The paper, "From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability," examines feature vectors: directions in a language model's activation space that correspond to human-interpretable concepts. The authors focus on concepts like honesty and deception, which are central to alignment concerns.
Existing unsupervised methods, such as sparse autoencoders, attempt to recover these features without supervision. The new work appears to take a different route, using stories as a scalable source of concept vectors. The abstract is brief, so the exact method is not fully detailed here, but the framing suggests a supervised or semi-supervised alternative to purely unsupervised feature discovery.
The significance lies in the potential to make interpretability tools more reliable by grounding them in narrative data, which naturally contains examples of complex human concepts. However, because the abstract is truncated, the specific scaling technique and evaluation results are not yet available from this source alone.