feature-mining

This paper introduces Mining via Activation Geometry (MAG), an unsupervised framework that extracts reasoning features from LLM activations using natural-language instructions, enabling activation steering and effective training data selection for classifier probes.