Intent Mismatch in Multi-Turn Conversations

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

Key findings used in wiki

• The paper reframes "lost in conversation" as an intent alignment gap between how users express needs and how models interpret them, not only as a raw capability shortfall.
• It shows how models can treat partial follow-ups as confirmation of earlier bad assumptions, silently drifting away from the user's actual goal.
• The paper is useful methodological support for scenarios that reveal intent incrementally and test whether the model realigns over time.
• Its Mediator-Assistant architecture is a design proposal, not a direct GiveCare product claim; the wiki uses the paper mainly for failure analysis and benchmark rationale.