],"review_version":1}