ProMAS：基于马尔可夫转移动态的多智能体系统主动错误预测

The integration of Large Language Models into Multi-Agent Systems (MAS) has facilitated the resolution of complex, long-horizon tasks through collaborative reasoning. However, this collective intelligence is inherently fragile, as a single logical fallacy can rapidly propagate and lead to system-wide failure. Current research predominantly relies on post-hoc failure analysis, which limits real-time intervention capabilities. To address this issue, the ProMAS framework is proposed, utilizing Markov transitions for predictive error analysis. ProMAS extracts Causal Delta Features to capture semantic displacement and maps them to a quantized Vector Markov Space, modeling reasoning as probabilistic transitions. By integrating a Proactive Prediction Head with a jump mechanism, ProMAS enables preemptive alerts before errors occur, thereby enhancing the robustness of multi-agent systems.