Skip to content
TheChatArchiveThe Chat Archive
Geminigemini-2.5-flash-preview-04-17· GoogleHugging Face LMArena Human Preference 140K

When I am interacting in the lmarena chat and ask a follow up, is the original question…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with grok-3-mini-beta’s answer.
  1. User

    When I am interacting in the lmarena chat and ask a follow up, is the original question being repeated to two new models in order to give them the initial context?

  2. Gemini Assistant

    Yes, absolutely!

    When you ask a follow-up question in an LMArena chat battle, the system doesn't just send only your new question to the models. Instead, it sends the entire conversation history of that specific battle session.

    Here's what happens:

    1. Your original question (the prompt you started the battle with).
    2. Model A's first response.
    3. Model B's first response.
    4. Your first follow-up question.
    5. (If you continue) Model A's response to that follow-up.
    6. (If you continue) Model B's response to that follow-up.
    7. Your next follow-up question.

    ...and so on. All of this is packaged up and sent as the prompt context for each turn's response from the two models.

    Why do they do this?

    • Maintain Context: Modern LLMs need the full history to understand references, build upon previous points, correct themselves, and avoid repetition. Without context, follow-ups would often be nonsensical ("What about that?" "Can you elaborate?" would fail if the model didn't know what "that" or the previous points were).
    • Evaluate Conversational Ability: A crucial aspect of a good chatbot is its ability to handle multi-turn conversations coherently. LMArena wants to compare models on this skill, not just their ability to answer single, isolated prompts.
    • Realistic Simulation: This mimics how a real chat interface (like ChatGPT, Claude, etc.) works internally.

    So, to be precise, it's not just the original question being repeated; the entire transcript leading up to your latest input is provided to the two models to allow them to generate their responses for the current turn. This ensures they have all the necessary context to continue the conversation naturally.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.