Report
ICA Lens brings a decades-old statistical method to interpreting LLM activations
Its creators say ICA was competitive with public sparse autoencoders across multiple evaluations, using 7–43× less parameter memory.
TLDR
The ICA Lens researchers say Independent Component Analysis (ICA) was competitive with public sparse autoencoders (SAEs) on interpretability, sparse reconstruction, concept prediction and targeted probe perturbation across GPT-2 Small, Gemma 2 2B and Qwen 3.5 9B Base. One researcher says a bug in an influential early SAE paper’s evaluation code had underestimated ICA’s interpretability. Their open-source tool lets users inspect and annotate LLM activations without training another dictionary.
Combined views
15.4K
10 Sources, first seen ago
