Alexandre Saunier and I are presenting a paper co-authored with Maurice Jones about the behaviour generator based on large language models (LLMs) that we used during the Wilding AI workshop at MONOM, part of CTM Festival 2025. Some highlights of what we’re presenting:
- LLM Behavior Generator that transforms textual prompts into expressive sound modulations and motions in space.
- Chain of thoughts involving a “recipe” LLM cascading into a “code” LLM.
- Max for Live implementation for artistic use in Ableton Live with immersive audio systems such as 4D SOUND.
- Artistic experimentations and user feedbacks suggests such LLM-based system should be contextualized in the history of sound practices employing aleatoric, emergent, and open-ended strategies.
Here are the title and abstract:
Large Language Models to generate sonic behaviors: the case of Wilding AI in exploring creative co-agency
Large Language Models (LLMs) and foundational models play a central role in multimodal text-to-sound systems, such as text-to-speech and text-to-music, as well as in recent musical agent systems designed to automate music production tasks. We explore an alternative approach that employs LLM-based sonic agents as spatial composition techniques. This method, emerging from the Wilding AI research-creation project, integrates LLMs into Max/MSP, Ableton Live, and spatial sound environments. LLMs are used to generate step-by-step sequences controlling spatial audio parameters, including sound motion in 3D space and other assignable sound properties.
This paper details the artistic framework of Wilding AI, a LLM behavior generator, a live performance at the CTM Festival 2025, and discussions on composition, improvisation, time, and agency. Our work repositions LLMs as tools for shaping sonic experiences rather than merely generating finished audio material.
Here is the full paper: https://zenodo.org/records/16946605
This is one outcome of the collective research-creation lab Wilding AI.
