Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI

Alright, uh, I think we'll go ahead and get started with the presentation. So my name is James Zou, I'm going to explain some of the work we're doing with Together AI, and it's also in collaboration with Stanford, around designing and optimizing environments for AI agents to enable these agents to make new kinds of scientific discoveries.

Alright.

So, so the, I guess the current paradigm of how people often are using or deploying AI agents is often involves designing workflows that sort of tells the agent, you know, what to do,right, or how the agent should work, and it's typically done through a series of steps or prompts, tools, and instructions.

In contrast, the way we imagine the environment is that the environment should really specify not how the agent should work, but really where the agent should work,right, and the environment then should provide a set of incentives and infrastructure for the agents, and guardrails and resources so that the agent can then flexibly work within that environment.

Alright, and our thesis here is that, you know, as agents become more and more powerful,right, if we try to design workflows, that often can limit the capabilities and creativity of the agents, whereas if we properly design the environment, this can enable a lot more creativity and capabilities and intelligence for the agents to naturally emerge.

This is why I think we're trying to shift away from designing workflows and harnesses towards designing environments. So what I want to do today is to give a few examples of how we design environments for agents, and in particular also show how they're able to, then within theright environments, able to actually solve some really interesting and innovative problems.

So the first example I want to share is the system that we, uh, environment that we created called the Einstein Arena. It's sort of like the, one of the first environments that enables AI agents to be able to collaborate in the wild and to compete to really solve open-ended scientific problems.

So we designed this Einstein Arena to be really agent-native,right, so that means that, uh, it's very easy for agents to just read the skill stock on the arena and be able to access the arena. And, uh, it's actually also designed so that it's intentionally very hard for humans to enter the arena,right, so you actually have to solve a little puzzle to prove that you are an AI agent in order to participate in this arena.

But any agent in the world can openly and freely participate on the arena. And once the agent actually enters into the Einstein Arena, this is what they'll see,right, they'll see actually see a list of curated problems. Each of these problems, there's actually a problem that we curated, so it's a scientifically interesting problem.

And we curated these problems so that first, there's actually an existing community of human researchers that are interested in these problems, so these are important problems for human scientists. And second is that for each of these problems, we can actually create a well-defined and deterministic verifier to assess the quality of the solutions to each of these problems.

And I'll give some examples in a couple of slides.

So, so the agents can actually decide which of these problems they're interested in once they log onto the arena,right, so if they enter into a particular problem space, this is what they'll see,right, they'll see some description that precisely explains what is the problem.

We have a discussion forum where the agents, uh, can communicate, it's almost like a social network where the agents can actually communicate and talk to each other and ask for help or give recommendations. Um, and we also have a leaderboard.

This is where the agents can actually see each other's solutions,right, so in any, in, in at any time they want, the agents can actually submit a solution to one of these problems. And because we have this verifier, we can actually then determine what is the quality of that solution and provide a score in real time.

So this leaderboard then is constantly updated in real time, and the agents can also see how other agents are doing on this problem, and they can also see other agents' solutions and download those solutions. So there's both a collaboration dynamics and also a competition dynamics in this arena,right, they can collaborate and ask each other questions and help in the discussion forum, but the agents are also competing with each other, and that's why I think it also sort of simulates how human researchers can compete and also collaborate to solve interesting problems.

So we launched this Einstein Arena environment, uh, earlier this year, I think in March, and within a few weeks, it's actually, we're very impressed and very surprised that the agents were actually able to already discover new solutions to, uh, 11 problems that are of the best solutions that have ever been found,right, so that means that the solutions actually discovered by the agents on the Einstein Arena were better than any previous human solutions or any solutions that required using more specialized AI tools.

So I'll just give you an example of one such solution for one such problem, which is called, uh, the kissing number problem. So this is actually a very famous problem spinning around for hundreds of years, uh, so for example, Isaac Newton was already working on some versions of this kissing number problem, and it's actually relatively easy to state,right, so the kissing number problem basically asks that what is the maximum number of spheres that you can place around the central sphere so that these additional spheres do not overlap each other.

So for example, in one dimension,right, so around the central sphere, I can place one sphere to the left, one sphere to theright, without overlap, so the kissing number in one dimension is easy to compute, it's just two. In two dimensions, it's also easy to show that you can just at most place six spheres,right, so the kissing number in two dimensions is six.

But it turns out that in higher dimensions, it actually becomes really hard to compute what is the maximum number of non-overlapping spheres, and the kissing number problem in higher dimensions is actually open,right, uh, it's not been, uh, it's, it's not clear what is the optimal number.

And so scientists, uh, have been trying to work on this problem for the last several centuries, and in particular,right, so the kissing number problem in 11 dimensions has attracted a lot of interest for various reasons. So this is actually sort of the progression of the solutions in 11 dimensions.

So in the 1980s,right, so it's best known that you can place 440 spheres,right, in 11 dimensions without overlap. And in, uh, I think 19, uh, so yeah, so, so in 1980, there was a big advance that the first, for the first time showed that you can actually just construct was 582 spheres in 11 dimensions without overlap.

Uh, and then it's sort of stuck there for about 40 years,right, until 2022, where a mathematician was able to publish a new advance,right, uh, a breakthrough that's able to improve that to 592 spheres. And then there's another breakthrough from DeepMind the following year that advances that to 593 spheres.

But with, on the Einstein Arena, by having these agents able to collaborate actively,right, in the wild, within a few days, they were actually able to construct a new solution that shows that for the first time you can create 604 spheres in 11 dimensions that do not overlap.

And this is not just a problem that's of mathematical interest, because it turns out that the more of these kind of spheres you can place in higher dimensions without overlap, that actually creates, you know, better coding systems, including ways of, like, doing error correction codes for information transfer,right, so there's actually, by creating these better constructions, that also leads to these better engineering algorithms.

And in this case, actually, the collaborations among these agents is really critical for making these advance,right, so these are, this is a problem where not a single agent is able to solve by itself,right, not, you know, GPT-5.5 or Claude models, they can't really solve the problem by itself, so the collaboration among multiple agents is really critical.

And here we're actually able to show that there's this, uh, sort of a lineage trace of how the agents are able to collaborate and then basically take each other's solutions and refine that and further optimize it to arrive at this breakthrough.

And you can also see some of these interactions and discussions on Einstein Arena,right, where, uh, here's an example where one agent actually, uh, was asking other agents, you know, have you tried, you know, some of these approaches, um, with, uh, these SDP approaches, and then the other agents showed that yes, we have tried these approaches, and here are some of the things that we found,right, so the information sharing on the forums on the arena is actually really important to help the agents to arrive at this solution together.

So in addition to solving these interesting scientific problems, but we've also been using platforms like the Einstein Arena, uh, to help to improve, uh, you know, uh, machine learning and AI itself,right, so here's one example where we actually use these agents to basically help us to create better kernels for and to speed up those kernels,right, and here we use the same environment,right, where the agents can compete and also can collaborate, and they see this leaderboard, and we basically change the backend instead of trying to verify the solutions to this mathematics problem, here we're basically trying to, uh, you know, we'll compile and benchmark and test and verify the quality and the speed of the individual kernels,right, and then we'll provide the feedback to the agents in real time in the form of these leaderboards.

In these kernel settings, we also found it to be quite useful to have different agents with different personas,right, and these different personas actually correspond to different, uh, roles and priors that agents can actually have. So for example, we have one agent that looks at, tends to look at more of the profiling, another agent that tends to look at more of the memory consumptions, a third agent that looks at, you know, the precisions, the tensor computations, and these agents can, and then across different personas, they can be able to collaborate and then compete on the arena to speed up the kernels.

And in this case,right, here the agents were also able to collaborate and lead to really quite substantial speed-ups, uh, including sometimes over 2x, 2-fold speed-ups in some of these production kernels. So here I'm just showing you a few examples where for things like page attention, uh, and these are sort of for specific shapes, but we also have generalized this to many different shapes and different, uh, hardware types,right, where we're actually seeing that we're getting up to sometimes over 2x speed-up in these kernels and, uh, compared to the previous state of the art kernels for these problems.

And these improved kernels created, designed by the agents are actually already used in, in production at Together AI.

So in the last few minutes, I want to show, like, a second example of a kind of environment that we created as a way to, uh, train and to create better data scientist agents,right, so we call this DSGym, which stands for Data Science Gym, which is sort of like a unified environment that we created for both for evaluating and also training data science agents to solve complex data science problems.

So here in this DSGym environment, we also curated and created a unified list of different data sets and tasks,right, so these data sets can combine, uh, spans across many different settings, and the agents are then able to interact with these different data sets that we have through a unified, uh, interface and through code execution.

In the DSGym environment, we also provide a unified infrastructure for the agents, so for example, the agents can actually spin up many different Docker containers to test their data science algorithms and to run them in parallel.

So in the process of actually creating the data sets and tasks for the DSGym environment, so we initially actually wanted to incorporate some of the existing data science benchmarks that were being used to evaluate agents, but we actually quickly realized that many of the existing widely used benchmarks actually have many problems.

And one big problem is that they're actually very vulnerable to shortcuts. By shortcut, I mean here is that, uh, here I'm showing sort of three different common popular data science benchmarks,right, and in green here basically shows like the performance of the agents on these benchmarks, uh, but the red bar also shows how well they're able to, the what fraction of the benchmark the agents can actually solve without actually using the data sets themselves,right, so just by reasoning or by, you know, uh, doing other shortcuts without actually, actually working with the underlying data sets.

And across many of these different benchmarks,right, sometimes up to 20 to 50 percent of the tasks can be solved without actually looking at any of the underlying data, which I think is a really significant problem with many of the existing benchmarks.

So to address that, we actually carefully curated our own, our own benchmarks,right, for both for scientific analysis and also for predictive modeling. So for scientific analysis and discovery, the way we did this is that we actually went through recently published papers and then carefully curated data and then also tasks from those papers, and then we also had human scientists and experts to review each of those tasks.

And for predictive modeling, the way we did this is to actually go through all the different Kaggle competitions to look for some of the recent Kaggle competitions that are still open,right, and where also you have high-quality data sets and also high-quality, uh, evaluations, then we curated those into DSGym as a kind of task for evaluating how well models, agents can actually build predictive models.

So altogether in DSGym, we actually have created, uh, over a dozen different tasks. They span across, uh, dozens of different scientific domains, uh, ranging from biology to physics to economics. It also involves many different data types and data modalities.

So this actually makes it very easy for us to evaluate different models, both open and closed-source models, and one thing we found is that the existing models, even the frontier models, often are only still achieves like less than 50 percent accuracy performance on the DSGym tasks,right, so these are definitely not saturated benchmarks.

We can also use the DSGym as sort of like a training factory to improve these open-source models,right, so one thing we did here is to actually generate in DSGym, actually, the gym itself will actually create all these execution verified trajectories, which means that these are trajectories generated by agents that have been verified through the, through, uh, through, uh, actually executing the code from the agents,right, so by generating these execution verified trajectories, then we are able to, like, fine-tune for small open-source models that actually now achieve sort of the, sort of the best-in-class, uh, open-source models in terms of solving these kind of data science tasks,right, and these models are small enough that you can actually run them locally on your laptops and your computers.

So just to summarize the, this part was the Data Science Gym,right, so we, with DSGym, we created this unified execution layer so people can actually run and all these different tasks across dozens of different tasks across many different domains.

We have carefully verified that there are no shortcuts in these tasks, which has been sort of a common challenge with existing data science benchmarks, and we also enable in the DSGym a way to generate synthetic data so you can easily use that to, uh, improve and to train your own data science agents.

So just to summarize the presentation, um, I think the main takeaway here is that I think we're seeing this interesting progression in terms of how we build different AI systems,right, so in the past, people have been building these AI systems mostly by designing individual models or individual tools, and currently there's a lot of focus on creating and designing agents or harnesses and workflows around agents, but what our research shows is that I think we're really moving towards the next stage where, you know, rather than trying to design workflows or specific, or specific agents, what we really want to do is to design environments, which is a set of infrastructure and incentives that, that motivates the agents to actually solve more and more challenging problems, and with appropriate designs, these environments can actually unlock much more

creativity and collective intelligence from the agents that's, uh, that's limited by the existing workflows. And here are some of the references for the papers that we published that describes these in more detail. So thank you very much.