AbsTopK, a sparse autoencoder that keeps both positive and negative activations, lets one unit capture opposite concepts. Posted on arXiv (2510.00404) Oct 2025. Read more: https://getnews.me/abstopk-improves-sparse-autoencoders-for-bidirectional-features/ #abstopk #sparseautoencoders

Xudong Zhu, Mohammad Mahdi Khalili, Zhihui Zhu: AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features https://arxiv.org/abs/2510.00404 https://arxiv.org/pdf/2510.00404 https://arxiv.org/html/2510.00404