Provable Limits and Certified Deferral for Verbalized Uncertainty in Small Language Models

За 30 секунд

Что произошло

Research proves small language models cannot reliably defer to humans using verbalized confidence alone, even with calibration techniques.

Почему это важно

Engineers deploying small models offline or in cost-sensitive settings need to know when deferral to humans is mathematically safe.

На что обратить внимание

Only three of twenty-two model-task pairs achieved certified safe autonomy at twenty percent risk; results assume independent and identically distributed deployment.