Diminishing Returns of Data in Sparse Autoencoders: Functional Behavior does not Guarantee Feature Reproducibility
This work studies how much activation data SAEs need by training four SAE architectures on three language models with budgets ranging from 5M tokens to a 500M-token reference, and measures two notions of convergence: functional behavior on standard evaluations and one-to-one reproducibility of learned features.
Sambit Ghosh, Madelyn Mathai, R. Babu et al.
· 0 citations