From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability
One sentence per concept is enough to extract a feature vector: 1,036 alignment-relevant concepts, two million synthetic stories, and vectors that read and steer six language models.
Dmitry Scherbakov, Oleg Chernikov, Victoria Khamash, Kirill Afendulev, Alexey Dontsov, Ivan Oseledets, Elena Tutubalina, Anton Korznikov
From Privacy Mechanisms to Model Behavior in Language-Model Adaptation
Privacy mechanisms change both what a model learns and the context it learns it in; separating the two predicts training cost and explains behavior after fine-tuning.
