Clément Dumas comments on Latent Adversarial Training (LAT) Improves the Representation of Refusal

[Clément Dumas]

Oh so when steering the LAT model at layer 4, the model actually generate valid response without refusal?