RecursiveMAS/RecursiveMAS

What it solves

RecursiveMAS addresses the inefficiency of treating multi-agent systems (MAS) as collections of isolated LLM modules. It aims to improve coordination and performance across various collaboration styles—such as sequential, mixture, distillation, and deliberation—by allowing agents to communicate and refine information more effectively than traditional discrete interactions.

How it works

Instead of isolated modules, the framework treats the entire multi-agent system as a unified recursive computation. It uses lightweight "RecursiveLink" modules to connect heterogeneous agents, enabling them to exchange and evolve latent states across multiple recursion rounds.

The system is trained using an Inner-Outer Loop paradigm:

1. Inner Loop: Provides a model-level warm start for each individual agent.
2. Outer Loop: Trains the RecursiveLink modules at the system level to optimize the coordination between agents.

Who it’s for

This project is designed for AI researchers and developers working on multi-agent coordination, LLM-based system optimization, and those looking to implement complex agent collaboration patterns (like distillation or deliberation) for tasks in mathematics, science, medicine, and coding.

Highlights

• Latent-Space Recursion: Scales collaboration by refining shared latent states rather than just exchanging text.
• Diverse Collaboration Patterns: Supports Sequential, Mixture, Distillation, and Deliberation styles.
• Inner-Outer Loop Training: A two-phase co-optimization strategy for individual agents and their connections.
• Broad Benchmark Success: Demonstrated improvements across 9 benchmarks including Math500, GPQA, and LiveCodeBench.