To interpret a direction in a language model's residual stream, we can look for text that strongly activates it.

Excerpt from arXiv AI

GETSECURE.ai hasn't published its own summary of this story yet. The complete report is at arXiv AI:

Read the full story at arXiv AI →