From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability

Researchers created a practical method to identify and measure specific concepts in AI models by generating pairs of stories that contrast concepts (e.g., honest vs. dishonest narratives) and using the difference in how the model processes them as a "feature vector." This targeted approach proved more effective than existing methods that try to map the entire feature space at once.

The team successfully scaled this approach to over 1,000 safety-relevant concepts using 2 million synthetic stories across multiple languages and genres, and publicly released these resources. This represents a significant expansion in our ability to systematically study alignment-related concepts in language models.

These feature vectors reliably detected their corresponding concepts across different models and real-world text, suggesting they capture something fundamental about how language models represent ideas like honesty, deception, and other safety-relevant behaviors—not just artifacts of a single model or training approach.

The features can actually change how models behave when injected into their internal processing, demonstrating that understanding these concept directions isn't merely academic but could enable practical interventions to steer model behavior toward safer outcomes.

The method revealed interpretable patterns in how different models represent concepts, opening new avenues for analyzing model vulnerabilities (such as during jailbreak attacks) and understanding reasoning processes—suggesting this framework could extend to studying millions of features across many domains beyond safety.