Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

360$^\circ$REA: Towards A Reusable Experience Accumulation with 360° Assessment for Multi-Agent System

Shen Gao, Hao Li, Chengrui Huang, Quan Tu, Zhiliang Tian, Minlie Huang, Shuo Shang

TL;DR

The paper tackles the limited improvement of LLM-based multi-agent systems that rely on self-reflection or pruning, by introducing 360$^\circ$REA, a hierarchical framework that combines a 360$^\circ$ performance assessment with dual-level experience pools. A leader-crew architecture enables task decomposition, iterative multi-turn responses, and evaluation from self, peer, and supervisory sources, translating assessments into local and global reusable experiences. Empirical results on creative writing and travel planning show superior performance and robust human evaluation compared to strong baselines, with ablations highlighting the importance of local experiences and multi-level assessments. The work offers a practical pathway to enhance generalization and efficiency of LLM-based multi-agent systems, with potential for multimodal extensions and broader applications.

Abstract

Large language model agents have demonstrated remarkable advancements across various complex tasks. Recent works focus on optimizing the agent team or employing self-reflection to iteratively solve complex tasks. Since these agents are all based on the same LLM, only conducting self-evaluation or removing underperforming agents does not substantively enhance the capability of the agents. We argue that a comprehensive evaluation and accumulating experience from evaluation feedback is an effective approach to improving system performance. In this paper, we propose Reusable Experience Accumulation with 360$^\circ$ Assessment (360$^\circ$REA), a hierarchical multi-agent framework inspired by corporate organizational practices. The framework employs a novel 360$^\circ$ performance assessment method for multi-perspective performance evaluation with fine-grained assessment. To enhance the capability of agents in addressing complex tasks, we introduce dual-level experience pool for agents to accumulate experience through fine-grained assessment. Extensive experiments on complex task datasets demonstrate the effectiveness of 360$^\circ$REA.

360$^\circ$REA: Towards A Reusable Experience Accumulation with 360° Assessment for Multi-Agent System

TL;DR

The paper tackles the limited improvement of LLM-based multi-agent systems that rely on self-reflection or pruning, by introducing 360REA, a hierarchical framework that combines a 360 performance assessment with dual-level experience pools. A leader-crew architecture enables task decomposition, iterative multi-turn responses, and evaluation from self, peer, and supervisory sources, translating assessments into local and global reusable experiences. Empirical results on creative writing and travel planning show superior performance and robust human evaluation compared to strong baselines, with ablations highlighting the importance of local experiences and multi-level assessments. The work offers a practical pathway to enhance generalization and efficiency of LLM-based multi-agent systems, with potential for multimodal extensions and broader applications.

Abstract

Large language model agents have demonstrated remarkable advancements across various complex tasks. Recent works focus on optimizing the agent team or employing self-reflection to iteratively solve complex tasks. Since these agents are all based on the same LLM, only conducting self-evaluation or removing underperforming agents does not substantively enhance the capability of the agents. We argue that a comprehensive evaluation and accumulating experience from evaluation feedback is an effective approach to improving system performance. In this paper, we propose Reusable Experience Accumulation with 360 Assessment (360REA), a hierarchical multi-agent framework inspired by corporate organizational practices. The framework employs a novel 360 performance assessment method for multi-perspective performance evaluation with fine-grained assessment. To enhance the capability of agents in addressing complex tasks, we introduce dual-level experience pool for agents to accumulate experience through fine-grained assessment. Extensive experiments on complex task datasets demonstrate the effectiveness of 360REA.
Paper Structure (24 sections, 10 equations, 1 figure, 6 tables)

This paper contains 24 sections, 10 equations, 1 figure, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. $360^\circ$REA Methodology
  4. Overview
  5. Main Workflow of $360^\circ$REA
  6. $360^\circ$ Performance Assessment
  7. Dual-level Experience Pool
  8. Local-level Experience Construction
  9. Global-level Experience Construction
  10. Experimental Setup
  11. Tasks and Datasets
  12. Evaluation Metrics
  13. Baselines
  14. Implementation Details
  15. Experimental Results
  16. ...and 9 more sections

Figures (1)

  • Figure 1: Our proposed $360^\circ$REA. There are two main parts in our $360^\circ$REA, including the $360^\circ$ performance assessment and dual-level experience pool. $360^\circ$ performance assessment can assist the agent in obtaining evaluations from multiple aspects. These evaluations will facilitate agents in refining their results and accumulating reusable experiences for accomplishing tasks better. Then, we store low-level specific and higher-level experiences in local and global experience pools separately.