Researchers at Nvidia and the University of Hong Kong have released Orchestrator, an 8-billion-parameter model that coordinates different tools and enormous language models (LLMs) to resolve complex problems. Of their experiments, Orchestrator achieved higher accuracy at a lower cost than much larger models in tool-use benchmarks, while also aligning with user preferences on which tools to make use of for a given query.
The model was trained through ToolOrchestra, a brand new reinforcement learning (RL) framework for training small models to act as intelligent coordinators. The approach is predicated on the concept a small "orchestrator" managing a various team of specialised models and tools may be simpler and efficient than a single, monolithic AI system.
The findings suggest that this composite approach could pave the way in which for more practical and scalable AI reasoning systems within the enterprise.
The bounds of current LLM tool use
Giving LLMs access to external tools is a promising strategy to extend their capabilities beyond their training data and into agentic tasks. By calling on resources like search engines like google and code interpreters, AI agents can improve their accuracy and perform in-app tasks.
Nevertheless, within the accompanying paper, the researchers argue that the present approach to constructing tool-using agents doesn't harness the complete potential of this paradigm. Most systems equip a single, powerful model with a set of basic tools like an internet search or a calculator.
They argue that humans, when reasoning, “routinely extend themselves by calling upon resources of greater-than-human intelligence, from domain experts to stylish processes and software systems.” Accordingly, LLMs should have the ability to interact with a big selection of tools in numerous capacities.
The tool orchestration paradigm
The paper proposes a shift from a single-model system to a composite one, managed by a light-weight "orchestrator" model. The orchestrator's job is to research a fancy task and break it down, invoking the suitable tools in the suitable order to reach at an answer.
This toolset includes not only standard utilities like web search and code interpreters, but other LLMs of assorted capabilities that function as "intelligent tools." For instance, the orchestrator can delegate a quantitative query to a math-focused model or a programming challenge to a code-generation model. As a substitute of placing the whole cognitive load on one large, generalist model, the orchestrator delegates narrowed-down sub-problems to specialized intelligent tools.
Based on this idea, the researchers developed ToolOrchestra, a way that uses RL to coach a small language model to act as an orchestrator. The model learns when and call upon other models and tools, and mix their outputs in multi-turn reasoning. The tools are defined in an easy JSON format, specifying their name, description and parameters.
The RL training process is guided by a reward system that produces an economical and controllable agent. The reward balances three objectives: The correctness of the ultimate answer, efficiency in cost and latency and alignment with user preferences. For instance, the system is penalized for excessive compute usage, and is rewarded for selecting tools that a user has marked as preferred, comparable to favoring an open-source model over a proprietary API for privacy reasons. To support this training, the team also developed an automatic data pipeline that generated hundreds of verifiable training examples across 10 different domains.
A small model with big results
Using ToolOrchestra, the researchers trained Orchestrator, an 8-billion-parameter model based on Qwen3-8B. They evaluated its performance on three difficult benchmarks: Humanity’s Last Exam (HLE), FRAMES and Tau2-Bench. It was compared against several baselines, including large, off-the-shelf LLMs each with and without tools.
The outcomes showed that even powerful models struggled without tools, confirming their necessity for complex reasoning. While adding tools improved performance for big models, it often got here with a steep increase in cost and latency.
In contrast, the 8B Orchestrator delivered impressive results. On HLE, a benchmark of PhD-level questions, Orchestrator substantially outperformed prior methods at a fraction of the computational cost. On the Tau2-Bench function-calling test, it effectively scheduled different tools, calling a big model like GPT-5 in just about 40% of the steps and using cheaper options for the remaining, while still beating an agent that used the massive model for each step.
The researchers noted that the RL-trained Orchestrator adapted its technique to recent challenges, showing a "high degree of general reasoning ability." Crucially for enterprise applications, Orchestrator also generalized well to models and pricing structures it hadn't seen during training. This flexibility makes the framework suitable for businesses that depend on a mixture of public, private and bespoke AI models and tools. The lower cost, higher speed and customizability make it a practical approach for constructing sophisticated AI agents that may scale.
As businesses look to deploy more advanced AI agents, this orchestration approach offers a path toward systems that aren’t only more intelligent but more economical and controllable. (The model weights are currently available under a non-commercial license, but Nvidia has also released the training code under the permissive Apache 2.0 license.)
Because the paper concludes, the long run may lie in much more advanced versions of this idea: “Looking ahead, we envision more sophisticated recursive orchestrator systems to push the upper sure of intelligence [and] also to further enhance efficiency in solving increasingly complex agentic tasks.”
