Helpers in Learning Systems

Recent advancements in AI have reshaped how AI agents interact with other agents, transitioning beyond static, repetitive processes to adaptive, multiagent collaborations. With the growing importance of AI agents in the era of Large Language Models (LLMs), we present two distinct lines of research involving helper agents in a learning system. In the first line, we address a collaborator scenario in which the AI agent collaborates with a bounded-rational helper it has never encountered before. The AI agent deploys a Stackelberg game solution without explicitly modeling the follower’s behavior, enabling rapid and efficient teaming. In the second line, we propose Learning to Help (L2H) for backup helpers, allowing the AI agent to autonomously manage tasks it can solve while delegating unsolvable tasks to backup helpers. Therefore, the AI agent only needs to focus on training for simple, specific tasks. In contrast, backup helpers only need to focus on those tasks the AI agent cannot solve due to legacy devices or limited computational power.