acceptodds
Under review as a conference paper at ICLR 2027

Bridging Stochastic Flow Maps and Boltzmann Generators with Normalizing Flows

Abstract

Generating independent, equilibrium samples of molecular systems at scale remains a central obstacle in computational statistical mechanics. Boltzmann Generators address this by pairing a generative model with importance sampling to obtain consistent samples from the target distribution. We introduce Normalizing Flow Flow Maps (NF²M), which combines the strengths of recent stochastic flow maps with the tractability of classic normalizing flows to build a Boltzmann Generator. Unlike most methods, which correct the generative model only at the end, NF²M reweighs each denoising transition as generation proceeds, avoiding wasted compute on trajectories that are ultimately discarded. At each denoising step, a conditional normalizing flow proposes clean configurations given the current noisy state (a simpler task than sampling directly from the target) and its exact likelihood enables correcting each proposal toward the true denoising transition of the target Boltzmann distribution. This is in contrast to most existing methods, whose likelihoods are approximate or expensive to evaluate, undermining the statistical reliability of the correction. We establish consistency of the corrected transitions and bound how approximation errors propagate through the sampling chain. We evaluate NF²M on peptide systems, demonstrating improved sampling efficiency and sample quality.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Boltzmann generators, diffusion, flow maps, normalizing flows, sampling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QjQ4OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes NF$^2$M, a model somewhere between Stochastic Normalizing Flows, Flow Maps and Diffusion models. In detail the core idea is to replace the Gaussian in the denoising step of a diffusion model with a conditional normalizing flow that predicts the clean sample x0x_0 given the current noised state xtx_t. They then go on and show how to use this model for sampling -- similar to annealed importance sampling --, but following a diffusion transport path and applying local SNIS corrections at each diffusion step.

They show good performance of the proposed model & sampling scheme on molecular systems.

Critical strengths

The derivation of the SNIS-corrected transition kernel is novel and sound. The resulting local SNIS correction together with the theoretical guarantees on the error are a nice package.

It's great that this framework provides exact proposal densities instead of relying on approximations, which gives us consistent estimators.

Critical weaknesses

The experimental section (+ appendix) is very cluttered. Since the proposed "NF2M-(M)IS" method+model has many parts it's not immediately clear what effect the NF2M actually has. While -- or maybe because -- there are over 40 pages of appendix, I don't see a clear cut experiments that shows the clear theoretic advantages actually translate into the real world.

I would have liked some simple dataset where we can see e.g. the error correction and variance scaling as a function of integration steps. I don't see a clear advantage of using MIS since for Chignolin it is worse the IS.

Questions for authors
  • If you want to optimize with the Reverse KL, you'd want a different non AR flow, right? Otherwise training would take way longer.
  • The Inner ESS in Fig 4. So this is ESS is for a given pool of up to 20 samples? An estimated ESS of 1 is simply the minimum. How is this informative then?
  • How do you explain that there is no consistent behaviour between Ala-2 to 6? I'd broadly expect the different models to perform similar relative to each other, because this is essentially a very similar problem.
7 comments
AuthorsOpenReview ↗

Answer to reviewer QjQ4 [1/2]

We thank the reviewer for their feedback and for finding the derivation of the SNIS-corrected transition kernel “novel and sound,” with the local correction and theoretical guarantees forming “a nice package.” We are also glad they value that our framework “provides exact proposal densities instead of relying on approximations, which gives us consistent estimators.” We address each point in detail below.

The experimental section (+ appendix) is very cluttered. [...] I don't see a clear cut experiments that shows the clear theoretic advantages actually translate into the real world.

We respectfully note that the main experimental content is concentrated in a single section, Sec. 5.1, with two tables. Tab. 1 compares NF²M against all baselines on five peptide systems under a matched protocol, and Tab. 2 reports their training and inference costs. Together, they directly show how the method performs under realistic conditions.

NF²M-IS itself has few moving parts: it is the SNIS-corrected transition of Eq. (12) applied in a loop (Alg. 1), which takes a few lines of code. The appendix is organized by component:

  • App. A.1 contains the theory and proofs.
  • App. A.2 details the MIS scheme, an optional extension of SNIS that shares candidates across trajectories.
  • App. A.3 describes MAExL, a simple posterior-invariant MCMC step (Eq. (62) in a loop) that is orthogonal to the SNIS correction.
  • App. A.4–A.5 cover the more technical parts: the centre-of-mass correction and the conditional architectures with their caching, both variants of standard tools.
  • App. B documents the experimental setup and our baseline reimplementations in full, for transparency.

We have added an appendix roadmap that explains each section’s role and highlights the relevant results and takeaways, including the candidate-budget comparisons and refinement ablations (p. 17, “Appendix roadmap,” lines 878–889).

I would have liked some simple dataset where we can see e.g. the error correction and variance scaling [...] don't see a clear advantage of using MIS since for Chignolin it is worse the IS.

Controlled experiment. We focused on the standard peptide benchmark, where ALA-2 already serves as a small, well-understood system. Our bounds predict consistency as MM grows, and both metrics indeed improve monotonically with MM on all systems (Fig. 5). Fig. 2 shows how the schedule-dependent quantities governing error propagation, as,ta_{s,t} and γt,s\gamma_{t,s}, vary with the number of steps KK.

MIS on chignolin. The gap in Tab. 1 reflects different selected configurations rather than MIS itself. The IS model uses K = 5 steps with M = 20 candidates per trajectory, while MIS was restricted to K = 8 and uses a smaller pool (C=2C = 2, M=6M = 6, i.e., 12 candidates). At matched configurations the two perform on par (Fig. 5), and MIS is 20% faster on chignolin at the same refinement setting (Tab. 4). As noted in our answer to Reviewer wD31, MIS is a secondary, optional component whose benefit should grow with system size.

If you want to optimize with the Reverse KL, you'd want a different non AR flow, right? Otherwise training would take way longer.

Indeed, on-policy reverse-KL training with an autoregressive flow is expensive, since every step requires sequential sampling and backpropagation through it. Off-policy objectives avoid this. For instance, samples produced by the NF²M sampler itself can be reused and reweighted, so that training only requires density evaluations, which take a single parallel pass. Moreover, reverse KL would be used to fine-tune the flow after forward-KL pretraining (Sec. 3.2), so only a limited number of steps would be needed.

The Inner ESS in Fig 4. So this is ESS is for a given pool of up to 20 samples? An estimated ESS of 1 is simply the minimum. How is this informative then?

Yes, the inner ESS is computed per transition over one trajectory’s candidate pool (M≤20M \leq 20 for IS, CMCM for MIS) and averaged over trajectories. It is not the ESS of the final samples, which carry equal weights and are evaluated directly by the Wasserstein metrics.

Despite its small range, it serves three purposes. It shows that MIS with CC components matches IS with CMCM candidates. It locates where the proposal is weakest, namely at low noise and on chignolin, which is where MAExL is most accurate. And ESS/M is the standard estimate of 1/ρt1/\rho_t, the χ2\chi^2 quantity in our bounds.

AuthorsOpenReview ↗

Answer to reviewer QjQ4 [2/2]

How do you explain that there is no consistent behaviour between Ala-2 to 6? [...]

The relative behavior is in fact consistent and follows two clear trends:

  • Energy. NF²M's raw samples achieve the best E-W2 on every system (Tab. 24) except on ALA-6.
  • Torsions. The ranking follows system size. SNIS's T-W2 grows steadily with dimension (0.24 → 0.36 → 0.78 → 1.77 → 3.86), while NF²M's grows much more slowly and overtakes it from ALA-6 onward. This is expected: global SNIS reweights whole configurations, which works well in low dimension but degenerates as dimension grows, whereas NF²M corrects easier per-step posteriors.

The one outlier, NF²M's T-W2 on ALA-2, stems from its training split (Klein & Noé, 2024), which deliberately over-represents a rare state. Correcting this requires a large reweighting of mode populations. Global SNIS with 10⁶ proposals performs it directly, while NF²M's small per-step pools (M = 14) leave residual bias. This is consistent with our bounds, whose finite-M bias grows with the χ2\chi^2-divergence (Thm. 7). The remaining differences among flow-map baselines mostly fall within their large seed variance.

We thank the reviewer again for their constructive feedback. We hope our responses address their questions, and we would be grateful if they would consider raising their score. We are happy to answer any further questions.

Dear authors, I have received your answers, thanks for clarifying.

I still feel like my question on the inner ESS estimator still hasn't been answered (at least for me to understand): If your estimator for Inner ESS is 1, then it's just at its minimum. To me this simply indicates the absence of signal showing that the Effecting Sampling Ratio is larger 0.

Does that make sense?

Reviewer_a8EPOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces NF$^2$M, a Boltzmann generator that learns a conditional normalizing flow to recover clean molecular configurations from noisy inputs, trained via maximum likelihood. At inference, proposals are re-weighted using Boltzmann energies, forward noising likelihoods, and tractable proposal densities to construct Gaussian-mixture reverse transitions. The approach also incorporates importance sampling for candidate sharing and posterior-specific Metropolis-adjusted Langevin refinement. Experiments on five single-system peptides (ranging from alanine dipeptide to chignolin) are conducted to compare performance across a range of baselines that comprise both flow-map methods and one-step normalizing flows.

Critical strengths
  1. The conditional-density formulation is a nice approach to addressing one of the main limitations of existing flow maps by enabling tractable proposal evaluation for energy-based importance sampling. Combined with an analytic Gaussian bridge, it supports expressive, non-Gaussian proposals without requiring the final marginal density.
Critical weaknesses
  1. The evaluation lacks comparisons with state-of-the-art Boltzmann generators. In particular, Autoregressive Boltzmann Generators (ArBG) report substantially stronger results on relevant benchmarks. Despite citing ArBG, the paper excludes it from all empirical comparisons. Restricting baselines to flow-map methods is difficult to justify, especially since SBG, a normalizing-flow-based BG, is included. Moreover, SBG has since been substantially improved upon by Prose. Although Prose focuses on transferable sampling, it remains unclear whether its architectural improvements would also yield substantially stronger results in the single-system setting, making the older SBG baseline potentially unrepresentative of current capabilities. Comparisons with these more recent BG methods are essential to establish NF$^2$M's competitiveness.

  2. The experiments do not establish scalability or transferability. While improvements over FALCON and SCALLOP on individually trained systems are valuable, recent BG methods such as Prose and ArBG demonstrate transferable sampling across molecular systems. Prose, in particular, extends the SBG approach toward transferable sampling. In contrast, NF$^2$M requires separate training for each system and provides no evidence of generalization to larger or unseen systems. Without competitive performance against leading BGs or evidence of transferable scalability, the experimental contribution remains insufficient to justify acceptance.

  3. The existing baseline comparisons do not establish a consistent advantage. FALCON and SCALLOP use modified training procedures without validation against their original implementations. NF$^2$M also improves some metrics while worsening others, often at greater computational cost. The SBG-SMC comparison lacks clear budget matching, and the claimed superiority on both metrics for larger systems contradicts the ALA-6 energy results. More rigorous, matched-budget comparisons are needed.

  4. The practical sampler's statistical reliability remains insufficiently established. Finite candidate pools introduce importance-sampling bias, while adaptive MAExL refinement and approximate terminal initialization fall outside theory-based exactness guarantees. Sharing a terminal noisy state may also induce correlations across samples. Near-unit inner ESS on Chignolin and strong dependence on Langevin refinement make me question the effectiveness of the learned proposal and importance correction. Independent-start experiments are needed to address these claims.

Questions for authors
  1. Comparison with state-of-the-art Boltzmann generators: Given that ArBG is already cited, could you provide direct comparisons with ArBG and Prose on shared benchmarks, using consistent reference distributions, evaluation metrics, and computational budgets? These comparisons are particularly important because both methods report strong results and extend beyond the older SBG baseline. Note, while transferable training would ideally be evaluated to establish scalability, I recognize that implementing it may be infeasible within the rebuttal timeline.

  2. Accuracy versus computational cost: Can you extend the existing refinement-budget analysis to include varying proposal counts and flow steps, alongside SBG-SMC and ideally ArBG as well? Reporting energy and torsion errors against total inference time, with training costs provided separately, would clarify whether NF$^2$M offers a consistent accuracy-efficiency advantage rather than improvements at selected operating points.

  3. Statistical reliability: Can you assess how finite candidate counts, approximate terminal initialization, and adaptive MAExL affect sampling bias? In particular, comparisons between independent and shared terminal states, along with repeated-run estimates of observable variance and effective sample size, would help establish whether increasing the number of generated samples improves statistical reliability.

  4. Attribution of performance gains: Can you disentangle the contributions of the learned conditional proposal, importance weighting, and Langevin refinement? Comparing weighted and unweighted sampling with and without refinement under matched computational budgets would clarify how much of the improvement comes from the proposed correction mechanism rather than MCMC refinement.

4 comments
AuthorsOpenReview ↗

Answer to reviewer a8EP [1/3]

We thank the reviewer for their feedback and for recognizing that our conditional-density formulation addresses “one of the main limitations of existing flow maps by enabling tractable proposal evaluation,” supporting “expressive, non-Gaussian proposals without requiring the final marginal density.” We address each weakness and question in detail below.

In particular, Autoregressive Boltzmann Generators (ArBG) report substantially stronger results on relevant benchmarks [..] Although Prose focuses on transferable sampling, it remains unclear whether its architectural improvements would also yield substantially stronger results [...] Comparison with state-of-the-art Boltzmann generators [...]

We respectfully disagree that our baselines do not reflect the state of the art. We compare against the normalizing-flow and flow-map Boltzmann Generators that NF²M builds on, all on the same peptide benchmark:

  • SBG (Tan et al., 2025, ICML 2025): its TarFlow/StarFlow proposal with SNIS (Tab. 1), its full SMC sampler (Tab. 26), and the ECNF++-style baseline (Tab. 22);

  • FALCON (Rehman et al., 2026a, ICLR 2026, oral);

  • F2D2 (Ai et al., 2026, ICLR 2026);

  • SCALLOP (OuYang et al., 2026, NeurIPS 2026).

These methods differ precisely in how they obtain the proposal density needed for importance weighting (exact NF Jacobian, exact flow-map Jacobian, or distilled likelihood head), which is the question NF²M addresses. Neither FALCON nor SCALLOP had released code at the time of submission, so rather than copying published numbers, we reimplemented every baseline and evaluated all methods under a single protocol, with matched training wall-clock (App. B.2) and target-energy budgets (Tab. 25, Figs. 8–10), so that differences reflect the methods themselves. Note that our reimplementation sometimes even outperforms the original paper.

ArBG (Rehman et al., 2026b) departs from this family on both axes. Its density is a coordinate-wise binned model with uniform dequantization rather than a flow, and its correction relies on twisted SMC with temporarily capped energies (Sec. 4). This correction is complementary to the choice of proposal studied here rather than an alternative to it.

ArBG also operates at a very different computational scale. We profiled its released code on our hardware, with the same training data and splits.

Training wall-clock (ArBG with its released configuration: 2000 epochs, H100):

ALA-2ALA-3ALA-4ALA-6Chignolin
ArBG14.7 h16.0 h19.4 h34.5 h≈ 70 h (4 GPUs)
NF²M4.3 h6.0 h5.3 h12.8 h21.7 h (4 GPUs)

Inference wall-clock (one A100-80GB; ArBG at its fastest batch size, as the released batch size is up to 3.3× slower; SNIS and NF²M include all refinement):

ALA-2ALA-3ALA-4ALA-6Chignolin
ArBG, 10⁶ proposals (sampling only)9 min17 min27 min56 min10.2 h
SNIS, 10⁶ proposals2.0 min2.1 min2.8 min6.2 min41 min
NF²M-IS / MIS2.3 / 2.1 min2.7 / 2.5 min3.1 / 2.7 min7.5 / 6.8 min47 / 38 min

At the 10⁶-proposal budget our protocol gives every SNIS-based baseline, ArBG's proposal sampling alone is 4–15× slower than SNIS with a TarFlow/StarFlow proposal, and 4–13× slower than NF²M's full inference. Even at its own budget of 2×10⁵ proposals, chignolin takes about 2 h.

The released code also departs from the paper on several settings, including learning rates, discretization range, number of bins and parameter counts. A faithful comparison would therefore require resolving these choices, about 365 GPU-hours of training with the released configuration, and over 30 GPU-hours of sampling for chignolin alone across three seeds.

In conclusion, ArBG is a recent Boltzmann Generator evaluated on similar data, but its mechanism is unrelated to NF²M or to any other method covered in this paper. It is also not concerned by the limitation we identify and address, namely the lack of exact proposal densities in flow-map Boltzmann Generators. It further operates at a substantially different computational scale, in both training and inference, from the methods studied here. Combined with the non-trivial effort required for a faithful reproduction, we believe a direct comparison falls outside the scope of this paper. We will nonetheless keep ArBG in our related work, as it is a relevant contribution to the field nonetheless.

Regarding PROSE, it targets transferable sampling and reports no single-system results, so evaluating it here would require designing and tuning a single-system variant ourselves. More importantly, its contributions are architectural improvements to TarFlow, on which NF²M’s conditional flow is itself built. They are therefore orthogonal to our contribution and would benefit NF²M as much as the SNIS baseline. Extending NF²M to the transferable setting is a natural next step, as noted in our conclusion.

AuthorsOpenReview ↗

Answer to reviewer a8EP [2/3]

While improvements over FALCON and SCALLOP on individually trained systems are valuable, recent BG methods such as Prose and ArBG demonstrate transferable sampling across molecular systems [...] the experimental contribution remains insufficient to justify acceptance.

We respectfully disagree that the experimental contribution is insufficient. Per-system training remains the standard setting for this line of work. FALCON (ICLR 2026, oral), F2D2 (ICLR 2026) and SCALLOP (NeurIPS 2026) were all accepted at top venues with per-system evaluations on the same benchmark we use. Within this setting, our evaluation extends to chignolin (10 residues which is 138 atoms), which the flow-map papers did not evaluate. NF²M’s advantage is also most pronounced on the largest systems, where it achieves the best T-W2 on ALA-6 and chignolin (Tab. 1).

Transferability is moreover orthogonal to our contribution. The NF²M sampler only requires a conditional flow with exact density and is agnostic to how that flow is trained. A transferable backbone such as Prose’s, additionally conditioned on xtx_t, would therefore directly yield a transferable NF²M.

FALCON and SCALLOP use modified training procedures without validation against their original implementations [...] More rigorous, matched-budget comparisons are needed [...] Can you extend the existing refinement-budget analysis [...] clarify whether NFM offers a consistent accuracy-efficiency advantage rather than improvements at selected operating points.

Baseline implementations. FALCON and SCALLOP did not release code (at time of submission), so we reimplemented them carefully rather than copying published numbers. Every departure from the papers is documented and justified in App. B.2: each was either dictated by released code, required for stability and verified on diagnostics, or a value the papers leave unstated. Our setup is, if anything, favorable to these baselines. They warm-start from an independently trained teacher and, at matched training wall-clock, see 2.7–9× more passes over the data than NF²M. Our reimplementations match or improve on the originally reported numbers.

Accuracy and cost. We do not claim NF²M dominates every metric. It achieves the best E-W2 on all systems except ALA-6 and the best T-W2 on the two largest systems, while SNIS remains best in T-W2 on the smaller ones (Tab. 1). Without post-hoc refinement, NF²M is best in E-W2 everywhere (Tab. 24). Its inference time is within 30% of SNIS, and NF²M-MIS is faster than SNIS on chignolin. It is slightly slower than F2D2/SCALLOP on small systems but 1.5–1.9× faster on chignolin, and 10–130× faster than FALCON (Tab. 2).

Budget matching. All methods use the same number of target-energy evaluations (Tab. 25), traced across three budgets (Figs. 8–10). NF²M's candidate budget matches the wall-clock of SNIS, and the flow-map pipelines receive NF²M's training wall-clock. Training and inference costs are reported separately (Tab. 2).

SBG-SMC. We agree that our "both metrics" statement is imprecise. It holds on chignolin, but on ALA-6 SBG-SMC has lower E-W2 and higher T-W2 (Tab. 26), and we have corrected this statement in the revised submission (App. B.2.2, “Comparison with SBG,” p. 54, lines 2909–2911). SBG-SMC uses the same proposal budget as NF²M but roughly 2× more wall-clock on ALA systems and 4–9× more on chignolin, so the mismatch favors it. Consistent with Tan et al. (2025), SMC does not consistently improve over SNIS while costing more, which is why SNIS is our main baseline.

Proposal count and steps. Figs. 4-5 report the effect of M and C on ESS, both metrics, and inference time. Both metrics improve monotonically with M. K was selected from {5, 8} on validation (Tab. 10). For ArBG, see the first answer.

AuthorsOpenReview ↗

Answer to reviewer a8EP [3/3]

Finite candidate pools introduce importance-sampling bias [...] Can you assess how finite candidate counts, approximate terminal initialization, and adaptive MAExL affect sampling bias? [...]

Finite candidates. This is precisely what our theory quantifies. The per-step bias and MSE decay as O(1/M)O(1/M) (Thm. 7) and accumulate additively along the chain (Cor. 16). Both metrics improve monotonically with MM (Fig. 5).

Terminal initialization and adaptive MAExL. The Gaussian initialization is standard in diffusion models. Its error is bounded in Remark 18, decays as 1/σT1/\sigma_T, and is independent of MM. Our analysis covers any state-independent step-size schedule (Remark 8).

Shared terminal state. At σT=40\sigma_T = 40, the first denoising kernel barely depends on xTx_T (Eq. 17), so the induced correlation is weak.

Dependence on refinement. Post-hoc MALA barely affects NF²M, unlike the baselines (Tab. 24). In-loop MAExL is part of the method: it is p0∣tp_{0|t}-invariant, covered by our guarantees (Cor. 17), and both refinements together cost only 4–11% of inference time (Tab. 23). MAExL improves E-W2 but leaves T-W2 unchanged (Tab. 4). NF²M's lead in metastable-state occupancy on the largest systems therefore comes from the learned proposal and importance correction.

Inner ESS. Inner ESS is computed over the M≤20M \leq 20 candidates of a single transition, not over the final samples. We have revised the main-text discussion to explain the purpose and definition of inner ESS before summarizing the results, including the low values at low noise, especially on chignolin (Sec. 5.1, “NF²M: IS versus MIS,” p. 9, lines 480–486). Its impact is limited because posterior errors are attenuated by as,ta_{s,t} (Prop. 1), and MAExL is most accurate precisely in that regime. NF²M still achieves the best T-W2 on chignolin. Repeated runs. All results use three independent seeds, and Figs. 4–5 report error bars over repeated runs as M grows.

Can you disentangle the contributions of the learned conditional proposal,importance weighting, and Langevin refinement? [...]

The refinement components are already ablated. Tab. 4 varies the number of in-loop MAExL steps (0, 512, 1024), and Tab. 24 reports every method with and without post-hoc MALA. Figs. 8–10 compare all methods at the matched energy budget of each MAExL setting. Refinement mainly improves E-W2, while T-W2 is essentially unchanged in both cases. NF²M’s advantage in metastable-state occupancy therefore comes from the learned proposal and the importance correction.

Informativeness: 1: I learned little or nothing new from this paper.

We respectfully disagree. To our knowledge, the following contributions are new:

  • SNIS-corrected stochastic flow maps. A conditional normalizing flow trained by maximum likelihood serves as an exact-density proposal for the denoising posterior, so each transition is corrected toward the Boltzmann target (Sec. 3.1–3.2). The reviewer notes this addresses a main limitation of existing flow maps, and Reviewer QjQ4 finds the derivation “novel and sound.”
  • Error analysis. We give non-asymptotic per-step and path-level bias/MSE bounds for SNIS-corrected transitions, showing how posterior errors are damped by the Gaussian bridge, including with MCMC rejuvenation (Sec. 3.4, App. A.1).
  • Candidate sharing via MIS. Trajectories share proposal candidates at each denoising step through mixtures whose components are selected by posterior proximity (Sec. 3.3, App. A.2).
  • Posterior-adapted MCMC. MAExL adapts exponential-integrator MALA to diffusion posteriors, with step sizes that transfer across noise levels (App. A.3).
  • Conditional molecular flows. We introduce conditional TarFlow/StarFlow architectures for molecules, conditioned on (t,xt)(t, x_t), and a conditional centre-of-mass correction showing that SBG’s radial term is unnecessary under the Gaussian lift (App. A.4–A.5).

We thank the reviewer again for their feedback. We hope our responses and additional experiments address their concerns, and we would be grateful if they would reconsider their assessment and substantially raise their score. We are happy to answer any further questions.

I would like to thank the authors for their comprehensive response to the review. Further, the addition of the transferable sampling formulation is thoroughly appreciated in the short rebuttal window.

My concerns about the relative performance ARBG have been addressed. However, my concerns about scaling the approach remain. On the large single peptide benchmarks with ALA-6 and Chignolin, the best framework is inconclusive. Further, on the transferable benchmark, although the numbers shown outperform SBG in some settings, the results still remain inconclusive. If the obtained results are also compared to those from the original Prose paper as well, the performance is worse across all peptide sizes 2, 4, and 8. As a result of these observations, I increase my score to a 2.

Reviewer_wD31OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Normalizing Flow Flow Maps (NF²M), a Boltzmann Generator that uses a conditional normalizing flow to propose clean configurations from noisy states along a diffusion path, and corrects each denoising transition via self-normalized importance sampling (SNIS). The key claimed advantage over stochastic flow maps is the tractability of the conditional flow density, which enables exact per-step correction.

Critical strengths

The paper demonstrates adequate familiarity with the immediate technical lineage: it correctly positions itself relative to Sequential Boltzmann Generators (SBG), stochastic flow maps, and classical BG frameworks. The discussion of why existing flow-map methods lack tractable densities is accurate.

Critical weaknesses

The theoretical guarantees rest on unverified assumptions. Proposition 1 assumes finite ll-th moments of the approximate posterior; Proposition 2 assumes sub-Gaussian tails and finite χ2\chi^2-divergence between proposal and posterior. These conditions are never empirically checked for the molecular systems considered. The path-error bounds (Corollaries 16–17) assume independence of the SNIS kernels across denoising steps—an assumption that is explicitly violated in NF²M-MIS, where trajectories share candidate pools (Section 3.3). The paper does not discuss how this dependence affects the bounds.

The MIS contribution is marginal in practice. Table 23 in the submission reports that on Chignolin, NF²M-MIS inference time is 2356s versus 2830s for IS—a 17% reduction. For this modest gain, the paper introduces a complex machinery of shared candidate pools, cross-likelihood caching, and component selection. The claim that MIS is “more expressive” is not supported by a corresponding improvement in distributional metrics at matched budgets (Table 1 shows IS and MIS are within noise of each other on most systems).

Questions for authors

See Critical Weaknesses.

1 comment
AuthorsOpenReview ↗

Answer to reviewer wD31 [1/1]

We thank the reviewer for their feedback and for noting that our paper “correctly positions itself relative to Sequential Boltzmann Generators (SBG), stochastic flow maps, and classical BG frameworks,” and that our discussion of why existing flow-map methods lack tractable densities “is accurate.” We address each weakness in detail below.

The theoretical guarantees rest on unverified assumptions [...] The paper does not discuss how this dependence affects the bounds.

Moment and tail assumptions. These are standard assumptions in non-asymptotic importance-sampling analysis. Our per-step bounds build directly on Agapiou et al. (2017), where the χ2\chi^2-divergence is the governing quantity, and our MCMC bounds on the EI-MALA analysis of Durmus & Moulines (2015). They are also mild in our setting. Finite moments are needed only for the Wasserstein distance in Prop. 1 to be defined. The target posterior p0∣t(x0∣xt)∝exp⁡(−E(x0))N(xt;x0,σt2I)p_{0|t}(x_0|x_t) \propto \exp(−E(x_0)) N(x_t; x_0, \sigma_t^2 I) inherits Gaussian tails from the likelihood factor whenever the energy is bounded below.

Moreover, the quantity governing our bounds is monitored empirically. The inner ESS in Fig. 4 is the standard estimate of 1/ρt1/\rho_t (App. A.4.5). While finiteness of a divergence cannot be verified from samples, this diagnostic and the monotone improvement with MM (Fig. 5) are consistent with the regime our bounds describe. We updated the text on page 9.

Independence in NF²M-MIS. The reviewer is right. The path-level bounds (Cors. 16–17) apply to NF²M-IS, where each trajectory draws fresh candidates at every step, but not directly to NF²M-MIS, where candidate sharing makes successive kernels dependent. We have clarified this scope in the revised submission: Sec. 3.4, Proposition 3 and the paragraph immediately below it (p. 7), and App. A.1.3, “Scope of the path bounds” (p. 26). Extending the bounds would require interacting-particle arguments, as in SMC theory, which are beyond our scope. Empirically, sharing does not degrade the sampler. NF²M-MIS with CC components matches the inner ESS of NF²M-IS with CMCM independent candidates (Fig. 4) and performs on par at matched budgets (Fig. 5, Tab. 1).

The MIS contribution is marginal in practice [...] The claim that MIS is “more expressive” is not supported by a corresponding improvement in distributional metrics at matched budgets

We agree that MIS is a secondary contribution, which is why it occupies a short paragraph in the main text, with details deferred to App. A.2. It is an inference-time option that requires no retraining.

We do not claim that MIS improves distributional metrics. Our claim (App. A.2) is that its proposal is at least as expressive as NF²M-IS: each trajectory draws on a pool of CMCM candidates, while the ensemble pays for only between MM and CMCM flow draws per trajectory. Fig. 4 confirms this effect: MIS with CC components matches the inner ESS of IS with CMCM independent candidates, and Fig. 5 shows the distributional metrics follow.

The wall-clock gain is modest at our current scale because, for molecules this small, sampling and density evaluation cost about the same. MIS trades flow samples, which require sequential autoregressive steps, for density evaluations, which take a single parallel pass (App. A.5.4). The gain should therefore grow with system size. It is already largest on chignolin, our largest system, where MIS takes 20% less inference time than IS(Tab. 4).

We thank the reviewer again for their feedback. We hope our responses address their concerns, and we would be grateful if they would consider raising their score. We are happy to answer any further questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global answer

We thank Reviewers wD31, a8EP and QjQ4 for their time and feedback. We are glad that the contribution was well understood. All three summaries accurately capture the core idea of NF²M: replacing the Gaussian denoising step with a conditional normalizing flow, whose exact density enables a local SNIS correction at each denoising transition. We are also happy that the reviewers acknowledged its value:

  • Reviewer QjQ4 finds the derivation of the SNIS-corrected transition kernel “novel and sound,” with the local correction and theoretical guarantees forming “a nice package.” They also note that our framework “provides exact proposal densities instead of relying on approximations, which gives us consistent estimators.”
  • Reviewer a8EP describes our conditional-density formulation as “a nice approach to addressing one of the main limitations of existing flow maps by enabling tractable proposal evaluation.” They add that it “supports expressive, non-Gaussian proposals without requiring the final marginal density.”
  • Reviewer wD31 notes that the paper “correctly positions itself relative to Sequential Boltzmann Generators (SBG), stochastic flow maps, and classical BG frameworks,” and that our discussion of why existing flow-map methods lack tractable densities “is accurate.”

Reviewers a8EP and QjQ4 also report no substantive concerns about soundness. We address each reviewer’s remaining points in our individual responses.

AuthorsOpenReview ↗

Transferable experiment

To address the points on transferability and PROSE, we trained NF²M in the transferable setting of PROSE and compared it with the two samplers used by Prose, SNIS and SBG, on unseen peptides. Preliminary results show that NF²M transfers to unseen systems and is competitive with PROSE, which we believe addresses Reviewer a8EP’s concerns on transferability and on the comparison with PROSE.

Data and protocol. We use PROSE’s ManyPeptidesMD dataset, MD references and evaluation code. Every model is trained once on peptides of 2 to 8 residues and evaluated on the 30 test peptides of 2, 4 and 8 residues (90 peptides, none seen during training). We report E-W2, T-W2 and TICA-W2 on 10410^4 samples per peptide, averaged over the peptides of each length (± s.e.m.). Architectures and K were selected on 10 validation peptides of 2 and 4 residues, disjoint from the test set.

Models. All flows have about 290M parameters and share one training recipe: 500 epochs of 1000 batches of 512, AdamW with learning rate 10−410^{-4}, 4 H100 GPUs.

  • Baselines: PROSE's TarFlow (384 channels, 8 blocks of 8 layers, 288M parameters). It outperformed a StarFlow of the same size on validation, for both SNIS and SBG.
  • NF²M: our conditional flows (App. A.4) on the PROSE backbone, with xtx_t entering by cross-attention. Among seven conditional TarFlow and StarFlow variants, validation selected a StarFlow with 384 channels, blocks of [2, 2, 2, 2, 28, 28] layers, cross-attention to xtx_t every 4th layer of the deep blocks and Prose's atom-identity conditioning on every block (292M parameters).

Inference. Flows run in bf16 and every method returns 10410^4 samples per peptide.

  • SNIS: 10610^6 proposals with Prose's 0.2% log-weight filter.
  • SBG: Prose's SMC settings, 10410^4 particles, 100 annealed Langevin steps, systematic resampling.
  • NF²M-IS: K = 5, M = 16 for 2 residues and 15 otherwise, 512 MAExL steps per level.
  • NF²M-MIS: K = 8, 2 mixture components of M/2 candidates (M = 10 for 2 residues, 9 otherwise), 512 MAExL steps per level.

Best value per length in bold.

LengthMethodE-W2T-W2TICA-W2
2SNIS0.317 ± 0.0530.151 ± 0.0140.033 ± 0.007
2SBG1.081 ± 0.0820.240 ± 0.0160.051 ± 0.010
2NF²M-IS0.399 ± 0.0590.250 ± 0.0150.044 ± 0.010
2NF²M-MIS0.388 ± 0.0560.250 ± 0.0140.044 ± 0.009
4SNIS2.279 ± 0.2640.649 ± 0.0280.361 ± 0.028
4SBG3.317 ± 0.2611.093 ± 0.0770.488 ± 0.066
4NF²M-IS1.956 ± 0.2490.711 ± 0.0280.389 ± 0.029
4NF²M-MIS1.967 ± 0.2590.712 ± 0.0260.364 ± 0.029
8SNIS18.97 ± 1.502.007 ± 0.0780.966 ± 0.094
8SBG17.86 ± 1.213.299 ± 0.1371.181 ± 0.081
8NF²M-IS14.44 ± 1.252.181 ± 0.1010.937 ± 0.059
8NF²M-MIS15.08 ± 1.242.237 ± 0.0960.893 ± 0.068
MethodTrainingInference, 2 res.4 res.8 res.
SNIS69.2 h, 4 H10010.2 min22.4 min62.8 min
SBG-10.7 min20.7 min31.1 min
NF²M-IS47.9 h, 4 H10015.4 min40.6 min137.1 min
NF²M-MIS-14.1 min32.1 min108.1 min

Inference times are per peptide on one H100.

Discussion. We set up this transferable version of NF²M within the rebuttal period and kept it deliberately simple. We reused Prose’s backbone, training recipe and evaluation protocol, and kept our single-system sampler settings, apart from M (set from timings) and K (selected on validation). The noise grid, candidate budget and conditioning on xtx_t were not tuned for this setting, leaving clear room for improvement.

Budgets are likewise not exactly matched, but they are of the same order. NF²M trains in less time than the Prose TarFlow (47.9 h vs 69.2 h on 4 H100), while its inference takes 1.4–2.2× longer than SNIS.

Even so, a single NF²M model, trained once, samples 90 unseen peptides of up to 8 residues. It achieves the lowest E-W2 on 4- and 8-residue peptides and the lowest TICA-W2 on 8-residue peptides. Both NF²M samplers improve on SBG in 8 of the 9 (length, metric) pairs. Prose’s SNIS retains the lowest T-W2 and the best results on dipeptides. As in our single-system experiments, NF²M’s relative performance improves with system size on all three metrics. NF²M therefore transfers to unseen systems and is competitive with Prose, which we believe addresses Reviewer a8EP’s concerns on transferability and on the comparison with Prose.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.