N-Gram Induction Heads for In-Context RL: Improving Stability and Reducing Data Needs

Open in new window