Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Parallelized Planning-Acting for Efficient LLM-based Multi-Agent Systems

Yaoru Li, Shunyu Liu, Tongya Zheng, Mingli Song

TL;DR

The paper tackles the bottleneck of serialized decision-making in LLM-based multi-agent systems operating in dynamic environments. It proposes a parallelized planning-acting framework with interruptible dual-thread execution, a centralized memory system, and a comprehensive skill library to enable real-time planning and execution. Through Minecraft benchmarks (Resource Collection, Boss Combat, Adversarial PvP), it demonstrates notable improvements in responsiveness, coordination, and adaptability, aided by RTDM and a real-time memory/communication mechanism. This approach offers practical benefits for real-time multi-agent collaboration and adversarial tasks, reducing latency and enabling dynamic re-planning. The work also acknowledges limitations such as computational costs, hallucination risks, and multimodal fusion considerations, outlining directions for efficiency and robustness improvements.

Abstract

Recent advancements in Large Language Model(LLM)-based Multi-Agent Systems(MAS) have demonstrated remarkable potential for tackling complex decision-making tasks. However, existing frameworks inevitably rely on serialized execution paradigms, where agents must complete sequential LLM planning before taking action. This fundamental constraint severely limits real-time responsiveness and adaptation, which is crucial in dynamic environments with ever-changing scenarios. In this paper, we propose a novel parallelized planning-acting framework for LLM-based MAS, featuring a dual-thread architecture with interruptible execution to enable concurrent planning and acting. Specifically, our framework comprises two core threads:(1) a planning thread driven by a centralized memory system, maintaining synchronization of environmental states and agent communication to support dynamic decision-making; and (2) an acting thread equipped with a comprehensive skill library, enabling automated task execution through recursive decomposition. Extensive experiments on challenging Minecraft demonstrate the effectiveness of the proposed framework.

Parallelized Planning-Acting for Efficient LLM-based Multi-Agent Systems

TL;DR

The paper tackles the bottleneck of serialized decision-making in LLM-based multi-agent systems operating in dynamic environments. It proposes a parallelized planning-acting framework with interruptible dual-thread execution, a centralized memory system, and a comprehensive skill library to enable real-time planning and execution. Through Minecraft benchmarks (Resource Collection, Boss Combat, Adversarial PvP), it demonstrates notable improvements in responsiveness, coordination, and adaptability, aided by RTDM and a real-time memory/communication mechanism. This approach offers practical benefits for real-time multi-agent collaboration and adversarial tasks, reducing latency and enabling dynamic re-planning. The work also acknowledges limitations such as computational costs, hallucination risks, and multimodal fusion considerations, outlining directions for efficiency and robustness improvements.

Abstract

Recent advancements in Large Language Model(LLM)-based Multi-Agent Systems(MAS) have demonstrated remarkable potential for tackling complex decision-making tasks. However, existing frameworks inevitably rely on serialized execution paradigms, where agents must complete sequential LLM planning before taking action. This fundamental constraint severely limits real-time responsiveness and adaptation, which is crucial in dynamic environments with ever-changing scenarios. In this paper, we propose a novel parallelized planning-acting framework for LLM-based MAS, featuring a dual-thread architecture with interruptible execution to enable concurrent planning and acting. Specifically, our framework comprises two core threads:(1) a planning thread driven by a centralized memory system, maintaining synchronization of environmental states and agent communication to support dynamic decision-making; and (2) an acting thread equipped with a comprehensive skill library, enabling automated task execution through recursive decomposition. Extensive experiments on challenging Minecraft demonstrate the effectiveness of the proposed framework.

Paper Structure

This paper contains 30 sections, 3 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. LLM-based Multi-Agent Framework
  3. Parallelized Planning-Acting Framework
  4. Centralized Memory System
  5. Comprehensive Skill Library
  6. Experiments
  7. Experiment Setup
  8. Benchmark Task Scenarios
  9. Main Results
  10. Resource Collection Task
  11. Boss Combat Task
  12. Adversarial PVP Task
  13. Ablation Study
  14. Human-Agent Interaction
  15. Related Works
  16. ...and 15 more sections

Figures (4)

  • Figure 1: An overview of the proposed multi-agent system
  • Figure 2: An overview of the proposed parallelized planning-acting architecture
  • Figure 3: Comparison of parallelized vs. serialized frameworks in PVP tasks. All results are shown with the mean and standard deviation over 10 trials.
  • Figure 4: Ablation study in boss combat tasks across 12 trials in three scenarios: Elder Guardian (EG), Wither (Wi), and Ender Dragon (ED), where efficiency is defined as the inverse of completion time in minutes.