{"id":"cc6905c1-4014-47a3-add9-e60588cfb8a8","arxiv_id":"2607.11631","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"MAD-Path uses forward–backward diffusion paths as Metropolis proposals so multimodal targets stay invariant and mode weights are preserved better than under tempering.","lead":"This paper introduces MAD-Path, an MCMC sampler that proposes moves by running a noising diffusion forward then an approximate reverse diffusion back, and corrects bias with a Metropolis step on the whole path. It keeps the target distribution exact even with bad score estimates, and mixes better than tempering on asymmetric multimodal posteriors.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The reader's strongest claim is exactly the pair of results that the manuscript delivers: exact invariance under approximate scores/discretization (Theorem 3.1) and a weight-preserving, dimension-free ideal gap for the unequal-variance mixture (Corollary 2.6). Both are supported by complete proofs in the appendices and by coherent experiments that isolate the MH correction. The practical fragility of reverse-KL score learning with annealed-Langevin mode discovery is correctly identified by the reader as the weakest assumption, but it is not load-bearing for the central theoretical claims: the paper never asserts that the learned score is accurate, only that the Metropolis adjustment restores invariance and that acceptance is governed by the symmetrized path-space KL (Theorem 4.2). The missing spectral-gap bound for the adjusted chain is likewise an open direction stated in Section 7, not a contradiction. Because no stronger concern lands, the ACCEPT verdict with high confidence should stand unchanged.","tokens_in":39881,"tokens_out":601,"duration_ms":5106,"concrete_test":"Reproduce the unequal-variance mixture experiment of Section 6.1 (d=20, σ_{+}^{2}=0.5, σ_{-}^{2}=0.2) with the released code, using a deliberately mode-collapsed score (e.g., a single-Gaussian reference instead of the two-component GMM). Confirm that MAD-Path still recovers both modes with Max/Min weight ratio near 1 while the unadjusted reverse diffusion remains biased; if MAD-Path also collapses, the practical claim that MH alone corrects weight error would need re-examination.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claims hold under the paper's own scope. Theorem 3.1 establishes reversibility of the MAD-Path kernel for arbitrary approximate scores and Euler–Maruyama discretizations via the path-reversal involution on the augmented measure Π; the argument is self-contained and does not rely on score accuracy. The ideal-kernel spectral-gap results (Theorem 2.3, Proposition 2.4, Theorem 2.5, Corollary 2.6) correctly isolate the weight-preserving advantage of the diffusion path over tempering for Poincaré mixtures, and the unequal-variance Gaussian case is dimension-free under α∼d^{-1}. The reader's weakest assumption—that practical efficiency depends on reverse-KL score learning with a mode-aware GMM reference—is real but is already scoped by the paper: invariance is unconditional, acceptance is quantified by E_score (Theorem 4.2), and experiments show the MH correction recovers mode weights when the unadjusted diffusion is biased. Incomplete mixing theory for the adjusted chain is likewise acknowledged (Section 7) and does not contradict the stated claims. No internal inconsistency or load-bearing gap that would overturn the strongest claim was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes MAD-Path, an MCMC method that uses forward–backward diffusion paths as nonlocal proposals and corrects them with a Metropolis–Hastings step on an augmented path space. The ideal continuous diffusion-path kernel is shown to be reversible with respect to the target and to enjoy favorable spectral gaps: under a Poincaré inequality (Theorem 2.3), and for mixtures of Poincaré components with an explicit two-component bound that is insensitive to relative mode weights and scales (Theorem 2.5, Corollary 2.6). In practice, intermediate scores are learned variationally and the SDEs are discretized; Algorithm 1 and Theorem 3.1 establish that the resulting Metropolis-adjusted kernel remains reversible for any approximate score and any Euler–Maruyama step size. Theorems 4.1–4.2 quantify how discretization and score error enter the acceptance probability and yield a high-dimensional tuning guideline. Experiments on unequal-variance Gaussian mixtures, skew-normal mixtures, Bayesian GMMs, sensor localization, and a multimodal SUR profile likelihood show improved mode exploration and weight recovery relative to tempering-based MCMC and unadjusted diffusion samplers.","tokens_in":40198,"tokens_out":1103,"duration_ms":14607,"significance":"The work cleanly separates three contributions that matter for multimodal MCMC: (i) a weight-preserving interpolating path with a spectral-gap analysis that explains why it can succeed where tempering fails on asymmetric modes; (ii) a path-space Metropolis correction that restores exact invariance without requiring tractable endpoint densities; and (iii) quantitative acceptance expansions that give practical step-size and score-accuracy guidance. Theorem 3.1 is particularly useful: correctness is unconditional on score quality, so future improvements in score learning plug in directly. The ideal-kernel analysis (maximal correlation / data-augmentation viewpoint, mixture decomposition, dimension-free gap for α∼d^{-1}) is a genuine addition relative to proximal-sampler literature. Code and detailed experimental protocols are provided. The main limitation—that mixing theory for the adjusted chain and practical efficiency still hinge on mode-aware score learning—is scoped honestly in Sections 5 and 7 and does not undercut the stated claims.","major_comments":[{"comment":"The spectral-gap results (Theorems 2.3–2.5, Corollary 2.6) apply only to the ideal continuous kernel P, not to the Metropolis-adjusted discretized kernel used in practice. Section 7 correctly flags this as open, but the abstract and introduction sometimes read as if the favorable mixing properties transfer immediately to MAD-Path. A short, explicit caveat in the abstract and at the end of §2 (that the gap bounds motivate the path choice, while acceptance and mixing of the adjusted chain remain separate) would prevent over-reading.","section":null},{"comment":"Practical performance is tightly coupled to the two-stage score pipeline in §5 / Appendix C.1 (annealed Langevin mode hunt → GMM reference → reverse-KL control). Theorem 4.2 already shows that acceptance is governed by the symmetrized path KL E_score; the experiments demonstrate MH recovering weights when the unadjusted sampler is biased. Still, the paper would be stronger with one controlled failure-mode experiment (or a clear negative result) when mode discovery is incomplete or the reverse-KL collapses a mode, so readers can see how acceptance and ESS degrade rather than only success cases.","section":null}],"minor_comments":[{"comment":"Notation for the reverse process switches between Yt, bYt, and the Euler iterates yk; a short notation table early in §3 would help.","section":null},{"comment":"Figure 2 caption and Remark 4.1: the coincidence of the 0.234 acceptance optimum with RWM is interesting but easy to misread as implying random-walk behavior; the remark already clarifies this—consider elevating one sentence into the main text near the figure.","section":null},{"comment":"In §6.1–6.2, computational budgets are matched by target-score evaluations, which is appropriate for amortized MAD-Path, but wall-clock times are only in the appendix; a one-line summary in the main experimental section would aid comparison.","section":null},{"comment":"Typo/consistency: “Poincar´ e” appears with a broken accent in several places (e.g., §2.2); standardize to “Poincaré”.","section":null},{"comment":"Corollary 2.6 sets α=κ/d; a brief remark that T=Θ(log d) follows from α=e^{-T} would make the dimension-free claim more immediately usable for practitioners choosing the horizon.","section":null}],"recommendation":"minor_revision","confidential_remarks":"I agree with the reader’s high-confidence accept lean: the central invariance and ideal-kernel arguments are sound, and the tempering contrast is well motivated. I recommend minor revision rather than straight accept only to force the ideal-vs-adjusted scoping language and one more transparent discussion of score-learning failure modes; neither issue threatens correctness. Fit for a serious computational statistics / MCMC venue is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that MAD-Path is a genuine Metropolis correction on the full forward–backward diffusion trajectory, so the chain is exactly invariant for any learned score and any Euler step size. That is not a small fix; unadjusted diffusion samplers have been stuck without a clean density, and this path-space involution argument (Theorem 3.1) closes it cleanly.\n\nWhat is new is the combination: diffusion-path proposals (weight-preserving, unlike tempering), the maximal-correlation spectral-gap analysis of the ideal kernel (including a dimension-free lower bound for unequal-variance Gaussian mixtures under α ~ d^{-1}), and the acceptance expansions that separate discretization O(h^{1/2}) from integrated score error. The proximal-sampler connection is acknowledged; the Gibbs/maximal-correlation view is the useful re-framing. Experiments are the right ones—unequal Gaussians, skew-normals, Old Faithful GMM, sensor localization, SUR—and they show the MH step recovering mode weights when the unadjusted reverse path is biased. Code is linked.\n\nSoft spots are real but scoped. Efficiency still leans on reverse-KL score learning with a mode-aware GMM reference from annealed Langevin; if mode finding fails, acceptance can tank even though invariance holds. Mixing theory stops at the ideal kernel; the adjusted chain is left open (they say so in the discussion). Free parameters (T, N, GMM schedule, MLP budget) matter in practice. None of that overturns the central claims.\n\nThis is for people who actually run multimodal Bayesian MCMC and for the diffusion-sampler crowd that needs exactness. Math, experiments, and citations look solid. I would send it to referees without hesitation and would cite the invariance result and the unequal-mode gap myself.","headline":"Solid, usable MCMC paper: exact path-space MH on diffusion proposals plus a clean spectral-gap argument that tempering lacks for unequal modes.","tokens_in":40769,"tokens_out":463,"would_cite":true,"duration_ms":5782,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","65C05","60J22"],"pacs":[],"model":"grok-4.5","headline":"Diffusion-path proposals with a Metropolis correction sample multimodal targets without distorting mode weights.","keywords":["Markov chain Monte Carlo","diffusion path","multimodal sampling","Metropolis–Hastings","score estimation","spectral gap","tempering","Bayesian posterior"],"falsifier":"On the unequal-variance two-Gaussian mixture in growing dimension, replace tempering with MAD-Path using a correctly learned score and α ∼ 1/d: if the empirical spectral gap (or mode-crossing rate) still decays exponentially rather than staying order-one, the claimed advantage is false.","tokens_in":40776,"feed_emoji":"🔀","tokens_out":677,"duration_ms":6137,"temperature":0.7,"pith_summary":"Classical MCMC struggles with well-separated modes, and the usual fix—tempering—can exponentially warp the relative weights of modes that differ in scale, so the chain mixes torpidly. This paper replaces the tempering bridge with the marginal path of a noising diffusion that carries the target to a Gaussian; that path keeps mixture weights intact and has a spectral gap controlled by inter-mode distance and per-mode Poincaré constants rather than scale disparity. Because the intermediate scores are unknown, they are learned approximately and the continuous path is discretized; the resulting bias is removed by a Metropolis–Hastings step on the whole forward–backward trajectory, which leaves the true target invariant no matter how inaccurate the score or the step size. Acceptance rates are then governed by discretization error and an integrated squared score error, giving concrete tuning rules. On mixtures with unequal variances, skew-normals, Bayesian mixture models, sensor localization, and multimodal profile likelihoods, the corrected sampler recovers mode weights more accurately and crosses modes more often than tempering-based competitors or the unadjusted diffusion sampler.","feed_headline":"Diffusion paths fix multimodal MCMC without warping modes","feed_subtitle":"A Metropolis correction on forward–backward trajectories keeps the target exact even with approximate scores","key_machinery":"The MAD-Path proposal: evolve the current state forward under a discretized noising SDE, then reverse under an approximate score, and accept or reject the entire path via a Metropolis–Hastings ratio on the augmented trajectory space (a volume-preserving involution), guaranteeing detailed balance with the target.","core_discovery":"The Metropolis-adjusted diffusion path (MAD-Path) transition is reversible with respect to the target for any approximate intermediate score and any Euler–Maruyama discretization, while the ideal continuous diffusion-path kernel preserves mixture weights and, for unequal-variance Gaussian mixtures with diffusion horizon scaling as log d, has a spectral gap bounded away from zero independently of dimension.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["MAD-Path: exact multimodal MCMC via diffusion path proposals","Diffusion paths preserve mode weights better than tempering","Metropolis correction makes diffusion paths exact for targets","Sampling multimodal posteriors with unbiased diffusion trajectories","MAD-Path improves mode exploration without distorting weights"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"Practical acceptance rates depend on first finding the modes and learning intermediate scores via a reverse-KL path-space objective built around a Gaussian-mixture reference; if mode discovery fails or the objective collapses modes, efficiency collapses even though invariance still holds.","fun_headline_variants_meta":{"raw":{"variants":["MAD-Path: exact multimodal MCMC via diffusion path proposals","Diffusion paths preserve mode weights better than tempering","Metropolis correction makes diffusion paths exact for targets","Sampling multimodal posteriors with unbiased diffusion trajectories","MAD-Path improves mode exploration without distorting weights"]},"model":"grok-4.5","effort":"low","cost_usd":0.004892,"raw_usage":{"total_tokens":1364,"prompt_tokens":775,"num_sources_used":0,"completion_tokens":56,"cost_in_usd_ticks":48920000,"prompt_tokens_details":{"text_tokens":775,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":533,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":775,"tokens_out":56,"duration_ms":6461,"temperature":1.0,"reasoning_tokens":533,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T04:14:51.401348+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the unequal-variance two-Gaussian mixture in growing dimension, replace tempering with MAD-Path using a correctly learned score and α ∼ 1/d: if the empirical spectral gap (or mode-crossing rate) still decays exponentially rather than staying order-one, the claimed advantage is false.","supporting_citations":[],"review_version":1}