Back to AI Research

AI Research

FlowAlign uses training paths to align data across modalities

Key Takeaways

  • FlowAlign uses a learned training signal to preserve paired observations and target geometry when mapping one data modality into the space of another.
  • FlowAlign uses flow matching during training to learn how a source representation should sit in a fixed target space.
  • A FlowAlign preprint describes a method for mapping one kind of data into an existing space defined by another.
  • The method aims to preserve two things at once: which observations belong together and how observations are arranged in the target space.
  • The authors test the idea on synthetic data, paired single-cell measurements and image-to-text alignment.

A FlowAlign preprint describes a method for mapping one kind of data into an existing space defined by another. The method aims to preserve two things at once: which observations belong together and how observations are arranged in the target space. The authors test the idea on synthetic data, paired single-cell measurements and image-to-text alignment. The preprint was submitted to arXiv on November 6, 2026.

What a fixed target space changes

In the paper's single-cell example, measurements are paired at the cell level. A mapping into RNA space should put each source measurement near the representation of the corresponding cell while respecting the neighborhood structure already present in the RNA data. Matching the broad distribution alone can leave the wrong individual cells together. Optimizing only for the paired endpoint leaves the geometry of the target space implicit.

How the method learns the mapping

FlowAlign uses conditional flow matching to learn a velocity field on training paths between paired representations. The field then scores candidate paths as the source-side model learns where to place an observation in the target space. A candidate is rewarded when the path fits the geometry the field learned from the paired data.
FlowAlign does not use the velocity field or sample paths during inference. The trained source-side model produces a target-space representation in one forward pass. That makes the learned field a way of teaching the mapper rather than a separate transport procedure used every time it runs.

What the experiments establish

The authors report improved instance-level correspondence while preserving target structure in their experiments. Their theoretical analysis connects the training criterion to endpoint accuracy and neighborhood stability. For readers building systems from paired measurements, the useful distinction is between evidence that a mapping preserves individual pairs and evidence that it preserves the broader structure. These experiments cover settings with paired training examples and an already defined target space; they do not establish the same result when the necessary pairs are absent.

Comments