LMArena and The Future of AI Reliability

Alongside this funding announcement, we’re excited to relaunch LMArena with a whole new UI! Our community’s feedback is the foundation of today’s launch and everything that follows. The backing will help us accelerate even further, but the heart of it all remains the same. LMArena will always be a neutral, open, community-driven space for evaluating and improving AI reliability. That’s not just our strategy, it’s our conviction, built on scientific rigor, fairness, and transparency in AI evaluation. Another reason why we chose to move forward with long-term partners who support the mission, as much as the science behind it all.

We started as a research project under UC Berkeley / LMSys. Then we became Chatbot Arena, publishing our first leaderboard fueled by community-driven model evaluation. Today, we officially begin our new chapter as LMArena, with a whole new look and a broader community-driven platform spanning multiple modalities, shaped by how people actually use AI. This shift reflects a deeper commitment to understanding not just which models perform best, but for whom, in what context, and why that matters.

The original Chatbot Arena project was born from the research community, and scientific curiosity still drives everything we do. We believe better evaluations lead to better models. Better models mean more reliable AI for everyone. Over the past few months, we’ve been rebuilding from the ground up, directly with our community. From the early Alpha invite in our Discord server to the open Beta, thousands of you have tested the limits, shared detailed feedback, and helped us shape the platform into what it is today. Thank you! Here’s a quick look at what the community accomplished during Alpha & Beta:

• Over 40,000 votes cast during Beta
• Over 1,000 polish and feature suggestions reviewed
• 40+ bugs reported and fixed

Today, this community-built, battle-tested version becomes the official LMArena at lmarena.ai.

It brings big improvements across the board:

• Better performance and reduced error rates
• A more intuitive layout and cleaner voting experience
• Mobile optimization and smoother conversation flow
• More transparency around what’s happening behind the scenes

Beyond the LMArena user experience, we believe that advancing the science of human preference evaluation is just as important as advancing the models themselves. That’s why we publish our evaluation methods, model sampling rules, and platform metrics. If we want trustworthy AI, the evaluations themselves must be trustworthy too. Our company will continue to support fundamental research and development in this area. We welcome the larger research community to come together with us to uphold evaluation integrity.

Features on LMArena like “Style Control” demonstrate the difference in human preference when the style and formatting of a model’s response is filtered. Prompt-to-Leaderboard (P2L) is another example, which generates a customized leaderboard of AI models specific to a prompt. Both are just the beginning of understanding the diversity of human preference. We’re actively working on more, but we’re not here just to build features. We want to lay the groundwork for human preference evaluation methods that are reproducible, robust, and truly reflective of how AI is used in the real world.

This launch and funding news marks a huge milestone, but LMArena is still very much a work in progress. We’re continuing to ship improvements from faster interfaces to better sampling methods, all grounded in our belief that great science requires open data, reproducible methods, and direct feedback from our community. Some features from the legacy site, like advanced model settings aren’t live yet, but they’re on the way.

Because of this, the legacy site will still be available at legacy.lmarena.ai while we continue to evolve and bring more features to parity. Going forward, all updates and improvements will be happening on the new LMArena, and we’re excited to keep building it with your input.

We’ll keep building in the open, because it’s the community that will help shape what comes next. If something feels off, confusing, or just missing, let us know. Every vote, comment, and bug report brings us one step closer to the platform you’ve asked for. Here are a few things you’ve already requested that are on the way:

• More Modalities
• Data Explorer
• User Login

The future of LMArena is about making the platform faster and easier to access for everyone. We’re especially excited to see if these foundational UI improvements help bring in more people from around the world, because the more diverse the voices, the more valuable the signal we generate for AI progress. Every perspective helps move the ecosystem forward, not just by ranking models, but by revealing how each model succeeds or fails in real-world use. When more people participate, our evaluations become not only more representative, but more powerful for shaping the future of AI.

We’re grateful to our community, and to our backers, for believing in a future where AI evaluation is open, community-led, and based in rigorous, transparent science. This launch isn’t the finish line, it’s the next step in making AI more reliable for everyone.

We’d love to hear what you think!

Try the new LMArena → https://lmarena.aihttps://lmarena.ai
Join the conversation →Join the conversation → DiscordDiscord
Join the team → Join the team → Openings
Follow us →Follow us → X/TwitterX/Twitter