Large Learning Agents: Towards Continually Aligned Robots with Scale in RL

Bram Grooten

Abstract: In the field of deep reinforcement learning significant progress has been made, but it seems we are missing the power of the scaling laws evident in large language models. This research aims to pioneer the development of large learning agents (LLAs) that can take advantage of efficient scaling. We focus on creating agents that generalize strongly, quickly adapt to continuously changing environments, and integrate the reinforcements received through human feedback. We believe that this is a key step towards the long-term vision for continually aligned and intelligent agents.