ASE 2026 paper: Singapore Management University et al. propose ProbGuard, predicting LLM agent risk and intervening in advance
ASE 2026|新加坡管理大学:别等出事再拦截!ProbGuard 给 LLM Agent 装上「风险预判雷达」
Researchers from Singapore Management University and other institutions propose ProbGuard (an ASE 2026 paper), an active runtime monitoring method for LLM agents that predicts the probability of an agent heading toward danger before danger occurs and intervenes in advance. The method offline trains a discrete-time Markov chain (DTMC) model using predicate abstraction and historical trajectories and provides PAC statistical guarantees; online, it computes in real time the probability P_safe that continued execution remains safe, and when this falls below a threshold θ, it triggers a re-prompt (reflect) or direct stop (stop) intervention. In autonomous driving experiments, under a zero-false-positive threshold it can warn of traffic violations 15.84 seconds in advance, and under a stricter safety threshold it can alert of collision risk up to 38.66 seconds in advance; in a household embodied agent experiment, the baseline unsafe behavior proportion was 40.63%, and at θ=0.9 the re-prompt mode reduced unsafe behavior by a relative 65.37% while retaining 80.4% of the baseline task completion rate, and the stop mode reduced it by a relative 93.60% but the task completion rate dropped to 10.42%.