Sequential Controlled Langevin Diffusions

• SCLD is a class of methods that use time-inhomogeneous, controlled Langevin diffusions to transport particles from an easy prior to a complex target distribution.
• It integrates diffusion-based sampling with sequential Monte Carlo techniques by partitioning trajectories, computing importance weights, and applying resampling and MCMC refinements.
• The approach enhances sampling efficiency and stability in high-dimensional or multimodal problems while mitigating issues like mode collapse and excessive training iterations.

Sequential Controlled Langevin Diffusions (SCLD) are a class of sampling and control constructions in which a time-inhomogeneous Langevin diffusion is used to transport particles from an easy prior toward a target distribution, while sequential corrections such as importance weighting, resampling, optional Markov chain Monte Carlo refinement, or state-dependent guidance are applied along the trajectory. In the explicit formulation introduced by "Sequential Controlled Langevin Diffusions" (Chen et al., 2024), SCLD combines Sequential Monte Carlo (SMC) and diffusion-based samplers in a continuous-time, path-space framework. Closely related later works interpret diffusion-path samplers and general annealed Langevin schemes through the same lens (Young et al., 29 Jan 2026, Habring et al., 29 Jan 2026), and a related line on controllable generative diffusion recasts classifier-guided reverse diffusion as sequential control of a Langevin-type process under a KL-regularized objective, although it does not use the term SCLD explicitly (Oertell et al., 27 May 2025).

1. Definition and conceptual placement

The core sampling problem is to draw from a target distribution

$\pi(x)\propto \tilde{\pi}(x),$

or, in an equivalent potential form used in related analyses,

$\pi(x)\propto e^{-V_\pi(x)},$

when the normalizing constant is unknown and samples from the target are unavailable (Chen et al., 2024, Young et al., 29 Jan 2026). The explicit SCLD formulation positions itself between two established paradigms. Classical SMC or annealed importance sampling transports a population of particles through intermediate tempered distributions using prescribed Markov kernels and resampling; it is asymptotically exact and robust, but its hand-crafted transitions can mix slowly. Diffusion-based samplers instead learn a drift for an SDE that transports an easy prior to the target; they are flexible and target-adaptive, but can be expensive or unstable to train, and can be prone to mode collapse (Chen et al., 2024).

SCLD is defined by combining these two viewpoints in continuous time and on path space. The method uses a controlled Langevin diffusion as the forward transport mechanism, splits trajectories into subtrajectories, computes path-space importance weights through a Girsanov-type Radon–Nikodym derivative, and interleaves resampling and optional MCMC steps at chosen times (Chen et al., 2024). In that formulation, SCLD is an SMC algorithm whose Markov kernels are learned controlled Langevin diffusions, trained end-to-end with a path-space variational loss.

A longer-range theoretical precursor appears in the variational characterization of Langevin–Smoluchowski diffusions, where time reversal followed by drift control yields a sequence of stochastic control problems whose values decrease along the relative entropy to the Gibbs measure (Karatzas et al., 2020). This earlier result does not define SCLD algorithmically, but it places sequential drift control of Langevin dynamics on a path-space, entropy-based foundation.

2. Controlled Langevin formulation on path space

The canonical forward SDE in the explicit SCLD framework is

$$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$

with control $u(x,t)\in\mathbb{R}^d$ and scalar diffusion coefficient $\sigma(t)$ (Chen et al., 2024). The method uses a Langevin-like parameterization

$$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$

where $\widetilde{u}_\theta$ is a neural network and $\pi(\cdot,t)$ is a prescribed path of intermediate distributions,

$$\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.$$

If $\widetilde{u}_\theta\equiv 0$, this reduces to annealed Langevin dynamics along the prescribed intermediates (Chen et al., 2024).

The corresponding reverse-time construction is tied to the same path $\pi(x)\propto e^{-V_\pi(x)},$0. With

$\pi(x)\propto e^{-V_\pi(x)},$1

the reverse SDE is

$\pi(x)\propto e^{-V_\pi(x)},$2

The forward and backward laws are measures on path space, and the central object is the Radon–Nikodym derivative between them on a subinterval $\pi(x)\propto e^{-V_\pi(x)},$3: $\pi(x)\propto e^{-V_\pi(x)},$4 On a full trajectory,

$\pi(x)\propto e^{-V_\pi(x)},$5

This factorization is the continuous-time analogue of incremental importance weights in SMC (Chen et al., 2024).

Related work broadens the same controlled-diffusion perspective. One line studies the time-inhomogeneous Langevin diffusion

$\pi(x)\propto e^{-V_\pi(x)},$6

where the control is an annealing schedule $\pi(x)\propto e^{-V_\pi(x)},$7 selecting an intermediate target $\pi(x)\propto e^{-V_\pi(x)},$8 at each time (Habring et al., 29 Jan 2026). Another line defines a diffusion path $\pi(x)\propto e^{-V_\pi(x)},$9 between a Gaussian base and the target, and runs diffusion-annealed Langevin dynamics

$$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$0

so that the controlled drift is the time-varying score of the path (Young et al., 29 Jan 2026).

3. Sequential structure: subtrajectories, weights, and resampling

The sequential aspect of SCLD is explicit. A time grid

$$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$1

partitions the diffusion into subtrajectories. On each interval $$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$2, the sampler simulates the forward SDE as a Markov kernel, computes the path-space subtrajectory weight $$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$3, updates cumulative weights via

$$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$4

and optionally resamples particles and applies an MCMC refinement kernel invariant for $$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$5 (Chen et al., 2024).

The practical implementation uses Euler–Maruyama. If $$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$6 discretization steps are used within each subtrajectory and $$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$7, then

$$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$8

The forward and backward one-step transition densities are Gaussian, and the discrete subtrajectory weight is approximated by

$$dX_t^u = u(X_t^u,t)\,dt + \sigma(t)\,dW_t,\qquad X_0^u\sim p_{\text{prior}},$$9

Effective sample size is monitored through

$u(x,t)\in\mathbb{R}^d$0

and if $u(x,t)\in\mathbb{R}^d$1 with $u(x,t)\in\mathbb{R}^d$2, multinomial resampling is performed and the weights are reset to uniform. After resampling, one Hamiltonian Monte Carlo step with 10 leapfrog updates can be applied as an additional refinement (Chen et al., 2024).

This yields a standard SMC structure with unusual forward kernels: the transport between resampling times is no longer a fixed MCMC transition but a learned stochastic flow. The same sequential viewpoint appears in diffusion-path samplers based on auxiliary SMC. There, at each main Langevin step, auxiliary particles approximate a conditional distribution $u(x,t)\in\mathbb{R}^d$3, and their weighted empirical average yields the score estimate needed for the next controlled Langevin update (Young et al., 29 Jan 2026). This suggests that “sequential” in SCLD can refer either to particle-level importance-resampling across path segments or to online estimation of the feedback drift along a prescribed path.

4. Training objectives, control learning, and theoretical guarantees

In the explicit SCLD sampler, three objects are learned jointly: the control $u(x,t)\in\mathbb{R}^d$4 through $u(x,t)\in\mathbb{R}^d$5, the annealing schedule $u(x,t)\in\mathbb{R}^d$6, and the Gaussian prior

$u(x,t)\in\mathbb{R}^d$7

The annealing schedule is parameterized on the discrete grid by

$u(x,t)\in\mathbb{R}^d$8

Training minimizes a sum of subtrajectory divergences,

$u(x,t)\in\mathbb{R}^d$9

with $\sigma(t)$0 (Chen et al., 2024).

Two divergence choices are discussed. A KL-based objective,

$\sigma(t)$1

is natural but problematic in high dimension when importance sampling is needed. The default choice is the log-variance divergence,

$\sigma(t)$2

estimated empirically by

$\sigma(t)$3

and summed over subtrajectories: $\sigma(t)$4 Trajectories are detached from the computation graph, and gradients are taken only through $\sigma(t)$5, not through the SDE integrator (Chen et al., 2024).

The theory attached to this construction has several layers. The path-space identity implies unbiased importance sampling with exact weights, while self-normalized estimators are biased but consistent as the number of particles $\sigma(t)$6. If all subtrajectory divergences vanish, then the global divergence between forward and backward path measures is zero. For KL training with importance sampling, the relative error of the Monte Carlo estimator grows exponentially with dimension through a $\sigma(t)$7-divergence bound, whereas for the LV divergence one has

$\sigma(t)$8

and the variance of its Monte Carlo estimator is $\sigma(t)$9 (Chen et al., 2024).

A complementary convergence theory applies to the broader class of time-inhomogeneous Langevin diffusions

$$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$0

Under uniform smoothness, dissipativity, and time-regularity assumptions on the intermediate potentials, the forward KL to the final target satisfies

$$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$1

and if $$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$2 with $$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$3, then

$$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$4

For the Euler–Maruyama discretization,

$$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$5

if $$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$6 and $$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$7, then

$$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$8

under the corresponding step-size and regularity conditions (Habring et al., 29 Jan 2026). These results make the control-versus-tracking tradeoff explicit: better intermediate log-Sobolev constants improve mixing, while rapid schedule changes contribute tracking error.

5. Related formulations and neighboring uses of the SCLD viewpoint

One major adjacent development is the diffusion-path sampler implemented via auxiliary SMC (Young et al., 29 Jan 2026). There the path between a Gaussian base $$u(x,t) = \sigma(t)^2\,\widetilde{u}_\theta(x,t) \;+\; \frac{\sigma(t)^2}{2}\,\nabla\log \pi(x,t),$$9 and the target $\widetilde{u}_\theta$0 is defined by

$\widetilde{u}_\theta$1

with law $\widetilde{u}_\theta$2. The controlled dynamics are diffusion-annealed Langevin dynamics,

$\widetilde{u}_\theta$3

and the score $\widetilde{u}_\theta$4 is estimated from an auxiliary conditional posterior

$\widetilde{u}_\theta$5

The resulting estimator uses SMC particles targeting $\widetilde{u}_\theta$6 and score identities such as

$\widetilde{u}_\theta$7

and

$\widetilde{u}_\theta$8

Control variates are introduced through matrices $\widetilde{u}_\theta$9 that combine these identities. The scalar optimum is

$\pi(\cdot,t)$0

and the matrix optimum is

$\pi(\cdot,t)$1

where $\pi(\cdot,t)$2. The score estimation error bound

$\pi(\cdot,t)$3

feeds into a final KL guarantee for the approximate sampler (Young et al., 29 Jan 2026).

A different but related interpretation appears in controllable generative diffusion. The KL-regularized objective

$\pi(\cdot,t)$4

has optimal solution

$\pi(\cdot,t)$5

The reverse SDE is controlled by a guidance field $\pi(\cdot,t)$6,

$\pi(\cdot,t)$7

and the optimal guidance is

$\pi(\cdot,t)$8

The SLCD procedure learns an approximation to the reward distribution with DAgger-style online data aggregation and obtains a no-regret convergence guarantee in KL to the optimal tilted target. This work does not use the label SCLD explicitly, but it treats the reverse-time diffusion as a sequentially controlled Langevin-type process whose control is applied at every diffusion step (Oertell et al., 27 May 2025).

6. Empirical behavior, limitations, and significance

The explicit SCLD sampler was evaluated on 11 targets spanning Bayesian statistics and synthetic multimodal or high-dimensional distributions (Chen et al., 2024). On the ELBO tasks, it achieves the best ELBO on Seeds, Sonar, Credit, and Brownian, and is competitive on LGCP. On the Sinkhorn tasks, the reported distances include GMM40 at approximately $\pi(\cdot,t)$9 versus the best baseline at approximately $$\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.$$0, MoS at approximately $$\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.$$1 versus the best baseline at approximately $$\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.$$2, and Robot1/4 at approximately $$\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.$$3 versus baselines at least $$\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.$$4. The same study reports that SCLD uses many fewer training iterations than CMCD, with examples of $$\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.$$5 versus $$\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.$$6 gradient steps, while often outperforming or matching it, and the abstract states that improved performance is reached on multiple benchmark problems, in many cases using only $$\pi(x,t) \propto p_{\text{prior}}(x)^{1-\beta(t)} \,\rho_{\text{target}}(x)^{\beta(t)},\quad \beta(0)=0,\ \beta(T)=1.$$7 of the training budget of previous diffusion-based samplers (Chen et al., 2024).

The reported ablations show that increasing the number of SMC subtrajectories is typically beneficial in training and evaluation, that even adding SMC steps only at test time improves diffusion-only methods, that removing MCMC or resampling degrades performance though the method remains competitive, and that replay buffers improve training stability and sample quality, especially in high dimension (Chen et al., 2024). In the diffusion-path formulation, control variate schedules reduce score-estimation mean squared error across the path, improving stability for anisotropic targets, and tempering of auxiliary posteriors helps exploration when initialization is weak (Young et al., 29 Jan 2026). In the forward-KL analysis of annealed Langevin diffusions, convolutional and Moreau-envelope paths dominate plain ULA, geometric tempering, and dilation on multimodal Gaussian mixtures, with faster KL decay and better mode coverage (Habring et al., 29 Jan 2026).

The limitations are correspondingly structural. The explicit SCLD sampler requires gradients of the prior and unnormalized target, simulating many particles with fine time discretization can be computationally heavy, and tuning of noise schedules, step sizes, the number of subtrajectories, and HMC parameters remains necessary (Chen et al., 2024). The diffusion-path formulation requires many auxiliary particles for accurate score estimates, uses a fixed path design, and its non-asymptotic bounds are conservative in practice (Young et al., 29 Jan 2026). The forward-KL theory is restricted to gradient drifts and assumes intermediate targets satisfy uniform smoothness, dissipativity, and log-Sobolev control (Habring et al., 29 Jan 2026).

Taken together, these results place SCLD at the intersection of annealed Monte Carlo, stochastic control, path-space variational inference, and modern diffusion sampling. In its most explicit form, it is an SMC method whose transition kernels are learned controlled Langevin diffusions and whose correction mechanism is path-space importance weighting with resampling. In broader usage, it also names a control perspective in which a time-dependent Langevin drift is designed, learned, or estimated sequentially so that the induced law tracks a path of easier intermediate distributions and converges to the target.