From Embeddings to Exploration: Engineering Interactive Latent Space Visualizations for AI Model Sensemaking
Our paper "From Embeddings to Exploration: Engineering Interactive Latent Space Visualizations for AI Model Sensemaking" (PDF) has been accepted at EICS 2026 and will appear in the EICS issue of Proceedings of the ACM on Human-Computer Interaction. This is work by Sebe Vanbrabant together with Jarne Thys, Gilles Eerlings, Gustavo Rovelo Ruiz, Davy Vanacken, and myself.
Machine learning practitioners routinely inspect model internals through 2D projections of high-dimensional representations, using techniques such as t-SNE or UMAP. These visualizations give a useful overview of clustering and similarity, but they are inherently static: they show only the data points already in the dataset and offer no way to explore what lies between them or in sparsely populated regions. I am very excited about this work as it provides an interactive and visual insight in how an AI black box makes decisions.
This paper presents an interactive and visual exploration system that uses a Variational Autoencoder (VAE) as a generative proxy over a model’s training distribution. The VAE turns the latent space into a navigable workspace: because it provides an explicit decoding path from latent coordinates to inputs, users can continuously sample new points, interpolate between anchors, and probe regions that static embeddings leave opaque. These capabilities are exposed through a set of interactive probes that augment a familiar scatter-plot overview with generative overlays for comparing class transitions and examining sparse or uncertain regions.
A within-subject formative study (N=16) compared the interactive VAE-based system to a static t-SNE baseline. Generative interaction substantially improved counterfactual reasoning and shaped how participants assessed model behavior in uncertain areas. Static embeddings sometimes provided clearer boundary perception. From these findings, the paper derives concrete design guidelines and architectural considerations for building interactive AI model exploration systems using generative latent representations.
Citation
@article{vanbrabantembeddings2026,
author = {Vanbrabant, Sebe and Thys, Jarne and Eerlings, Gilles and Luyten, Kris and {Rovelo Ruiz}, Gustavo and Vanacken, Davy},
title = {From Embeddings to Exploration: Engineering Interactive Latent Space Visualizations for {AI} Model Sensemaking},
journal = {Proc. {ACM} Hum. Comput. Interact.},
number = {{EICS}},
year = {2026},
month = {jun},
publisher = {ACM},
address = {New York, NY, USA}
}