Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Investigate-Consolidate-Exploit: A General Strategy for Inter-Task Agent Self-Evolution

Cheng Qian, Shihao Liang, Yujia Qin, Yining Ye, Xin Cong, Yankai Lin, Yesai Wu, Zhiyuan Liu, Maosong Sun

TL;DR

ICE introduces a general strategy for inter-task agent self-evolution by disentangling planning and execution experiences and organizing them into reusable workflows and pipelines. Through Investigate-Consolidate-Exploit, ICE records, normalizes, and retrieves past experiences to accelerate new-task planning and automate execution, demonstrated within the XAgent framework. Empirical results show substantial reductions in API calls (up to 80%), lower demands on backbone model capabilities, and competitive performance with lighter models when reusing consolidated experiences and pipelines. The work suggests a paradigm shift toward experience-driven, memory-augmented agent design with scalable, community-friendly memory of past task executions.

Abstract

This paper introduces Investigate-Consolidate-Exploit (ICE), a novel strategy for enhancing the adaptability and flexibility of AI agents through inter-task self-evolution. Unlike existing methods focused on intra-task learning, ICE promotes the transfer of knowledge between tasks for genuine self-evolution, similar to human experience learning. The strategy dynamically investigates planning and execution trajectories, consolidates them into simplified workflows and pipelines, and exploits them for improved task execution. Our experiments on the XAgent framework demonstrate ICE's effectiveness, reducing API calls by as much as 80% and significantly decreasing the demand for the model's capability. Specifically, when combined with GPT-3.5, ICE's performance matches that of raw GPT-4 across various agent tasks. We argue that this self-evolution approach represents a paradigm shift in agent design, contributing to a more robust AI community and ecosystem, and moving a step closer to full autonomy.

Investigate-Consolidate-Exploit: A General Strategy for Inter-Task Agent Self-Evolution

TL;DR

ICE introduces a general strategy for inter-task agent self-evolution by disentangling planning and execution experiences and organizing them into reusable workflows and pipelines. Through Investigate-Consolidate-Exploit, ICE records, normalizes, and retrieves past experiences to accelerate new-task planning and automate execution, demonstrated within the XAgent framework. Empirical results show substantial reductions in API calls (up to 80%), lower demands on backbone model capabilities, and competitive performance with lighter models when reusing consolidated experiences and pipelines. The work suggests a paradigm shift toward experience-driven, memory-augmented agent design with scalable, community-friendly memory of past task executions.

Abstract

This paper introduces Investigate-Consolidate-Exploit (ICE), a novel strategy for enhancing the adaptability and flexibility of AI agents through inter-task self-evolution. Unlike existing methods focused on intra-task learning, ICE promotes the transfer of knowledge between tasks for genuine self-evolution, similar to human experience learning. The strategy dynamically investigates planning and execution trajectories, consolidates them into simplified workflows and pipelines, and exploits them for improved task execution. Our experiments on the XAgent framework demonstrate ICE's effectiveness, reducing API calls by as much as 80% and significantly decreasing the demand for the model's capability. Specifically, when combined with GPT-3.5, ICE's performance matches that of raw GPT-4 across various agent tasks. We argue that this self-evolution approach represents a paradigm shift in agent design, contributing to a more robust AI community and ecosystem, and moving a step closer to full autonomy.
Paper Structure (26 sections, 2 equations, 5 figures, 3 tables)

This paper contains 26 sections, 2 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Problem Formulation.
  4. ICE Self-Evolution Strategy
  5. Planning ICE
  6. Execution ICE
  7. Experiments
  8. Experimental Settings
  9. Data
  10. Settings
  11. Baseline
  12. Metrics
  13. Results
  14. Case Study
  15. Ablation Study
  16. ...and 11 more sections

Figures (5)

  • Figure 1: An overview of the inter-task agent self-evolution. ICE automatically identifies re-utilizable plans and tool execution trajectories as past experiences for agent self-evolution. The human effort may also be involved in crafting experiences for learning.
  • Figure 2: The detailed illustration of Planning ICE. Planning ICE dynamically tracks the plan of the agent system during Investigate stage, prunes and linearizes the plan into a workflow during Consolidate stage, and re-utilizes the workflow as a reference for new goals during Exploit stage.
  • Figure 3: The detailed illustration of Execution ICE. Execution ICE identifies successful execution trajectories of the agent system during Investigate stage, extracts potential useful pipelines during Consolidate stage, and re-utilizes relevant pipelines for new goals during Exploit stage.
  • Figure 4: Case study on the Investiagte and Consolidate stages for the planning (left) and execution (right) self-evolution.
  • Figure 5: Case study on the Exploit stage for the planning (up) and execution (down) self-evolution.