{"id":"86f502ed-7348-43ed-a8a3-8c529354e277","arxiv_id":"2501.15522","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DASTR adaptively samples points from |∇q|² e^{-βV} with a normalizing flow, producing more accurate neural committor approximations on the tested high-dimensional problems.","lead":"DASTR is a training scheme that uses a deep generative model to create extra training points near the transition state when learning committor functions with neural networks. It matters because committor functions are central to rare-event simulations in chemistry and biology, and transition-region samples are normally scarce and expensive to obtain.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (11) equates densities on latent CVs with densities on atomic coordinates with no Jacobian correction; this leaves the autoencoder variant's stated target distribution unjustified, though the heuristic may still work empirically.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: Eq. (11) treats a pullback of a density under a dimension-reducing map as if it were the density itself, with no Jacobian or many-to-one corrections. My reading agrees and sharpens the consequence: Eq. (14) optimizes a cross-entropy whose target is not the pushforward of p_{V,q}, and Algorithm 2 then decodes from a KRnet that has learned a different distribution. This does not make the synthetic 10D/20D results invalid, because those experiments do not use the latent-space construction. It does, however, leave the alanine dipeptide latent-CV results without a correct mathematical explanation of the adaptive-sampling mechanism. The paper still contains a plausible heuristic pipeline, and the empirical evidence in Section 4.3.2 could be reproducible even if the stated density identification is false. Therefore the reader's conditional verdict is unchanged: the paper should not be accepted without either fixing the density identification, supplying a corrected derivation, or presenting a direct empirical check that the missing Jacobian does not change the outcome. No ad hominem or fatal-sounding language is needed; the issue is a concrete, testable gap between Eq. (11) and the probability calculus for dimension-reducing encoders.","tokens_in":23158,"tokens_out":4156,"duration_ms":43262,"concrete_test":"Run a controlled version of Algorithm 2 on the 10D rugged Mueller potential, where a reference committor is available. Train a deterministic autoencoder on metadynamics-like samples, then compare two variants: (A) Algorithm 2 exactly as written, using Eq. (14) with p_{V,q}(s(x_i)) = p_{V,q}(x_i); and (B) a corrected variant that targets the pushforward density (p_{V,q})_#(s), using the decoder Jacobian in the likelihood of decoded samples and the proper importance weights in Eq. (10). Compare the final L2 relative error against the reference and the histogram of decoded samples near the transition region. If variant B changes the error by more than the run-to-run standard deviation reported in Table 1, Eq. (11) is materially wrong for the claimed mechanism; if the errors agree within noise, the heuristic survives despite the missing Jacobian.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 introduces p_{V,q}(s(x)) = p_{V,q}(x) in Eq. (11). Since p_{V,q} in Eqs. (6)-(7) is a density with respect to Lebesgue measure on x in R^d, and s is a many-to-one encoder R^d -> R^m with m << d, the object p_{V,q}(s(x)) is not a density in the latent variable. The correct object is the pushforward density (p_{V,q})_#(s) = integral over {x : s(x)=s} of p_{V,q}(x) / sqrt(det(J_s J_s^T)) dσ(x), which includes a Jacobian factor and accounts for the fiber of the encoder. Equation (14) instead uses p_{V,q}(s(x_i)) / p_{KRnet}(s(x_i)) in the cross-entropy loss for KRnet, and Algorithm 2 decodes KRnet samples through the pretrained autoencoder. The variational problem actually solved by KRnet is therefore not the intended one: the decoded samples are not drawn from a distribution proportional to |∇q|^2 e^{-βV} on the molecular configuration space, and the importance weights in Eq. (10) omit both the encoder pushforward and the decoder Jacobian. The reported 97% physical-validity rate and 95% sampling-time reduction may still hold as empirical facts, but the stated mechanism for the latent-space DASTR variant is not justified by the equations as written. This is the most load-bearing weakness for the paper's realistic-system claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DASTR, an iterative deep adaptive sampling method for training neural-network approximations of the committor function. Starting from a standard variational formulation, DASTR defines a sampling density proportional to |∇q_θ|^2 e^{-βV}, approximates it with a normalizing flow (KRnet), and uses newly generated samples to refine the training set. For high-dimensional molecular systems the algorithm is applied in a reduced latent space, either with hand-picked collective variables combined with umbrella sampling or with autoencoder-learned latent variables decoded back to atomic coordinates. Numerical experiments cover a 10D rugged Mueller potential, a 20D Brownian-motion problem, and alanine dipeptide (d=66), reporting improved L2 errors relative to SDE, uniform, and artificial-temperature baselines, as well as a 97% molecular-validity rate and more than 95% sampling-time reduction for the autoencoder variant.","tokens_in":23537,"tokens_out":7702,"duration_ms":72105,"significance":"If the claims hold, DASTR addresses a real bottleneck in neural committor estimation: the scarcity of transition-region data. The paper has clear strengths: the variational loss is standard and correctly derived; the synthetic benchmarks use external reference solutions (FEniCS and an analytic formula); implementation details in Appendix B are unusually complete; and the KRnet sampling procedure is fully specified. The comparison of sampling-time in Table 3 is a useful practical metric. The main technical risk is the latent-space density identification in Section 3.4, which is load-bearing for the alanine-dipeptide claim. The evaluation on the 1/2-isosurface is informative but not sufficient to establish global accuracy. The absence of comparisons with the closest adaptive committor methods [11,14] limits the strength of the claimed improvement. These issues are fixable, so the correct disposition is major revision rather than rejection.","major_comments":[{"comment":"The identity p_{V,q}(s(x)) = p_{V,q}(x) is not a valid density transformation. Since the encoder s: R^d -> R^m is many-to-one with m < d, the density induced by p_{V,q} on the latent coordinates is the pushforward ρ(s) = ∫_{x: s(x)=s} p_{V,q}(x) (det(J_s(x) J_s(x)^T))^{-1/2} dσ(x), not p_{V,q}(x) evaluated at a single preimage. As a result, Eq. (14) and Algorithm 2 train the latent KRnet on an inconsistent objective, and the decoded samples are not, in general, samples from a distribution proportional to |∇q|^2 e^{-βV}. The reported alanine-dipeptide results may still hold empirically, but the stated mechanism for the autoencoder variant is not justified by the equations as written. Please derive the correct pushforward and the associated importance weights, or explicitly reformulate the latent-space objective as an approximation and validate it, for example by comparing weighted moments or effective sample sizes against the intended target.","section":"Section 3.4, Eq. (11)"},{"comment":"The realistic-system evaluation rests entirely on the 1/2-isosurface of q_θ: test points are selected by the condition |q_θ(x) - 0.5| ≤ 5 × 10^{-5}, and their committor values are then estimated by short SDE simulations. This is a necessary condition but not a sufficient one: a systematically biased q_θ can have a level set on which the empirical PB concentrates near 0.5 while the function is inaccurate elsewhere, and the filtering step uses q_θ itself to select the test points. To support the claim that the latent autoencoder variant provides a good approximation of the committor, the authors should add a more global check, such as evaluating PB on multiple isosurfaces (for example 0.3 and 0.7) with points selected independently of q_θ, or comparing against a reference obtained by an independent method.","section":"Section 4.3, Figures 11 and 15"},{"comment":"The closest adaptive committor methods — active importance sampling for variational objectives [11] and deep adaptive committor sampling [14], and arguably the committor-to-committor method [13] — are cited but never used as baselines. The numerical comparisons are limited to SDE sampling, uniform sampling, artificial temperature, and umbrella/metadynamics. Since the paper's central claim concerns the advantage of adaptive sampling over non-adaptive and potential-modification strategies, at least one direct numerical comparison with [14] or [11] is needed to calibrate the significance of the reported gains. Without such a comparison, the improvement over 'baseline methods' is not benchmarked against the state of the art.","section":"Section 1.1 and Section 4"}],"minor_comments":[{"comment":"Typographical errors such as 'commttor', 'standaredd-dimensional', and 'committer' should be corrected.","section":"Throughout"},{"comment":"The proposal density is written as p_IS(s(x_i)) = e^{-βV_modified(x_i)}, which is not a normalized density on the collective variable s. If unnormalized weights are intentional because constant normalization factors cancel in the cross-entropy minimization, this should be stated explicitly.","section":"Eq. (12)"},{"comment":"In the DASTR row with 20 neurons and |S| = 5 × 10^3, the reported mean is 0.0266 with standard deviation 0.1396, so the standard deviation is larger than the mean. This indicates high variance across runs and should be discussed or addressed with more seeds.","section":"Table 2"},{"comment":"The support or bounding box of the KRnet in the latent space is not specified; since the decoder may map out-of-support latent points to unphysical states, a sentence on how the latent range is determined would improve reproducibility.","section":"Section 4.3.2"}],"recommendation":"major_revision","confidential_remarks":"The original-space DASTR empirical story is strong and the code-level details are unusually complete. The main obstacle is the latent-space density derivation in Section 3.4: Eq. (11) is incorrect as a change of variables and the fix requires either a new derivation or an explicit reframing of the autoencoder variant as a heuristic with additional validation. Given that the core idea is sound and the issue is local, major revision is more appropriate than rejection. I would also ask the authors to include a comparison with Ref. [14] before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"DASTR is a plausible and mostly well-executed adaptive sampling method for neural committor estimation. The new part is the q-dependent sampling target |∇q|^2 e^{-βV}, the KRnet resampling loop, and the autoencoder latent-space variant. That combination is not in the cited papers verbatim, and the synthetic evidence is genuinely good: across the 10D rugged Mueller and 20D Brownian tests, DASTR consistently beats the SDE, uniform, and artificial-temperature baselines, often by an order of magnitude, with multiple runs and different network widths. The alanine dipeptide study is also a real attempt at a realistic system, and the 97% physically valid configurations and >95% sampling-time reduction are striking empirical claims. I also credit the authors for validating the 1/2-isosurface with independent SDE hitting probabilities rather than relying only on the training loss.\n\nThe soft spots are real. The most load-bearing one is Eq. (11): p_{V,q}(s(x)) = p_{V,q}(x) is not a valid density identity for a many-to-one encoder from R^d to R^m with m << d. The correct latent density is a pushforward with a Jacobian factor and an integral over the encoder fiber. Because Eq. (14) and Algorithm 2 use this identification, the KRnet in latent space is not solving the stated target distribution, and the decoded samples are not provably drawn from anything proportional to |∇q|^2 e^{-βV} on configuration space. The empirical results might still hold if the heuristic works, but the stated mechanism for the realistic-system claim is not justified as written. This needs to be fixed or reframed.\n\nTwo more issues. The closest adaptive committor solvers—Rotskoff et al. [11] and Lin and Ren [14]—are cited but not benchmarked, which weakens the novelty claim for the method. And there is no code or data release, so the KRnet and molecular dynamics pipeline details can't be checked; that's partly why I'd keep confidence at moderate rather than high. The validation procedure uses qθ to pick isosurface points, but because the SDE simulations provide an external check, I don't see that as a fatal circularity.\n\nWho is this for? Researchers working on neural committor functions and adaptive sampling for rare events will get value from the synthetic results and the algorithmic recipe. It deserves a serious referee. I'd send it to review with the expectation of major revision, not desk reject, but I'd ask the authors to address Eq. (11), add the missing baselines, and release code and data before acceptance.","headline":"Solid adaptive sampling scheme with convincing synthetic gains, but the latent-space variant's stated target distribution is unjustified and needs work before the molecular claims stand.","tokens_in":24032,"tokens_out":4014,"would_cite":true,"duration_ms":33421,"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":"DASTR breaks the transition-data bottleneck in neural committor estimation by sampling from the gradient-weighted Boltzmann density, and its autoencoder variant makes the scheme usable on molecules.","keywords":["committor function","deep adaptive sampling","rare event","transition path","KRnet","autoencoder","collective variables","normalizing flows"],"falsifier":"Replace the latent-space identification with the exact change-of-variables formula, multiplying the target density by the encoder Jacobian determinant, and compare the decoded samples and committor error on alanine dipeptide against the reported results; if the corrected sampler does not match or improve the accuracy, the stated mechanism is not what produces the gains. A cheaper check is to compare histograms of $|\\nabla q_\\theta|^2 e^{-\\beta V}$ at decoded versus directly sampled configurations.","tokens_in":22947,"feed_emoji":"🎯","tokens_out":11711,"duration_ms":102703,"temperature":0.7,"pith_summary":"The paper addresses a bottleneck in neural committor estimation: transition-state configurations are rarely visited by straightforward molecular dynamics, so training sets are starved of exactly the points that matter. It proposes DASTR (Deep Adaptive Sampling for TRansition paths), an iterative scheme in which the current committor network's gradient magnitude times the Boltzmann weight, $|\\nabla q_\\theta(x)|^2 e^{-\\beta V(x)}$, is treated as an unnormalized density and approximated by a normalizing-flow model whose new samples are added to the training set. In the 10- and 20-dimensional benchmark problems the resulting estimators are reported to be up to an order of magnitude more accurate than sampling from biased dynamics at equal sample counts. For the alanine dipeptide system, running DASTR on autoencoder-learned latent variables yields molecular configurations that pass a 150 kJ/mol energy filter over 97% of the time and cuts molecular sampling time to under 4% of the umbrella-sampling alternative. If these claims hold, DASTR gives a general route to high-quality transition data for committor networks without hand-picked collective variables or repeated SDE simulation.","feed_headline":"Adaptive sampler makes committor nets 10 times more accurate","feed_subtitle":"DASTR targets the transition region directly; its autoencoder variant keeps molecules physically valid.","key_machinery":"The load-bearing object is the target density $p_{V,q}(x)=|\\nabla q_\\theta(x)|^2 e^{-\\beta V(x)}/C$ (or with a bias potential added), which assigns more weight where the committor changes fastest, that is, in the transition-state region. The second object is a normalizing-flow density model (the paper uses KRnet) that approximates this target by minimizing cross-entropy with importance weights and generates new training points through the inverse flow. The third is a trained autoencoder whose latent variables serve as automatically selected collective variables, so that flow samples can be decoded into molecular geometries and then filtered by an energy threshold. Together these pieces turn the variational loss into an adaptive importance-sampling loop: train $q_\\theta$, locate where its gradient is large, generate fresh samples there, retrain.","core_discovery":"The committor $q(x)$ is the probability that a trajectory starting at $x$ reaches $B$ before $A$; it solves a high-dimensional linear PDE. DASTR's central move is to define an adaptive target $p_{V,q}(x) \\propto |\\nabla q_\\theta(x)|^2 e^{-\\beta V(x)}$, optionally with a bias potential that lowers the barrier, approximate this target by a normalizing-flow model, draw new training points from it, retrain $q_\\theta$, and repeat. Because the gradient of the committor concentrates near the transition state, the generated points land preferentially on rare transition paths. For realistic molecular systems the same loop is run on low-dimensional latent collective variables: an autoencoder maps atomic coordinates to latent variables, the flow samples in latent space, and a decoder reconstructs molecular configurations, which are then filtered by potential energy. The paper reports that this latent-space DASTR produces physically valid alanine dipeptide configurations with high probability while reducing molecular sampling time by more than 95% relative to DASTR combined with umbrella sampling.","pith_inferences":["The paper's identification $p_{V,q}(s(x))=p_{V,q}(x)$ in Eq. (11) drops the encoder's Jacobian and many-to-one structure; a corrected target that accounts for the change of variables is a natural extension and could be tested on the same benchmarks.","Because the sampling target is driven by the current network's gradient, DASTR is a form of self-training; a stopping rule or convergence analysis based on successive changes of the flow target would clarify when the iterative loop has extracted the available transition information.","The energy threshold is applied as a post-processing filter rather than being built into the sampling density; including a truncated or reweighted target might raise the fraction of usable samples further.","The same 'gradient as density' recipe applies to any variational loss with a nonnegative integrand, so a similar adaptive collocation principle should transfer to other high-dimensional PDE solvers beyond committor estimation."],"forward_implications":["At fixed sample size, DASTR reduces the $L^2$ error of the estimated committor by roughly an order of magnitude on the rugged Mueller test, with faster error decay as the sample budget grows.","The autoencoder variant removes the need to hand-pick collective variables and avoids SDE-based umbrella sampling for generating transition-state molecules, making the pipeline more general for molecular systems.","In the alanine dipeptide test, committor values on the 0.5 level set produced by DASTR cluster around 0.5, whereas a network trained on metadynamics samples alone spreads widely, indicating better resolution of the transition-state ensemble.","In the latent-space setting, over 97% of decoded molecular configurations pass the 150 kJ/mol energy filter, while direct flow sampling on heavy-atom coordinates yields almost no physically valid configurations, so the latent space is what makes the construction physical."],"supporting_citations":[{"why":"Supplies the variational loss functional whose integrand DASTR turns into a sampling density, plus the artificial-temperature baseline used for comparison.","marker":"[6]"},{"why":"Introduces the KRnet normalizing-flow model used to approximate the target density and generate new collocation points.","marker":"[16]"},{"why":"Provides the umbrella sampling procedure used to convert generated collective variables into physical molecular samples.","marker":"[30]"},{"why":"Supplies metadynamics, used both to create the initial training set and to define the bias potential in the sampling density.","marker":"[31, 32]"},{"why":"Powers the molecular dynamics simulations that produce the alanine dipeptide data and umbrella-sampling trajectories.","marker":"[55]"},{"why":"Provides the finite-element reference solution used to measure errors on the 10-dimensional rugged Mueller benchmark.","marker":"[44, 45]"}],"fun_headline_variants":["DASTR homes in on transition states to train committor nets","Adaptive deep sampling finds rare transition paths for committor","Deep generative model sharpens committor estimates via targeted sampling","DASTR cuts molecular sampling time 95% while improving committor accuracy","Targeted sampling in latent space speeds committor learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Eq. (11), which treats the sampling density on the autoencoder's latent variables as equal to the density on atomic coordinates without correcting for the fact that many atomic configurations can map to the same latent point; if that identification fails, the autoencoder version of DASTR samples from a distribution that is not the intended importance target.","fun_headline_variants_meta":{"raw":{"variants":["DASTR homes in on transition states to train committor nets","Adaptive deep sampling finds rare transition paths for committor","Deep generative model sharpens committor estimates via targeted sampling","DASTR cuts molecular sampling time 95% while improving committor accuracy","Targeted sampling in latent space speeds committor learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000732,"raw_usage":{"total_tokens":3301,"prompt_tokens":994,"completion_tokens":2307,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":2218}},"tokens_in":610,"tokens_out":2307,"duration_ms":15406,"temperature":1.0,"reasoning_tokens":2218,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:12:55.304303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the latent-space identification with the exact change-of-variables formula, multiplying the target density by the encoder Jacobian determinant, and compare the decoded samples and committor error on alanine dipeptide against the reported results; if the corrected sampler does not match or improve the accuracy, the stated mechanism is not what produces the gains. A cheaper check is to compare histograms of $|\\nabla q_\\theta|^2 e^{-\\beta V}$ at decoded versus directly sampled configurations.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variational loss functional whose integrand DASTR turns into a sampling density, plus the artificial-temperature baseline used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the KRnet normalizing-flow model used to approximate the target density and generate new collocation points."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the umbrella sampling procedure used to convert generated collective variables into physical molecular samples."},{"cited_title":"Eastman, J","cited_arxiv_id":null,"evidence_quote":"Powers the molecular dynamics simulations that produce the alanine dipeptide data and umbrella-sampling trajectories."}],"review_version":1}