MixAgent splits music mixing into steps that an engineer can read and edit

A mixing plan, not a black box

Mixing is the step in music production where separate recordings, such as vocals, guitar and drums, are balanced and shaped into one finished song. As arxiv.org shows in a newly posted paper, a team has built MixAgent, a system that automates this work by copying the way human audio engineers organize it.

The system is an "agentic pipeline," meaning several AI components each handle one job and pass results along. The output is a mixing plan that a person can read and edit, guided by a reference song whose sound the user wants to approach.

That design choice is the main point of the paper. Instead of returning a single finished audio file, the system lays out which effects go where, so an engineer could step in, change one setting and keep the rest.

Four roles, one workflow

The first piece is a grouping agent. It sorts related input tracks into subgroups, for example putting all the drum microphones together so they can be treated as one instrument.

Next comes a planning agent. It looks up similar examples and uses them to choose which audio effects to apply and in what order. The effects named in the paper include volume control and reverb, which adds a sense of room echo. Choices are made at three levels: single tracks, subgroups and the full mix.

Then the refinement stage begins. A detector flags what needs correcting, and a critic proposes specific numerical changes to settings, working from individual tracks up to the whole mix. The authors report that the retrieved examples improved how well the system chose effects.

Why "listen and fix" prompts fall short

One finding stands out. The authors tried simply prompting recent multimodal models, which can take in audio as well as text, to listen to a mix and revise it. In their experiments this often made the mix worse. They describe reliable correction as the central challenge.

Their answer was to build training data. MixCritic contains 45,657 repair critiques, derived from settings that were fitted to professional mixes. In effect, the data teaches the critic what a fix toward a professional result looks like in numbers.

The critic was trained in two stages. First came supervised fine-tuning, which means learning from worked examples. Then came preference optimization, which nudges the model toward changes whose rendered audio turned out better. The outcome was judged on actual rendered sound, not just on whether the suggested numbers looked plausible.

What the results show, and what they don't

On MedleyDB, the collection of music the authors tested on, learned refinement reduced error by 19% relative to the initial mix. Preference optimization accounted for an 8.4% reduction over supervised training alone.

A human listening test also compared MixAgent with mixes made by Qwen3-Omni, Gemini 3.8 Flash and GPT-4o-audio. Listeners gave MixAgent higher quality ratings than all three.

There are limits to what can be said from the summary available. It does not state how many people listened, what kinds of songs were used beyond the named dataset, or how the system handles genres far from its training material. This is also a preprint, so it has not necessarily gone through peer review. Treat the results as the authors' own claims until others test them.

Who should pay attention

For working engineers and home producers, the draw is control. A plan expressed as named effects and numeric settings is something you can audit, adjust or reject, which is closer to how a human assistant would work.

For everyone else, the paper is a small but clear example of a wider pattern. General-purpose models are not automatically good judges of specialized output, and carefully built training data from experts can matter more than a bigger model. It is also a reminder that making an AI's steps visible can be as valuable as making its results better.