🤖 machine learning

Multi-Marginal Flow Matching with Adversarially Learnt Interpolants

This paper introduces ALI-CFM, a novel multi-marginal flow matching method that employs an adversarial loss to learn smooth, unique interpolants between discrete time-point snapshots, thereby enabling accurate inference of underlying dynamics without ground-truth trajectories and outperforming existing baselines on biological datasets.

Original authors: Oskar Kviman, Kirill Tamogashev, Nicola Branchini, Víctor Elvira, Jens Lagergren, Nikolay Malkin

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Oskar Kviman, Kirill Tamogashev, Nicola Branchini, Víctor Elvira, Jens Lagergren, Nikolay Malkin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out the story of a journey, but you only have a few snapshots of the traveler at different stops along the way. You have a photo of where they started, a photo of where they ended, and maybe a few blurry pictures of where they were in the middle. Your goal is to draw the smooth, continuous line that connects all these dots to show exactly how they moved.

This is the problem scientists face when studying things like how cells grow, how diseases spread, or how tumors evolve. They have "snapshots" of data at different times, but they don't have the video of the actual movement.

This paper introduces a new method called ALI-CFM to solve this puzzle. Here is how it works, using simple analogies:

The Old Way: The "Rigid String" Problem

Previous methods tried to connect these snapshots by drawing straight lines or rigid curves between the points.

The Problem: If the traveler took a winding, complex path (like a knot), a straight line or a rigid spline (a stiff plastic ruler) would look nothing like the real journey. It would be jagged, awkward, and might even break the rules of physics.
The Limitation: These old methods tried to force the path to hit every single data point exactly. If the data was a bit noisy (like a blurry photo), the path would get wobbly and chaotic.

The New Way: The "Adversarial Dance" (ALI-CFM)

The authors propose a smarter way to draw the path using a technique inspired by GANs (Generative Adversarial Networks). Think of this as a game between two characters:

The Forger (The Interpolant): This is a neural network trying to draw a smooth, beautiful path between the start and end points. It wants to make a path that looks real.
The Art Critic (The Discriminator): This is another neural network that looks at the Forger's path and compares it to the real "snapshots" (the data) taken at the middle stops. The Critic asks: "Does this path actually pass through the crowd of people seen in the middle photo?"

How they play:

The Forger draws a path.
The Critic checks if the path matches the crowd in the middle photos.
If the path misses the crowd, the Critic says, "No, that's fake!" and the Forger tries again.
They keep playing this game until the Forger draws a path that is so smooth and accurate that the Critic can't tell the difference between the drawn path and the real data snapshots.

Why This is Special

It's Smooth: Unlike the old "rigid ruler" methods that get jagged when there are many stops, this method learns to draw a naturally flowing, smooth curve, like a river flowing around rocks.
It Handles Noise: Because the Forger is trying to match the general shape of the crowd (the distribution) rather than hitting every single person's exact location, it works great even if the photos are blurry or the data is messy.
It Works in High Dimensions: The paper shows this works not just for 2D drawings, but for complex, multi-dimensional data (like thousands of gene measurements in a cell).

Where They Tested It

The authors tested this "Forger vs. Critic" game on three real-world scenarios:

A "Knot" in the Air: A synthetic test where data points formed a complex knot shape. The old methods failed to trace the knot smoothly, but ALI-CFM nailed it.
Cell Tracking: Watching a single cell move and change shape under a microscope. The old methods made the cell's path jump around wildly; ALI-CFM drew a smooth, realistic path that followed the cell's actual movement.
Tumor Mapping: Using spatial data from breast cancer tissue to guess where tumor cells are in a slice of tissue that wasn't scanned. ALI-CFM was much better at predicting the tumor's location than previous methods.

The Bottom Line

This paper presents a new tool that learns to connect the dots of a process by playing a game of "guess the path." Instead of forcing a rigid connection between data points, it learns a smooth, flexible path that naturally fits the data, making it much better at predicting how complex systems (like cells or diseases) change over time.

Technical Summary: Multi-Marginal Flow Matching with Adversarially Learnt Interpolants

Problem Statement
The paper addresses the challenge of learning the time-dependent dynamics of a system given only sampled observations (snapshots) at discrete time points, without access to ground-truth trajectories. This is a central task in fields such as biology (e.g., single-cell RNA sequencing, cell tracking) and medicine. Formally, the goal is to infer a time-dependent vector field $v_t$ governing an Ordinary Differential Equation (ODE) $dx_t = v_t(x_t)dt$ , such that the integration map pushes an initial marginal distribution $q_0$ to a target $q_1$ while satisfying intermediate marginal constraints $p_{t_i} = q_{t_i}$ for a set of time steps $0 = t_1 < \dots < t_K = 1$ .

Existing approaches, such as applying standard Flow Matching (FM) to consecutive pairs of marginals or using piecewise linear/cubic spline interpolants, often result in non-smooth interpolation curves. These non-smoothness issues lead to high gradient variance in the learning objective and difficulties in marginalizing the resulting vector fields, particularly in high-dimensional settings or when data is noisy.

Methodology: ALI-CFM
The authors propose ALI-CFM (Adversarially Learnt Interpolants with Conditional Flow Matching), a two-stage method that decouples the learning of interpolants from the learning of the vector field.

Adversarial Learning of Interpolants (ALI):
Instead of forcing interpolants to pass through specific data points (pointwise matching), ALI learns a neural interpolant $G_\phi(x_0, x_1, t)$ that matches the distribution of intermediate marginals. The interpolant is parameterized as a linear interpolation plus a neural correction term:
$G_\phi(x_0, x_1, t) = (1-t)x_0 + tx_1 + t(1-t)f_\phi(x_0, x_1, t)$
To enforce that the pushforward distribution of the interpolant matches the observed intermediate marginals $q_{t_i}$ , the authors employ a GAN-inspired adversarial objective. A discriminator $D_\gamma$ distinguishes between samples from the true marginal $q_{t_i}$ and samples generated by the interpolant $G_\phi$ . The generator (interpolant) is trained to minimize the Jensen-Shannon divergence between these distributions.

To ensure the uniqueness of the interpolants and prevent arbitrarily curved paths, the authors introduce regularizers:
- Linear Reference: Penalizes deviation from a straight line between coupled end-points.
- Piecewise Linear Reference: Regresses the interpolant to a piecewise linear path defined by Markov-chained Optimal Transport (OT) couplings.
- Second Derivative Norm: Enforces smoothness by penalizing the curvature of the interpolant.
Marginalization via Conditional Flow Matching (CFM):
Once the interpolants $G_\phi$ are trained, they are fixed. A second neural network $u_\theta$ is trained to approximate the vector field that generates these interpolants. This is done using the standard Conditional Flow Matching objective, where the target vector field is the time derivative of the learned interpolant:
$L_{CFM} = \mathbb{E}_{t, (x_0, x_1) \sim \pi} \| u_\theta(G_\phi(x_0, x_1, t)) - \frac{d}{dt}G_\phi(x_0, x_1, t) \|^2$
The coupling $\pi$ between $q_0$ and $q_1$ can be independent or an Optimal Transport plan.

Key Contributions

Novel Interpolation Technique: The paper introduces a method to learn interpolants using an adversarial objective that matches distributions rather than points. This allows for smooth trajectories that are unique under mild assumptions (proven in the appendices).
Robustness to Noise and Complexity: The method is designed to handle noisy datasets and complex, time-varying geometries where pointwise interpolation methods (like splines or piecewise linear) fail or become unstable.
Scalability and Versatility: The authors demonstrate that ALI-CFM scales to high-dimensional data and large numbers of marginal distributions ( $K$ ), outperforming existing baselines in specific complex scenarios while remaining competitive in others.

Experimental Results
The authors evaluate ALI-CFM on four distinct tasks:

Synthetic "Knot" Distribution: On a synthetic dataset with 1,200 marginal distributions forming a knot, ALI-CFM was the only method capable of accurately capturing the time-dependent geometry. Competing methods (OT-MFM, cubic splines, piecewise linear) produced non-smooth interpolants with high gradient variance.
Cell Tracking (Glioblastoma): Using microscopy data of a moving cell, ALI-CFM successfully learned spatio-temporal dynamics from noisy, subsampled data. In contrast, OT-CFM trajectories diverged, and time-independent OT-MFM failed to capture temporal dependencies.
Single-Cell Trajectory Inference (scRNA-seq): On Embryoid Body (EB), Cite-seq, and Multiome datasets, ALI-CFM performed on par with state-of-the-art baselines (OT-MFM, OT-CFM) in terms of Earth Mover's Distance (EMD). The authors note that while adversarial training makes it difficult to strictly outperform methods that overfit to pointwise data in these specific metrics, the method remains highly competitive.
Tumour Coordinate Inference (Spatial Transcriptomics): On a breast cancer dataset with spatially ordered tissue sections, ALI-CFM significantly outperformed all baselines (including OT-CFM, OT-MMFM, and OT-MFM) in predicting the distribution of tumor regions in held-out sections. The authors attribute this success to the method's ability to handle multimodal distributions and the challenges of aligning tissue sections.

Significance and Claims
The paper claims that ALI-CFM offers a conceptually novel approach to multi-marginal flow matching by shifting from strict pointwise assignment to distribution matching. This shift enables the learning of smooth, unique interpolants that are robust to noise and capable of modeling complex dynamics where previous methods fail. The authors highlight that their method is particularly effective in scenarios with hundreds or thousands of marginal distributions and noisy data, a regime where no other flow matching algorithm has demonstrated comparable performance. While the method performs on par with existing state-of-the-art on standard single-cell benchmarks, its superior performance on spatial transcriptomics and synthetic complex geometry tasks underscores its versatility and scalability.