Code a Soul, Soul a Code: All the chips are stages, and all the Agents merely players

ACL ARR 2025 May Submission4366 Authors

Abstract:

Large language models (LLMs) have transitioned agent systems from theory to practice, highlighting the critical challenges of safety and controllability in multi-agent systems, particularly in complex environments. Current research often overlooks the need for dynamic trust mechanisms in multi-agent collaborations, especially in real-world applications with cross-modal interactions and dynamic adaptation.
To address these issues, we propose the Security-Oriented Multi-Agent System (SOMAS), a novel framework that uses reinforcement learning to enable trusted and secure interactions among agents. SOMAS integrates real-time task execution with simulation training, creating a supervised closed loop of "execution - simulation - optimization." This design ensures policy stability and decision traceability across domains, enhancing the safety and reliability of multi-agent systems in emergency management.
Our experiments show that SOMAS optimizes policy stability and decision traceability in cross-domain tasks, improving system safety and reliability. We also release the first fine-tuned multi-modal safe large language model, with training data and an evaluation dataset for multimodal security outputs. Our dynamic security validation approach improves assistance by 11% and reduces risk response rates to 18%-48% compared to traditional methods.
SOMAS represents a significant step toward secure and trustworthy multi-agent systems, offering a robust solution for complex real-world applications.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: Multi-Agent System, ​​Security Interaction, Reinforcement Learning

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English, Chinese

Submission Number: 4366