Paper summary: Recursive Multi-Agent Systems

Problem

Looped and recursive models (Ouro, recurrent depth, TRM on this blog) scale reasoning by repeating computation in latent space. Can agent collaboration scale the same way?

Method

RecursiveMAS treats the entire multi-agent system as one latent-space recursive computation:

• a lightweight RecursiveLink module connects heterogeneous agents in a loop, so they generate latent thoughts and transfer latent states across agents instead of text
• an inner–outer loop learning algorithm co-optimises the whole system, with shared gradient-based credit assignment across recursion rounds
• theoretical analysis argues it’s more efficient than text-based multi-agent systems and keeps gradients stable

Key results

From the abstract: across 4 collaboration patterns and 9 benchmarks (maths, science, medicine, search, code), an average +8.3% accuracy with 1.2×–2.4× end-to-end inference speed-up over advanced single-agent, multi-agent and recursive baselines. Figure 1 shows a clean scaling trend as recursion deepens for sub-1.5B agents.

Limitations

There’s no dedicated limitations section in the extracted text. Latent communication is harder for people to inspect than text messages between agents.