Paper recorded by Signals 4 on 2026-10-16 in cs.AI. Abstract reproduced from arXiv; link to the original below.
Published 2026-10-16 on arXiv · recorded by Signals 4 on 2026-10-19
Category: cs.AI · 人工智能 · first seen 2026-10-19
Sparse autoencoders (SAEs) explain foundation model representations by decomposing activations into sparse combinations of learned dictionary atoms. However, the learned dictionaries can vary across training runs, limiting the reliability and reproducibility of SAE-based analyses. We propose Consistent Sparse Dictionary Learning (ConSDL), which replaces the standard linear encoder with per-sample