Meta AI introduces MIRA architecture for long-horizon research agents

In an arXiv paper, Meta AI researchers proposed MIRA, an architecture designed to help research agents decide what to investigate next during long-horizon tasks. To overcome the difficulty of learning from sparse decisions in long execution traces, MIRA splits the agent into an outer meta-reasoner that reads a persistent research record to issue work orders, and a fresh executor that carries out each order.

Because decisions occur exclusively at work-order boundaries, the researchers train a critic at those points to forecast remaining return, followed by a unified actor-critic model (MIRA-AC) that assesses partial progress and chooses the next task. The structural separation improves theorem proving and open-ended architecture research without training, while MIRA-AC trained on the model's proxy signals improves gold scores across four autoresearch environments.