Physis Systems BlogLumina AI Companion and Inventi Matrix
AI Research

What is ICA Lens?

ICA Lens is a workflow for discovering non-Gaussian directions in language models. The study is a preprint and has not been peer reviewed.

Illustration of a futuristic lens with abstract patterns, representing the ICA Lens workflow for discovering non-Gaussian directions in language models

Illustration: generated for Physis Systems

Short answer: Researchers found that non-Gaussianity provides a signature for discovering interpretable directions in language models. The study was conducted by Sida Liu, Feijiang Han, Sarah Wiegreffe, and Furong Huang. The paper was submitted on June 10, 2026, and last revised on November 17, 2026.

The primary source for this article is arXiv cs.AI's ICA Lens: Non-Gaussianity Reveals Interpretable Directions in Language Models. It is a preprint, which means it has not yet been peer reviewed.

What is the purpose of ICA Lens?

The purpose of ICA Lens is to provide a practical workflow for efficient discovery of non-Gaussian directions in language models. The study used GPT-2 Small, Gemma 2 2B, and Qwen 3.5 9B Base models.

What are the key findings of the study?

The recovered non-Gaussian directions captured coherent lexical, syntactic, and semantic patterns. Non-Gaussianity is positively associated with human-label confidence and automated interpretability.

How does ICA Lens compare to other methods?

The resulting coordinates are competitive with SAEs across interpretability and sparse reconstruction. The ICA Lens workflow fits in minutes per layer and uses 7-43 times less memory.

What do we know, and what is still open?

What is confirmed: The study found that non-Gaussian directions tend to have non-Gaussian projection distributions.

What is still open: The preprint has not been peer reviewed, and further research is needed to confirm the findings.

Why does Physis Systems cover this?

Physis Systems follows research at the frontier, from space to fundamental science. For more stories like this one, visit the Physis Systems blog.

About Physis Systems

Physis Systems (Physis Defense Systems, Inc.) builds Inventi Matrix, an invention engineering platform available in the United States and Canada at inventimatrix.com, and Lumina AI Companion, a personal AI companion you can meet on the Lumina page.

Frequently asked questions

What is non-Gaussianity?

Non-Gaussianity provides a signature for discovering interpretable directions in language models.

Who conducted the study?

The study was conducted by Sida Liu, Feijiang Han, Sarah Wiegreffe, and Furong Huang.

What is the current status of the paper?

The paper was submitted on June 10, 2026, and last revised on November 17, 2026.

Watch

Inventi Matrix, the film Watch on YouTube

Get the weekly frontier roundup

One email a week from Physis Systems with the research stories worth knowing, each linked to its primary source.

We will email you a link to confirm. We never share your address.