Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Annealed neural samplers train better when the proposal distribution sweeps a continuously varying temperature coordinate, letting samples cross energy barriers where standard annealed Langevin proposals get stuck.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

CTDS trains neural samplers with a controlled Langevin dynamics over both position and a continuous temperature coordinate, and reports improved sampling on a 40-mode Gaussian mixture.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A promising NETS extension with a real W2 gain on a single 2D mixture, but the 'extended exploration' story is unmeasured and eq. (19) has a likely typo. the 3 major comments →

arxiv 2509.00316 v1 pith:FGBITTFC submitted 2025-08-30 cs.LG cs.AI

Continuously Tempered Diffusion Samplers

classification cs.LG cs.AI
keywords continuously tempered diffusion samplersannealed neural samplingPINN lossproposal distributionJarzynski equalitycontinuous temperingLangevin dynamicsmultimodal sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a neural sampler trained with a physics-informed (continuity-equation) loss produces better final samples when the training proposal is a continuously tempered diffusion sampler: a Langevin dynamics run over positions together with a continuous inverse-temperature coordinate. The temperature coordinate lets proposal samples rise to high temperature, where energy barriers are low, wander between modes, and return to the target temperature; the standard proposals, controlled annealed Langevin dynamics, get trapped near isolated modes. The authors support this with two theoretical pieces: a multi-temperature PINN loss whose unique minimizer is the joint free energy, and a continuously tempered controlled Jarzynski equality that reweights the proposal's samples toward the target distribution. Empirically, on the challenging 40-mode Gaussian mixture, the CTDS-trained sampler reaches a 2-Wasserstein distance of 12.87 versus 20.11 for the best NETS baseline, and also improves the evidence lower and upper bounds. If the claim holds, it means the exploration machinery of molecular-dynamics tempering transfers directly to the training phase of amortized neural samplers, where exploration, not the learned transport, has been the bottleneck.

Core claim

The proposal distribution that feeds a PINN sampler loss can itself be a tempering scheme. The paper extends the time-indexed annealing path to a density continuum over (t, β), reparameterizes β via a coordinate ξ with plateaus at both endpoints, and runs a controlled underdamped Langevin dynamics on the augmented state (x, ξ). A multi-temperature PINN loss trains the control and a learned free-energy estimate over this continuum; Theorem 3.2, a controlled Jarzynski equality for the tempered dynamics, justifies reweighting samples. On the 40-mode Gaussian mixture, the CTDS sampler reaches W2 = 12.87 ± 0.20 versus 20.11 ± 0.23 for the best baseline, and the reweighted variant gives the best u

What carries the argument

The continuously tempered diffusion sampler (eq. 17): a controlled underdamped Langevin dynamics over the augmented state (x, ξ) with a non-separable Hamiltonian whose kinetic energy weights the position momentum by β(ξ). The reparameterization β(ξ) (eq. 30) maps R onto [β_min, 1] with plateaus at both endpoints, so the temperature marginal spends real time at the target temperature and at the high-temperature end, while a confining potential (eq. 32) keeps ξ in range. Two further pieces carry the argument: the multi-temperature PINN objective (eq. 13), whose unique minimizer over F is the joint free energy (Thm 3.1), and the continuously tempered controlled Jarzynski equality (Thm 3.2), use

Load-bearing premise

The method succeeds only if the temperature coordinate ξ mixes quickly enough that, at every curriculum stage, the proposal spends real time at both the target temperature β=1 and the high-temperature end β_min; if the ξ marginal pools away from β=1 early in training, a failure the authors observe and blame on a poor free-energy estimate, the learned control at β=1, the only temperature used at evaluation, is undertrained.

What would settle it

Record the empirical marginal of ξ from the CTDS proposal at several early curriculum stages on a multimodal target with the free-energy network at initial weights. If the fraction of training samples at β(ξ)=1 is negligible through the first half of training while post-training W2 at β=1 matches the non-tempered baselines, the exploration claim fails. Alternatively, freeze the learned free energy, simulate eq. (17), and check whether any trajectory that reaches the β_min plateau returns to the β=1 plateau within the barrier-crossing time scale; if none do, the mechanism is absent.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Post-training, the CTDS-trained control spreads samples across the modes of the mixture, while controls trained with the ODE-only, overdamped, and underdamped proposals collapse onto fewer modes; the W2 gap (12.87 vs 20.11) is the paper's headline number.
  • Because the loss is defined over the whole time–temperature continuum, training is amortized across temperatures: one control network carries the transport for every β, with β=1 recovered at evaluation.
  • The Jarzynski-reweighted CTDS reaches the lowest evidence upper bound, indicating the learned density covers the target's mass without leaving regions underweighted.
  • The construction only needs energies that scale with β (U^β ∝ βU), so any target whose barriers shrink at higher temperature, the standard tempering assumption, fits the same proposal design, not just Gaussian mixtures.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper reports that early in training the ξ marginal pools away from β=1 (fig. 2), yet does not measure ξ mixing time or the fraction of training samples at β=1; that fraction is a cheap, direct diagnostic of whether the method's exploration is actually operating at each curriculum stage.
  • The non-reweighted CTDS beating the reweighted CTDS on W2 suggests the Jarzynski weights, computed with an early free-energy estimate that is still poor, inject noise; a variance-reduced or progressively annealed weight estimator might combine the best of both rows.
  • A natural extension the paper does not explore: run several ξ trajectories with replica-exchange swaps as the proposal instead of one confining-potential trajectory, which could decorrelate the β=1 particles without relying on returns to the plateau.
  • The 2D validation leaves high-dimensional behavior open; in higher dimensions the joint free-energy learning over (t, ξ) becomes harder, so whether the exploration gain survives the harder estimation problem is a testable open question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes continuously tempered diffusion samplers (CTDS), an extension of PINN-based annealed neural samplers (NETS) to a time-and-temperature-indexed density continuum. The authors introduce a multi-temperature PINN objective over (x, t, ξ), a continuously tempered underdamped Langevin proposal in the augmented state (x, ξ), and a controlled Jarzynski equality for reweighting the proposal during training. They evaluate CTDS on a 40-mode Gaussian mixture, reporting lower W2 and better ELBO than baseline, overdamped NETS, and underdamped NETS proposals. The paper also provides a proof of the multi-temperature PINN minimizer and a Crooks/Jarzynski derivation in appendices, with code available.

Significance. If the theoretical statements are corrected, the paper makes a useful conceptual contribution: it transfers continuous-tempering ideas from molecular dynamics to amortized neural sampling, and it shows a plausible way to use multi-temperature proposals in PINN-based samplers. The multi-temperature PINN loss and the controlled Jarzynski reweighting are nontrivial extensions of existing single-temperature results, and the public code is a strength. However, the empirical support is limited to one 2D benchmark, and the central mechanistic claim—that improvements are driven by extended exploration—is not directly measured or isolated by an ablation.

major comments (3)
  1. [Section 3.4, Eq. (19); Appendix C.3, Eqs. (78)/(90)] The work functional is printed with ∂_t log \tilde U^θ_t (and ∂_t log U_t in the appendix). The derivation at Appendix C.3, Eq. (83) yields ∂_t U_t, and the 'log' is not a harmless notation variant: \tilde U is an energy and may be negative or zero. Additionally, Theorem 3.2 defines F_t as ∫ e^{-\tilde U^θ_t} dx dz, which is a partition function Z_t, while the displayed identity E[e^{A_T}] = exp(F_0 − F_T) requires F_t = −log Z_t; as printed, the identity becomes exp(Z_0 − Z_T). Since Eq. (22) relies on this theorem for reweighting, the theorem needs correction before the method is well-defined.
  2. [Section 4.1, Fig. 2] The abstract claims 'improved sampler performance driven by extended exploration,' but the paper does not measure exploration or the training-time population at β=1. Fig. 2 shows only the ξ marginal at t=1.0 before/after training; it does not report the fraction of training proposals with |ξ|<Δ (i.e., β=1) as the curriculum progresses, nor the mode coverage of the proposal. Because CTDS also adds a temperature input and Fourier features, the W2 gain over NETS could result from the extra capacity or multi-task regularization rather than from the proposed tempering mechanism. Section 3.4 itself notes that with a poor free-energy estimate samples pool away from β=1. Please add an ablation with ξ dynamics frozen at β=1 (or ξ drawn from a fixed distribution) and, if possible, report per-epoch β=1 occupancy and proposal mode coverage.
  3. [Section 3.1 and Appendix D.1] The formal continuum definition sets π^β_0 = N(0, (1/β)I), i.e., U^β_0 = (β/2)||x||², but the experiments use π0 = N(0, 5I₂). It is ambiguous whether the continuum in the experiments is defined by U^β_0 = β U^1_0 (so the source variance becomes 25/β) or by the canonical U^β_0 = (β/2)||x||² with a different β=1 source. This affects the temperature scaling of the proposal and should be clarified for reproducibility.
minor comments (5)
  1. [Theorem 3.2 and Eq. (22)] The theorem writes h(x, ξ) but then evaluates h(Q) on a path; Eq. (22) similarly uses P_t as if h were a terminal function. Please state explicitly that h is evaluated at the terminal state q_t and that P_t is the path measure on [0,t].
  2. [Appendix C.2, Eq. (59)] The term log(dπ0/dE0) should be log(dπ0/dΛ); 'E0' appears to be a typo.
  3. [Theorem C.2] In the displayed Girsanov formula, the drift terms are evaluated at Z_t but the statement writes ϕt(Xt); this should be ϕt(Zt).
  4. [Figure 2] The caption does not specify the training epoch at which the 'after training' marginal is taken, nor how the before/after histograms were normalized. Adding axes and a seed description would help.
  5. [Figure 3] The results in Figure 3 are printed as a table, not a figure; the caption should say 'Table' and the table should be referenced consistently in the text.

Circularity Check

0 steps flagged

No significant circularity: Theorems 3.1 and 3.2 are proved from the continuity equation and a controlled Crooks identity, and the reported W2 / ELBO / EUBO values are post-training measurements rather than fitted quantities.

full rationale

The derivation chain is self-contained. Theorem 3.1 (unique free-energy minimizer of the multi-temperature PINN loss) is proved in Appendix B.3 by applying the Maté–Fleuret argument pointwise in ξ: if L_PINN(F*, μ*; π~)=0, then eq. (34) gives ∂t U_t^ξ = ∂t F*_t − e^{U}(∇·(μ* e^{−U})); substituting into ∂t F_t(ξ) = ∫ ∂t e^{−U}/Z and using the divergence theorem yields ∂t F*_t = ∂t F_t(ξ), and with the boundary condition F*_0 = F_0 this identifies the true free energy. This does not assume the conclusion. Theorem 3.2 is derived in Appendix C from a Radon–Nikodym computation (Theorem C.1) and a controlled Crooks fluctuation theorem (Theorem C.3) for the non-separable Hamiltonian of eq. (16); the CTDS specialization in Appendix C.4 is substitution of the specific Γ_t, E_t, K(q,p) into the general result, and the work functional eq. (19) is the corresponding A_T. The proof uses standard Girsanov/Crooks machinery, not the target equality as an input. The empirical results (Figure 3) are measured on a 40-mode Gaussian mixture with a fixed source N(0,5I) and the learned path eq. (11); CTDS changes the training-time proposal, while post-training metrics are obtained by simulating eq. (23) with μθ_t(x, β=1). No fitted parameter is renamed as a prediction. The paper itself flags in Section 3.4 and Figure 2 that a poor free-energy estimate Fθ_t pools samples away from β=1 and biases Jarzynski reweighting toward particular ξ values; this is an acknowledged limitation and a possible explanation for the non-reweighted CTDS result, but it does not make any equation reduce to its own input. There are no load-bearing self-citations. Accordingly, no circularity is present.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central results rest on standard stochastic calculus and Langevin theory plus two unmeasured domain assumptions: fast temperature-coordinate mixing and sufficient accuracy of the learned free energy during training. The hand-chosen hyperparameters (beta_min, Delta, Delta', eta, dynamics coefficients) are free parameters for the single reported experiment. No new entities beyond the auxiliary temperature coordinate (borrowed from CT-HMC) are introduced.

free parameters (3)
  • CTDS dynamics coefficients (gamma^x, epsilon^x, gamma^xi, epsilon^xi, M_x, M_xi) = (50.0, 2.0, 5.0, 2.0, 1.0, 1.0)
    Chosen by hand for the 40-mode example (Appendix D.1); no sensitivity analysis, and the empirical comparison depends on these values.
  • temperature reparameterization parameters (beta_min, Delta, Delta') = (0.2, 0.25, 1.9)
    Chosen by hand to define the plateau regions of eq. (30); the proposal's temperature coverage depends on them.
  • confining potential (eta, Delta-tilde) = (10.0, 2.0)
    Chosen to keep xi near the interesting interval; too strong a potential could suppress mixing, too weak could let xi drift.
axioms (5)
  • domain assumption The Langevin dynamics in eq. (17) has the joint density pi-dagger_t proportional to e^{-H^theta_t} as an invariant or quasi-stationary measure when the control satisfies the continuity equation.
    Standard underdamped Langevin result invoked implicitly to justify using marginals of eq. (17) as PINN proposals (Section 3.3).
  • standard math The Girsanov theorem for degenerate diffusions (Theorem C.2, from Liptser & Shiryaev) applies to the CTDS diffusion matrix where noise is injected only into momentum.
    Used in the proof of Theorem C.1 to derive the Radon-Nikodym derivative; requires mild regularity conditions on drifts and a finite integrability condition.
  • domain assumption Fast mixing of the xi (temperature) coordinate during training.
    The empirical claim that CTDS improves exploration presumes the temperature coordinate explores beta in [beta_min, 1] quickly relative to the 1250-epoch training budget; not measured.
  • domain assumption Learned free energy F^theta_t(xi) is sufficiently accurate during training so that the joint proposal pi-hat^theta in eq. (15) assigns nontrivial mass near beta=1 at all times t.
    The authors note samples pool at high/low temperatures when F^theta is poor (Section 3.4, Fig. 2), so the method depends on this bootstrap being stable.
  • standard math Boundary terms vanish at infinity in the proof of Theorem 3.1 (divergence theorem applied to the integral of div(mu e^{-U}) dx).
    Required for the step from eq. (36) to eq. (35) in Appendix B.3; assumes sufficient decay of mu times pi, not stated in the paper.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Continuously Tempered Diffusion Samplers." pith.science (2026). https://pith.science/paper/FGBITTFC

@misc{pith2026250900316,
  author       = {Pith},
  title        = {Pith review of: Continuously Tempered Diffusion Samplers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FGBITTFC}},
  note         = {Machine review of arXiv:2509.00316}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Annealing-based neural samplers seek to amortize sampling from unnormalized distributions by training neural networks to transport a family of densities interpolating from source to target. A crucial design choice in the training phase of such samplers is the proposal distribution by which locations are generated at which to evaluate the loss. Previous work has obtained such a proposal distribution by combining a partially learned transport with annealed Langevin dynamics. However, isolated modes and other pathological properties of the annealing path imply that such proposals achieve insufficient exploration and thereby lower performance post training. To remedy this, we propose continuously tempered diffusion samplers, which leverage exploration techniques developed in the context of molecular dynamics to improve proposal distributions. Specifically, a family of distributions across different temperatures is introduced to lower energy barriers at higher temperatures and drive exploration at the lower temperature of interest. We empirically validate improved sampler performance driven by extended exploration. Code is available at https://github.com/eje24/ctds.

Figures

Figures reproduced from arXiv: 2509.00316 by Bowen Jing, Ezra Erives, Peter Holderrieth, Tommi Jaakkola.

Figure 1
Figure 1. Figure 1: A comparison of samples generated by simulating the ODE given by a control [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of sample inverse tempera￾tures at t = 1.0 obtained by simulating eq. (17) before and after training for the non-Jarzynski￾reweighted CTDS proposal as in section 4.1. Details can be found in appendix C.4. We con￾clude from eq. (22) that we may reweight sam￾ples from the forward process eq. (17) so as obtain the proposal π θ t (x, ξ), which as we re￾call from eq. (15), converges to π ξ t (x)πt(… view at source ↗
Figure 3
Figure 3. Figure 3: Results for samplers trained on the 40-mode Gaussian mixture target as in section 4.1. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An illustration of the inverse temperature reparameterization given in eq. (30). [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Three flavors of stochastic integration. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: fig. 6. For each, we parameterize the learned control [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: The proposals and their accompanying density paths for the 40-mode Gaussian mixture [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 18 canonical work pages · 7 internal anchors

  1. [3]

    URL https: //arxiv.org/abs/2406.07423. Gavin E. Crooks. Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences. Physical Review E, 60(3):2721–2726, September

  2. [4]

    (15) as ˆπθ t (x, ξ) = e−U ξ t (x)+F θ t (ξ)+ψ′ t(ξ) = e− ˜U θ t (x,ξ)

    B.2 B IASING AND CONFINING POTENTIALS Recall the joint density ˆπt(x, ξ) given in eq. (15) as ˆπθ t (x, ξ) = e−U ξ t (x)+F θ t (ξ)+ψ′ t(ξ) = e− ˜U θ t (x,ξ). (31) In practice, and to encourage the ξ-component to stay close to the critical interval [−∆′, ∆′] (see eq. (30)), we may introduce choose ψ′ t(ξ) so as to include a confining potential ψconf(ξ). In...

  3. [6]

    For each, we parameterize the learned control µθ t (x) (µθ t (x, β(ξ)) for CTDS), the free energy F θ t (F θ t (β(ξ)) for CTDS), and the learned potential U θ t as feed-forward neural networks with width 256 and depth three, and using the SiLU non-linearity (Elfwing et al., 2017). We train using a replay buffer, re-sampling once per epoch for 1250 epochs ...

  4. [7]

    MCMC Variational Inference via Uncorrected Hamiltonian Annealing

    URL https://arxiv.org/abs/2107.04150. Tomas Geffner and Justin Domke. Langevin diffusion variational inference,

  5. [9]

    Ulrich HE Hansmann

    URL https://arxiv.org/abs/1704.03338. Ulrich HE Hansmann. Parallel tempering algorithm for conformational studies of biological molecules. Chemical Physics Letters, 281(1-3):140–150,

  6. [10]

    Hiroshi Kunita

    URL https://arxiv.org/abs/1412.6980. Hiroshi Kunita. Stochastic flows and jump-diffusions. Springer,

  7. [13]

    Radford M

    URL https://arxiv.org/abs/2301.07388. Radford M. Neal. Annealed importance sampling,

  8. [15]

    Solving high-dimensional Hamilton-Jacobi-Bellman PDEs using neural networks: perspectives from the theory of controlled diffusions and measures on path space

    URL https://arxiv.org/abs/2005.05409. Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378:686–707,

  9. [16]

    Lorenz Richter, Ayman Boustati, Nikolas N¨usken, Francisco J

    URL https: //arxiv.org/abs/2307.01198. Lorenz Richter, Ayman Boustati, Nikolas N¨usken, Francisco J. R. Ruiz, and ¨Omer Deniz Akyildiz. Vargrad: A low-variance gradient estimator for variational inference,

  10. [19]

    Continuously-Tempered PDMP Samplers

    URL https://arxiv.org/abs/2205.09559. Robert H Swendsen and Jian-Sheng Wang. Replica monte carlo simulation of spin-glasses.Physical review letters, 57(21):2607,

  11. [20]

    Achille Thin, Nikita Kotelevskii, Arnaud Doucet, Alain Durmus, Eric Moulines, and Maxim Panov

    URL https: //arxiv.org/abs/2006.10739. Achille Thin, Nikita Kotelevskii, Arnaud Doucet, Alain Durmus, Eric Moulines, and Maxim Panov. Monte carlo variational auto-encoders,

  12. [21]

    Monte Carlo Variational Auto-Encoders

    URL https://arxiv.org/abs/ 2106.15921. Yifeng Tian, Nishant Panda, and Yen Ting Lin. Liouville flow importance sampler,

  13. [22]

    Suriyanarayanan Vaikuntanathan and Christopher Jarzynski

    URL https://arxiv.org/abs/2405.06672. Suriyanarayanan Vaikuntanathan and Christopher Jarzynski. Escorted free energy simulations: Im- proving convergence by reducing dissipation. Physical Review Letters, 100(19):190601,

  14. [24]

    Denoising diffusion samplers,

    10 Published in the Frontiers in Probabilistic Inferenceworkshop at ICLR 2025 Francisco Vargas, Will Grathwohl, and Arnaud Doucet. Denoising diffusion samplers,

  15. [25]

    Sifan Wang, Shyam Sankaran, Hanwen Wang, and Paris Perdikaris

    URL https://arxiv.org/abs/2302.13834. Sifan Wang, Shyam Sankaran, Hanwen Wang, and Paris Perdikaris. An expert’s guide to training physics-informed neural networks,

  16. [26]

    Hao Wu, Jonas K ¨ohler, and Frank No ´e

    URL https://arxiv.org/abs/2308.08468. Hao Wu, Jonas K ¨ohler, and Frank No ´e. Stochastic normalizing flows,

  17. [27]

    Guodong Zhang, Kyle Hsu, Jianing Li, Chelsea Finn, and Roger B Grosse

    URL https:// arxiv.org/abs/2002.06707. Guodong Zhang, Kyle Hsu, Jianing Li, Chelsea Finn, and Roger B Grosse. Differentiable annealed importance sampling and the perils of gradient noise.Advances in Neural Information Processing Systems, 34:19398–19410,

  18. [28]

    Adrianne Zhong, Ben Kuznets-Speck, and Michael R

    URL https://arxiv.org/abs/2111.15141. Adrianne Zhong, Ben Kuznets-Speck, and Michael R. DeWeese. Time-asymmetric fluctuation the- orem and efficient free-energy estimation. Physical Review E , 110(3), September

  19. [29]

    doi: 10.1103/physreve.110.034121

    ISSN 2470-0053. doi: 10.1103/physreve.110.034121. URL http://dx.doi.org/10.1103/ PhysRevE.110.034121. A A B RIEF OVERVIEW OF PINN O BJECTIVES We briefly derive the PINN objective from eq. (8). Recall the the continuity equation ∂tπt(xt) = −∇ · µθ t (xt)πt(xt) , ∀(xt, t) ∈ Rd × T, (24) from eq. (5). Dividing both sides by πt(xt) yields ∂t log πt(xt) = −∇µθ...

  20. [31]

    (35) However, by eq

    (34) Now, observe that the true free energy Ft(ξ) satisfies ∂tFt(ξ) = −∂t log Zt(ξ) = − R Rd ∂te−U ξ t (x) dx Zt(ξ) = R Rd (∂tU ξ t (x))e−U ξ t (x) dx Zt(ξ) . (35) However, by eq. (34), we have that ∂tU ξ t (x, ξ) = ∂tF ⋆ t (ξ) − ∇x · µ⋆ t (x, ξ) + (∇xU ξ t )T µ⋆ t (x, ξ) = ∂tF ⋆ t (ξ) − eUt(x)∇x h µ⋆ t (x, ξ)e−U ξ t (x) i . (36) Plugging eq. (36) into eq...

  21. [32]

    First, as is directly observed from the discretizations shown in fig

    for a more technical treatment. First, as is directly observed from the discretizations shown in fig. 5, we may relate the forward and backward Itˆo integrals to a corresponding Stratonovich integral via Z T 0 Xt(dYt + ← −d Yt) = 2 Z T 0 Xt ◦ dYt. (39) Second, and less intuitively obvious, is the fact that for X = − →X from eq. (37), Z T 0 µt(Xt)(← −d Xt ...

  22. [33]

    (59) and eq

    We may therefore plug in eq. (59) and eq. (67) into eq. (68) to obtain log dQT dPT (X) = log dπT dΛ (XT ) − log dπ0 dΛ (X0) (69) + Z T 0 1 2 (−Ct + νt)T (ΓtEt)−1(−← −d Pt− 1 2 (−Ct − νt) dt) (70) − Z T 0 1 2 (−At − νt)T (ΓtEt)−1(dPt− 1 2 (−At + νt) dt). (71) 17 Published in the Frontiers in Probabilistic Inferenceworkshop at ICLR 2025 We may now simplify ...

  23. [35]

    Additionally, we follow the lead of (Wang et al., 2023; Albergo & Vanden-Eijnden,

    with learning rate 1 × 10−3, reducing by a factor of γ = 0 .97 every 1000 iterations, and after an initial burn-in period of 15000 training iterations. Additionally, we follow the lead of (Wang et al., 2023; Albergo & Vanden-Eijnden,

  24. [36]

    For CTDS, we reparameterize using eq

    in utilizing curriculum-based training whereby at T ∈ {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9} (in that order) we spend {1000, 1000, 1000, 1000, 2000, 2000, 2000, 3000, 3000, 3000} iterations respectively, and the remaining iterations at T = 1.0. For CTDS, we reparameterize using eq. (30) with βmin = 0.2, ∆ = 0 .25, and ∆′ = 1.9, and additionally set...

  25. [37]

    Sampling Details

    to encode position, time, and temperature with 100, 20, and 20 features respectively, drawn from isotropic Gaussians with standard deviations 0.1, 5, and 1, respectively. Sampling Details. In practice, we discretize using an Euler solver with with ∆t = 0 .004 (250 timesteps). Letting ϕθ : Rd → Rd, ϕθ : X0 7→ X1 denote the corresponding unit-time flow of e...

  26. [1992]

    doi: 10.1209/0295-5075/19/6/002

    ISSN 1286-4854. doi: 10.1209/0295-5075/19/6/002. URL http://dx.doi.org/10.1209/0295-5075/19/6/002. 9 Published in the Frontiers in Probabilistic Inferenceworkshop at ICLR 2025 Alexander G. D. G. Matthews, Michael Arbel, Danilo J. Rezende, and Arnaud Doucet. Continual re- peated annealed flow transport monte carlo,

  27. [1998]

    Radford M Neal

    URL https://arxiv.org/abs/ physics/9803008. Radford M Neal. Mcmc using hamiltonian dynamics. arXiv preprint arXiv:1206.1901,

  28. [1999]

    Physics-informed neural networks for sampling

    Jingtong Sun, Julius Berner, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural networks for sampling. In ICLR 2024 Workshop on AI4DifferentialEquations In Science. Matthew Sutton, Robert Salomone, Augustin Chevallier, and Paul Fearnhead. Continuously- tempered pdmp samplers,

  29. [2008]

    Transport, variational inference and diffusions: with appli- cations to annealed flows and schr\” odinger bridges

    Francisco Vargas and Nikolas N¨usken. Transport, variational inference and diffusions: with appli- cations to annealed flows and schr\” odinger bridges. arXiv preprint arXiv:2307.01050,

  30. [2017]

    Mingzhou Fan, Ruida Zhou, Chao Tian, and Xiaoning Qian

    URL https://arxiv.org/abs/ 1702.03118. Mingzhou Fan, Ruida Zhou, Chao Tian, and Xiaoning Qian. Path-guided particle-based sam- pling. In Forty-first International Conference on Machine Learning ,

  31. [2019]

    Thermostat-assisted continuously-tempered Hamiltonian Monte Carlo for Bayesian learning

    URL https: //arxiv.org/abs/1711.11511. E Marinari and G Parisi. Simulated tempering: A new monte carlo scheme. Europhysics Letters (EPL), 19(6):451–458, July

  32. [2020]

    Yuji Sugita and Yuko Okamoto

    URL https:// arxiv.org/abs/2010.10436. Yuji Sugita and Yuko Okamoto. Replica-exchange molecular dynamics method for protein folding. Chemical physics letters, 314(1-2):141–151,

  33. [2021]

    Julius Berner, Lorenz Richter, and Karen Ullrich

    URL https://arxiv.org/abs/2102.07501. Julius Berner, Lorenz Richter, and Karen Ullrich. An optimal control perspective on diffusion-based generative modeling. arXiv preprint arXiv:2211.01364,

  34. [2022]

    Score-Based Diffusion meets Annealed Importance Sampling

    URL https://arxiv.org/ abs/2208.07698. Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,

  35. [2023]

    Langevin Diffusion Variational Inference

    URL https: //arxiv.org/abs/2208.07743. Gianpaolo Gobbo and Benedict J Leimkuhler. Extended hamiltonian approach to continuous tem- pering. Physical Review E, 91(6):061301,

  36. [2024]

    8 Published in the Frontiers in Probabilistic Inferenceworkshop at ICLR 2025 Michael Arbel, Alexander G

    URL https://arxiv.org/abs/2410.02711. 8 Published in the Frontiers in Probabilistic Inferenceworkshop at ICLR 2025 Michael Arbel, Alexander G. D. G. Matthews, and Arnaud Doucet. Annealed flow transport monte carlo,

  37. [3787]

    URL http://dx.doi.org/10.1103/PhysRevE

    doi: 10.1103/physreve.60.2721. URL http://dx.doi.org/10.1103/PhysRevE. 60.2721. Pierre Del Moral, Arnaud Doucet, and Ajay Jasra. Sequential monte carlo samplers. Journal of the Royal Statistical Society Series B: Statistical Methodology, 68(3):411–436,

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.