{"id":"02e8f3d9-051a-4a7c-85ca-3cffa8937749","arxiv_id":"2507.15138","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"An adaptive-kappa, uncertainty-penalized GP-UCB is claimed to outperform fixed-parameter baselines, but the supporting theory is sketched and the empirical evidence is not shipped.","lead":"A paper proposes letting the exploration-exploitation dial in Gaussian-process optimization adjust itself as data arrive, and adds a penalty for uncertain, high-curvature regions. It claims this always beats fixed dial settings, especially in high-dimensional or noisy cases, but offers no code, no data, and only a sketch of the main proof.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's regret bound is not derived; its constants and assumptions are unverified, and Eq. (20) can let lambda_t exceed any lambda_max.","rationale":"The reader identifies unverified boundedness assumptions (Assumptions 3 and 4) as the weakest assumption. I agree: the update Eq. (20) has no upper bound, and the finite-difference Hessian complexity factor of Eq. (17) is not proved uniformly bounded, with only 10-100 evaluations in d = 20. These are exactly the conditions on which Theorem 1's regret bound depends. The stress-test adds that even if boundedness were granted, the proof of Theorem 1 is only sketched and the stated constants C1, C2 are not derived; the text simply says the proof follows from extending Srinivas. The paper also reports all empirical results in prose with no data, code, or numerical tables, so the 'always outperform' claim is not checkable. The paper describes a plausible algorithm and gives enough detail for re-implementation, but as submitted it does not contain a complete theoretical derivation nor the claimed empirical evidence. The central claim is therefore unsupported; REJECT (not CONDITIONAL) is appropriate because the missing proof and missing data are not minor fixes but the core evidence for the headline claim.","tokens_in":28401,"tokens_out":1663,"duration_ms":15254,"concrete_test":"Re-derive Theorem 1 from Srinivas-type regret analysis while explicitly tracking the acquisition function alpha_t(x) = mu_t(x) + kappa_t sigma_t(x) - lambda_t U_t(x) of Eq. (15), and verify whether Eq. (24) with C2 = 2*sqrt(T) beta_T lambda_max C_max / kappa_min actually follows once kappa_t, lambda_t, and C_t(x) vary with t. In parallel, run Algorithm 1 with the stated update Eq. (20) for 100 iterations on a d = 20 Ackley function at sigma_n = 0.05 and record max_t lambda_t and max_{x,t} C_t(x); if lambda_t exceeds any stated lambda_max or C_t(x) exceeds C_max, Assumptions 3-4 fail and the theorem gives no bound for the implemented algorithm.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim—that adaptive strategies always outperform fixed ones, with an enormous advantage in difficult settings—rests on Theorem 1, which is stated as a loose extension of the fixed-parameter GP-UCB analysis of Srinivas. However, Section 3.2.1 offers only a one-sentence proof sketch and does not actually derive the bound, so the bound's constants (C1 and C2 in Eq. (24)) and its formal conditions are not established. In particular, Assumption 3 requires kappa_min <= kappa_t <= kappa_max and 0 <= lambda_t <= lambda_max, and Assumption 4 requires C_t(x) <= C_max for all t; Eq. (20) updates lambda multiplicatively by a factor 1 + gamma*(I_t - Ibar_t)/Ibar_t with no cap, so nothing prevents lambda_t from growing without bound or from crossing a prescribed lambda_max. Similarly, Eq. (28) approximates the Hessian by finite differences, but no uniform bound on its eigenvalues for t <= 100 in d = 20 is proved, so the complexity factor C_t(x) in Eq. (17) is not shown to satisfy Assumption 4. A second, more basic gap is that the paper reports detailed empirical results in Section 5 without providing any data, code, or even tables of numbers, only prose and figure captions; the 'always outperforms' conclusion is therefore not verifiable from the manuscript. For these reasons, the central theoretical and empirical assertions are unsupported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adaptive extension of GP-UCB in which the exploration parameter κ_t and an uncertainty-penalty coefficient λ_t are updated online from prediction errors and integrated posterior variance, and an extra term λ_t U_t(x) is subtracted from the acquisition function. It states a regret bound (Theorem 1), a simple-regret bound (Theorem 2), and reports a large empirical comparison claiming that adaptive strategies always outperform fixed-parameter baselines, especially in high-dimensional and noisy settings. The theoretical section gives only a one-sentence proof sketch for Theorem 1, and the empirical section reports only prose summaries and figure captions, without tables of raw results, error bars, code, or reproducibility data.","tokens_in":28774,"tokens_out":3858,"duration_ms":46216,"significance":"If the regret bound and the empirical superiority claim were fully established, the adaptive update heuristics could be a practically useful contribution to Bayesian optimization. The paper does provide a clear algorithmic description (Algorithm 1), a broad literature review, and a plausible mechanism for adapting exploration. However, the central theoretical result is not derived, and the central empirical claim is not verifiable from the manuscript as written. The paper therefore does not currently meet the evidentiary standard for its headline conclusions.","major_comments":[{"comment":"","section":"Section 3.2.1, Theorem 1 and Assumptions 3-4"},{"comment":"The paper's central empirical claim—that adaptive strategies 'always outperform' fixed ones and do so by large margins in difficult settings—is reported only through prose and figure captions (e.g., '23% higher', '37% in 20 dimensions', '31% at σ_n=0.05'). No tables of mean values, standard deviations, standard errors, or per-trial data are provided, and the figures themselves are not included in the manuscript text. This makes the claim non-verifiable. The robustness analysis in Section 5.0.7 mentions a 40% lower coefficient of variation but again gives no supporting numbers. For a 'comprehensive empirical study', the absence of raw data or a reproducibility artifact is a serious deficiency.","section":"Section 5, Results and Analysis"},{"comment":"The update-rate constants β=0.1, γ=0.05, and η=0.1 are described as set 'based on preliminary experiments' (Section 3.1.3) and the experiments in Section 5 are run on the same test functions and noise settings used for that informal tuning. The fixed-κ baselines are drawn from a small hand-picked set {0.1, 0.5, 1.0, 2.0} rather than a systematic sweep or a theoretically motivated schedule. Consequently, the reported 'always outperform' conclusion may reflect tuning advantage rather than a general property of adaptive strategies, and the paper provides no cross-validation or sensitivity analysis to rule this out.","section":"Sections 3.1.3 and 4.4, parameter choices and baselines"}],"minor_comments":[{"comment":"The abstract contains the word 'strategery' and the phrase 'show conclusively', which overstates the evidence presented in the manuscript.","section":"Abstract"},{"comment":"The phrase 'usefully useful' appears in the discussion of GP assets; this should be corrected.","section":"Section 2.2"},{"comment":"There is a time-index inconsistency: Algorithm 1 computes α_{t-1}(x) = μ_{t-1}(x) + κ_t σ_{t-1}(x) - λ_t U_{t-1}(x), while Eq. (15) defines α_t(x) with matching time subscripts. This should be harmonized.","section":"Algorithm 1, line 5"},{"comment":"Reference [17] is cited for meta-learning and transfer-learning approaches, but the cited paper (Feurer et al., hyperparameter optimization via spectral approach) is not a transfer- or meta-learning method; a more appropriate citation is needed.","section":"Section 6.2.4"},{"comment":"No code or data availability statement is provided, which further hinders reproducibility of the empirical claims. A statement with a repository link or supplemental material should be added if a revised version is produced.","section":"General"},{"comment":"The spelling 'Matérn' is inconsistent in several places, and the section numbering in Section 5 uses third-level subsections such as 5.0.1, which should be renumbered.","section":"Section 4.2"}],"recommendation":"reject","confidential_remarks":"The manuscript's two main pillars—the regret theorem and the 'always outperform' empirical claim—are both unsupported as written: the theorem's proof is a one-sentence sketch that never verifies its own boundedness assumptions, and the empirical section contains no raw data, tables, or error bars. These are not presentation issues; they are load-bearing deficiencies that would require a substantial rewrite with a full proof and a reproducible empirical study. I therefore recommend rejection rather than major revision, although I acknowledge that the underlying algorithmic idea is not unreasonable and could be resubmitted after such work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The algorithm is plausible and the paper is readable, but the central claim—adaptive strategies always beat fixed ones, hugely in high-dimensional or noisy settings—rests on a one-sentence proof sketch and on empirical results that are described in prose only. There are no tables, no code, no raw numbers, so the main assertions are not verifiable from the manuscript.\n\nWhat is new: the acquisition function in Eq. (15) is UCB plus an uncertainty penalty λ_t U_t(x), with U_t = σ²_t C_t and C_t a Hessian-eigenvalue complexity factor. The update rules for κ_t and λ_t are multiplicative adjustments driven by prediction error and integrated variance. That is a reasonable heuristic; it could become a useful contribution if the theory and experiments were solid. The experimental design is broad—four test functions, d ∈ {2,5,10,20}, several noise levels, 30 trials—and the implementation details are specific enough for someone to re-implement the method.\n\nWhere it falls down. Theorem 1 is not derived. The proof is one sentence: 'follows from extending the analysis of Srinivas.' That does not establish the constants C1 and C2, nor does it show how the update rules (18)–(20) satisfy Assumptions 3 and 4. In particular, Eq. (20) for λ_t has no cap, so λ_t can exceed any prescribed λ_max and Assumption 3 can fail. The finite-difference Hessian in Eq. (28) has no uniform eigenvalue bound, so Assumption 4 on C_t is unverified. Without those bounds, the regret bound has no basis. The empirical 'always outperform' claim is reported as percentages with no raw data, confidence intervals, or code; it cannot be checked. Also, the update constants β=0.1, γ=0.05, η=0.1 were chosen on the same test problems, and the fixed-κ baselines are a small hand-picked set, so the universal conclusion is not supported.\n\nThe paper is not a waste. The heuristic is worth a proper study, and the author clearly knows the GP-UCB literature. But as it stands, this is a working-paper sketch rather than a publishable claim.\n\nWho it is for: someone wanting a concrete adaptive-κ heuristic to try, or a referee looking for a cautionary example of why evidence matters. I would not cite it in its current form, and I would desk reject it rather than send it to peer review. If the author supplies code, raw results, and a real proof of the boundedness assumptions, it could become a solid conference paper.","headline":"A plausible adaptive-GP-UCB heuristic, but the central 'always outperform' claim rests on a one-sentence proof sketch and on empirical results that exist only as prose.","tokens_in":29249,"tokens_out":1960,"would_cite":false,"duration_ms":20974,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive Gaussian-process search beats fixed strategies, with the largest gains in high-dimensional noisy problems.","keywords":["Gaussian process optimization","Bayesian optimization","adaptive exploration-exploitation trade-off","uncertainty quantification","regret bounds","high-dimensional optimization","noisy black-box optimization","upper confidence bound"],"falsifier":"Run the algorithm on the Ackley function in $d=20$ with $\\sigma_n = 0.05$, record $\\lambda_t$ and $C_t$ at each of 100 evaluations, and compare observed regret with Theorem 1's bound; if $\\lambda_t$ or $C_t$ exceeds the assumed $\\lambda_{\\max}$ or $C_{\\max}$ while regret stays sublinear, the boundedness assumption rather than the adaptive rule is carrying the theoretical guarantee.","tokens_in":28189,"feed_emoji":"🎯","tokens_out":5300,"duration_ms":56988,"temperature":0.7,"pith_summary":"This paper tries to establish that Gaussian-process optimization improves when the exploration-exploitation parameter and an uncertainty-penalty weight are updated online instead of fixed. The proposed acquisition function combines the posterior mean, an adaptive exploration term, and a penalty that grows with posterior variance and local curvature. The paper derives a sublinear cumulative-regret bound for this scheme and reports that adaptive strategies beat fixed-parameter UCB in every tested configuration, with the biggest margins at high dimension and high noise. If correct, this gives a practical route to removing manual parameter tuning in expensive black-box optimization.","feed_headline":"Adaptive GP tuning beats fixed strategies; hardest cases gain most","feed_subtitle":"Largest gains are at high dimension and high noise, with a sublinear regret bound to back it.","key_machinery":"The carrying mechanism is the uncertainty-aware acquisition function of Eq. (15): $\\alpha_t(x) = \\mu_t(x) + \\kappa_t \\sigma_t(x) - \\lambda_t U_t(x)$, where $U_t(x) = \\sigma_t^2(x) C_t(x)$ and $C_t(x) = \\sum_{i=1}^{d} \\max(|\\lambda_i|, \\epsilon)$ uses the eigenvalues of the Hessian of the posterior mean to estimate local curvature. Adaptive update rules move $\\kappa_t$ up when recent prediction error exceeds its moving average and move $\\lambda_t$ up when integrated posterior variance exceeds its moving average. These rules are what the regret analysis extends; the maximum information gain $\\gamma_T$ enters through $\\beta_T$ and through kernel-dependent bounds on $\\gamma_T$ for linear, squared-exponential, and Mat\\'ern kernels.","core_discovery":"On the paper's own terms, the central claim is that uncertainty quantification can serve as a control signal: raise exploration when recent prediction errors exceed their moving average, and raise the penalty on uncertain regions when global integrated variance exceeds its moving average. Formally, for the acquisition function $\\alpha_t(x) = \\mu_t(x) + \\kappa_t \\sigma_t(x) - \\lambda_t \\sigma_t^2(x) C_t(x)$, where $C_t(x)$ is the sum of absolute Hessian eigenvalues of the posterior mean, the paper derives a cumulative-regret bound $R_T \\le \\sqrt{C_1 T \\beta_T \\gamma_T} + C_2$ holding with probability at least $1-\\delta$, under boundedness assumptions on $\\kappa_t$, $\\lambda_t$, and $C_t$. Empirically, the paper reports consistent outperformance of fixed-parameter GP-UCB across dimensions $d \\in \\{2,5,10,20\\}$ and noise levels $\\sigma_n \\in \\{0.001, 0.005, 0.01, 0.05\\}$, with advantages growing from roughly 12-15% in easy settings to 31-37% in hard ones.","pith_inferences":["Editorial inference: the paper compares against fixed-$\\kappa$ UCB, EI, random search, CMA-ES, and BOHAMIANN, but it never ablates its own components; an experiment with only the $\\kappa$ update active and $\\lambda = 0$ would reveal which mechanism drives the reported gains.","Editorial inference: the Hessian-based complexity factor requires finite-difference estimates after only 10-100 evaluations in $d=20$; a cheaper proxy such as local nearest-neighbor variance or a GP gradient norm could test whether the complexity term is essential or merely suppresses exploration.","Editorial inference: the proof assumes bounds on $\\lambda_t$ and $C_t$ that the update rules do not explicitly enforce; if those quantities drift, the empirical success may rest on a mechanism other than the stated regret theorem."],"forward_implications":["Practitioners could drop manual tuning of $\\kappa$ in GP-UCB: the adaptive rule reproduces or beats the best fixed $\\kappa$ in every tested configuration.","The advantage of adaptivity grows with both dimension and noise, so the method is most valuable exactly where Bayesian optimization is hardest.","The sublinear regret bound transfers to linear, squared-exponential, and Mat\\'ern kernels, so convergence to the global optimum is preserved despite parameter adaptivity.","Simple regret converges at $O(\\sqrt{\\gamma_T / T})$ up to logarithmic factors, matching known rates for fixed-parameter GP optimization.","Lower variance across random initializations means fewer repeated runs are needed to obtain a reliable solution."],"supporting_citations":[{"why":"Supplies the GP-UCB regret framework and the information-gain quantity $\\gamma_T$ that the paper's Theorem 1 extends to adaptive parameters.","marker":"[66]"},{"why":"Provides information-theoretic regret bounds and the Mat\\'ern-kernel $\\gamma_T$ bound used in the corollaries.","marker":"[73]"},{"why":"Establishes kernelized-bandit regret tools that underpin the sublinear-regret argument.","marker":"[10]"},{"why":"Motivates adaptive parameter updates for GP bandits under heteroscedastic noise, the lineage of the paper's update rules.","marker":"[38]"},{"why":"Shows that no-regret Bayesian optimization is possible with unknown hyperparameters, the premise that adaptation need not destroy regret guarantees.","marker":"[4]"},{"why":"Defines the RKHS-norm assumption and kernel classes on which the regret bound rests.","marker":"[23]"}],"fun_headline_variants":["Adaptive GP beats fixed in hard settings, gains 31-37%","Uncertainty-guided GP tuning: adaptive wins, especially when hard","Adaptive κ in GP-UCB: biggest wins at high noise and dimension","For GP, adapt to uncertainty: huge gains in difficult problems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The regret guarantee presumes that the adaptive parameters and the curvature-based complexity factor stay inside fixed bounds, but the paper does not prove its update rules enforce those bounds.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive GP beats fixed in hard settings, gains 31-37%","Uncertainty-guided GP tuning: adaptive wins, especially when hard","Adaptive κ in GP-UCB: biggest wins at high noise and dimension","For GP, adapt to uncertainty: huge gains in difficult problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00033,"raw_usage":{"total_tokens":1903,"prompt_tokens":1070,"completion_tokens":833,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":686,"completion_tokens_details":{"reasoning_tokens":754}},"tokens_in":686,"tokens_out":833,"duration_ms":9278,"temperature":1.0,"reasoning_tokens":754,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:39:40.686471+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on the Ackley function in $d=20$ with $\\sigma_n = 0.05$, record $\\lambda_t$ and $C_t$ at each of 100 evaluations, and compare observed regret with Theorem 1's bound; if $\\lambda_t$ or $C_t$ exceeds the assumed $\\lambda_{\\max}$ or $C_{\\max}$ while regret stays sublinear, the boundedness assumption rather than the adaptive rule is carrying the theoretical guarantee.","supporting_citations":[{"cited_title":"Kakade, and Matthias Seeger","cited_arxiv_id":null,"evidence_quote":"Supplies the GP-UCB regret framework and the information-gain quantity $\\gamma_T$ that the paper's Theorem 1 extends to adaptive parameters."},{"cited_title":"Information-theoretic regret bounds for Gaussian process optimization in the Bandit setting","cited_arxiv_id":null,"evidence_quote":"Provides information-theoretic regret bounds and the Mat\\'ern-kernel $\\gamma_T$ bound used in the corollaries."},{"cited_title":"Cyclic hyperbolic Veech groups in finite area","cited_arxiv_id":"1706.06290","evidence_quote":"Establishes kernelized-bandit regret tools that underpin the sublinear-regret argument."},{"cited_title":"Adaptive Gaussian process bandits with heteroscedastic noise","cited_arxiv_id":null,"evidence_quote":"Motivates adaptive parameter updates for GP bandits under heteroscedastic noise, the lineage of the paper's update rules."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the RKHS-norm assumption and kernel classes on which the regret bound rests."}],"review_version":1}