{"id":"aa4ca277-242f-4681-addc-bfde3c5f4e48","arxiv_id":"2412.07186","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MCTS-transfer uses Monte Carlo tree search to partition the search space and adaptively reweight source-task data for faster Bayesian optimization on new tasks.","lead":"This paper describes an algorithm that borrows data from previously solved optimization problems to help solve a new, similar problem faster. It uses a tree-search method to divide the search space, adaptively weighting past problems that look most like the current one.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Adaptive-weighting claim rests on a similarity measure that the paper tunes on its own test benchmarks; unless the choice transfers to held-out tasks, the central claim is not established.","rationale":"The reader's weakest assumption identifies the reliability of the task-similarity proxy, and my reading agrees: this is the single most load-bearing condition for the central claim. The paper's own appendices show that similarity measures disagree and that the chosen measure was tuned on the very benchmarks used for the main comparison, so the experimental evidence does not yet show that the adaptive weighting mechanism generalizes. This is not an accusation of bad faith; the released code and detailed appendices make the method checkable, which is why the verdict remains conditional rather than reject. The UCB formula inconsistency is a real but secondary issue: it affects reproducibility and the stated exploration-exploitation trade-off, but the core scientific question is whether the adaptive reweighting actually delivers the claimed acceleration. A held-out ablation with uniform versus adaptive weights, using the released code, directly tests that question without circularity. If the adaptive weights do not provide a clear advantage on unseen tasks, the paper's main contribution reduces to MCTS-based space transfer without demonstrated adaptive source selection, which would weaken the strongest claim. If they do provide an advantage, the conditional concerns about measure choice are resolved. The recommended verdict is unchanged: the paper should be accepted only after the authors either release the exact code path for Eq. (1) and the chosen similarity measure, or add a held-out validation of the adaptive weighting component.","tokens_in":27590,"tokens_out":5392,"duration_ms":61710,"concrete_test":"Run the released MCTS-transfer implementation on a held-out set of transfer problems not used in Appendix D or the main evaluation (e.g., 10 additional BBOB functions or held-out HPOB search spaces), with two arms: adaptive weights via best-5 distance (Eq. (4)) and uniform all-one weights. If uniform weights match or beat adaptive weights in average final rank across the held-out tasks, the central adaptive-weighting claim is not supported; if adaptive weights win consistently, the concern is resolved. Also inspect the released UCB code to confirm whether n_m divides p_m a second time as printed in Eq. (1).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MCTS-transfer adaptively identifies similar source tasks and weights them to reconstruct the search space. This hinges entirely on the similarity measure of Section 3.2: Distance(mean of best N target samples, mean of best N source samples), used to compute ranks and weights in Eq. (4). For the claim to hold, this proxy must rank source tasks by transfer usefulness well enough that Eq. (3) biases node potentials toward genuinely promising regions. The paper's own evidence undercuts this. Appendix C.1 defines five competing similarity measures, and Appendix D shows their relative quality varies by problem: on mixed real-world problems, KL divergence is described as more appropriate, while best-5 distance is selected for the main experiments. Because these sensitivity analyses are conducted on the same benchmarks (LunarLander, RobotPush, Rover) that appear in the main evaluation, the choice of best-5 distance and of the hyperparameters γ, α, Cp, θ, and the classifier is partly selected by looking at the test data. If the measure is misspecified, the adaptive weights can mis-rank sources, the potential in Eq. (3) can be pulled toward bad regions, and the reconstruction step in Algorithm 2 can delete useful subtrees. The released code can resolve the implementation details, but the experimental protocol as written does not demonstrate that the adaptive weighting generalizes. A secondary concern is the apparent inconsistency in Eq. (1): p_m is already an average, so dividing by n_m again is not standard UCB, and the paper should clarify whether the code follows Eq. (1) or the usual average form.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MCTS-transfer, a search-space transfer method for Bayesian optimization. In a pre-learning stage it recursively partitions the search space by clustering source-task data and training binary classifiers, building an MCTS tree whose nodes correspond to subspaces. During optimization, node potentials are updated as a discounted, similarity-weighted combination of source and target sample means, UCB selects a leaf subspace for BO, and a Treeify procedure rebuilds subtrees that violate the left-better-than-right invariant. The authors evaluate MCTS-transfer-GP against non-transfer baselines (GP, LA-MCTS), three search-space transfer baselines (Box-GP, Ellipsoid-GP, Supervised-GP), and the surrogate-transfer method PFN, on BBOB, three real-world problems, Design-Bench, and HPOB, under both similar and mixed transfer settings, and report improved mean ranks. The paper includes a limitations section, hyperparameter sensitivity analyses in appendices, and a public code link.","tokens_in":27895,"tokens_out":10368,"duration_ms":97781,"significance":"If the empirical claims hold, MCTS-transfer would be a useful contribution to transfer Bayesian optimization: it is more flexible than existing search-space transfer methods, can be combined with any BO optimizer, and is evaluated on a wide range of benchmarks with code released. The adaptive weighting idea is interesting and the runtime analysis is helpful. However, the main claims are empirical, and the current evidence does not yet establish that the adaptive similarity weighting generalizes: the similarity measure and several hyperparameters are selected on the same benchmarks used for the headline results, and the weight curves used to illustrate 'identification' of similar tasks are produced by the very distance rule being validated. The UCB formula in Eq. (1)-(3) is also internally inconsistent. These issues are fixable, but they need to be addressed before the central claims can be accepted.","major_comments":[{"comment":"Section 3.1 states that v_m in Eq. (1) is replaced by the potential value p_m from Eq. (2). Since p_m is already an average of objective values, and Eq. (3) later defines p_m as a weighted average of per-task means, the exploitation term in UCB becomes p_m/n_m, i.e., an average divided by the visit count, rather than the standard total-reward/n_m. This changes the exploration-exploitation balance and makes the node-selection rule ambiguous. Please clarify whether p_m is meant to be a total or an average and adjust Eq. (1) accordingly (e.g., use p_m directly, or define the total value as p_m·n_m).","section":"Section 3.1 / Eq. (1)-(3)"},{"comment":"The sensitivity analysis in Appendix D is conducted on LunarLander, RobotPush, and Rover under the mixed setting, and the same three problems appear in the main evaluation (Section 4.3, Figure 3b, and Appendix E, Figure 16). Appendix D further states that 'In mixed transfer real-world problems, KL divergence is more appropriate to the problem,' yet Appendix B.1 reports that the main experiments adopt the best-5-solutions distance. This means the similarity measure and several hyperparameters (γ, α, Cp, θ, classifier) are effectively selected on the test benchmarks, undermining the claim that MCTS-transfer can 'adaptively identify' similar source tasks in a general setting. Please provide a held-out validation of the similarity-measure choice or clearly restrict the claims to the tuned configuration.","section":"Appendix D / Section 4.3 / Appendix B.1"},{"comment":"The evidence for the claim that MCTS-transfer 'can extract the most similar source tasks and give them higher weights' consists of weight-change curves (Section 4.1, Figure 2(b); Appendix G). These weights are computed by Eq. (4) from exactly the best-N mean-distance rule whose validity is at issue, so the curves largely confirm the definition rather than demonstrating that the weighting improves transfer. An independent validation would compare optimization performance with adaptive weights against fixed/equal weights, or measure the correlation between assigned weights and task transfer usefulness.","section":"Section 4.1 / Appendix G / Eq. (4)"},{"comment":"Section 3.4 states that 'for node expansion, only samples of the target task are considered,' and Algorithm 2 reconstructs subtrees with the same expansion process. If Treeify also uses only target samples, then any subtree rebuilt during optimization discards the source-task data that shaped the original partition, which appears to contradict the paper's claim that the tree can 'retain historical information.' Please state explicitly whether source samples are included in the clustering/classification during Treeify, and if so, how they are combined with target samples.","section":"Section 3.4 / Algorithm 2"}],"minor_comments":[{"comment":"There are numerous typos: 'filed' should be 'field' (Section 2.4), 'ajusts' should be 'adjusts' (Section 4.1), 'pre-learnned' should be 'pre-learned' (Section 3.4), 'splitale' should be 'splittable' (Algorithm 1), 'tranfer' should be 'transfer' in several figure legends, 'Specailly' should be 'Especially' (Appendix B.3), 'he result' should be 'the result' (Appendix D), 'influencial' should be 'influential' (Appendix D), 'stategy' should be 'strategy' (Appendix G), 'distribition' should be 'distribution' (Appendix C.1), and 'Simialr' should be 'Similar' in Appendix E titles.","section":"Throughout"},{"comment":"The sentence 'The first three methods are point-based measures and the last three are distribution-based methods' is inaccurate, as only five methods are listed and only the last two (Kendall coefficient and KL divergence) are distribution-based; please correct the count.","section":"Appendix C.1"},{"comment":"The parameter N in the definition of Distance(x_i^*, x_T^*) is not specified in the main text; it first appears as '5 best solutions distance' in Appendix B.1. Please state the default value in the main text.","section":"Section 3.2"},{"comment":"The per-task means y_i,m and y_T,m are undefined when a node contains no samples from a source task or from the target, which can occur after subtree reconstruction and for nodes not on the selected path. Please specify the default handling (e.g., skip, zero, or smoothing).","section":"Section 3.2, Eq. (3)"},{"comment":"The text says 'Higher weights are assigned to datasets D(5,5) and D(-5,5)', but the three source datasets listed at the start of the section are D(5,5), D(5,-5), and D(-5,-5); D(-5,5) is not among them. Please correct the dataset labels.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is borderline between major revision and reject; the core idea is interesting and the code is available, but the experimental protocol needs an independent validation of the similarity measure, and the UCB and tree-reconstruction descriptions need to be made consistent. I recommend requiring the authors to clarify these points and add a held-out experiment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful paper, not a landmark. The core idea—wrap LA-MCTS-style space partitioning with source-task similarity weighting and live tree reconstruction—is a sensible combination that fills a real gap in search space transfer for BO. The implementation is broad: BBOB, three real-world problems, Design-Bench, HPOB, plus runtime analysis and released code. The Sphere2D motivating example clearly shows the intended behavior, including weight curves that favor similar sources. That part works.\n\nThe soft spots are mostly in the evaluation, not the method.\n\nFirst, the UCB formula in Eq. (1) is inconsistent with the text. Eq. (1) is the standard UCB with v_m the total value, so v_m/n_m is an average. Section 3.1 says replace v_m with p_m from Eq. (2), which is already an average. Then the exploitation term becomes p_m/n_m, which shrinks with visits and would make the search degenerate. Either the text or the formula is wrong; the code will tell, but the paper as written is internally inconsistent.\n\nSecond, the central claim that the method 'adaptively identifies and leverages similar source tasks' rests on a similarity measure that is chosen by looking at the same test problems. Appendix D's sensitivity analysis runs on the same real-world benchmarks used in the main evaluation, and it explicitly says KL divergence is more appropriate for mixed real-world problems, yet the main experiments use best-5 distance. That is tuning on the test set, and it weakens the claim that the adaptive weighting generalizes. The weight-change curves in the motivating example are encouraging, but they are on a toy where similarity is obvious by construction.\n\nThird, the HPOB results are described as 'clear advantages' when the paper itself says the final-value gaps are tiny. That is overstated.\n\nFourth, the error bars are never defined (std vs SEM, number of replicates), which makes it hard to judge the significance of the rank plots.\n\nThese are not fatal. The method is plausible and the code is available, so a referee can check the actual UCB implementation and rerun the experiments. But the published version overstates what the experiments establish, and the similarity-measure selection needs stronger justification, ideally a held-out validation set.\n\nBottom line: worth a serious referee, and I would cite it as related work, but I would not rely on the headline numbers until the code confirms the formula and the experimental protocol is tightened.","headline":"MCTS-transfer is a reasonable, well-tested extension of LA-MCTS to transfer BO, but the main empirical claim is undercut by tuning the similarity measure on the test benchmarks and a UCB formula inconsistency.","tokens_in":28441,"tokens_out":4189,"would_cite":true,"duration_ms":44589,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MCTS-transfer grows a search-space tree from source tasks and adaptively rewires it during optimization, and the experiments report better rankings than other transfer methods on BBOB, real-world problems, Design-Bench, and HPOB.","keywords":["search space transfer","Bayesian optimization","Monte Carlo tree search","black-box optimization","transfer learning","adaptive task similarity","hyperparameter optimization","search space partition"],"falsifier":"Construct a synthetic family of target functions whose true similarity ordering is known in advance but where the best-$N$ sample-mean distance mis-ranks the sources, for example by shifting the optimum along a direction where the best samples of a dissimilar source are closer to the target's early samples than a truly similar source's samples. If MCTS-transfer then assigns higher weights to the misleading source and its tree reconstruction moves the chosen subspace away from the true optimum, the central adaptive-weighting claim is contradicted.","tokens_in":27347,"feed_emoji":"🌳","tokens_out":8192,"duration_ms":77401,"temperature":0.7,"pith_summary":"Fresh Bayesian optimization runs are usually started from scratch even when related tasks have already been solved. This paper proposes MCTS-transfer, a method that transfers the search space itself from source tasks to a new target task by growing a Monte Carlo tree whose nodes are subspaces. The tree is first built from source-task data to warm-start the target, then rebuilt during optimization so that source tasks judged more similar to the target get larger weights in the node-potential calculation. The authors aim to show that this adaptive weighting lets the method keep the benefits of transfer even when some source tasks are dissimilar, and their experiments on synthetic functions, real-world problems, Design-Bench, and hyperparameter optimization report better rankings than existing search-space transfer methods. If correct, the method makes transferable Bayesian optimization more reliable when task similarity is unknown ahead of time.","feed_headline":"Adaptive tree search reuses past tasks to speed black-box optimization","feed_subtitle":"Similar source tasks get higher weight as the tree rebuilds, so mixed or even misleading data still helps the search.","key_machinery":"The central object is the MCTS search tree, where the root is the entire domain and each node is a subspace produced by k-means clustering of the samples inside its parent followed by a binary classifier that separates good from bad regions. Node value is the potential $p_m$ of Eq. (3): a decay-weighted blend of the average target-task value in the subspace and a similarity-weighted average of source-task values, and node selection follows the UCB score of Eq. (1), the empirical mean potential plus an exploration bonus that grows with the logarithm of the parent's visit count. The load-bearing update is the adaptive weight $w_i$ of Eq. (4), computed from the rank of each source task's distance to the target, which changes with every new target sample and feeds both the node potentials and the reconstruction decision.","core_discovery":"On its own terms, the paper's central claim is that search-space transfer for Bayesian optimization can be made adaptive by representing the partition of the search space as a Monte Carlo tree. Each node of the tree is a subspace; its potential is a weighted combination of the average objective values of the source-task samples and target-task samples inside that subspace, with source weights recomputed after every new target evaluation according to how close each source task is to the target (distance between means of the best few samples). Because the left child of every node is supposed to be the better subspace, a violation of that ordering triggers a tree-reconstruction step that prunes and regrows the subtree, letting the search space track the target task as observations accumulate. The paper reports that this two-stage procedure, an offline pre-learned tree for warm-start followed by online adaptive partition, outperforms the compared search-space transfer and transfer-BO baselines across BBOB, real-world, Design-Bench, and HPO-B benchmarks in both similar and mixed transfer settings.","pith_inferences":["Editorial inference: because the decay factor $\\gamma$ controls how quickly source influence fades, MCTS-transfer is effectively a continuous interpolation between full transfer and from-scratch optimization; setting $\\gamma$ per task, rather than fixing it at 0.99, could extend the method to settings where the target drifts over time.","Editorial inference: the similarity measure is chosen by sensitivity analysis on the same benchmarks used for evaluation, so an independent check would be to fix one distance and one weight rule on a held-out family of tasks and see whether the adaptive weights are still better than fixed equal weights.","Editorial inference: the tree represents the search space generically rather than through a geometric box, so the same recipe could be applied to categorical or conditional search spaces; a concrete test would be to run it on the hierarchical spaces the paper discusses as future work."],"forward_implications":["A user can take any Bayesian optimizer and run it inside the promised leaf subspace, so improving the partition immediately improves any downstream acquisition function.","Because weights are recomputed online, the method can identify which source tasks resemble the target and progressively ignore misleading ones, which is the key to working under mixed transfer.","The tree's reconstruction gives a built-in warm-start for the first evaluation and a continuing correction mechanism, so the method does not need to know task similarity in advance.","The paper's runtime analysis indicates that back-propagation and reconstruction add only a small overhead relative to evaluation on expensive real-world problems, making the transfer mechanism affordable where evaluations dominate.","The paper's MCTS-transfer-PFN combination shows that the search-space transfer is orthogonal to the surrogate model, so its gains stack with stronger BO models."],"supporting_citations":[{"why":"Supplies the MCTS-based space-partition backbone that MCTS-transfer extends by adding source-task weighting and tree reconstruction.","marker":"[43]"},{"why":"Introduces search-space pruning for hyperparameter optimization, the line of transfer work this paper builds on.","marker":"[49]"},{"why":"Defines the Box-GP and Ellipsoid-GP search-space transfer baselines that MCTS-transfer is compared against.","marker":"[23]"},{"why":"Proposes Supervised-GP, a similarity-guided search-space design baseline, and its voting aggregation is a direct comparison.","marker":"[17]"},{"why":"Provides the PFN surrogate-model transfer baseline that MCTS-transfer aims to match or surpass.","marker":"[21]"},{"why":"Supplies the BBOB synthetic benchmark functions used to evaluate the method.","marker":"[9]"},{"why":"Supplies the HPO-B hyperparameter optimization benchmark used in the experiments.","marker":"[24]"},{"why":"Supplies the Design-Bench problems used for the high-dimensional experiments.","marker":"[39]"}],"fun_headline_variants":["MCTS space transfer adapts search to similar tasks","Tree search rebuilds space using source-task data","Adaptive tree search learns from similar source tasks","Monte Carlo tree partitions space for transfer BO","Reconstruct search space adaptively via MCTS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-picked task-similarity measure, distance between the means of the best few sampled points in each task, ranks source tasks the way a genuinely useful transfer would; the adaptive weights, node potentials, and tree reconstructions all inherit whatever errors that ranking makes.","fun_headline_variants_meta":{"raw":{"variants":["MCTS space transfer adapts search to similar tasks","Tree search rebuilds space using source-task data","Adaptive tree search learns from similar source tasks","Monte Carlo tree partitions space for transfer BO","Reconstruct search space adaptively via MCTS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000639,"raw_usage":{"total_tokens":2950,"prompt_tokens":962,"completion_tokens":1988,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":1916}},"tokens_in":578,"tokens_out":1988,"duration_ms":13057,"temperature":1.0,"reasoning_tokens":1916,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:03:38.124959+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a synthetic family of target functions whose true similarity ordering is known in advance but where the best-$N$ sample-mean distance mis-ranks the sources, for example by shifting the optimum along a direction where the best samples of a dissimilar source are closer to the target's early samples than a truly similar source's samples. If MCTS-transfer then assigns higher weights to the misleading source and its tree reconstruction moves the chosen subspace away from the true optimum, the central adaptive-weighting claim is contradicted.","supporting_citations":[{"cited_title":"Learning search space partition for black-box optimization using monte carlo tree search","cited_arxiv_id":null,"evidence_quote":"Supplies the MCTS-based space-partition backbone that MCTS-transfer extends by adding source-task weighting and tree reconstruction."},{"cited_title":"Hyperparameter search space pruning - A new component for sequential model-based hyperparameter optimization","cited_arxiv_id":null,"evidence_quote":"Introduces search-space pruning for hyperparameter optimization, the line of transfer work this paper builds on."},{"cited_title":"Learning search spaces for Bayesian optimization: Another view of hyperparameter transfer learning","cited_arxiv_id":null,"evidence_quote":"Defines the Box-GP and Ellipsoid-GP search-space transfer baselines that MCTS-transfer is compared against."},{"cited_title":"Transfer learning based search space design for hyperparameter tuning","cited_arxiv_id":null,"evidence_quote":"Proposes Supervised-GP, a similarity-guided search-space design baseline, and its voting aggregation is a direct comparison."},{"cited_title":"Pfns4bo: In-context learning for Bayesian optimization","cited_arxiv_id":null,"evidence_quote":"Provides the PFN surrogate-model transfer baseline that MCTS-transfer aims to match or surpass."},{"cited_title":"Real-parameter black-box optimization benchmarking 2009: Noiseless functions definitions","cited_arxiv_id":null,"evidence_quote":"Supplies the BBOB synthetic benchmark functions used to evaluate the method."},{"cited_title":"Jomaa, Martin Wistuba, and Josif Grabocka","cited_arxiv_id":null,"evidence_quote":"Supplies the HPO-B hyperparameter optimization benchmark used in the experiments."},{"cited_title":"Design-bench: Bench- marks for data-driven offline model-based optimization","cited_arxiv_id":null,"evidence_quote":"Supplies the Design-Bench problems used for the high-dimensional experiments."}],"review_version":1}