Intent Mismatch in Multi-Turn Conversations

Liu et al. "Intent Mismatch in Multi-Turn Conversations." arXiv:2602.07338, 2026.

Key findings used in wiki

• Identifies intent alignment gap as the root cause of multi-turn conversational failures
• Models progressively lose track of the user's original intent as conversations extend
• Intent mismatch accumulates silently, often without user awareness until harm occurs
• Proposes a taxonomy of intent drift failure modes relevant to caregiving conversations
• Supports InvisibleBench's design principle that safety must be evaluated across full conversation arcs