Smurfs: Multi-Agent System using Context-Efficient DFSDT for Tool Planning

TL;DR

The paper introduces Smurfs, a multi-agent system that improves tool planning efficiency using a context-efficient, training-free extension of DFSDT, outperforming baselines in complex reasoning tasks with reduced token usage.

Contribution

Smurfs enhances DFSDT with a modular, context-efficient design for multi-agent tool planning, addressing previous limitations and enabling high performance without training.

Findings

Smurfs reduces token usage by 60.9% compared to DFSDT.

Smurfs achieves performance comparable to GPT-4-DFSDT on HotpotQA.

Extensive ablation studies validate the effectiveness of Smurfs' core components.

Abstract

Teaching large language models (LLMs) to use tools for solving complex problems can grant them human-like reasoning abilities. ReAct and its variants are popular frameworks for tool use in both single-agent and multi-agent systems. To address issues like error propagation and limited exploration in ReAct, the Deep First Search Decision Tree (DFSDT) was proposed, but it faces challenges such as rollback instability, redundant context, and premature termination in single-agent settings. We introduce "Smurfs," a novel multi-agent system (MAS) that enhances DFSDT with a modular, context-efficient, and training-free design. Smurfs surpasses baseline methods in both the open-ended StableToolBench and the closed-ended HotpotQA tasks, reducing token usage by 60.9\% compared to DFSDT and enabling Mistral-7b to perform on par with GPT-4-DFSDT. Extensive ablation studies confirm the effectiveness of Smurfs' core components, offering valuable insights for the construction and interpretation of MAS, and paving the way for future exploration.