erwinmsmith/SOMAS

A Trusted Human-Multi-Agent Reinforcement Learning Interaction Framework

What it solves

SOMAS addresses the challenge of deploying multi-agent systems in high-stakes crisis response scenarios where safety and reliability are paramount. It provides a framework that balances task completion with strict safety guardrails, ensuring that AI agents act responsibly when collaborating with humans.

How it works

The framework operates in two primary modes: an online execution system and an offline simulation system. The online system uses a planning-execution pipeline integrated with GPT-4-based risk assessment to monitor and guide agent actions in real-time. The offline system focuses on training and optimization, using synthetic task generation and experience replay to refine reinforcement learning policies for safer decision-making.

Who it’s for

This project is designed for researchers and developers building safety-critical AI applications, particularly those focused on human-machine collaboration in emergency or crisis management environments.

Highlights

  • Dual-mode architecture supporting both real-time task execution and offline simulation training.
  • Includes a fine-tuned safe LLM and a specialized dataset for emergency scenarios.
  • Demonstrates a 15% improvement in helpfulness and a 40% reduction in risk response rates compared to baseline models.
  • Implements a dynamic trust mechanism that weighs task utility against safety constraints using reinforcement learning.

Related

  • Project
    TsinghuaC3I/MARTI

    An open-source framework for training LLM-based multi-agent systems using reinforcement learning, featuring tree search augmentation for complex reasoning and code generation.

  • Project
    liquidos-ai/AutoAgents

    A high-performance, modular multi-agent framework written in Rust that supports structured tool calling, local and cloud LLM backends, and secure agent execution.

  • Project
    RecursiveMAS/RecursiveMAS

    A multi-agent framework that improves agent collaboration by treating the system as a unified recursive computation in latent space, reducing token usage and speeding up inference.

  • Project
    fan-ziqi/rl_sar

    A framework for simulation verification and physical deployment of reinforcement learning policies for quadruped, wheeled, and humanoid robots.

  • Project
    langwatch/scenario

    An agent testing framework that uses simulations, user simulators, and judge agents to evaluate the behavior and robustness of AI agents in multi-turn conversations.