Audio-language models usually take single-channel sound, so they can say what a noise is but not where it comes from. Direction needs multi-channel recordings, and a useful assistant must also handle follow-up questions and lookups.
The system trains an agent, meaning a model plus a loop that lets it act, on ambisonic audio, a format whose channels encode direction. It writes out reasoning and a scene description before answering, keeps that description in the conversation, and can call a tool and use the result.
It scored 44.9 percent on a new spatial benchmark, 7.4 points above the best earlier system. Removing the reasoning training cost far more accuracy on an outside test, so the step-by-step reasoning carries real weight.