Researchers tune prompts and model choice together to cut multi-agent AI costs

What the paper proposes

A multi-agent system is a group of AI agents, each with a role, working together on a task. Each role gets its own instructions (a prompt) and runs on a particular language model.

As arxiv.org reported, existing prompt-tuning methods usually leave the model choice fixed. That means differences in what each role needs, and what each costs to run, are never part of the search. The authors' framework, UniMASO, tunes both at once.

How UniMASO works

The system tries candidate changes and scores them using three kinds of feedback: local, downstream and global. It then keeps a set of options that balance quality against cost, rather than chasing quality alone.

Updates are coordinated according to each agent's role and how it depends on the others. A memory of past successes and failures also guides later rounds of tuning.

What the results say

The authors ran experiments on five benchmarks and five execution topologies, meaning different ways of wiring the agents together. They report competitive task performance at much lower deployment cost than the strongest single-model setups.

In the centralized case, the average score moved from 69.79 to 70.40. Cost, weighted by tokens (the chunks of text a model processes), dropped 85.2% compared with DeepSeek-V4-Pro paired with MASPO.

How much to trust it

This is a v1 preprint, so it has not been through peer review. The figures come from the authors' own benchmarks, and the abstract gives the headline comparison for one setup only.

The code is public, which lets others check the claims. Still, the idea is simple enough to borrow without the paper: not every step in a workflow needs the most powerful model.