What is ica-lens-paper?

This repository contains the code released alongside an academic paper called ICA Lens. The paper is about understanding what is happening inside large AI language models, a field called interpretability research. When an AI model processes text, it produces large tables of numbers at each internal layer, called activations. This work applies a mathematical technique called Independent Component Analysis (ICA) to find patterns in those numbers that correspond to interpretable concepts, such as topics, grammatical structures, or factual associations. ICA is a method that tries to separate a mixed signal into independent sources. In this context, it is applied to the activations of three publicly available AI models: GPT-2 (a smaller, older model), Gemma 2 2B, and Qwen 3.5 2B Base. The paper compares ICA against other approaches used in interpretability research, including sparse autoencoders (SAEs), which are currently the more popular tool for the same goal. The core claim is that ICA is fast, compact, and produces results comparable to or better than those alternatives on several benchmarks. The repository includes a browser-based visual explorer that lets researchers look at the ICA components: what text patterns activate each component most strongly, how components compare across layers, and how well they correlate with known features. There are two versions of the database: a small one (around 250 MB) for quick browsing, and a full one (around 7 GB) with richer information. Precomputed model files and databases are hosted on Hugging Face and downloaded on first use. For researchers who want to reproduce the paper's results, the repository provides numbered workflow scripts that walk through capturing activations, fitting ICA models, computing analysis metrics, building the explorer database, and running the benchmark comparisons. A mini reproduction mode runs the full pipeline on a small token sample and completes in a reasonable time on a single GPU. Full-scale reproduction requires substantially more GPU time and storage. Python 3.10 or later is required, along with PyTorch. The package manager used is called uv. A live demo of the explorer is available on Hugging Face Spaces.