@_Vassim@froger_romain@maxencelsb Thanks! I think models build linear representations in very local regions of their activation. In our experiments, some steered generations became incoherent at ~4 steering strength, though it's hard to tell whether that's due to a bad estimate or out-of-distribution activations
Did you ever wonder if your AI model trusted you ?
We just release a new paper TrustMI where we finds a linear trust direction in 6 LLMs.
📄 https://t.co/tzNKy1OnB1
🤗 https://t.co/I44dMClXF4
🚨Our new paper, TrustMI, is out! Watch Théo's video 👇
We study how LLMs represent trust in their users—and how to steer it.
TL;DR: more trust often meant less safety; lower trust reduced harm in our tests.
🧵1/X
Huge thanks to my co-authors @Blyzi1, @froger_romain, and @zehavoc for doing such an amazing job, especially @Blyzi1 and @froger_romain for leading this work!!
📄 Paper: https://t.co/12qjNlfqY2
💻 GitHub: https://t.co/H1xrkCXF9r
🤗 Hugging Face: https://t.co/YGBf97tGQM