Researchers from Nvidia and the University of Hong Kong have developed Orchestrator, an 8-billion-parameter model that coordinates various tools and immense language models (LLMs) to solve intricate problems. In its experiments, Orchestrator achieved higher accuracy at lower cost than much larger models in tool usage benchmarks, while adapting to user preferences for which tools to employ for a given query.
The model has been trained ToolOrchestraa modern reinforcement learning (RL) framework for training diminutive models to act as knowledgeable coordinators. The approach is based on the assumption that a diminutive “orchestrator” managing a diverse set of specialized models and tools can be more effective and proficient than a single, monolithic AI system.
The findings suggest that this intricate approach could pave the way for more practical and scalable AI reasoning systems in enterprises.
Limits of the current employ of LLM tools
Giving LLM access to external tools is a promising way to extend their capabilities beyond training data and into agent-based tasks. By referencing resources such as search engines and code interpreters, AI agents can improve their accuracy and perform tasks within the application.
However, in paper includedresearchers argue that the current approach to creating tool-based agents does not realize the full potential of this paradigm. Most systems equip a single, powerful model with a set of basic tools, such as a web search engine or calculator.
They argue that when people reason, they “routinely expand their capabilities by drawing on resources of intelligence beyond human measure, from domain experts to sophisticated processes and software systems.” Therefore, LLMs should be able to interact with a wide range of tools with different capabilities.
Tool orchestration paradigm
The article proposes a transition from a single-model system to a intricate system managed by a lightweight “orchestrator” model. The orchestrator’s job is to analyze a intricate task and break it down into its parts and invoke the right tools in the right order to find a solution.
This toolkit includes not only standard tools such as web search and code interpreters, but also other LLM tools with different capabilities that act as “smart tools”. For example, a coordinator might delegate a quantitative question to a math model or a programming challenge to a code generation model. Instead of putting all the cognitive load into one immense, general model, the orchestrator delegates narrowed sub-problems to specialized knowledgeable tools.
Based on this concept, researchers developed ToolOrchestra, a method that uses RL to train a diminutive language model to act as an orchestrator. The model learns when and how to refer to other models and tools and how to combine their results in multi-reciprocal reasoning. Tools are defined in a straightforward JSON format, providing their name, description and parameters.
The RL training process is based on a reward system that creates a profitable and controllable agent. The reward balances three goals: correctness of the final answer, efficiency in terms of cost and latency, and compliance with user preferences. For example, the system is penalized for excessive employ of processing power and rewarded for choosing tools that the user has marked as preferred, such as preferring an open source model over a proprietary API for privacy reasons. To support this training, the team also developed an automated data pipeline that generated thousands of verifiable training examples across 10 different domains.
A diminutive model with gigantic results
Using ToolOrchestra, researchers trained Orchestrator, a model based on 8 billion parameters Qwen3-8B. They evaluated its performance in three challenging benchmark tests: The final test of humanity (hle),, FRAME AND Tau2 bench. It was compared to several baselines, including immense commercially available LLMs, both with and without tools.
The results showed that even powerful models performed without tools, confirming their need for intricate reasoning. While the addition of tools improved the performance of immense models, it often resulted in skyrocketing costs and delays.
Meanwhile, 8B Orchestrator delivered impressive results. For HLE, a benchmark for PhD-level questions, Orchestrator significantly outperformed prior methods at a fraction of the computational cost. In the function calling test, Tau2-Bench successfully scheduled various tools by calling a immense model such as GPT-5 in only about 40% of the steps and using cheaper options in the rest, while still beating an agent that used a immense model in every step.
The researchers noted that the RL-trained Orchestrator adapted his strategy to modern challenges, demonstrating a “high degree of general reasoning ability.” Most importantly for enterprise employ cases, Orchestrator also generalized well to models and pricing structures that he had not seen during training. This flexibility makes the platform suitable for companies that rely on a combination of public, private and customized AI models and tools. Lower cost, greater speed, and customizability make it a practical approach to creating sophisticated AI agents that can scale.
As companies look to deploy more advanced AI agents, this orchestration-based approach opens the door to systems that are not only smarter, but also more cost-effective and easier to control. (T model scales are currently available under a non-commercial license, but Nvidia also made the file available training code under the liberal Apache 2.0 license.)
As the article summarizes, the future may lie in even more advanced versions of this concept: “Looking to the future, we envision more sophisticated recursive orchestrator systems that will push the upper bounds of intelligence [and] also to further increase efficiency in solving increasingly complex agentic tasks.”