1 paper · 1 filter
Mathis Immertreu, Achim Schilling, Thomas Kinfe +1
Language models are typically trained to predict the next token in a sequence. Here, we explore an alternative predictive principle from reinforcement learning: Successor Represent…