Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects

In 30 Sekunden

Was passiert ist

Researchers show sparse autoencoders encode both static concepts and context-dependent functions, with most features lacking stable steering directions.

Warum es zählt

Engineers using SAEs for model interpretability and steering should reconsider whether activation descriptions reliably predict causal effects on outputs.

Achtung

Features can be interpretable and causally relevant without providing consistent, reusable directions for steering across different contexts or prompts.