Latent space steering

Can we improve the performance of LLMs or LVLMs by manipulating the Latent space (Embedding)?

e.g., VIT by Liu2024reducing.