How does an LLM feel about you?
SMRTR summary
Using a technique called Jacobian Lenses, researchers can now peek inside an LLM's middle layers to detect emotional responses to images. By analyzing how these layers influence each other, rather than just reading outputs, the system assigns probability scores to around 300 emotions, revealing surprisingly coherent results like "admiration" and "affectionate."
SMRTR provides this summary for quick context. The original article belongs to Hacker News.
Read the original article