MinionsLLM: a Task-adaptive Framework For The Training and Control of Multi-Agent Systems Through Natural Language

TL;DR

MinionsLLM is a framework that combines large language models with behavior trees and formal grammars to enable natural language control of multi-agent systems in customizable environments, improving task performance and syntactic validity.

Contribution

It introduces a novel integration of LLMs with behavior trees and formal grammars, along with synthetic dataset generation methods for fine-tuning, to enhance multi-agent system control.

Findings

Method B achieves 92.6% syntactic validity.

33% mean task performance improvement over baseline.

Smaller models benefit most from fine-tuning.

Abstract

This paper presents MinionsLLM, a novel framework that integrates Large Language Models (LLMs) with Behavior Trees (BTs) and Formal Grammars to enable natural language control of multi-agent systems within arbitrary, user-defined environments. MinionsLLM provides standardized interfaces for defining environments, agents, and behavioral primitives, and introduces two synthetic dataset generation methods (Method A and Method B) to fine-tune LLMs for improved syntactic validity and semantic task relevance. We validate our approach using Google's Gemma 3 model family at three parameter scales (1B, 4B, and 12B) and demonstrate substantial gains: Method B increases syntactic validity to 92.6% and achieves a mean task performance improvement of 33% over baseline. Notably, our experiments show that smaller models benefit most from fine-tuning, suggesting promising directions for deploying compact, locally hosted LLMs in resource-constrained multi-agent control scenarios. The framework and all resources are released open-source to support reproducibility and future research.