Jan Dubiński(@jan_dubinski_):"Inoculation Midtraining with Learned Neologisms". They introduce a token that tells the model where misalignment should occur. Also very cool that the authors test vulnerability to conditional misalignment (some triggers still increase misalignment a bit). A nice idea! Authors: Kyle O'Brien, Edward Young, @RadmardPuria, @n_lie_k, @cam

Jan Dubiński

@jan_dubinski_

TruthfulAI | PhD in AI | Warsaw University of Technology and NASK, working on AI safety and generative models

参加 July 2022

1.4K フォロー中 686 ファン

Jan Dubiński@jan_dubinski_

2026.09.16 21:25

"Inoculation Midtraining with Learned Neologisms". They introduce a token that tells the model where misalignment should occur. Also very cool that the authors test vulnerability to conditional misalignment (some triggers still increase misalignment a bit). A nice idea! Authors: Kyle O'Brien, Edward Young, @RadmardPuria, @n_lie_k, @cam_tice, @tomekkorbak, @DavidDAfrica