A team of researchers has built a dimensionality-reduction tool that tries to explain itself while it works.
The method, called DMT-ME, pairs a mixture-of-experts architecture with a hyperbolic-geometry mapper to compress high-dimensional data such as images, tables, and text into visualizations. Sparse expert routing handles most of the accuracy gains, sending different slices of the data to specialized sub-models, while the hyperbolic component adds refinement for data with complicated underlying structure. Because the MoE routing explicitly links raw inputs, embedding outputs, and the features driving them, the researchers say the model's decisions are traceable rather than opaque. The code is posted on GitHub for anyone who wants to check the claims.
Dimensionality-reduction tools like t-SNE and UMAP are everywhere in data science, but they have always forced a choice: tune for accuracy and the output becomes a black box, or tune for interpretability and the results get worse. DMT-ME's pitch is that mixture-of-experts models, the same architecture powering efficient large language models, can resolve that trade-off by making the routing itself a form of explanation. If the claimed gains hold up outside the paper's own benchmarks, this would matter most in fields like genomics and medical imaging, where analysts need to justify why a data point landed where it did, not just that it did.
For now, the evidence lives entirely in the researchers' own experiments, worth revisiting once outside labs put the public code through its paces.