{"id":"4fc38baf-3653-43f1-98d5-4b4d8bb3b75d","arxiv_id":"2501.09900","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SBAMDT is a Bayesian additive tree model with hard-or-soft and univariate-or-multivariate splits, reporting better spatial predictions than existing tree ensembles.","lead":"This paper introduces SBAMDT, a Bayesian tree model whose splits can be sharp or fuzzy and can use several spatial features at once, aiming at data on complex-shaped domains. It reports lower prediction error than BART, SBART, and BAMDT on simulated U-shaped data and New York City education data, though one variant underperforms in a benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'consistently outperforms' claim is undercut by tuning asymmetry: SBAMDT's q is tuned per dataset via CV while baseline tree depth/softness are not equivalently tuned, and S2-BAMDT loses on the square benchmark.","rationale":"The reader's verdict of CONDITIONAL is appropriate. The reader's weakest_assumption focused on the manifold embedding dimension k and the kinked derivative at Voronoi boundaries; my analysis agrees on the differentiability gap but identifies the tuning asymmetry and the S2-BAMDT failure on the square benchmark as more directly load-bearing for the paper's headline claim of 'consistent' superiority. These concerns are concrete, empirical, and addressable with a controlled re-run; they do not invalidate the core algorithmic idea. I therefore keep CONDITIONAL rather than moving to ACCEPT or REJECT, and note partial agreement because the reader did not emphasize the tuning asymmetry or the fact that Table 2 already offers a counterexample to the unqualified conclusion.","tokens_in":26527,"tokens_out":1491,"duration_ms":13458,"concrete_test":"Re-run the U-shape, square, and NYC benchmarks with a fixed q (e.g., q=8) for both SBAMDT variants and with a comparable tuning protocol for BART/SBART/BAMDT (e.g., tune their tree depth and SBART bandwidth by the same CV criterion). If Sk-BAMDT no longer leads on all three datasets, or if S2-BAMDT remains worse than BART on the square domain, the 'consistently outperforms' conclusion should be weakened to 'Sk-BAMDT, with per-dataset tuning, improved most benchmarks.' Additionally, evaluate the derivative in Eq. 8 at points on two sides of a Voronoi boundary using the exact nearest-neighbor procedure; if the directional derivatives differ, the mean-square differentiability assertion in Theorem 3.1(1) requires qualification.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Section 7) asserts SBAMDT 'consistently outperforms BART, SBART, and BAMDT.' The experiments support this only for the Sk-BAMDT variant, and only when its smoothness grid multiplier q is tuned per dataset: q=8 for U-shape and square, q=10 for NYC (Tables 1-3). No equivalent tuning of BART/SBART/BAMDT hyperparameters is reported, so part of the reported gain may reflect tuning effort rather than the adaptive soft split. More directly, the S2-BAMDT variant—the version closest to SBART because it learns a tree-specific smoothness scale—underperforms BART in the square example (RMSPE 11.35 vs 10.76) yet is bundled into the 'SBAMDT consistently outperforms' statement. The theoretical support is also weaker than claimed: in Appendix L, the soft multivariate derivative (Eq. 8) differentiates through nearest-neighbor assignments even though the nearest knots s_L*, s_R* change at Voronoi boundaries; the displayed expression is the derivative only on the open cell where the nearest knots are fixed. The GP-connection result (Theorem 3.1, part 1) asserts mean-square differentiability 'in the absence of hard splits' from that expression, but the kink at Voronoi boundaries is not addressed, so the smoothness conclusion is not established. These are correctness risks in the evidence base, not merely differences from consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SBAMDT, an additive Bayesian decision-tree regression model that combines hard and soft splits at each internal node, uses graph-Laplacian embeddings of structured multivariate features for multivariate splits based on minimum spanning trees, and assigns observations to children via distances to reference knots. Two variants are introduced: Sk-BAMDT, with a small discrete set of smoothness levels, and S2-BAMDT, with a tree-specific smoothness parameter. The model is evaluated on two synthetic examples and a New York City education dataset against BART, SBART, and BAMDT, with a claimed consistent improvement; a Gaussian-process connection is stated in Theorem 3.1 and proved in Appendix L.","tokens_in":26831,"tokens_out":6861,"duration_ms":60726,"significance":"The modeling idea is useful and the paper provides substantial implementation detail: full MCMC proposals, conditional distributions, and appendices for both variants. The GP connection is a nice theoretical observation, though the differentiability part is not fully established. If the empirical claim were supported, SBAMDT would offer a practical default for spatial nonparametric regression with complex domains. However, the current evidence base contains a contradictory result and a tuning asymmetry that prevent the strong 'consistently outperforms' claim from being accepted as stated.","major_comments":[{"comment":"Table 2 shows S2-BAMDT with RMSPE 11.35 (sd 3.46) on the square example, worse than BART at 10.76 (1.39) and BAMDT at 10.34 (1.06); yet Section 7's concluding sentence asserts that 'SBAMDT consistently outperforms BART, SBART, and BAMDT'. The conclusion therefore overstates what the experiments show: at best only Sk-BAMDT supports the claim, and only on the U-shape and NYC examples. Please restrict the claim to Sk-BAMDT or provide an explanation for the discrepancy and show that S2-BAMDT's underperformance does not undermine the central message. Additionally, the Sk-BAMDT RMSPE in Table 2 has a standard deviation of 3.76, which is large relative to the difference from BART, so the reported advantage for the winning variant may not be stable.","section":"Table 2 and Section 5.2"},{"comment":"The smoothness multiplier q is tuned per dataset for the SBAMDT variants (q=8 for U-shape, q=8/10 for square, q=10/6 for NYC, and q=6/4 in Appendix K) using cross-validation/WAIC, but no equivalent hyperparameter tuning is reported for BART, SBART, or BAMDT, which are presumably run with default settings. Under this asymmetry, part of the reported gain may be attributable to tuning effort rather than to the adaptive soft-split mechanism. Please add tuned baselines or a sensitivity analysis showing the results are robust to baseline hyperparameters.","section":"Section 3.3.0.2 and Tables 1-3"},{"comment":"The derivative of the soft multivariate split probability in Eq. (8) is computed as if the nearest left and right reference knots s_L* and s_R* are fixed, but these nearest knots change at Voronoi boundaries; at those boundaries the expression is not the derivative and the soft assignment has a kink. Consequently, the mean-square differentiability assertion in Theorem 3.1(1), which relies on this expression in 'the absence of hard splits', is not established. Please either prove differentiability across the Voronoi boundaries or weaken the statement to differentiability on the open cells where the nearest knots are constant.","section":"Appendix L, Eq. (8)"},{"comment":"The graph-Laplacian embedding uses the k eigenvectors corresponding to the k smallest non-zero eigenvalues, but the value of k is never specified or tuned in any of the experiments (U-shape, square, NYC, Appendix K). Since the multivariate split geometry and the faithfulness of Euclidean distance in the embedded space depend on k, the experimental protocol is not fully reproducible and the manifold-embedding premise is untested. Please report the chosen k for each dataset and provide a sensitivity check.","section":"Section 3.1"}],"minor_comments":[{"comment":"The labels 'Approach 1' and 'Approach 2' are swapped relative to their definitions in Section 3.2: in Section 4, Approach 1 (Multinomial/Dirichlet) corresponds to Sk-BAMDT, while Section 3.2 calls this Approach 2; similarly, Approach 2 in Section 4 (Gamma/Bernoulli) is Section 3.2's Approach 1 (S2-BAMDT). This will confuse readers trying to map the algorithm to the model variants.","section":"Section 4"},{"comment":"In the univariate split description, 'x∗ j(η) is the jth structured feature' should read 'unstructured feature', since the univariate split uses x.","section":"Section 3.1"},{"comment":"The text says 'we consider a noise-free environment' but then generates data with σ=0.1; please reconcile this inconsistency.","section":"Section 5.2"},{"comment":"The text at the start of Section 3.2 says 'Figure 1 compares hard and soft decision trees', but Figure 1 earlier shows the spanning tree bipartition; the comparison figure is actually Figure 2.","section":"Figure numbering"},{"comment":"'Watanable-Akaike information criterion' is a typo; also the Gelman et al. (2014) citation is repeated for both WAIC and cross-validation, but the latter is standard and could use an explicit reference.","section":"Section 3.3.0.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read SBAMDT. The core idea is genuinely new and useful: replacing BAMDT's hard multivariate splits with a node-specific mixture of hard and soft logistic gates based on distances to nearest reference knots, and using a graph-Laplacian embedding to approximate manifold distances. That is a reasonable upgrade for spatial problems with complex domains, and the Sk-BAMDT variant does well in the reported experiments. The MCMC details are spelled out more thoroughly than in many BART papers, which is a plus.\n\nThe problem is the conclusion. \"Consistently outperforms\" is not supported by their own tables. S2-BAMDT, the variant closest to SBART, loses to BART on the square domain (RMSPE 11.35 vs 10.76), and its MAPE and CRPS edges there are modest. Sk-BAMDT leads, but its smoothness multiplier q is tuned per dataset via cross-validation, and the baselines are not given equivalent tuning. That asymmetry alone could explain part of the reported gain. The authors should report a fairer comparison: tune all methods on the same criterion, or fix defaults for all.\n\nThe GP-connection theorem (3.1) is largely a repackaging of the known BART-as-GP result, with a proof gap that matters. Appendix L differentiates the logistic soft split as though the nearest knots were fixed, but those knots change at Voronoi boundaries. So the mean-square differentiability claim does not follow from the displayed expression. This is fixable by stating the derivative holds on open cells and addressing the kinks separately, or by proving differentiability of the composition more carefully.\n\nNo code or data is released, which limits reproducibility, especially since the empirical claims rest on a handful of simulations.\n\nI would send this to peer review: the core method is a genuine contribution, and the issues are all addressable in revision. The paper is for spatial statisticians and Bayesian nonparametrics researchers interested in tree-based models that handle soft transitions and complex geometry. Not yet a practical default, but a solid starting point.","headline":"A promising BART extension for spatial data whose empirical claims overreach, but the core idea deserves a close look.","tokens_in":27382,"tokens_out":2164,"would_cite":true,"duration_ms":22968,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62F15","62M30"],"pacs":[],"model":"deepseek-v4-flash","headline":"SBAMDT's node-level mixture of hard and soft splits beats BART, SBART, and BAMDT on spatial regression benchmarks.","keywords":["Bayesian additive regression trees","soft decision rules","semi-multivariate splits","spatial nonparametric regression","graph Laplacian embedding","minimum spanning tree bipartition","Gaussian process covariance","MCMC backfitting"],"falsifier":"On a noiseless square-domain regression surface that is exactly piecewise constant with axis-aligned jumps, run BART and Sk-BAMDT with the paper's default settings but a larger tree ensemble; if BART's RMSPE or CRPS falls below Sk-BAMDT's, the claim that SBAMDT consistently outperforms BART is false. The paper's square example already edges toward this regime but includes smooth pieces, so the sharp piecewise-constant version is the decisive check.","tokens_in":26312,"feed_emoji":"🗺️","tokens_out":13372,"duration_ms":119758,"temperature":0.7,"pith_summary":"The paper proposes SBAMDT, a Bayesian additive tree ensemble for nonparametric regression in which every internal node chooses between a hard split and a soft, probabilistic split, and in which soft splits can use multivariate structured features as well as single unstructured features. Its aim is to show that this adaptive hard-soft rule lets one ensemble represent sharp discontinuities and smooth transitions simultaneously, and it reports that SBAMDT consistently outperforms the BART, SBART, and BAMDT baselines on U-shaped, square-domain, and New York City education data. A careful reader would care because spatial regression on complex domains—road networks, city boundaries, brain surfaces—has typically forced a choice between axis-aligned trees that ignore geometry and rigid hard multivariate splits that cannot capture smooth variation; SBAMDT is offered as a single tool that does both. The claim is carried by a probabilistic leaf-assignment rule based on distances to nearest reference knots, with multivariate split geometry supplied by a graph-Laplacian embedding and spanning-tree bipartitions.","feed_headline":"Mixed hard-soft tree splits beat BART on complex spatial data","feed_subtitle":"A single Bayesian ensemble handles sharp boundaries and smooth transitions on U-shaped, square, and city domains.","key_machinery":"The central object is the adaptive soft semi-multivariate split rule at each internal node $\\eta$. It is a categorical choice $A_\\eta \\in \\{0,1,\\dots,k\\}$: $A_\\eta=0$ is a hard split that sends a point $d$ to the nearer child with probability one, while $A_\\eta=c$ is a logistic gate $z_{\\eta L}(d)=\\left(1+\\exp(-\\alpha_c(d_{\\eta R}(d)-d_{\\eta L}(d))/C_\\eta)\\right)^{-1}$, where $d_{\\eta L}(d)$ and $d_{\\eta R}(d)$ are the distances from $d$ to the nearest reference knots in the left and right children and $C_\\eta$ is a normalizing constant. Univariate splits use a single unstructured feature; multivariate splits first embed the structured-feature reference knots through the normalized graph Laplacian and then remove an edge from a minimum spanning tree, so the split boundary follows the geometry of the domain. The node probabilities multiply along a root-to-leaf path into leaf-assignment probabilities $\\phi_{hl}(d)$, and the ensemble $f(d)=\\sum_{h=1}^{m}\\sum_{l=1}^{L_h}\\mu_{hl}\\phi_{hl}(d)$ with Gaussian leaf weights is the model being fit. Theorem 3.1 identifies the induced covariance $C(f_i,f_j)=\\frac{\\beta_\\mu}{\\alpha_\\mu-1}\\sum_{h=1}^{m}\\sum_{l=1}^{L_h}\\phi_{hl}(d_i)\\phi_{hl}(d_j)$, which turns these probabilities into coherent uncertainty quantification.","core_discovery":"The central claim, stated in the paper's conclusion, is that SBAMDT consistently outperforms BART, SBART, and BAMDT for nonparametric regression with structured and unstructured features. In the U-shape simulation, the Sk-BAMDT variant attains the lowest root mean square prediction error, mean absolute prediction error, and continuous ranked probability score of the four methods; in the square-domain example, which was designed to favor axis-parallel BART, Sk-BAMDT still wins; and on the NYC education data both SBAMDT variants produce the lowest metrics while following spatial boundaries more closely than BART and SBART and adapting to smooth regions that BAMDT's hard splits cannot. The paper attributes these wins to the mixture of hard and soft decisions at each node rather than soft splits everywhere (SBART) or hard splits everywhere (BAMDT). It also proves in Theorem 3.1 that, conditional on trees and decision types, the prior and posterior of the fitted function are valid Gaussian processes, so predictive intervals come from a coherent stochastic-process model rather than an ad hoc covariance.","pith_inferences":["An implication the paper leaves implicit is that the covariance in Theorem 3.1 is a natural, data-driven kernel for kriging-style interpolation or Bayesian optimization on manifolds, since the same leaf-assignment products define similarity without a hand-chosen metric.","The hard/soft indicators could be summarized across MCMC iterations as a posterior 'discontinuity map' of the fitted surface, giving applied users a spatially explicit measure of where transitions are abrupt; the paper computes importance metrics but not this summary.","A testable extension is to put a hyperprior on the Laplacian embedding dimension $k$, which the paper never specifies or tunes, or to compare the embedded distances against true geodesic distances on the structured domain before fitting; this would show whether the multivariate split quality depends on that silent choice.","A controlled experiment with all splits univariate, and structured features excluded, would isolate whether the hard-soft mixture alone reproduces SBART's smoothness adaptation; the paper's comparisons always involve structured multivariate splits."],"forward_implications":["If the reported comparisons hold, SBAMDT becomes a better default than BAMDT for spatial regression on complex domains, because multivariate soft splits capture smooth within-cluster variation while hard splits preserve sharp boundaries.","The ensemble can represent functions smoother than the piecewise-constant surfaces of ordinary BART, since the logistic gates contribute smooth basis functions; this lifts a practical ceiling on hard-tree ensembles.","Posterior draws of the decision indicators $A_\\eta$ provide a per-node description of whether a learned boundary is sharp or gradual, so the model output also says where the fitted function changes suddenly.","Because the fitted function is a valid Gaussian process conditional on trees and decision types, users get predictive intervals without having to specify a covariance kernel.","Replacing geodesic $k$-nearest-neighbor construction with a graph-Laplacian embedding reduces the computational cost of MST-based multivariate splits, which matters for scaling to larger spatial datasets."],"supporting_citations":[{"why":"Defines BART, the baseline additive tree ensemble whose axis-parallel hard splits and prior structure SBAMDT extends.","marker":"Chipman et al., 2010"},{"why":"Introduces soft BART (SBART) and the soft logistic split rule that SBAMDT generalizes to a node-specific hard-soft mixture.","marker":"Linero and Yang, 2018"},{"why":"Introduces BAMDT and the semi-multivariate decision tree with MST-based multivariate splits, the direct baseline SBAMDT outperforms.","marker":"Luo et al., 2022"},{"why":"Surveys geodesic path and distance algorithms, motivating the replacement of geodesic k-NN construction with graph-Laplacian embedding.","marker":"Crane et al., 2020"},{"why":"Supplies the normalized graph Laplacian used to embed structured-feature reference points before forming multivariate splits.","marker":"Shi and Malik, 2000"},{"why":"Provides the Laplacian eigenmaps rationale for using Euclidean distance in the embedded space to approximate manifold geometry.","marker":"Belkin and Niyogi, 2003"},{"why":"Defines the continuous ranked probability score, one of the three metrics used to compare predictive performance.","marker":"Gneiting and Raftery, 2007"}],"fun_headline_variants":["Adaptive soft-split trees beat BART on spatial data","Soft splits that adapt beat BART on sharp or smooth regions","Bayesian tree with adaptive soft splits beats BART on complex data","Soft-split Bayesian trees outdo BART on spatial regression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's central gamble is that replacing true distances along the surface with ordinary distances in a graph-based coordinate embedding does not distort the geometry that determines the splits, and that the soft-split formulas stay smooth where the nearest reference knot changes; both assumptions are untested in the paper.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive soft-split trees beat BART on spatial data","Soft splits that adapt beat BART on sharp or smooth regions","Bayesian tree with adaptive soft splits beats BART on complex data","Soft-split Bayesian trees outdo BART on spatial regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001078,"raw_usage":{"total_tokens":4506,"prompt_tokens":937,"completion_tokens":3569,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":3497}},"tokens_in":553,"tokens_out":3569,"duration_ms":24241,"temperature":1.0,"reasoning_tokens":3497,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:33:56.299017+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a noiseless square-domain regression surface that is exactly piecewise constant with axis-aligned jumps, run BART and Sk-BAMDT with the paper's default settings but a larger tree ensemble; if BART's RMSPE or CRPS falls below Sk-BAMDT's, the claim that SBAMDT consistently outperforms BART is false. The paper's square example already edges toward this regime but includes smooth pieces, so the sharp piecewise-constant version is the decisive check.","supporting_citations":[{"cited_title":"Bamdt: Bayesian additive semi-multivariate decision trees for nonparametric regression","cited_arxiv_id":null,"evidence_quote":"Introduces BAMDT and the semi-multivariate decision tree with MST-based multivariate splits, the direct baseline SBAMDT outperforms."}],"review_version":1}