# ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration

**Authors:** Hongjin Su, Shizhe Diao, Ximing Lu, Mingjie Liu, Jiacheng Xu, Xin Dong, Yonggan Fu, Peter Belcak, Hanrong Ye, Hongxu Yin, Yi Dong, Evelina Bakhturina, Tao Yu, Yejin Choi, Jan Kautz, Pavlo Molchanov  
**arXiv:** [2511.21689](https://arxiv.org/abs/2511.21689)  
**Date:** 2025-11-26  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2511.21689v1/

## Key points

- Study titled "ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration"
- Introduces ToolOrchestra as a novel approach for training compact orchestrators to coordinate intelligent tools
- Develops Orchestrator, an 8B model incorporating reinforcement learning with outcome-, efficiency-, and user-preference-aware rewards
- Orchestrator outperforms GPT-5 on tasks like HLE, tau2-Bench, and FRAMES while operating at 30% of the cost
- Demonstrates optimal balance between performance and cost, with robust generalization capabilities towards unseen tools
- Integration of diverse tools with lightweight orchestration model enhances efficiency and effectiveness in problem-solving scenarios

## Summary

In the study "ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration," authored by Hongjin Su, Shizhe Diao, Ximing Lu, Mingjie Liu, Jiacheng Xu, Xin Dong, Yonggan Fu, Peter Belcak, Hanrong Ye, Hongxu Yin, Yi Dong,
Evelina Bakhturina,Tao Yu,Yejin Choi ,Jan Kautz,and Pavlo Molchanov,the researchers delve into the realm of leveraging small orchestrators to manage various intelligent tools in order to enhance problem-solving capabilities. The authors introduce ToolOrchestra as a novel approach for training compact orchestrators that coordinate intelligent tools. By incorporating reinforcement learning with outcome-, efficiency-, and user-preference-aware rewards into ToolOrchestra's framework,
they develop Orchestrator—a remarkable 8B model that not only achieves higher accuracy at a reduced cost compared to previous tool-use agents but also aligns with user preferences regarding tool selection for specific queries. Through rigorous experimentation and evaluation on tasks such as HLE (Humanity's Last Exam), tau2-Bench,and FRAMES (Flexible Reasoning Architecture for Multi-Modal Execution Systems), Orchestrator outperforms GPT-5 by significant margins while operating at approximately 30% of the cost. The comprehensive analysis conducted demonstrates Orchestrator's ability to strike an optimal balance between performance and cost across multiple metrics while exhibiting robust generalization capabilities towards unseen tools. The findings underscore the efficacy of integrating diverse tools with a lightweight orchestration model in enhancing efficiency and effectiveness in problem-solving scenarios. This research paves the way for the development of practical and scalable tool-augmented reasoning systems that can tackle complex challenges with heightened intelligence and resource optimization.