Training Strategy for a Central Coordination Unit in LLM-Based Agents

Determine an effective training methodology for a central coordination unit that manages interactions among heterogeneous components in LLM-based agents, including whether to optimize the coordination unit end-to-end jointly with the rest of the agent or to train it using supervised labels, reinforcement learning signals, or meta-learning strategies.

Background

The paper argues that current LLM-based agents struggle to coordinate heterogeneous components (e.g., different backbone LLMs, memory systems, perception modules, and tool units) due to mismatched specifications and representations. To address this, the authors propose a dedicated coordination mechanism to mediate configuration mismatches, translate representations, and optimize information flow across components without re-customizing the backbone LLM.

They further suggest a central coordination unit as a concrete instantiation of this mechanism but highlight that the appropriate training approach is unresolved, posing options such as end-to-end optimization with the entire agent, supervised training on configuration labels, reinforcement learning, or meta-learning.