Shivam Raval, Kevin Dunnell, and Andrew Lippman. 2026. Latent manipulator: using concept-guided embedding manipulation to steer embedding visualizations. In Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA '26). Association for Computing Machinery, New York, NY, USA, Article 425, 1–6. https://doi.org/10.1145/3772363.3798816

Abstract

Embedding visualization systems reduce complex data to projection maps that support exploration and sensemaking, but only support a single static view of the data. Although changing DR methods and hyperparameters yields different projections, these views remain constrained by the concepts most salient in the upstream embeddings. We introduce concept-guided embedding manipulation, which enables users to steer visualizations by amplifying or suppressing user-defined concepts directly in embedding space. The method derives concept vectors from contrastive exemplars and reweights embeddings to generate alternative projections without retraining models. Through case studies with a collection of CHI research papers and a corpus of literary text, we show how steering can surface cross-domain methodological structure and reveal narrative patterns that are muted in default views of the datasets. Our approach helps users align embedding visualizations with their goals and exploration of multiple meaningful facets of the same dataset.