{"id":"40fed3b1-da36-41ee-a6b3-eecd6966f8f2","arxiv_id":"2602.03729","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Off-policy log-dispersion regularization—generalizing the log-variance objective—improves Boltzmann generator training across unbiased, biased, and variational settings, with gains of up to 10× in sample efficiency.","lead":"This paper introduces a new training regularizer, log-dispersion regularization (LDR), that boosts the data efficiency of Boltzmann generators by exploiting target energy values already available in simulation datasets. Across molecular benchmarks it improves final accuracy and sample efficiency, in some settings matching methods that use ten times more data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CMT+LDR lacks theoretical cover: at each annealing step the data term targets intermediate qi while LDR targets pX, so the combined objective's optimum is not defined by Prop. A.3.","rationale":"The reader's weakest assumption was overlap between the reference distribution and the target. My concern is related but distinct: even with good overlap, the theoretical consistency argument (Prop. A.3) does not apply to the CMT setting because the data-based loss at intermediate annealing steps is a divergence to qi, not to pX. This leaves the variational efficiency claim without a formal basis, though the empirical results are strong and the reader already assigned CONDITIONAL. I therefore do not move the verdict; I only add a specific technical condition that should be addressed. The proposed test directly checks whether LDR distorts the annealing path, which would determine whether the missing theory is a practical problem or merely an exposition gap.","tokens_in":33030,"tokens_out":11230,"duration_ms":131206,"concrete_test":"Run CMT with and without LDR on alanine dipeptide at 1e7 target evaluations and, at each annealing step, estimate the KL divergence (or ESS) between the current q_theta and the intended intermediate qi using importance sampling on the buffer. If LDR causes q_theta to deviate from qi substantially more than vanilla CMT during the first half of the schedule, the regularizer is corrupting the annealing path; then restrict LDR to only the final annealing steps (or to steps where qi is close to pX) and check whether the 10x gains persist. If the deviation is negligible, or the late-only variant retains the gains, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Prop. A.3 proves consistency only when the data term alone has pX as its unique minimizer. In the variational CMT setting (Section 4.4), the data-based loss at annealing step i is the forward KL to the intermediate distribution qi (Algorithm 1), so its zero is qi, not pX. The LDR term L_LD^(p) uses f_theta = -log q_theta - E/kBT with the buffer over qi as reference; its zero would require q_theta proportional to exp(-E/kBT) = pX on the buffer support. Thus the combined objective (Eq. 5) has no common zero at intermediate annealing steps; its optimum is an unspecified compromise that can pull q_theta away from the prescribed qi and, in principle, violate the trust-region overlap on which CMT relies. The paper offers no analysis of this interaction; the statement that CMT 'enforces sufficient overlap' explains stability of the importance weights but does not restore the consistency argument. The headline 10x efficiency gain on CMT therefore rests on an unproven dynamical assumption, not on the paper's theoretical framework.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes off-policy log-dispersion regularization (LDR) for training Boltzmann generators. The regularizer is L_LD^(p)=E_{r_X}[|f_theta - E_{r_X}[f_theta]|^p], with f_theta = -log q_theta - E/k_BT, added to a standard data-based objective (Eq. 5). LDR is evaluated on fixed/off-policy reference distributions and requires only energy labels already available in the dataset. The authors prove basic consistency for a combined objective when the data term alone has p_X as its unique minimizer (Prop. A.3), characterize gradient behavior (Prop. A.4), and empirically test LDR on a GMM, alanine dipeptide/hexapeptide in internal coordinates, biased-data refinement, Cartesian-coordinate training, and variational CMT training. Across these settings LDR improves NLL/ESS and sample efficiency, with claimed gains up to 10x in biased and variational settings.","tokens_in":33234,"tokens_out":7265,"duration_ms":73164,"significance":"If the findings hold, the contribution is practically valuable: a simple, broadly applicable regularizer that exploits energy labels already computed during dataset generation, with consistent improvements across multiple training regimes. The empirical study is thorough: seed-averaged results with standard deviations, significance testing, a robustness sweep over lambda_data (Fig. 4), additional KL/TICA/Wasserstein metrics in the appendix, and a new correction for the Cartesian augmentation scheme. The theoretical propositions are elementary but correct as stated. However, the central consistency theorem is proved only for data terms whose unique minimizer is the final target p_X; this does not cover the annealing-based CMT setting, where the data term at each step targets an intermediate q_i. The headline 10x efficiency claim for CMT therefore rests on an unproven dynamical assumption, and the CMT baseline itself is a simplified variant of the method it is claimed to improve.","major_comments":[{"comment":"The consistency result Prop. A.3 assumes L_data is a divergence whose unique minimizer is p_X. In CMT annealing, however, the data term at step i is the forward KL to the intermediate distribution q_i (Algorithm 1, App. B.2), so its unique minimizer is q_i, not p_X. The LDR term uses f_theta = -log q_theta - E/k_BT and is evaluated on a buffer associated with q_i; on that support, its minimizer is q_theta proportional to exp(-E/k_BT) = p_X. Thus the combined objective (Eq. 5) has no common zero at intermediate annealing steps, and Prop. A.3 does not apply. The statement in Sec. 4.4 that CMT's trust-region constraint 'enforces sufficient overlap' addresses the variance of importance weights, not the consistency of the combined objective. Since the claimed 10x efficiency gain on CMT is a headline result, this gap should be addressed -- either by providing an analysis of the intermediate-st","section":"Sec. 4.4 / Algorithm 1 / Prop. A.3"},{"comment":"The CMT baseline used in the variational experiments is not the original CMT algorithm: the authors discard the entropy constraint and adopt a manual geometric temperature schedule, writing that 'CMT is very similar to TA-BG with an additional trust-region constraint.' Since the contribution is framed as improving 'the current state-of-the-art method, CMT,' the comparison should include the original CMT (or a clear justification that the simplified variant is representative). Otherwise the 10x efficiency improvement may be relative to a weaker baseline than the actual state of the art.","section":"Sec. 4.4 / App. B.6"}],"minor_comments":[{"comment":"To report best-case LDR performance, lambda_data was tuned per experiment on a grid while fixing lambda_LD=1. Although Fig. 4 shows robustness on one system, a fixed lambda_data across all benchmarks would strengthen the claim that LDR is a plug-in regularizer. Please state whether results are sensitive to per-system tuning.","section":"App. B.6"},{"comment":"The reproducibility statement says code and data 'will soon be made publicly available.' For an archival version, please provide a repository link or an explicit release plan, since the empirical claims rely on a substantial experimental pipeline.","section":"Reproducibility Statement"},{"comment":"The phrase 'Details can be found in Appendix 6' appears to be an error; the biased-dataset generation is described in Appendix B.4.","section":"Sec. 4.3"},{"comment":"The sensitivity plot tunes the learning rate separately for each lambda_data value. This is reasonable, but the caption should state this explicitly, as it affects the interpretation of the flat robustness curve.","section":"Fig. 4 caption / App. D.1"},{"comment":"The sentence 'For 1e6, the model likely remains too far from the target, leading to high-variance importance weights and a noisy LD objective' is speculative. Consider supporting it with the reported metrics or removing the causal claim.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the CMT section: both the theoretical consistency claim and the baseline fidelity need work. If the authors can either extend the analysis to the annealing setting or reframe the CMT results as empirical, and include a comparison against the original CMT implementation, the paper would be close to publishable. The data-based experiments are solid and the proposed regularizer is a useful idea."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is genuinely new: taking the log-variance objective, generalizing it to log-dispersion, and using it off-policy as a shape regularizer on fixed energy-labeled datasets, on top of standard data-based divergences. That is a clean idea, and the paper executes it carefully. The theory (Props A.1–A.3) is elementary but correct for the setting it covers, and the experiments are unusually thorough: seed-averaged error bars, a robustness sweep over λdata, multiple coordinate representations, and a range of metrics in the appendix. I also credit the authors for stating the key limitation plainly — LDR alone is not a divergence without full support, and references far from the target can destabilize training. The empirical gains across unbiased, biased, and variational regimes are convincing enough to take seriously.\n\nThe soft spots are real but not fatal. The biggest one is the stress-test concern about the CMT experiments. Prop A.3 proves consistency only when the data term alone has pX as its unique minimizer. In the CMT annealing loop, the data term targets the intermediate qi, not pX, so the combined objective has no common zero at intermediate steps. The paper does not analyze this interaction; the 10x efficiency claim on CMT rests on an unproven dynamical assumption, plus the informal argument that CMT's trust-region overlap stabilizes the importance weights. That may well be true, but it is not the paper's theoretical framework.\n\nTwo smaller issues. First, λdata is tuned per experiment to report best-case performance; the sensitivity plot helps, but a fixed default or a clearer protocol would be better. Second, the abstract's claim that LDR improves both NLL and ESS across all benchmarks is contradicted by the hexapeptide 1e6-sample result for LDR-L2, where NLL is slightly worse than the baseline. The biased-data efficiency comparison also relies on interpolation between IS sample counts rather than a matched measurement — minor, but worth tightening.\n\nNone of this undermines the core contribution. The paper is honest about its limits, the main idea is reproducible in principle, and the experiments are among the more careful I've seen in this area. It deserves a serious referee, and I would send it to review with a request for code and data, plus a discussion of the CMT interaction. If the authors can close that theoretical gap or at least frame it honestly, this is a solid contribution to the Boltzmann generator literature.","headline":"A simple, well-tested regularizer that plugs into Boltzmann generator training and gives real data-efficiency gains, though the theory doesn't cover the strongest variational result and a few reporting choices overstate the abstract.","tokens_in":33821,"tokens_out":2402,"would_cite":true,"duration_ms":26674,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65C05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Boltzmann generator training can be made up to ten times more data-efficient by adding an off-policy log-dispersion regularizer to any standard training objective.","keywords":["Boltzmann generators","normalizing flows","log-dispersion regularization","log-variance objective","data efficiency","sampling from unnormalized densities","energy-based regularization","molecular dynamics"],"falsifier":"Run LDR on a fixed dataset drawn from a distribution with deliberately poor overlap with the target (e.g., a short MD trajectory trapped in one metastable basin, without any importance-sampling correction) and measure the effective sample size and NLL on a held-out target sample; if the model's performance degrades below the unregularized baseline, the claim of universal improvement fails. A more quantitative version: vary the bias of the reference distribution and show that performance monotonically worsens as overlap decreases.","tokens_in":1579,"feed_emoji":"⚛️","tokens_out":3231,"duration_ms":67068,"temperature":0.7,"pith_summary":"Boltzmann generators are generative models that produce independent samples from the equilibrium (Boltzmann) distribution of a physical system, but training them is expensive because both simulation data and target-energy evaluations cost time. This paper claims that adding one regularizer—the log-dispersion objective—to whatever data-based objective is already used cuts those costs dramatically. The regularizer penalizes the spread of the log-importance-weight function over a fixed reference dataset, using only the energy labels that are already computed when the dataset is built. Because it runs off-policy on existing data, it adds no extra energy evaluations and no on-policy samples. The authors prove that the combined objective still converges to the true target, and they report up to an order-of-magnitude gains in sample or energy-evaluation efficiency across unbiased, biased, and purely variational settings.","feed_headline":"Boltzmann generators get up to 10x more data-efficient","feed_subtitle":"One added term—the log-dispersion regularizer—uses energy labels already in the dataset to cut sample budgets.","key_machinery":"The central object is the log-dispersion objective, a generalization of the log-variance divergence: Lθ(p)_LD = E_{rX}[ |fθ − E_{rX}[fθ]|^p ], p ≥ 1, where fθ(x) = −log qθ_X(x) − (−log p̃_X(x)) is the unnormalized log importance weight of the model relative to the target. At the optimum qθ_X = p_X, fθ is constant and the term vanishes. The key move is to use it off-policy as a regularizer on top of a standard data-based objective: the data term supplies full-support normalization and correct convergence, while LDR uses energy labels to shape the proposal distribution where the reference data lives. The p = 1 and p = 2 variants are both tested, with the authors noting p = 1 is more outlier-ro","core_discovery":"The central claim is that minimizing the p-th central moment of the unnormalized log importance weights fθ(x) = −log qθ_X(x) − (−log p̃_X(x)) around their mean over a reference distribution rX—for any p ≥ 1—acts as a shape regularizer that aligns the learned proposal with the target energy landscape, and that this regularizer can be safely deployed off-policy on fixed datasets. Formally, Lθ(p)_LD = E_{rX}[ |fθ − E_{rX}[fθ]|^p ]. The paper shows (Propositions A.1–A.3) that LDR alone is not a divergence when the reference has limited support, but that adding it to a data-based divergence yields a consistent objective whose unique optimum is still the target distribution. Across experiments on","pith_inferences":["If the mechanism is shape alignment via log-importance-weight dispersion, LDR should transfer to any likelihood-based model class, including diffusion samplers trained over joint diffusion paths; the paper mentions this as out of scope, but the theory over arbitrary reference distributions suggests it.","The overlap requirement implies LDR is most effective when the reference and target already overlap; a natural heuristic that follows is to anneal or reweight the reference (as CMT's trust-region does) before applying LDR.","The gradient analysis suggests p = 1 and p > 1 have complementary regimes: L1 keeps a constant-magnitude gradient far from the optimum but introduces persistent noise near it, while p > 1 damps gradients near the optimum; an adaptive or scheduled p could get the best of both.","Because LDR needs only energy labels, it could be applied with approximate or surrogate energies as long as the labels correlate with the true Boltzmann weight, extending its reach to systems where exact energies are unavailable."],"forward_implications":["On unbiased equilibrium datasets, LDR with 1 million samples matches or beats unregularized training with 5 million samples on alanine dipeptide, a 5× data-efficiency gain.","On biased simulation data, LDR on importance-sampled refinements and the original biased set reaches the same performance as unregularized refinement with roughly 10× fewer IS samples.","In purely variational training without target samples, adding LDR to Constrained Mass Transport (CMT) matches its performance with 10× fewer target energy evaluations on alanine dipeptide and prevents mode collapse on alanine hexapeptide.","LDR transfers to Cartesian-coordinate autoregressive flows, improving effective sample size from 27% to 36% on alanine dipeptide.","The regularizer adds no extra energy evaluations and no on-policy samples, so it can be added to existing pipelines at no computational cost."],"fun_headline_variants":["One regularizer cuts training samples 10x for Boltzmann generators","Log-dispersion regularizer boosts Boltzmann generator efficiency","Off-policy trick makes Boltzmann generators 10x sample-efficient","Energy labels already in hand? Use them for 10x faster training","New regularization: train Boltzmann generators with far less data"],"cache_read_input_tokens":35072,"weakest_assumption_plain":"The regularizer only works well when the reference distribution over which the dispersion is measured significantly overlaps the target; if the reference is far from the target, LDR can destabilize training instead of helping—the authors state this limitation explicitly.","fun_headline_variants_meta":{"raw":{"variants":["One regularizer cuts training samples 10x for Boltzmann generators","Log-dispersion regularizer boosts Boltzmann generator efficiency","Off-policy trick makes Boltzmann generators 10x sample-efficient","Energy labels already in hand? Use them for 10x faster training","New regularization: train Boltzmann generators with far less data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2540,"prompt_tokens":733,"completion_tokens":1807,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":1737}},"tokens_in":477,"tokens_out":1807,"duration_ms":12649,"temperature":1.0,"reasoning_tokens":1737,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T04:51:28.090923+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run LDR on a fixed dataset drawn from a distribution with deliberately poor overlap with the target (e.g., a short MD trajectory trapped in one metastable basin, without any importance-sampling correction) and measure the effective sample size and NLL on a held-out target sample; if the model's performance degrades below the unregularized baseline, the claim of universal improvement fails. A more quantitative version: vary the bias of the reference distribution and show that performance monotonically worsens as overlap decreases.","supporting_citations":[],"review_version":1}