When Evidence is Sparse: Weakly Supervised Early Failure Alerting in Dialogs and LLM-Agent Trajectories

In three linesMethod for early failure detection in dialogs and LLM-agent trajectories. Two-stage approach: attention-based predictor identifying sparse failure evidence (4.7-11.3% of turns) and α-STOP policy selecting operating point at inference time. 3-42% Pareto-frontier improvement over state-of-the-art trigger policies.

Summary generated by Claude — human-verified