Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Embedding Orrery: Visual Exploration of Embedding Spaces and Sparse Autoencoder Features

AbstractEmbedding Orrery is an open-source platform for interactive 3-dimensional visualization of embedding spaces, with native support for Sparse Autoencoders (SAEs), built for data exploration and interpretability research. Given any textual, image, or vector dataset, a single pipeline - embedding, projection, clustering, automatic topic labeling and, optionally, the collection of per-document SAE feature activations - turns it into a 3D constellation that can be searched visually, semantically, lexically, or by SAE feature names; probes can be trained, configured and scored directly in the interface. SAEs are automatically downloaded and rendered as navigable constellations in which entering a prompt highlights the features it activates. Individual features can be inspected inside a dashboard and injected into the model for causal steering. The star-cartography-inspired interface sustains up to 500,000 points, including WebGL animations and volumetric nebula shaders, on 8 GB of RAM. Embedding Orrery is released under the Apache 2.0 license. Code: https://github.com/Giacomo-De-Luca/embedding-orrery. HF Spaces: https://huggingface.co/spaces/GiacomoDeLuca/orrery-demo.

More from EMNLP 2026

All EMNLP 2026 papers →