Multi Turn Eval

How do you evaluate a customer service chatbot that handles multi-turn conversations? What dimensions matter?

A single-turn eval shows your chatbot scores 4.2/5. But users complain it 'forgets what they said.' Which eval dimension is your single-turn setup missing?

What makes multi-turn dialog evaluation harder than single-turn, and how does MT-Bench address it?

Which quality dimension is uniquely challenging about multi turn dialog evaluation vs single turn?