TDAG: A multi-agent framework based on dynamic Task Decomposition and Agent Generation.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Large Language Models (LLMs) show promise for complex tasks but face limitations like error propagation and poor adaptability.
• Existing benchmarks lack the granularity to assess multi-step task execution and incremental progress.

Purpose of the Study:

• To propose a novel multi-agent framework, dynamic Task Decomposition and Agent Generation (TDAG), to enhance LLM agent adaptability.
• To introduce ItineraryBench, a new benchmark for evaluating LLM agents on complex, multi-step tasks, particularly in travel planning.

Main Methods:

• Developed a multi-agent framework (TDAG) that dynamically decomposes tasks into subtasks, assigning each to a generated subagent.
• Created ItineraryBench, a benchmark with interconnected, progressively complex travel planning tasks and a fine-grained evaluation system for memory, planning, and tool usage.

Main Results:

• The TDAG framework demonstrated significantly superior performance compared to established baseline methods.
• TDAG exhibited enhanced adaptability and context awareness in complex task execution scenarios.

Conclusions:

• The proposed TDAG framework effectively addresses the limitations of current LLM agents in handling complex, real-world tasks.
• ItineraryBench provides a valuable tool for assessing and advancing the capabilities of LLM agents in multi-step task execution.