Steered Generation via Gradient Descent on Sparse Features

Sumanta Bhattacharyya, Pedram Rooshenas

Large language models (LLMs) encode a diverse range of linguistic features within their latent representations, which can be harnessed to steer their output toward specific target…