AERIC: Anticipatory Hidden-State Monitoring for Implicit Harmful Dialogue 事件

AERIC: Anticipatory Hidden-State Monitoring for Implicit Harmful Dialogue arXiv:2605.23974v1 Announce Type: new Abstract: Current language models create two safety challenges: risk must be detected early enough to avoid exposing harmful continuation, and the harmfulness itself may be implicit rather than signaled by overtly toxic text. Existing response-level guards are strong at judging completed text, and native streaming guards move closer to token time, but both settings leave open whether