Beyond Leaderboards: LMArena’s Mission to Make AI Reliable

Yeah, that's a good way. That's a very well... Sometimes I get asked what's the last exam that AI should take, you know, for humanity. And it seems like that's the wrong question to ask. We should be asking what's the real-time exam you want your AIs to be taking before they get deployed, every hour, every second of the day, especially as we start to get... I think one of the things that's emerging for me is that the Arena is misunderstood partly because we're just early in AI. And so while benchmarks like MMLU and the idea of these static exams were useful three years ago, the future is about real-time evaluation, real-time systems, real-time testing in the wild.

I think that's one of the very reasons we wanted to create a company to support this project, to, you know, further scale the platform. Right, so right now we are at a million monthly users. Now, what if we scale it to five, to ten, or even more, to capture an even more diverse user base across different industries? And then in that case, we'll have the ability to really, like, zoom in into all these different kind of, like, areas that people really care about for critical mission tasks that it will be used to.

Many people have asked us for this already. So these would be sort of private arrivals. And it's worth saying, you know, I think when people have these mission-critical industries in mind, they often are thinking about the sort of, like, factual nature of the responses and so on and so forth. But in reality, even in such industries, the majority of questions that people ask are subjective. Okay, so the mythology that in hard sciences or in mission-critical industries people just have, like, cut-and-dried questions and they just need, like, a retrieval and a lookup, that's completely false. That's the very reason why these models are useful, is because they allow you to sort of, like, interpolate between these weird questions and answer questions that are not fully specified and give responses that are sort of, like, geared to answer the question but might not have a fully factual basis, right? And they might incorporate factual elements through RAG, let's say, but there's a subjective nature to the response. And that's a reality that everyone's going to have to live with. If these systems are going to be deployed in medicine and defense and so on and so forth, they're going to be deployed in places where the data is messy, because that's where they're useful. Okay, given that fact, how are you going to make sure that they're reliable? Well, you need something like Arena.

No, we work with model providers, small and large. Okay, so we work with basically anybody who wants to work with us, you know, within our constraints, try to be as, like, helpful as possible. The fact is that, you know, part of the reason to build a company is because we don't want Wei having to serve all of these requests from people like Demis himself. So it's been a challenge, but we try to scale as much as possible. And in fact, one of the things that we help everybody to do is pre-release testing of their models. Okay, so it's not just that, you know, we work together to evaluate the models that are released, but we also try to be their release partners and say, hey, can we help you guys pick the models that do best on our user base and use that as a guideline for which models they should actually release to the world. So that's the way our platform works, and that's getting us closer to something, you know, as we've talked about, like, reliability is so important, these subjective measurements are so important. How are we going to get to a world where there's a CI/CD pipeline where people can, you know, test their models pre-release and make sure that they're doing well for all sorts of different, diverse people? Well, you need something like Arena to do that. And that's part of what the company is geared to do. So what people do is they can come, they can test a bunch of different models. We do this with basically every provider that comes to us, and they can see, oh, which one's doing better or worse on the distribution of Arena users. And then they can use that information to help them decide which model to release. After they decide to release a model, it gets continually evaluated forever. So that's where you get the freshness of the data, that the model, you know, continues to be tested. And, you know, we're pushing towards a world where these subjective and human considerations are part of model developers' final release pipeline.

So it's great that people do these expert evals. It's totally fine. It's orthogonal from what we do. You know, I'm glad we have them. But at the same time, you have to ask yourself, what makes somebody an expert, right? What are they an expert on? I think the whole world is moving in a direction against, like, experts being the be-all end-all of everything. And, you know, everybody actually has their own opinions and everybody has their own point of view. And in fact, there's so many natural experts in the world on all sorts of topics that they don't necessarily need a PhD in order to be really, really intelligent and have high taste and have valuable opinions. And I think that's one of the things I'm proud of with the Arena, is it allows us to actually go and say, hey, where are the natural experts? And actually, can we find data-driven ways to identify them? What if we can go look and say, hey, this person here in this random part of the world is actually incredible at coding and math, right? Let's... their vote actually means so much, and their preferences are able to guide the future of AI. That's an amazing thing, and we hope to be able to scale it further.

So I think a few things here. So one, like Anastasia said, this alternative to have hard exams, to create hard exams, is valuable, no question about that. I think the other thing I want to point out, and I took to heart this kind of criticism about having expert labeling. Right now, I went to quite a few experts I know and respect and asked them, would you label? Almost everyone told me, no, I don't have time. Right? So there is a question there about, are you going to get really the experts, right? And I don't think so. You're getting some people from that area who are willing to do the labeling, but the best people are not willing. They fundamentally don't have time right now. But these people, if we offer to them, and we're not doing it now, I'm talking about the future, a platform where their community comes there and asks questions to push the boundaries, to help with their research and things like that, they are going to be, first of all, you are going to get these people, right? Because they do that in order to advance in their research. And we are going to get also their votes. So fundamentally, I do think that we can get in Arena, again, right now, I'm not talking about today, I'm talking about the future.

The other thing I want to say about the layman and so forth, you know, and you fund a lot of companies, you are funding a lot of companies in AI and A16Z, and you are on the board of many companies. What do they build? What do these companies build? What is their product? Who uses their product? Right? It's not the top experts. It's the layman. Right? These are their users. That's how OpenAI makes money and so forth, many others. Right? So then wouldn't the evaluation take into account the prevalence of these users? Right? The answer is obviously yes. Right? So how can something, again, like these exams are very important to understand the capabilities of these models, these benchmarks, again, no question about that. But they are not going to reflect, MMLU is not probably as good as reflecting the preferences of the users of these AI products and services.

I want to just dig one level deeper here, which is that one of the things that we are really excited about is the question of why do people vote the way they do? Who's voting left, right? Why are they doing it? On what kind of prompts do they vote for one model or another? On what kind of topics are models better? You know, basically, can we decompose human preference into its constituent components? Let's say you have a criticism. You say people vote based on slop. You know, emojis are driving votes, and response length, there's a huge response length bias, which is, you know, it's true that people vote for longer responses, you know, preferentially over shorter responses, even given the same content. So well-known human bias. This gets back into RL. Can we learn this bias and actually adjust for it and correct for it? And the answer is yes. That's why we're making style control default. So what we developed is this method called style control, which allows you to run not just, let's say, a Bradley-Terry regression, but also include certain covariates that model the effect of style and sentiment on how people vote. And what you get when you fit this model is not just a prediction of preference, but also an understanding of why people are voting the way they do. You know, we're trying to target a causal quantity, which is, you know, the causal effect of, let's say, response length or sentiment and so on and so forth. Okay, you know, there's always more work to do to get things closer to the actual causal estimate that we want. But if we continue to decompose human preference into its constituent components, what we're building is an ever richer evaluation that can tell us all the factors that go into response. How you can optimize people's preferences, you know, keeping style fixed. Let's say I want to remain concise, but I want to maximize your preferences. Okay, how am I supposed to do that? Well, you know, not a lot of people have that information, but we're building the methodology that allows you to do that.

I think that comes back to, I think, Wei's point a little bit, like when you build a product, an AI product, you want to know about how people use it. You want to understand why users prefer this over that. And in order to collect that kind of data, you have to build sort of like a product first, right? And then, as we know, over the past few years, people have been building very different kinds of applications on top of AI beyond just chatbot, right? So chatbot is one of the widely used interfaces right now for humans to interface with AI. But these days, people are applying these models to, you know, coding and then more like tool use, agentic behavior, that kind of stuff. Right? So as the first step, we were thinking like, okay, how can we capture all these use cases, right? And then the answer to that is we have to build something that people can use, again, in the same right, an environment that people can test in real time to give us real-world feedback. So that was the original like idea that was kind of like last summer, just at the beginning of this kind of text-to-web, text-to-app right trend. At the beginning, it was very beginning was Claude, Claude Artifacts. That was, that was the first, right? And then we saw that, we were amazed by that. And then how can you build eval for that, right? We have to give credit to Arian on the team. And then Arian basically was like, just joined the team, and then he was interning actually at Vercel. So we were saying, okay, why don't we build something new and time Claude effect? How can we do evaluation for that kind of application, text-to-web app? And then that was the idea.

One small parenthesis. So far, we, we talked, three of us, but eventually the, the team grew quite a bit. And right now, I don't know, it's like almost like twenty people, and both, you know, graduate and especially undergraduate students, and doing a lot of very exciting and interesting work to expand the abilities and capabilities and the reach of Chatbot Arena. Right? So I just want to make sure that other colleagues of mine are involved here, and like Joey and others. And I think the credit goes much more beyond the three of us. But that's what we want to provide. You want to, you know, whatever you want, we hope that you'll find the answer to. Fundamentally, there's, if you're looking at the leaderboard, there's only one thing really that matters, which is, do you care about the preferences of the community of people that come to vote on our platform? That's it. That's, that's what we measure. It's the only thing that we claim to measure. We don't claim to be an AGI benchmark. We are faithfully representing the preferences of our community. So that's why it's so important to us that we continue to grow our community and we get a diverse community of all different people, experts, non-experts, artists, scientists, different languages, you know, different languages, everybody under the sun, we want to come to this platform to express their preferences. Because if we can get that to happen, you know, it already happens to some extent, but if we can continue to grow it, what's going to happen is, again, in order for a model to do well, what needs to happen is new people need to come in and vote for it. And so if we can provide this lens into the preferences of the world...

Things are moving, right? It's like these are changing. This is, again, it's not like, it's like we, we have the study and you want, you want to talk about, about the freshness, right? About we all, all, you know, see fresh prompts, right? It's not like, oh, we are going to see all these prompts over and over again, like kind of saturated. That's not the case. That was the very beginning of why we believe Arena is fundamentally different, related to contamination. The very beginning of Arena is to try to solve the contamination problem, the overfitting problem, which is like people test models on a static benchmark or what people call overfitting. Yeah, what people call overfitting. So how do you overcome overfitting? You collect new data, right? That's how you overcome overfitting. And then Arena is designed to collect new data every second. So all the questions are, are new, like all the votes are new. And then we measure basically how different, what's the difference between all these prompts, right? What's the distribution look like and so on. And then we conservatively estimate like, you know, over eighty percent, something like that, prompt, yeah, different. Correct if I... And this study was done by another member of our team, Lisa, and basically measures out how many more, you know, fresh prompts you have in one day compared to what you've seen in the past three months, right? And by a similarity score of something like seventy, seventy-five percent, you have over seventy percent of these prompts are fresh.

Yeah, first, let me just say, I think that all of these arenas have signal in them. It's not just WebDev Arena. It's just people have more opinions on language, right? So WebDev is a little bit more objective. It's easier to see the website, you build it, and it's, you know, this one's better than the other. There's a lot of signal in it. Also, it shatters the models. And what I mean by that is, it can be very clear that one model is way better than another on WebDev Arena. Immediately, bam, you see it. And it's like, why is that? Just the capabilities of the models? Maybe Wei can say more.

I think it's just a much, much harder task because it's like from text, a description of a website, and you have to, you know, first understand the request and then build, like, write code, right? And then the code has to fit, maybe satisfy certain requirements, let's say, style requirements or, like, components, that kind of stuff. And then it has to compile, right? It has to be, we, we basically run it live in a browser and then with, connect to a sandbox, that kind of stuff. So there are a lot of, like, parts the model has to get right in order to build a website that people can, you know, really interact with. So fundamentally, it discriminates much better, right, across some models because it's a much harder problem. So very few get it right.

Yeah, for sure. And then a lot of, like, you know, every input from a user is for a real-world task, right? They're trying to build some website, real. And then that also measures something that's, like, beyond just, you know, academic benchmark. That's, we imagine what a user would do, right? This, like, really trying to approximate the user intent, user preferences directly.

I have to say, I also just, like, completely disagree with the foundation of the question. The, like, implicit assumption is that, like, chat is easy or that it's even easier than WebDev. That's completely false. It's a completely naive perspective that people have on this because it's, it's hard to build something that people love. People are good at chat, but you like some people way better than others. It's subjective and everybody has their own opinions and the landscape is very, very rich. You might like a very different model than me. Somebody else, you know, let's say a musician might like a different model much, you know, a much different model than I do. And understanding all of those differences is really hard. And anybody who thinks that that is gameable is deluding themselves.

Absolutely. Yeah. And we, we should be giving you the tools to do that. We're currently building them.

Absolutely. Absolutely. It should be personalized just for you. You should understand which models are best for you, and it's going to be for your task. It's person and task, right? Because for a different, if you want to do different things, you may have a different leaderboard, right? If you have a question about tax today, you go to different people than if you have a question about programming or whatever, right? So that's kind of also is going to depend on what task you want to accomplish, right? And I, I really want to go back to one thing. I think that because it's, you know, I tried to think quite a bit about all the criticism because on the face of it, intuitively for some people, it makes sense, right? That's why many people make the same criticism. And I think there is another thing going on. We as humans, we believe that, why do people say Arena is not good? Because people are fooled, right? That's fundamentally their kind of argument. They are fooled because long answers, more emoji, and things like that. And when I look from that perspective as a human, what I, in my mind is that I am not going to be fooled, right? That's what it's not relevant. These guys are going to be fooled by these things. No, I am not. That's why it's not good. That's why it's not a good proxy. The problem is that you are, we, I am fooled, right? Everyone is fooled. And this is actually, that's why the Chatbot Arena, it's like Anastasia said, it's, it's, it's a proxy. It provides a magnifying glass. But, you know, all of us, we have our own peculiarities, our own, you know, culture, our own, you know, history built on interaction with different people, right? That we have different preferences, right? That's fundamentally what it is, right? And these preferences are not fully objective, right? Because people say it's only on objective answers, but all of us are different, right? That's kind of the fundamental disconnect between the criticism and actually what we provide and actually what everyone, we believe, needs, right?

And that's why, again, it's like going back to the, to the, to the previous, that, you know, I think what you hear is that capturing, capturing the human preference is fundamental because we are building these AIs to interact with humans, right? I think that's kind of the foundation. But if you, like we have in the previous discussion, say people say, well, you know, like, yeah, but these other people are fooled, right? This is not me, right? Although I am fooled as well, but I don't believe, right? You always believe about yourself that you are better than you are. But, but then we can provide, like, style control, okay, so forth, you can remove that, adjust for that effect, right? And we are going to provide more and more, you can adjust for that effect and that effect and that effect, right? So you get your answers as well.

So it was started around two years ago in late April 2023, I think. And then at that time, before Arena, we were working on a project called Vicuna, which is like one of the first open models that's been released, like the ChatGPT kind of clone. Yeah. And at that time, LLaMA 1 was just released, which is a base model. It doesn't really know how to chat with humans, only gone through like a pre-training process. At that time, I don't think people call this post-training yet. People call it instruction fine-tune, that kind of stuff. So we were like in the lab exploring how do we reproduce this, right? How, how do we make an open-source version of ChatGPT? And then, you know, credits to Lian, he was having kind of like an idea that we could use some of the open data published on the internet that's kind of like user ChatGPT conversation. It was called ShareGPT, and it was like kind of like a high-quality set of dialogues that users share. So basically, we, we, we come as a group, a bunch of PhD students in the lab, set an ambitious goal, which is like we try to, you know, release this model, train this model in two weeks, that kind of stuff. So, and, and during that process, the result was surprisingly, at that time we were like playing with the model, and then we thought like we, we got to, you know, demo this to the world. So we basically just set up a website and then put that model on the website and then release it. At that time, it was like there's like a huge debate like internally, like how should we, when we release it, you know, how should we evaluate this model? How good is this model really? It vibe well, right? Because when we compare it to LLaMA, the base one, you can feel the difference, right? The model just learned how to chat, learned how to speak like a ChatGPT. So there was a huge debate like how, how do we evaluate this? At that time, we didn't have much time. So we were like, okay, we either do this kind of like labeling, come up with questions ourselves and label the data and then compare it with other models, or we do something like, something automatic. And then at that time, there was like GPT-4 just came out in March. People were like wondering what can it do, right? And then we were like, okay, why don't we just use it for evaluation? As we use GPT-4 as a judge to do this automatic evals at that time, and no one believed it. And then it was like, again, huge debate, but we didn't have time. So we just, you know, ended up doing it, and it worked surprisingly well, again. And we released it. But the, the huge problem after that is like still there's an open problem, which is like how do we evaluate these chatbots, right? So soon after, we, um, kind of like come up with that idea that why don't we let everyone in the community vote which model is better? Because at the time, we served that model, and then we also served some of the other open-source models at the time. Every week there's a new fine-tune. So we registered a website, we demoed them all of them, and then we come up with like kind of like a side-by-side UI that people can compare them. And then soon after, we say, okay, why don't we come up with a battle mode, which is like we anonymize their identity and that people vote. So that was the origin of Arena, which is basically trying to solve our problem, which is how do we evaluate the smart, understand difference.

Actually, the first time we did start, we did try to have some students, you know, buy some pizza, get them in a room, and label the replies from Vicuna and other models to compare them. And then obviously that didn't scale. And then it was this LLM as a judge we tried, and it worked surprisingly well. And then it was quite a bit of debate. Okay, we are going to build a platform to scale because the question is still, okay, it seems that anecdotally, anecdotally, like GPT-4 was released just two weeks before, before we started to use it as a judge. It's still the question, okay, yeah, it seems that it's doing well, but still, you know, it's like, how about, how does it compare with humans, right? And that was a question about, okay, how do we scale the human evaluation? And we discussed quite a bit about how to do it because it was not clear. Because if you think about it, before, you just ask people, I have a prompt, and the prompt is answered by all models, and then you label it, right, good and bad. That's kind of the typical way. How would you scale up the process? Then you need kind of to rank them, right? And you have n choices, answers of the same problem from n models. It's very hard to rank them, right? It's like, think about it, right, we get slightly different in tone and so forth, try to rank them. And then we thought quite a bit. I think the inspiration about was how humans in real life rate, say, players or teams in games, right? And obviously you have the tournament. That's one way to do it, where you have, so to speak, the players play with each other head-to-head, and then based on that, you are going to have some number of points to win or lose or to, and or tie, and then you, you, you are going to have a leaderboard, right? And, um, but that again, the problem with that is that if you have a tournament, typically the assumption in a tournament is that the number of players do not change across during the tournament. Right, right. And also in general, you know, most of the tournament, each player has to play with everyone. So it's kind of n squared problem or n, the number of players. And then we thought about, okay, there are other ways in real life how players or teams are ranked when they don't play with each other, or they don't have a chance either because the number of players is too large, or, you know, you need to also accommodate new players, right, entering the game. And, and that's why we were about, then we thought about, okay, there are disciplines where this is done, like chess, tennis, ATP rating, and many others. And that was the idea. And we start, okay, why don't you do something like an Elo score, okay? And for that, what do you need? Oh, you need only, you know, head-to-head, and you don't, not everyone needs to play in the same tournament. And that's how we adopted, and that's why Arena has this battle mode in which you have a prompt and answer from two randomized, anonymized language models, and you can pick which one is better, where there is a tie, and, and so forth.

Yeah, it turned out it was like deeper technically than we thought. So at that time, Jan was like, we need to find someone to back this up more like on the theory side to have a solid foundation, right, to rank all these models. It's not, at that time, it was no longer like really like a fun project anymore. It was, it was started as like a fun project, and people started to pay attention to it, right? So you better do something. And I, I, I went to Michael Jordan, my colleague, very famous machine learning, AI researcher and faculty here. And I've been working with him actually when I, when he built this kind of labs at Berkeley, cross-disciplinary labs. We were working with him in like 2005, 2006. He was joining the systems people, database people to work together on new, exciting projects. And he told me, oh, I know exactly, I have the, the guy for you, it's Anastasios.

I saw what was being built at the time. Arena was not still close to what it is today. It was, I think, not that much usage. I saw it and I thought, wow, what a great opportunity to do some interesting statistical modeling and theory, like being able to understand how do we optimally sample models, how do we perform this estimation. Okay, let's move from Elo to Bradley-Terry because we're actually performing an estimate here instead of just like, you know, in the Elo score moves over time, it doesn't converge, but Bradley models converge. And how do we then construct confidence intervals properly for this estimate and so on and so forth? All that stuff was super interesting to me. We were meeting like just next to, two, two doors down, and we wrote on the whiteboard like five or six different, you know, topics. I mean, we started working on them, and the rest is history.

Well, it certainly would have been worse if it came out of somewhere else. And the reason is because the fact that we come from Berkeley and from a university really speaks to our scientific approach and neutrality. I think if it came from an industrial lab, people would always have questions about, oh, well, these people, are they also training a model and what's their incentive and so on and so forth. But the reality is we were just students. We were doing this in order to evaluate models, and they came from a scientific perspective. That's it. And I think that's something that people can see when they look at us, and it builds a lot of trust in our business.

The other angle here is that in a lab like this one, what you get, you can get maybe in the, in industry, you can also get maybe interdisciplinary teams, but they are going to be large teams, right? Because you go, okay, you are going to do, you know, the team which is doing AI, the team is doing systems work together. So these are already large teams. But here what you get, you get small teams, a few people, which everyone comes from, can come from different areas. Right, right. We have people who are kind of systems early on. We have to build the systems to serve these open-source models. Right, we have to serve this Vicuna, like mentioned, right? Then you have to have people who are pretty good, actually, when we use this ShareGPT data, we are doing quite a bit of data preprocessing in order to, you know, to pick some data curation and so forth, right? Then when Anastasios joined, we have now machine learning experts, right? But the team still is like four or five people. And as you know, early on, small teams move very fast. So I think that's kind of the difference, that you have a very small team, but, and interdisciplinary and small. So I think you, you may get in industry, you may get in interdisciplinary teams, but they're going to be large.

I think that if I, if I may take a step back here, because, you know, I'm old enough, so I've seen a few of those. I remember when I was a student, I was doing systems work and networking, right? That was the internet days. So, and I was going to these conferences, and there were panels, are the operating systems dead? Research in the operating systems, that was the topic of the panel. And the reason for that was at that time, it was at the end of the previous century, it was Microsoft was dominating, and then Apple, and then of course there were some FreeBSD and so forth. But then, you know, although it didn't come from academia, it was Linux, right? Actually, Linux was preceded by Minix, which, which came from academia from the Netherlands. So that's one. Then in 2004, like when we come here and we started this lab, there was a question also, that was this kind of distributed system, because a lot of researchers on distributed systems, what can academia do, right? Because it was Google was doing all the research of query systems, MapReduce, Google File System, all of that happening at Google, right? Best people going at Google and so forth. And then we, we, we done here, then it come Spark, right? Which also come from academia, right? And I think that's when this started. It was actually the question people were surprised about Vicuna, just a bunch of students and actually their own initiatives, right? I knew almost after the fact this happened, pick this kind of data set from the internet, which is high quality, right? And use it, and people were so surprised about the quality, right? That was, that was, were people asking, is this real, right? Right. So I want evaluation. I was kind of, you just show me some, some stuff anecdotal, right? And, oh, okay, it look good, but how about, you know, is that, some people didn't believe it and then say this is a GPT-4 wrapper.

Again, since then, there are many other things we've done here in open source, like inference, LLM inference, like vLLM and Hugging Face. But I do think that what happens in this kind of, this sense in, in industry, I was actually on a panel yesterday, was the same, this kind of discussion, you know, okay, it's like, you know, what can maybe, you know, academia, I should do this thing and the industry do this thing, right? Like let's, the industry just, you know, can't do anything like pre-training and so forth. I think at the end of the day, it is about what resources you have and what problems you solve. And over and over again, I think through the example I gave, if academia has the resources, it's going to surprise you. Clearly, it's going to be at the very edge of innovation and creativity. And so that's always almost happened, right? Like in this case, in this case, of course, we, we, and we didn't need huge resources, right? It's just a group of smart, passionate students.

So when Chatbot Arena started, it was a lot of excitement and so forth. But then it was this thing, okay, it was a feeling, at least for some people in the group, that we are done here. We publish a paper and so forth, right? For a while, even if you look at the usage, it's kind of dropping a little bit, and it's almost, yeah, almost died, almost died. And then Wei, at that time, his main thrust of research was different, some, you know, graph neural networks, distributed graph neural networks and things like that. And I remember at some point, at one of our one-on-one meetings, he came to me and said, look, I, I really like, and I want to instead of doing this kind of work, I really am passionate about Chatbot Arena. I want really to, to do it right and to focus on it. And then when kind of then it started, and Wei is like one-man backend, right? Starting and marketing and, and marketing and marketing, and he started to, you know, to add more models to the leaderboard, market it and so forth. And very soon after that, Anastasios came, and then it was kind of magical, right? You have these people who are so passionate and they are working so well together. They are so complementary in skills and even personalities. Then it started to shut up, right? And I'm, I'm mentioning that because without that kind of inflection point, which came, you know, long after it started as a project, we wouldn't be here.