LM Arena

• Side-by-side comparison of two anonymous LLMs
• Crowd-voted battles to determine which model performs better
• Elo-based model rating and leaderboard system
• Support for a large variety of LLMs (open-source and commercial)
• Persistent user prompt tracking and history of evaluations
• Search Arena extension to evaluate search-augmented or retrieval-enabled models
• Community-driven platform with voting, feedback, and transparency
• Public model performance statistics and ranking data
• Role-based contribution and participation via community platform
• Secure and open infrastructure to ensure neutrality and fairness

LM Arena is a powerful and open web platform created to benchmark large language models (LLMs) in a way that truly reflects human preferences. As users, you submit a prompt and receive responses from two anonymous models running side-by-side. After reading both, you vote on which answer you prefer. Only after voting are the model identities revealed. This approach ensures unbiased feedback and captures real human judgments about what makes a response compelling, useful, or coherent.

By gathering massive amounts of human preference data from these pairwise battles, LM Arena computes global Elo ratings for models. These ratings feed into a public leaderboard, letting anyone see which models are most favored. The data grows richer over time, as new models are added and the community continues voting. This creates a dynamic evaluation ecosystem where performance is not just measured once, but constantly refined by real users.

LM Arena also supports specialized evaluations. For example, the Search Arena tests search-augmented LLM systems, helping determine which models integrate real-time web data in the most useful and accurate way. This makes LM Arena more than just a generic benchmark—it is a living, evolving platform for testing the latest innovations in model design.

The platform was born out of academic research and is deeply rooted in transparency. It places a strong emphasis on neutrality: the organization behind it remains dedicated to providing a fair, open space for evaluation. The community is at the heart of the process, contributing not just votes, but feedback shaping new features, such as improved user interfaces, mobile support, and better evaluation workflows.

LM Arena’s mission is to bridge the gap between model developers and real-world users, offering a tool where model quality is judged by people, not just by automated metrics. Its democratized, community-driven benchmarking contributes to greater model reliability, helps expose strengths and weaknesses, and drives meaningful innovation in the world of AI.