Summary RecursiveMAS is an open-source framework designed to scale agent collaboration within multi-agent systems (MAS) by employing latent-space recursion. It targets researchers and developers building complex AI applications, offering a novel approach where the entire MAS functions as a unified recursive computation, enabling iterative refinement and evolution of agent states for enhanced collective intelligence.

How It Works The core architectural choice is to treat the MAS as a single, recursive computational entity rather than a collection of independent modules. Heterogeneous LLM agents are interconnected through lightweight "RecursiveLink" modules. These links facilitate iterative exchange and refinement of agent latent states across multiple recursion rounds. This unified recursive computation approach aims to overcome limitations of traditional MAS designs by enabling deeper, more dynamic collaboration and state evolution.

Quick Start & Requirements

• Installation: Requires Python 3.10. Set up a conda environment (conda create -n recursivemas python=3.10 -y), activate it (conda activate recursivemas), and install dependencies (pip install -r requirements.txt).
• Prerequisites: For Deliberation-Style, a Tavily API key must be configured in a .env file. CUDA is recommended for accelerated inference (device="cuda").
• Setup: Users must download pre-trained model checkpoints from Hugging Face, organized by collaboration style and agent role.
• Links:
  • Model Checkpoints: Hugging Face (specific links detailed in README tables).
  • Code Repository: GitHub.
  • Research Paper: arXiv.

Highlighted Details

• Implements four distinct collaboration patterns: Sequential (Light & Scaled), Mixture, Distillation, and Deliberation.
• Offers a range of pre-trained agent checkpoints (1.5B to 9B parameters) for various roles and styles, hosted on Hugging Face.
• run.py provides a unified command-line interface for executing inference across different MAS configurations and datasets.

Maintenance & Community The provided README does not contain specific information regarding project maintainers, community support channels (e.g., Discord, Slack), or a public roadmap.

Licensing & Compatibility The software license is not explicitly stated in the README. Consequently, compatibility for commercial use or integration into closed-source projects remains undetermined.

Limitations & Caveats The project appears to be under active development, with announcements indicating that the "complete training/inference pipeline and additional features" are still forthcoming. The Deliberation-style requires an external API key for its tool-calling capabilities.