Bi-Predictability: A Real-Time Signal for Monitoring LLM Interaction Integrity

arXiv:2604.13061v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly deployed in high-stakes autonomous and interactive workflows, where reliability demands continuous, multi-turn coherence. However, current evaluation methods either rely on post-hoc semantic judges,...