Data scale requirements for Sparse Autoencoder feature reproducibility
October 13, 2026
Research shows that while reconstruction performance in SAEs approaches full-data levels at lower token budgets, feature reproducibility requires significant scale. For Gemma-2-2B, feature matching only rose from 2% at 5M tokens to 29% at 50M tokens.
HOW THIS AFFECTS YOU
●
researcherThis suggests that training SAEs on smaller datasets may lead to unstable or non-reproducible feature sets.