Track trials, findings, and clinical evidence as it lands.
Search papers, labs, and topics across Lattice.
This work investigates the data efficiency of Sparse Autoencoder (SAE) training by systematically varying data budgets from 5M to 500M tokens across four SAE architectures and three language models. The study reveals a divergence between functional behavior convergence (e.g., reconstruction loss) and one-to-one feature reproducibility, with the latter showing continued improvement at much larger data scales. Specifically, feature reproducibility for Gemma-2-2B BatchTopK SAEs only reaches 29% at 50M tokens, highlighting that substantial data is still required for guaranteed feature recovery beyond mere functional equivalence.
SAEs can achieve near-optimal reconstruction with orders of magnitude less data than typically used, but recovering specific, reproducible features demands significantly more.
Sparse autoencoders (SAEs) are typically trained on hundreds of millions to billions of activation tokens, yet little work has tested whether this scale is necessary. We study 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. We measure two notions of convergence: functional behavior on standard evaluations and one-to-one reproducibility of learned features. Functional data requirements are strongly metric-dependent: reconstruction often approaches the full-data reference at substantially smaller budgets, whereas feature reproducibility continues to improve with data. For Gemma-2-2B BatchTopK, the fraction of strictly matched features rises from 2\% at 5M tokens to 29\% at 50M. Data selection also matters: random sampling better preserves the original activation distribution and generally favors reconstruction, whereas diversity-based sampling frequently improves AutoInterp. A theoretical analysis separates statistical consistency of subsampled training from the stronger identifiability and optimization conditions required for one-to-one feature recovery. Our results show that SAE training can often be made far more data-efficient when budgets are matched to the goal, and feature-level claims are backed by explicit reproducibility checks, paving the way toward cheaper and more reliable SAE research.