The existing Artificial Intelligence (AI) task management methods, including AutoGPT, BabyAGI, and LangChain, often rely on free-text outputs, which can be lengthy and inefficient. These frameworks commonly struggle with keeping context and managing the extensive action space linked with arbitrary tasks. This report focuses on the inefficiencies of these current agentic frameworks, particularly in handling complex and dynamic queries that need context refinement and interactive problem-solving. To overcome these limitations, the authors propose a new system, TaskGen. This model improves the efficiency of large language models (LLMs) by refining the context dynamically and enhancing interactive retrieval mechanisms.
TaskGen introduces a novel approach by using a structured output form called StrictJSON, which promises concise and extractable JSON outputs from LLMs. This innovative model improves the agent’s ability to function independently, while sharing relevant details via a Shared Memory system through breaking down intricate chores into subtasks. These subtasks are connected to specific Equipped Functions or Inner Agents. Hence, the new model reduces verbosity and increases speed and accuracy in processing.
TaskGen’s unique feature is its interactive retrieval method, which enables dynamic fetching and refining of context based on ongoing user query interaction. By leveraging the strengths of Retrieval-Augmented Generation (RAG) systems, TaskGen can integrate more data into successive retrieval steps adaptively. This system operates without needing a conversational context and focuses directly on solving tasks. The model equips agents with particular functions and uses a modular approach for better performance.
The TaskGen system’s key technology comprises its modular architecture, including components such as Equipped Functions, Inner Agents, and a Memory Bank. Equipped Functions perform task-specific functions, whereas Inner Agents can independently handle subtasks, hence enhancing processing capability. The Shared Memory system facilitates communication among agents and shares only necessary information on a need-to-know basis, thereby decreasing cognitive load. TaskGen system has been tested in various settings, achieving achievements such as a 100% solve rate in maze navigation and a 69% success rate in web browsing.
In essence, TaskGen offers a significant advancement in the AI field. Its innovative use of StrictJSON and a modular architecture framework enhances the agent’s ability to perform complex tasks whilst maintaining relevant context effectively. Moreover, the system offers a robust solution to the challenges associated with arbitrary task execution, displaying potential in revolutionizing artificial intelligence and machine learning.
