Ricardo Herrmann

2/ The original 2023 SAE paper "Sparse Autoencoders Find Highly Interpretable Features in Language Models" (ICLR 2024) showed that SAEs outperform ICA and other methods on automated interpretability metrics. However when looking at the authors public code (https: //github.com/HoagyC/sparse_coding) we noticed that there was a bug with their evaluation: it didn't properly handle ICA's sign ambiguity and evaluated only the positive direction. This hugely underestimated ICA's interpretability! An SAE will have nonnegative features. Where as ICA will give you an axis that has positive and negative sides. We found that meaningful concepts often appear on only one side of an ICA axis. So if a direction responds strongly to a concept on its negative side, but the evaluation only checks the positive side, it can miss that concept entirely. As the figure shows, we first reproduced the original Pythia-70M experiment and recovered the reported gap between SAEs and ICA. Then we corrected the signs using the skewness of the activation distribution (ie finding which side the tail is on). Without retraining or changing the recovered components, ICA's automated interpretability scores improved substantially, making it much more competitive with SAEs! We also ran this on Qwen 3.5 9B and got similar results.