Recursive Flow Matching: Multi-Scale Modeling

• RecFM is a generative modeling framework that extends flow matching to multi-scale and sequential Bayesian filtering for forecasting complex spatiotemporal dynamics.
• It enforces cross-scale self-consistency by aligning velocity fields across discretization steps to control error and ensure high-fidelity, few-step sampling.
• RecFM achieves significant inference speedups and state-of-the-art performance on scientific benchmarks, making it ideal for real-time forecasting and online prediction.

Recursive Flow Matching (RecFM) refers to a family of generative modeling frameworks for forecasting complex spatiotemporal dynamics and implementing efficient sequential inference for stochastic systems. RecFM methods build on flow-matching (FM) techniques, enforcing either self-consistency across discretization or principled Bayesian filtering recursions to yield high-fidelity and computationally efficient generation or streaming prediction. The methodology is detailed in foundational works ["Recursive Flow Matching" (Huang et al., 26 May 2026)] and ["Accelerated Sequential Flow Matching: A Bayesian Filtering Perspective" (Huang et al., 5 Feb 2026)].

1. Core Principles and Mathematical Formulation

Recursive Flow Matching generalizes standard flow matching by aligning continuous-time trajectories not only across time but also across multiple scales or sequential prediction steps. In classic FM, a neural velocity field $v_t(x;\theta)$ parameterizes an ODE transporting samples from a data distribution $p_0$ to a prior $p_1$, minimizing

$$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$

with $x_t = (1-t)x_0 + t x_1$ and $u_t$ as the velocity under optimal transport.

RecFM introduces an additional "scale parameter" $\alpha \in (0,1]$, generating a family of secondary trajectories from $x_0$ to partial targets $x_\alpha = (1-\alpha)x_0 + \alpha x_1$ in rescaled time $\tau = t/\alpha$. The target velocities for these auxiliary flows are $p_0$0 at the same spatial interpolation point. The learning objective includes both recovery of the appropriate velocity and explicit cross-scale self-consistency constraints. The RecFM loss is

$p_0$1

where $p_0$2 is the recursion depth, $p_0$3, and $p_0$4. The first term supervises velocity recovery at each scale; the second enforces multiscale self-consistency, tightly constraining discretization error and trajectory curvature (Huang et al., 26 May 2026).

In the streaming/online setting, RecFM reframes sequential inference as learning a probability flow that transports the posterior from one step to the next. At each time $p_0$5, observations $p_0$6 define a Bayesian filtering recursion; RecFM learns a velocity field mapping the previous posterior directly to the updated posterior, using the prior sample as a "warm start" and occasionally "re-noising" to control error accumulation (Huang et al., 5 Feb 2026).

2. Algorithmic Workflow

The training iteration of RecFM in the scale-consistent formulation proceeds as follows:

1. Sample $p_0$7.
2. Compute $p_0$8.
3. For $p_0$9:
   • Set $p_1$0.
   • Compute $p_1$1, accumulate $p_1$2.
4. For $p_1$3, accumulate the consistency penalty $p_1$4.
5. Update $p_1$5 via gradient descent (Huang et al., 26 May 2026).

In the recursive Bayesian filtering context, the inference at step $p_1$6 consists of:

• Using the posterior sample from $p_1$7 as the base, possibly after "re-noising" (setting $p_1$8, $p_1$9).
• Solving the learned ODE $$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$0 from $$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$1 to $$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$2, where $$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$3, yielding the new posterior sample $$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$4 (Huang et al., 5 Feb 2026).

3. Theoretical Properties

RecFM provides theoretical error guarantees in both the scientific emulation and streaming contexts:

• The cross-scale PDE constraint $$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$5 controls the magnitude of temporal derivatives and, by the Euler method error bound, directly tightens the upper bound on terminal discretization error. Specifically, Theorem 1 in (Huang et al., 26 May 2026) gives

$$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$6

where $$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$7 is the trajectory acceleration.

• In the recursive filtering context, the law of total variance yields strictly lower one-step Wasserstein error for RecFM compared to sampling from a fixed base;

$$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$8

$$\mathcal{L}_{\rm FM}(\theta) = \mathbb{E}_{t, x_0, x_1} \| v_t(x_t; \theta) - u_t(x_t | x_0, x_1) \|^2,$$9

so recursive "warm-start" updates reduce generation error (Huang et al., 5 Feb 2026).

• Global minimizers of the RecFM loss recover the exact conditional velocity field; secondary trajectories preserve correct marginals. Each scale within the multiscale family yields a valid few-step generative sampler for the interpolant distribution (Huang et al., 26 May 2026).

4. Computational Complexity and Efficiency

RecFM training involves $x_t = (1-t)x_0 + t x_1$0 velocity evaluations per iteration (for $x_t = (1-t)x_0 + t x_1$1 recursion depths), but the number of training iterations is divided by $x_t = (1-t)x_0 + t x_1$2, yielding effective cost comparability to vanilla FM. At inference, RecFM produces high-fidelity results in as few as $x_t = (1-t)x_0 + t x_1$3 or 2 ODE steps, in contrast to $x_t = (1-t)x_0 + t x_1$4 for diffusion-based emulators.

On the VideoPDE benchmark, RecFM realizes up to $x_t = (1-t)x_0 + t x_1$5 end-to-end rollout speedup on NVIDIA L40S GPUs relative to leading diffusion models (Huang et al., 26 May 2026). In streaming applications such as state estimation, planning, and scientific forecasting, RecFM achieves performance competitive with full-step diffusion samplers but requires only one (or a few) ODE steps per update and thus delivers an order-of-magnitude reduction in inference latency (Huang et al., 5 Feb 2026).

The "re-noise level" $x_t = (1-t)x_0 + t x_1$6 in the recursive setting controls bias-variance trade-offs; $x_t = (1-t)x_0 + t x_1$7 retains full information from the previous sample but risks bias accumulation, while $x_t = (1-t)x_0 + t x_1$8 discards all prior information. Empirically, optimal $x_t = (1-t)x_0 + t x_1$9 lies in the interval $u_t$0, depending on system stochasticity (Huang et al., 5 Feb 2026).

5. Empirical Results and Practical Considerations

On scientific benchmarks (sea-surface temperature forecasting, Navier-Stokes flow, Helmholtz Staircase), RecFM achieves state-of-the-art metrics for one- and few-step generation:

Benchmark	RecFM MSE	Diffusion MSE	RecFM CRPS	Diffusion CRPS	RecFM Inference Time	FM Inference Time	VideoPDE Time
SST	0.162	0.161–0.177	0.217	n.a.	0.43 s	n.a.	n.a.
Navier-Stokes	0.0064	0.0076 (FM)	0.031	n.a.	1.59 s	6.9 s	72 s
Helmholtz S.	4.2×10⁻⁵	5.6×10⁻⁴	0.0034	n.a.	1.59 s	n.a.	19.8 s

RecFM reduces MSE by over 15% versus vanilla FM and can be run in single step per rollout (Huang et al., 26 May 2026). In the streaming paradigm, RecFM matches or exceeds multi-step diffusion and learning-based baselines across forecasting, planning, and state estimation domains, often achieving comparable accuracy with significantly reduced function evaluations (Huang et al., 5 Feb 2026).

Practical recommendations include:

• Selecting recursion depth $u_t$1 as optimal for performance versus memory consumption.
• Setting self-consistency penalty $u_t$2.
• Using transformer-based velocity backbones (e.g. HV-DiT) without substantial architectural changes.
• No warm-up is required for inference; few steps (often one) suffice.
• For filtering, managing the re-noise parameter $u_t$3 to balance information retention and bias mitigation.

6. Connections to Related Methodologies

RecFM generalizes flow matching to multi-scale and sequential settings, addressing core speed-vs-fidelity bottlenecks in generative scientific emulation. By embedding cross-scale self-consistency, RecFM merges the error control of multi-step ODE solvers with the sampling efficiency of one-step flow matching (Huang et al., 26 May 2026). It directly improves on previous fast-sampling flow baselines (such as consistency and MeanFlow models) and achieves acceleration over CL-Diffusion and autoregressive schemes.

The recursive, Bayesian formulation unifies streaming generative modeling with classical state-space filtering, enabling direct learning of belief update operators for online inference (Huang et al., 5 Feb 2026). The Wasserstein error reduction and sequential variance compounding further distinguish RecFM from earlier autoregressive and full-restart frameworks.

7. Theoretical Guarantees and Limitations

The foundational works provide both quantitative and qualitative guarantees. The cross-scale PDE constraint bounds discretization errors and theoretically stabilizes few-step sampling. Law-of-total-variance arguments establish provable advantages of recursive warm-start sampling over fixed-base alternatives. Multi-step propagation of error is controlled under Lipschitz continuity assumptions for the velocity field, with total error bounded by the sum of per-step Wasserstein distances (Huang et al., 5 Feb 2026).

No explicit convergence rates are established for the stochastic optimization in RecFM, but consistency follows as in standard flow matching when parameter limits and step size limits are taken jointly. Practical limitations include hyperparameter sensitivity in the scaling depth and noise parameter, as well as increased training memory due to multi-scale velocity evaluation.

RecFM thus constitutes a theoretically grounded and computationally scalable advance in flow-based generative and sequential modeling, aligning ODE-based simulation with high-throughput scientific and online inference requirements.