Three research papers published on arXiv address distinct challenges in multi-agent AI systems, spanning continual learning, security, and training methodologies.
According to arxiv.org, researchers introduced MasFACT, a framework addressing “topology forgetting” in multi-agent systems powered by large language models. The paper identifies that existing topology generation methods optimize for isolated tasks, causing systems to lose previously effective collaboration patterns when adapting to new tasks. MasFACT uses “geometry-aware posterior transfer” to preserve historical collaboration knowledge through Fused Gromov-Wasserstein optimal transport and PAC-Bayes-guided adaptation. Experiments demonstrated that MasFACT “consistently improves average accuracy while reducing topology forgetting” across different continual learning settings.
In a separate development, arxiv.org reports on an open-source Multi-Agent Firewall Architecture designed to protect privacy during LLM interactions. The system combines a browser extension and proxy to intercept HTTP(S) and WebSocket communications, implementing “data leakage prevention through a hybrid approach combining deterministic detectors with LLM-driven semantic analysis.” The framework achieved F1 scores up to 94.93% in optimal configurations.
Finally, according to arxiv.org, researchers proposed Visual Inspection of Policies (VIP), which uses Video Language Models to assess task difficulty for reinforcement learning curricula by “directly inspecting policy behavior via recorded episode videos.” Testing on the StarCraft Multi-Agent Challenge showed that VIP with VideoLLaMa2-7B “generate[d] more effective curricula than both its text-only ablation and methods that rely on scalar task scores.”