TUMIX: Multi-Agent Test-Time Scaling with Tool-Use Mixture

2025-09-30

AI Summary

Existing approaches for complex reasoning tasks suffer from low synergy among heterogeneous tools—such as textual reasoning, code execution, and web search—and lack practical, adaptive tool-use strategies. Method: This paper proposes TUMIX, a test-time multi-agent scaling framework that deploys heterogeneous agents in parallel, each employing distinct, proactive tool-selection policies. It introduces a tool-usage mixture mechanism and a confidence-based dynamic termination strategy to reduce computational overhead without compromising accuracy. Furthermore, TUMIX enables deep integration of multi-source capabilities via iterative answer sharing, LLM-driven self-optimizing agent design, and cross-agent collaborative optimization. Results: Evaluated on the Gemini family of models, TUMIX achieves an average accuracy gain of 3.55% with comparable inference cost; with dynamic termination enabled, computational cost drops to 49% of the baseline while preserving full performance.

Technology Category

• Multiagent Systems: Other Foundations of Multi Agent Systems
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

While integrating tools like Code Interpreter and Search has significantly enhanced Large Language Model (LLM) reasoning in models like ChatGPT Agent and Gemini-Pro, practical guidance on optimal tool use is lacking. The core challenge is effectively combining textual reasoning, coding, and search for diverse questions. In this paper, we propose Tool-Use Mixture (TUMIX), an ensemble framework that runs multiple agents in parallel, each employing distinct tool-use strategies and answer paths. Agents in TUMIX iteratively share and refine responses based on the question and previous answers. In experiments, TUMIX achieves significant gains over state-of-the-art tool-augmented and test-time scaling methods, delivering an average accuracy improvement of up to 3.55% over the best baseline on Gemini-2.5-Pro and Gemini-2.5-Flash across key reasoning benchmarks, with near-equal inference costs. We find that agent diversity and quality are crucial and can be enhanced by using LLMs to auto-optimize agent designs. Furthermore, TUMIX can halt refinement upon reaching sufficient confidence, preserving performance at only 49% of the inference cost. Further scaling can achieve higher performance, albeit at a greater cost.

Problem

• Optimizing tool combination for diverse question types
• Enhancing multi-agent collaboration through iterative refinement
• Balancing performance gains with computational cost efficiency

Innovation

• Parallel multi-agent ensemble with diverse tool-use strategies
• Iterative response sharing and refinement among agents
• Auto-optimized agent designs using LLMs for enhanced diversity