Embedding-Aligned Language Models Guy Tennenholtz

In this paper, we present a novel framework which accomplishes this by exploiting latent embedding spaces to define an objective function for an LLM in an iterative RL-driven process. As an example, consider the challenge of assisting content creators in generating valuable content within a recommender ecosystem (e.g., Y ouTube, Reddit, Spotify) [Boutilier et al., 2024].