Dynamic LLM-Agent Network: An LLM-agent Collaboration Framework with Agent Team Optimization

strategic team of agents communicating in a dynamic interaction architecture based on the task query. Specifically, we build a framework named Dynamic LLM-Agent Network (DyLAN) for LLM-agent collaboration on complicated tasks like reasoning and code generation. DyLAN enables agents to interact for multiple rounds in a dynamic architecture with inference time agent selection and an early-stopping mechanism to improve performance and efficiency. We further design an automatic agent team optimization algorithm based on an unsupervised metric termed Agent Importance Score, enabling the selection of best agents based on the contribution each agent makes

In addition, there is no systematic way to ensure that the LLM-agent collaboration systems have a sufficient number of LLM agents and, most importantly, an optimized team of agents.

The landscape of LLM-agent collaborations requires a systematic framework in order to improve generalizability, efficiency, and performance. Therefore, several properties should be exhibited: (1) Task Agnostic: Prior works have suffered from the difficulty of generalizing across various domains due to the dependence on task-specific tools. A task-agnostic system can facilitate the fast adaptation of existing approaches to new situations. (2) Efficient: Instead of assigning agents in a static pattern, dynamically removing agents with uninformative responses can prevent the creation of useless information as well as ensure accuracy in the process of reaching consensus. (3) Agent Team Optimization: With thousands of open-source and unlimited LLM-generated prompts that serve a variety of roles, it is difficult to identify what the optimal team of agents might be. Ideally, multiagent systems should be able to adapt their composition in response to the particular domain of a query with minimal supervision.

we view LLM agents at specific time steps as nodes in a network and the messages they exchange at different time steps as edges

LLM-empowered ranker (Qin et al., 2023) to rank different LLM agents and deactivate low-performing agents in the subsequent interaction (i.e., inference-time agent selection), thereby creating a dynamic architecture of interactions.

a three-step procedure where we first ask each agent to rate its predecessors on their solutions (propagation), and then for each agent, aggregate the ratings from their successors to quantify its contribution (aggregation). Finally, after summing up the ratings across all time steps, we derive an Agent Importance Score for each agent. Then, we can select the top-performing agents based on their importance scores to obtain the optimized team of agents (selection). In this way, DyLAN achieves the task-agnostic property by formulating LLM-agent collaboration into a feed-forward network to decouple the interaction architecture and task-specific design, exhibits efficiency through inference-time agent selection and early-stopping mechanism, and enables agent team optimization through a selection algorithm based on agent importance scores.

Capabilities Acquisition is a crucial process in LLM-MA, enabling agents to learn and evolve dynamically

2. Feedback from Agents Interactions means that the feedback comes from the judgement of other agents or from agents communications.