Time-Reversed Flow Matching

• Time-Reversed Flow Matching is a methodology that learns a reverse-time vector field to transport an unknown data distribution to a Gaussian reference, useful for anomaly detection and online reinforcement learning.
• The anomaly detection formulation uses reverse dynamics to distinguish normal from anomalous samples by measuring displacement magnitudes, achieving state-of-the-art efficiency on benchmarks like the MVTec dataset.
• The approach addresses theoretical challenges such as non-invertibility and high-dimensional triviality, and introduces methods like Worst Transport Flow to overcome these issues in practice.

Searching arXiv for papers on time-reversed/reverse flow matching to ground the article in the cited literature. Time-Reversed Flow Matching (rFM) denotes reverse-time or reverse-inference formulations of flow matching in which a time-dependent vector field is learned from observed data or intermediate noisy states so as to transport toward a specified reference or target distribution. In current arXiv usage, the term appears in two technically distinct but conceptually related settings: as a mapping from an unknown data distribution to a standard Gaussian for unsupervised anomaly detection and localization, and as a unified framework for training diffusion and flow policies in online reinforcement learning when direct samples from the target distribution are unavailable (Li et al., 7 Aug 2025, Li et al., 13 Jan 2026).

1. Core formulation

In the anomaly-detection formulation, rFM begins with an unknown data distribution $p^*(x)$ supported on $\mathbb{R}^d$ and fixes a standard Gaussian prior $p_1(x)=\mathcal{N}(0,I)$. The task is to learn a time-dependent vector field $v_t(x)$ such that the marginal density

$p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$

evolves from $p_0=p^*$ to $p_1=\mathcal{N}(0,I)$. The conditional path is a linear-interpolation Gaussian,

$$p_t(x\mid x_0)=\mathcal{N}(x;\mu_t(x_0),\sigma_t(x_0)^2I),$$

with $\mu_t(x_0)=(1-t)x_0$, $\sigma_t(x_0)=t$, and $\mathbb{R}^d$0. Reparameterizing as $\mathbb{R}^d$1 yields the conditional vector field

$\mathbb{R}^d$2

which becomes $\mathbb{R}^d$3 after substituting the specified $\mathbb{R}^d$4 and $\mathbb{R}^d$5. The learned network $\mathbb{R}^d$6 is trained by the conditional flow-matching loss

$\mathbb{R}^d$7

with $\mathbb{R}^d$8. At inference, a discrete ODE solver is applied from $\mathbb{R}^d$9 backward to $p_1(x)=\mathcal{N}(0,I)$0 to map data points into the Gaussian prior (Li et al., 7 Aug 2025).

In the online-RL formulation, the reverse perspective is not a map from data to Gaussian but a posterior-mean estimation problem. Standard conditional flow matching minimizes

$p_1(x)=\mathcal{N}(0,I)$1

In online RL, however, $p_1(x)=\mathcal{N}(0,I)$2 is an unnormalized Boltzmann target, so direct sampling from $p_1(x)=\mathcal{N}(0,I)$3 is unavailable. Reverse Flow Matching therefore treats $p_1(x)=\mathcal{N}(0,I)$4 as observed and infers latent $p_1(x)=\mathcal{N}(0,I)$5 or $p_1(x)=\mathcal{N}(0,I)$6 through the posterior

$p_1(x)=\mathcal{N}(0,I)$7

leading to a loss defined under posterior sampling rather than direct endpoint sampling (Li et al., 13 Jan 2026).

2. Reverse flow matching for unsupervised anomaly detection

The anomaly-detection version of rFM is presented as a new paradigm for unsupervised anomaly detection and localization using Flow Matching, with the explicit goal of addressing the model expressivity limitations of conventional flow-based methods. The formulation maps the unknown data distribution toward a standard Gaussian through vector field regression along a predefined probability path, and it is positioned as the first successful application of Flow Matching to unsupervised anomaly detection (Li et al., 7 Aug 2025).

A central operational idea is that normal and anomalous samples should behave differently under the learned reverse dynamics. When combined with the later Worst Transport construction, the learned dynamics are trained so that anomaly-free samples are associated with near-zero vector fields and remain clustered near the origin, whereas anomalous samples produce non-zero residuals and exhibit larger displacement under a single-step or few-step solver. This provides a separation mechanism in which anomaly scores are derived from displacement magnitude rather than from expensive likelihood computation or multi-scale training. The reported empirical outcome is state-of-the-art single-scale performance on the MVTec dataset, together with robustness across feature extractors and substantially lower computational cost because inference can use single-step ODE integration rather than 20–400 steps (Li et al., 7 Aug 2025).

The significance of this formulation lies in its use of flow-matching machinery for anomaly scoring without relying on an explicit likelihood proxy. In the paper’s presentation, the reverse map is not primarily a generative mechanism but a dynamical test of whether a feature lies on the anomaly-free training manifold.

3. Non-invertibility and trivial vector fields

The first theoretical observation is that Flow Matching with linear interpolation probability paths is inherently non-invertible in the reverse construction. With $p_1(x)=\mathcal{N}(0,I)$8 and $p_1(x)=\mathcal{N}(0,I)$9, one has $v_t(x)$0, so the factor $v_t(x)$1 diverges at $v_t(x)$2. Accordingly, $v_t(x)$3 is undefined by division-by-zero, the continuity equation breaks down at the start of the reverse mapping, and any attempt to start the reverse ODE there leads to uncontrolled approximation error and makes the mapping non-invertible in principle (Li et al., 7 Aug 2025).

The second theoretical observation concerns high-dimensional behavior. Even if the $v_t(x)$4 singularity is ignored numerically, reverse FM trajectories in high dimension collapse to a “mean-field” direction that is almost radially inward. The argument uses the Gaussian Annulus Theorem: in $v_t(x)$5 dimensions, most of $v_t(x)$6 lies in a thin shell of radius approximately $v_t(x)$7. For each data point $v_t(x)$8, random targets $v_t(x)$9 are almost uniformly distributed on that sphere. Because of the non-crossing property of flow trajectories, the learned vector field is forced toward the average direction of $p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$0 over all couplings, which by symmetry is essentially the radial direction $p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$1 (Li et al., 7 Aug 2025).

The manifold argument sharpens the consequence for anomaly detection. Real data are taken to lie on a lower-dimensional manifold inside the sphere, so the reverse FM ODE first pulls all points toward the manifold boundary before any OT correction can act. In the terminology of the paper, this trivial mean-field component dominates and makes the learned flow useless for density proxy or anomaly scoring. The broader implication is that reverse-time use of conventional linear-interpolation FM is not merely numerically delicate; its pathology is structural.

4. Worst Transport displacement interpolation and WT-Flow

To remove probabilistic coupling and avoid the trivial reverse field, the anomaly-detection framework introduces Worst Transport (WT) displacement interpolation. Let $p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$2 denote the batch-mean and batch-std of extracted features. For any feature $p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$3,

$p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$4

with

$p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$5

Under WT, the source distribution $p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$6 and the target $p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$7 lie in the same unit-variance space. The paper then considers the Kantorovich OT problem with constant cost $p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$8,

$p_t(x)=\int p_t(x\mid x_0)\,p^*(x_0)\,dx_0$9

Because every coupling $p_0=p^*$0 is optimal, the induced displacement field is identically zero, and no gradient of a nontrivial potential can be defined (Li et al., 7 Aug 2025).

WT-Flow training proceeds by extracting features $p_0=p^*$1, normalizing $p_0=p^*$2, sampling $p_0=p^*$3 and $p_0=p^*$4, and forming $p_0=p^*$5. The network is trained with

$p_0=p^*$6

In practice, the network is therefore trained to output near zero for normal WT-normalized features. At inference, a single-step or few-step discrete solver is applied from $p_0=p^*$7; since $p_0=p^*$8 for normal samples, they remain in a “degenerate potential well” around the origin, whereas anomalous samples cannot be WT-normalized as cleanly, produce non-zero residuals, and escape this well with measurably larger net displacement (Li et al., 7 Aug 2025).

Within the paper’s logic, WT converts a problematic probabilistic transport problem into a non-probabilistic constant-cost OT construction. This suggests that the anomaly detector’s discriminative power is tied less to accurate density transport than to the controlled failure of anomalous features to remain in the degenerate well.

5. Reverse Flow Matching in online reinforcement learning

In online reinforcement learning, Reverse Flow Matching is introduced as a unified framework for diffusion and flow policies when the target policy is an unnormalized Boltzmann distribution defined by the Q-function. The key difficulty is the absence of direct samples from the target distribution. rFM addresses this by adopting a reverse inferential perspective and formulating the training target as posterior mean estimation given an intermediate noisy sample. The reverse loss is

$p_0=p^*$9

Expanding the inner MSE shows that minimizing this loss is equivalent to regressing onto the conditional posterior mean $p_1=\mathcal{N}(0,I)$0, where $p_1=\mathcal{N}(0,I)$1 (Li et al., 13 Jan 2026).

In noise-posterior form, the target velocity becomes

$p_1=\mathcal{N}(0,I)$2

To estimate this posterior mean efficiently, the framework introduces Langevin Stein operators as zero-mean control variates. For a density $p_1=\mathcal{N}(0,I)$3 and function $p_1=\mathcal{N}(0,I)$4,

$p_1=\mathcal{N}(0,I)$5

and under mild boundary conditions $p_1=\mathcal{N}(0,I)$6. A vector generalization yields matrix-valued control variates for the posterior $p_1=\mathcal{N}(0,I)$7, which are inserted into self-normalized importance sampling estimators in order to reduce variance (Li et al., 13 Jan 2026).

A major consequence of this construction is unification. Existing noise-expectation and gradient-expectation methods are recovered as special cases under a diagonal constant test function. With isotropic constant $p_1=\mathcal{N}(0,I)$8, $p_1=\mathcal{N}(0,I)$9 recovers Q-weighted noise estimation, whereas $$p_t(x\mid x_0)=\mathcal{N}(x;\mu_t(x_0),\sigma_t(x_0)^2I),$$0 recovers iterated denoising energy matching and Diffusion Q-Sampling. The paper further derives optimal minimum-variance coefficients under both diagonal and isotropic ansätze, provides an actor–critic loop for training a flow policy $$p_t(x\mid x_0)=\mathcal{N}(x;\mu_t(x_0),\sigma_t(x_0)^2I),$$1, proves that rFM and standard CFM share the same global minimizer under mutual absolute continuity and a sufficiently rich model class, and reports improved continuous-control performance relative to diffusion policy baselines (Li et al., 13 Jan 2026).

6. Terminological scope, adjacent usage, and common confusions

The recent literature does not use the acronym “RFM” exclusively for time-reversed or reverse flow matching. “RFM-Editing: Rectified Flow Matching for Text-guided Audio Editing” uses RFM to denote a rectified flow matching-based diffusion framework for audio editing rather than Time-Reversed Flow Matching (Gao et al., 17 Sep 2025).

In that audio-editing setting, the method is built on the Rectified Flow Matching ODE framework, defines a straight-line interpolation in latent space between Gaussian noise and the edited-audio latent, and trains a velocity field with a deterministic ODE. The paper states that “rectified” refers to two things: the substitution of the usual score-based reverse SDE by a deterministic flow ODE, and the addition of the original latent $$p_t(x\mid x_0)=\mathcal{N}(x;\mu_t(x_0),\sigma_t(x_0)^2I),$$2 into the conditioning so that only the instructed parts are edited. Training uses an MSE between the predicted velocity and the target velocity, while inference integrates the drift field forward in time from a partially noised initialization (Gao et al., 17 Sep 2025).

This suggests that recent usage requires disambiguation by task and formulation. In anomaly detection, rFM is a reverse map from unknown data distributions toward a Gaussian reference, motivated by non-invertibility and high-dimensional degeneracy of naive reverse FM. In online RL, Reverse Flow Matching is a reverse-inference training principle for targets that cannot be sampled directly. In text-guided audio editing, “RFM” instead abbreviates rectified flow matching. A common confusion is therefore to treat “RFM” as a single canonical method family; the cited papers indicate that the shared acronym masks substantially different objectives, probability paths, and theoretical claims (Gao et al., 17 Sep 2025).