← AI Terminology
Multi-Turn Conversation
Multi-turn conversation is dialogue spanning multiple user–assistant exchanges where each turn conditions on prior messages in the thread.
It is the default UX of chatbots and the setting where context, memory, and instruction hierarchy matter.
It is the default UX of chatbots and the setting where context, memory, and instruction hierarchy matter.
Why It Matters in AI
Single-shot prompts miss clarification, correction, and collaborative work. Multi-turn systems must track state, resolve references (“do that again”), and stay consistent — the real deployment mode of LLMs.
Key Points
| Aspect | Description |
|---|---|
| Eval | Multi-turn benchmarks beyond single questions |
| State | Message history in context window or summarised |
| Safety | Jailbreaks over many turns; gradual policy erosion |
| Product | ChatGPT/Claude threads, support bots |
| Related | Chat templates, memory, context engineering |
| Challenges | Anaphora, goal drift, long-thread lost details |
Simple Analogy
A back-and-forth meeting rather than a single sticky note — later sentences only make sense if you heard earlier ones.
Common Usage Examples
- Chat APIs with
messages=[{role, content}, ...] - Thread summarisation after long chats
- Multi-turn red-teaming
- Customer-support dialogue flows
Summary
In short: Multi-turn conversation is multi-exchange dialogue with shared history — how people actually use chat AI.