CORE Method Mitigates Persona Drift in Multi-Turn Dialogue
September 14, 2026
The CORE framework separates turn-local evidence from persistent persona updates to prevent models from adopting transient user behaviors. By using uncertainty-aware belief revision, it maintains stable user profiles across long-term interactions while improving fidelity on the PERSIST benchmark.
HOW THIS AFFECTS YOU
●
builderYou can implement more stable long-term personalization in conversational agents without the risk of personality drift.
●
researcherThe PERSIST benchmark offers a new way to evaluate model robustness under sequential interaction stress.