Has there been a feature request for LMArena to support continuing the conversation…
- User
Has there been a feature request for LMArena to support continuing the conversation separately for the two LLMs being compared? I.e. the first prompt would always be the same for both LLMs, but subsequent messages wouldn't have to be.
Assistant
Based on the search results provided, there doesn't appear to be a specific feature request for LMArena (also called Chatbot Arena) to support continuing conversations separately for the two LLMs being compared, where the first prompt would be identical but subsequent messages could differ between models.
The closest relevant mention is in one of the articles about ChatBotArena which states: "LMSYS, you should maybe consider tweaking the interface to encourage multi-turn conversations; more on this need later"[3]. This suggests that someone has recognized the need for better support of multi-turn conversations in the platform, but it's not specifically requesting the ability to have diverging conversation paths for each model.
The current ChatBotArena design appears to focus on single-turn or identical multi-turn interactions, as evidenced by statistics showing "the prevalence of single-turn interactions"[3]. This aligns with its primary purpose as an evaluation platform that collects pairwise comparisons under identical conditions to ensure fair model assessment.
Other search results discuss various conversation continuity techniques, including:
- A prompt technique for maintaining conversation continuity across different chat sessions[1]
- Methods for keeping conversation context for multiple users separate[7]
- Follow-up responses in voice assistants without repeating wake words[4]
- A detailed product idea for branching conversations in AI systems[10]
However, none of these specifically mention implementing such features in LMArena for the purpose of evaluating and comparing different LLMs along separate conversation paths.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.