@repligate the RepE paper has a great example of this--the honesty representation drops (red) when the model is hallucinating.makes sense to me that the diff behaviors are learned in pretraining,…
Posters: voooooogel. voooooogel
3 likes recorded by the source corpus.
Topics: Consciousness, AI & LLMs
View Post: https://x.com/voooooogel/status/1878661648619340040