{"id":"c22b52c2-1598-4ffc-a323-50893aa6a01a","arxiv_id":"2505.19431","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Importance Weighted Score Matching trains diffusion samplers by reweighting score matching with self-normalized importance sampling to approximate the forward KL and improve mode coverage.","lead":"This paper proposes a training objective for neural samplers that reweights the score matching loss with importance sampling, aiming to cover all modes of a target distribution when only its unnormalized energy is available. It reports gains in mode coverage on Gaussian mixtures and particle systems, together with bias and variance bounds for the estimator.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-parent SNIS normalization in Eq. (15) estimates a different objective than Eq. (14), so Proposition 3 does not cover the implemented loss.","rationale":"The reader's weakest assumption identifies the same gap: the per-parent SNIS normalization in Eq. (15) is not the marginal-proposal SNIS estimator analyzed in Eq. (12) and Proposition 3. My reading of the full text confirms this is the most load-bearing weakness in the central claim. The paper wants to show that the implemented loss optimizes a forward-KL-type objective with principled importance weighting; that connection is broken by the per-parent normalization. I do not see a stronger objection. The empirical ablations (SNIS vs W/O IW in Section 5.3) are useful and suggest importance weighting helps, but they do not validate the specific theoretical objective because the implemented normalization differs from the one analyzed. The experiments are otherwise reasonably thorough, and the claim of improved mode coverage on GMM benchmarks is plausible as an empirical finding. Thus the conditional verdict remains appropriate: accept only if the authors either prove the per-parent estimator is consistent for Eq. (14), modify the algorithm to match Eq. (12), or substantially soften the theoretical claims to describe the per-parent loss as a heuristic.","tokens_in":34851,"tokens_out":5461,"duration_ms":50479,"concrete_test":"Modify Algorithm 1 line 7 so that the SNIS denominator in Eq. (12) is computed over the joint batch {x_t^{(b,s)}}_{b=1..B, s=1..S} drawn from pB_t (one x_t per parent, or all conditional draws pooled), instead of per-parent normalization in Eq. (15). Retrain GMM-40 and GMM-80 with identical hyperparameters and seeds, and compare W1 and E-TVD against Table 1. A material degradation (or improvement) would show that per-parent normalization is load-bearing and that the published numbers do not test the objective in Eq. (14). Additionally, re-derive Proposition 3 for the per-parent estimator to see whether a residual bias term E_{x0}[(1/Z(x0)-1)g(x0)] appears.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4's practical loss is not the estimator analyzed in Section 3.3. Eq. (12) defines SNIS weights over a batch {x_t^(s)} drawn from the marginal proposal pB_t, and Proposition 3 bounds that estimator. Algorithm 1 / Eq. (15) instead draws S conditional samples x_t|0 ~ p_{t|0}(·|x0) for one parent x0 and normalizes the weights within that per-parent conditional batch. For fixed x0, self-normalization estimates E_{pt|0(·|x0)}[w f] / E_{pt|0(·|x0)}[w], not E_{pt|0(·|x0)}[w f]; the denominator E_{pt|0(·|x0)}[w] is generally not 1 and varies with x0. Averaging over x0 therefore optimizes E_{x0~pB0}[ (∫ pt|0 w f) / (∫ pt|0 w) ], not E_{x0~pB0}[∫ pt|0 w f] = E_{xt~pt}[f], the ideal objective of Eq. (14). Since w = pt/pB_t, the per-parent normalizer is data-dependent and can be far from 1 whenever the buffer misrepresents a mode of pt. Thus the central theoretical claim that the trained loss is the importance-weighted forward-KL/score-matching objective does not hold for the implemented algorithm, and Proposition 3's bias/MSE bounds do not apply to Eq. (15). The reported improvements may still be real, but they are currently supported only empirically.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Importance Weighted Score Matching (IWSM), a data-free training objective for diffusion samplers that reweights a score-matching loss by the density ratio pt/pB_t, where pB_t is the marginal obtained by diffusing samples from a replay buffer of the model's own past samples. The paper derives a path-KL/score-matching equivalence (Proposition 1), proves bias/variance bounds for the proposed Monte Carlo estimators (Propositions 2 and 3), and reports experiments on GMM-40/80/120, DW-4, LJ-13, and LJ-55 that outperform PIS, FAB, DiKL, and iDEM on several distributional metrics. The central difficulty is that the theoretical estimator in Section 3.3 normalizes SNIS weights over a batch from the marginal proposal pB_t, whereas Algorithm 1 and Eq. (15) normalize weights within per-parent conditional batches {x_{t|0}^{(s)}} ~ p_{t|0}(·|x0); these two estimators are not the same, and the paper does not prove that the per-parent normalization preserves the intended expectation.","tokens_in":35279,"tokens_out":12242,"duration_ms":89338,"significance":"The underlying idea—using the sampler's own replay buffer to build an adaptive proposal and correcting the score-matching loss by importance weights—is attractive and, if properly analyzed, would be a meaningful step toward principled mode-covering samplers. The experiments cover a good range of benchmarks, and the ablations in Section 5.3 and Appendix H provide useful empirical evidence about the role of SNIS sample counts and the cost of direct estimated-score sampling. However, the manuscript's central theoretical claim does not currently cover the implemented loss: the per-parent normalization in Eq. (15) estimates a different functional that need not converge to the forward-KL score-matching objective. The theoretical support is therefore load-bearing and currently broken, although the empirical findings may survive a corrected analysis.","major_comments":[{"comment":"The loss actually optimized is not the estimator analyzed in Proposition 3. In Eq. (12) the SNIS weights are normalized over S samples {x_t^(s)} drawn from the marginal proposal pB_t, and Proposition 3 bounds that estimator. In Algorithm 1 (lines 4–7) and Eq. (15), for each buffer sample x0 the S samples are drawn conditionally from p_{t|0}(·|x0), and the weights are normalized inside that per-parent batch. For fixed x0, self-normalization estimates E_{p_{t|0}(·|x0)}[w f] / E_{p_{t|0}(·|x0)}[w], not E_{p_{t|0}(·|x0)}[w f]. Since E_{p_{t|0}(·|x0)}[w] is generally not 1 and varies with x0, the batch-averaged loss in Eq. (16) is not an estimator of Eq. (14) even as S grows; the per-parent normalizer is data-dependent through w = pt/pB_t whenever the buffer misrepresents a mode. The paper provides no argument that this denominator cancels or equals one, so Proposition 3's bias and MSE bounds do not apply to the trained objective. This gap is load-bearing because the abstract and Section 3.4 claim that the practical loss is theoretically analyzed.","section":"Section 3.4, Eq. (15); Algorithm 1"},{"comment":"The regularity assumptions behind Propositions 2 and 3 are not satisfied by the Lennard-Jones benchmarks. Proposition 2 assumes ∇E is Lipschitz with bounded norm, while the LJ energy (Appendix F.1, Eq. (33)) has unbounded gradients as the interparticle distance dij tends to zero; the score estimator SL in Eq. (7) contains terms of the form -exp(-E)∇E, so its variance is not uniformly bounded. Assumptions A1 and A2 in Appendix E.3 require uniform bounds on E_{S_L}[||sθ - S_L||^2] and on ||sθ - ∇log pt||, which are not verified and are implausible for LJ-13 and LJ-55. Since Table 2 and the claims in Section 5.2 include these systems, the theoretical analysis does not extend to the full experimental scope. The paper should either restrict its theoretical claims to settings where the assumptions hold or supply a separate argument for singular potentials.","section":"Appendix E.3, Assumptions A1–A2; Appendix F.1"},{"comment":"The sample-reuse strategy for the denominator estimate D_M is not covered by Proposition 3. In Algorithm 1, for a parent x0^(b), the denominator D_M in the weight \tilde w(x_{t|0}^{(b,s)}) sums p_{t|0}(x_{t|0}^{(b,s)} | x0^(j)) over j = 1,...,B including j = b, while the numerator N_K uses fresh samples from N(x0; x_t, σ_t^2 I). Proposition 3 treats the buffer draws for the denominator as independent of the state x_t, but here the state x_t is generated from one of the same buffer draws that appear in the denominator. This coupling can bias the estimated weights in a way that the current bounds do not control, and the paper does not analyze it. Please either sample the denominator from a separate buffer batch or prove that the reuse leaves the bias and MSE bounds unchanged.","section":"Section 3.4, Eq. (11) and Algorithm 1"}],"minor_comments":[{"comment":"The notation {x_{0|t}^{(i)}} is used for samples from N(x0; xt, σ_t^2 I), but the conditional distribution p_{0|t}(x0|xt) is never formally defined in the main text; please introduce it explicitly or use a neutral symbol such as q(x0|xt).","section":"Section 3.3, Eq. (9)"},{"comment":"The DiKL row for LJ-55 reads `∗ ∗ ∗ ∗iDEM` with no separator; this formatting glitch makes the table difficult to parse and should be corrected.","section":"Table 2"},{"comment":"The text alternates between 'SwES' and 'DwES' (Tables 3 and 4, Figure 7, and the surrounding discussion); please use one consistent name for the direct estimated-score sampling baseline.","section":"Appendix H"},{"comment":"Reference [78] contains a typo in the page range ('279–29 Q–learning'); please fix the citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The per-parent versus marginal SNIS normalization mismatch is, in my view, the main obstacle: it breaks the link between Proposition 3 and the implemented loss in Eq. (15). The proposed fix seems implementable—sample the S proposal points from the marginal pB_t rather than from per-parent conditional batches, or provide a new analysis of the per-parent estimator—so I do not recommend outright rejection. The second major issue, the inapplicability of the regularity assumptions to the Lennard-Jones benchmarks, further weakens the paper's theoretical claims. I would also note that no code is provided, so the state-of-the-art empirical claims rest entirely on the text descriptions; a public implementation would strengthen the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Main thing you should know: the paper shows a consistent and sometimes large empirical gain from reweighting the replay-buffer score matching loss, but the theoretical justification does not cover the implemented loss. The algorithm normalizes SNIS weights per parent x0 (Eq. 15, Algorithm 1), while Proposition 3 and Eq. (12)-(13) analyze weights normalized over a batch drawn from the marginal proposal pB_t. Those are different estimators; per-parent normalization estimates E_{p_t|0(·|x0)}[w f]/E_{p_t|0(·|x0)}[w], not E_{p_t|0(·|x0)}[w f]. The denominator is not 1 in general, so the trained loss is not the forward-KL/score-matching objective they claim. This is the central theoretical gap.\n\nWhat is genuinely new: the importance-weighted objective is new relative to iDEM (which uses w=1), and the ablation studies showing SNIS-5 improves both loss and Wasserstein distances are a useful empirical contribution. The method does achieve SOTA on the GMM benchmarks by a clear margin, and on particle systems the gains are modest but consistent. The visualizations are honest about LJ-55's energy shift, though the main text oversells it slightly.\n\nOther soft spots: Proposition 2's bounded-gradient assumption is false for Lennard-Jones potentials, which the appendix itself acknowledges. So the theory only applies to their GMM experiments, where it is also partly moot.\n\nVerdict: worth engaging with, not worth taking at face value. The empirical recipe may be right, but the paper's central claim—that this is a principled forward-KL objective—needs either an algorithm change to match the theory, or a revised theory for the per-parent estimator. I would send it to reviewers with instructions to focus on this gap. If the authors can fix or reframe the theory, it is a solid paper.","headline":"A useful empirical recipe for mode coverage in data-free diffusion samplers, but the theory section analyzes a different estimator than the algorithm actually uses.","tokens_in":35710,"tokens_out":2628,"would_cite":false,"duration_ms":24877,"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":"Data-free diffusion samplers can be trained to cover every mode of a multi-modal target by reweighting the score matching loss with self-normalized importance sampling estimates, turning a forward-KL mode-covering objective into a…","keywords":["importance weighted score matching","diffusion samplers","mode coverage","forward KL divergence","self-normalized importance sampling","data-free sampling","score matching","Lennard-Jones benchmarks"],"falsifier":"On a two-mode Gaussian mixture, compute the importance-weighted score matching loss three ways: with per-parent self-normalization exactly as in Eq. (15), with batch-level self-normalization over the marginal proposal as assumed in Proposition 3, and with a high-sample-count direct Monte Carlo estimate of the true forward-KL objective, then track all three through training. If the per-parent loss converges to a different value or yields different trained scores than the batch-level loss while the batch-level version tracks the true objective, the implemented estimator is not the one the theory analyzes.","tokens_in":34689,"feed_emoji":"🎯","tokens_out":13100,"duration_ms":110356,"temperature":0.7,"pith_summary":"Neural samplers trained from an unnormalized energy function, with no target samples available, tend to optimize reverse-KL objectives and therefore collapse onto a few modes. This paper claims that the objective choice is the root cause, and that targeting the forward KL between diffusion path measures, which penalizes putting zero probability where the target has mass, is the principled fix. The contribution is a method, Importance Weighted Score Matching, that makes such a forward-KL-type objective tractable in the data-free setting by reweighting the score matching loss with self-normalized importance sampling estimates computed from the sampler's own replay buffer and the known energy function. If the claim holds, mode coverage becomes a direct property of the training objective rather than a hoped-for byproduct of heuristics, and data-free samplers could be applied more reliably to multi-modal problems in molecular and statistical physics.","feed_headline":"Importance weighting makes diffusion samplers cover every mode","feed_subtitle":"Zero target samples needed: the reweighted loss targets forward KL and captures all 120 GMM modes.","key_machinery":"The load-bearing object is the importance-weighted score matching objective $\\mathbb{E}_{t}\\mathbb{E}_{x_t\\sim p_t^B}[w(x_t)\\lVert s_\\theta(x_t,t)-\\nabla\\log p_t(x_t)\\rVert^2]$ with $w=p_t/p_t^B$, which converts an expectation under the intractable true marginal into one under an accessible proposal that adaptively improves as the replay buffer is refreshed. Three estimator components carry the argument. First, the score target $S_L(x_t,t)=\\nabla_{x_t}\\log\\sum_{i=1}^{L}\\exp(-E(x_{0|t}^{(i)}))$ with $x_{0|t}^{(i)}\\sim\\mathcal{N}(x_0;x_t,\\sigma_t^2 I)$, whose validity rests on the kernel symmetry $\\mathcal{N}(x_t;x_0,\\sigma_t^2 I)=\\mathcal{N}(x_0;x_t,\\sigma_t^2 I)$ that turns the intractable marginal into a computable expectation. Second, the unnormalized weight estimator $\\tilde{w}=N_K/D_M$, combining averaged energy exponentials in the numerator with averaged forward-kernel evaluations over buffer samples in the denominator, which is proportional to the true weight up to the unknown normalizing constant $Z$. Third, self-normalization of these weights over a batch, which cancels $Z$ and yields the practical loss of Eq. (13) and Eq. (15). The theory rides on Proposition 1 (path forward KL equals score matching), Proposition 2 (bias and variance of $S_L$), and Proposition 3 (bias and MSE of the self-normalized loss).","core_discovery":"The paper's central claim is that the forward path-KL objective $\\mathrm{KL}(\\mathbb{P}_r \\,\\|\\, \\mathbb{P}_\\theta)$ between the true and learned reverse diffusion measures is both desirable and optimizable without target data. Proposition 1 identifies this objective with the score matching loss $\\mathbb{E}_{t\\sim U(0,1)}\\mathbb{E}_{x_t\\sim p_t}[\\lVert s_\\theta(x_t,t)-\\nabla\\log p_t(x_t)\\rVert^2]$, and the method then rewrites the expectation under the intractable true marginal $p_t$ as an expectation under the sampler-induced proposal $p_t^B$ with importance weight $w(x_t)=p_t(x_t)/p_t^B(x_t)$. The weight is estimated by Monte Carlo: the numerator uses the Gaussian symmetry of the variance-exploding (VE-SDE) kernel to write the unnormalized marginal as an expectation over $x_0\\sim\\mathcal{N}(x_0;x_t,\\sigma_t^2 I)$ of $\\exp(-E(x_0))$, and the denominator averages the forward kernel over replay-buffer samples; self-normalization over a batch then cancels the unknown normalizing constant. On these estimates the paper proves a score-estimator bias and variance of order $1/\\sqrt{L}$ (Proposition 2) and bias and mean-squared-error bounds for the self-normalized loss that shrink with the importance-weight variability and with the score-estimator sample size (Proposition 3). The paper reports that the resulting sampler matches or beats the compared baselines on every distributional metric across GMM-40/80/120, DW-4, LJ-13, and LJ-55, and that the reweighting itself, not the base score matching objective, is what drives the improvement.","pith_inferences":["Whether the per-parent normalization used in Eq. (15) is equivalent to the batch-level normalization the theory analyzes is an empirical question; a controlled comparison of the two on a low-mode GMM would show whether the implemented estimator matches the one whose bias and variance are bounded.","The importance weights themselves could serve as a diagnostic: tracking their variability over training would measure how far the replay-buffer proposal has drifted from the true marginal, giving a principled mode-collapse monitor and stopping rule that the paper does not propose.","The forward-KL mechanism suggests a stress test the paper does not run: freeze the replay buffer mid-training; if coverage relies on the adaptive proposal as the argument claims, a stale buffer should degrade performance toward the unweighted baseline, isolating the importance weighting as the active ingredient.","Because the objective is agnostic to the sampler architecture, the same importance weighting could attach to other score-based estimators, for instance proposals built from annealed or sequentially refined densities, extending mode coverage to settings where maintaining one replay buffer is impractical."],"forward_implications":["Mode coverage becomes a property of the training objective itself: regions of the target that the current sampler under-represents receive the largest importance weights and therefore the strongest gradient pressure, so training automatically concentrates on the modes it is missing.","The reported benchmarks imply the approach scales: 120-mode Gaussian mixtures and the 165-dimensional LJ-55 cluster are handled without the divergence or breakdown the compared baselines suffer at that complexity.","Small SNIS sample counts suffice, since the ablations show $S=5$ to $S=10$ essentially saturates the benefit at modest training-time cost, so the correction is cheap to add to an existing score-based sampler.","Direct sampling with the estimated score alone is shown in the paper's Appendix H to be orders of magnitude slower at matched quality, so amortized learning plus importance weighting, not a larger Monte Carlo batch, is the practical route to accurate sampling.","If Proposition 3's bounds are taken at face value, the bias and variance of the trained loss are controlled by the importance-weight variability, giving a concrete lever for future work: improve the proposal distribution and the bound tightens."],"supporting_citations":[{"why":"Establishes the path-measure KL and score-matching equivalence used in Proposition 1 as the objective's foundation.","marker":"[69]"},{"why":"Contributes the Monte Carlo score estimator $S_L$ of Eq. (7) and the replay-buffer, iterated denoising structure the method extends.","marker":"[2]"},{"why":"Supplies Theorem 2.1, the self-normalized importance sampling bias and MSE bounds used in the proof of Proposition 3.","marker":"[1]"},{"why":"Source of the self-normalized importance sampling estimator and its known finite-sample bias.","marker":"[57]"},{"why":"The strongest mode-coverage baseline; its $\\alpha$-divergence plus annealed importance sampling strategy is the main alternative the method compares against.","marker":"[47]"},{"why":"Baseline that also targets coverage but through a diffusive reverse KL; its one-step generator limits are contrasted with score-based diffusion samplers.","marker":"[26]"},{"why":"Stochastic-optimal-control baseline whose mode-collapse behavior on GMMs motivates the forward-KL objective.","marker":"[83]"}],"fun_headline_variants":["Reweighted score matching gives diffusion samplers full mode coverage","No target samples? Importance weighting still covers all modes","Forward KL without target data: importance-weighted score matching","Importance weighting unlocks every mode for diffusion samplers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training loss normalizes the estimated importance weights separately for each base sample drawn from the replay buffer, while the theory's bias and variance bounds are derived for normalization over a full batch of proposal samples; the paper offers no proof that per-base-sample normalization preserves the weighted expectation that the objective is claimed to estimate.","fun_headline_variants_meta":{"raw":{"variants":["Reweighted score matching gives diffusion samplers full mode coverage","No target samples? Importance weighting still covers all modes","Forward KL without target data: importance-weighted score matching","Importance weighting unlocks every mode for diffusion samplers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000802,"raw_usage":{"total_tokens":3634,"prompt_tokens":1165,"completion_tokens":2469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":781,"completion_tokens_details":{"reasoning_tokens":2402}},"tokens_in":781,"tokens_out":2469,"duration_ms":12672,"temperature":1.0,"reasoning_tokens":2402,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:14:37.883566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a two-mode Gaussian mixture, compute the importance-weighted score matching loss three ways: with per-parent self-normalization exactly as in Eq. (15), with batch-level self-normalization over the marginal proposal as assumed in Proposition 3, and with a high-sample-count direct Monte Carlo estimate of the true forward-KL objective, then track all three through training. If the per-parent loss converges to a different value or yields different trained scores than the batch-level loss while the batch-level version tracks the true objective, the implemented estimator is not the one the theory analyzes.","supporting_citations":[{"cited_title":"Monte Carlo theory, methods and examples, volume 1","cited_arxiv_id":null,"evidence_quote":"Source of the self-normalized importance sampling estimator and its known finite-sample bias."},{"cited_title":"Flow annealed importance sampling bootstrap","cited_arxiv_id":null,"evidence_quote":"The strongest mode-coverage baseline; its $\\alpha$-divergence plus annealed importance sampling strategy is the main alternative the method compares against."},{"cited_title":"Path integral sampler: a stochastic control approach for sampling","cited_arxiv_id":null,"evidence_quote":"Stochastic-optimal-control baseline whose mode-collapse behavior on GMMs motivates the forward-KL objective."}],"review_version":1}