SAEs Need Far Less Training Data Than Assumed — Unless You Need Feature Reproducibility
Aikyam-Lab · hf · 2026-10-17
A new paper from VAL and Aikyam Lab (IBM, Indian Institute of Science, University of Virginia) examines how much activation data sparse autoencoders actually need, training four SAE architectures on three language models with budgets from 5M to 500M tokens.
Key findings: convergence depends strongly on the metric. Functional behavior (standard evaluations, reconstruction) approaches the full-data reference at substantially smaller budgets, suggesting the typical hundreds-of-millions scale is overestimated. But feature-level reproducibility (one-to-one feature matching) keeps improving with more data — e.g., strictly matched features in Gemma-2-2B BatchTopK rise from 2% with scale. Practical upshot: SAEs can be trained with far fewer tokens when evaluation goals align with functional behavior, but reproducible features demand larger datasets.
More from Research
- SAE features are only partially compositional, new paper finds — OpenSAE · 2026-10-17
- Contact-UAN: MIT's residual world model fixes humanoid sim-to-real with 6.5 minutes of data — kevin_zakka · 2026-10-17
- ROS 2 LiDAR localization runs 390 m outdoors on handheld Livox MID-360 — rsasaki0109 · 2026-10-17
- New paper asks how seriously to take AI welfare, with constraints from consciousness epistemology — birchlse · 2026-10-17
- GPT-6 Astra fastest-rising method on Papers with Code, 27 papers in 30 days — NielsRogge · 2026-10-17
- Embodied Astra tops RoboMME at 92.1% — beating humans by just keeping full visual history in context — hankyang94 · 2026-10-17