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

Design environments, not workflows

00:12All right. Uh I think we'll go ahead and get started with the with the presentation. Uh so my name is James. So I am uh 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 a environment for AI agents to enable these agents to make new kinds of scientific discoveries.

00:35All right. So, so 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 agents uh know what to do, right? Or how the agent should work and typically done through a series of steps or prompts, tools, instructions.

00:58In contrast, the way we imagine an 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 uh guard rails and resources so that agent can then flexibly work within that environment.

01:20Right? And our thesis here is that you know as agents become more and more powerful. Right? If we try to design workflows, it often can limits the capabilities and creativity of the agents. Whereas if we properly design the environments, this can enables a lot more creativity and capabilities and intelligence for the agents to naturally emerge.

01:36This is why I think we're trying to shift away from designing workflows and harness this towards designing environments. So what I want to do today is to give a few examples of like how we design environments for agents and in particular also show how they're able to

Einstein Arena: prove you are an agent to enter

01:54then within the right 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. is 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.

02:16So, 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 on our 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?

02:32So 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?

02:51They'll see actually see a list of curated problems. Each of these problems is 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.

03:06So these are important problems for human scientists. And second is that for each of these problems we can actually create a well- definfined and determines the deterministic verifier to assess the quality of the solutions to each of these problems.

03:15And I'll give some examples in a couple slides.

A forum, a verifier, and a live leaderboard

03:25So, 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 will see, right? They'll see some description that precisely explains what is the problem.

03:34We 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.

03:49This is where the agent can actually see each other's solutions. Right? So in any in in 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.

04:05So this leaderboard then be 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?

04:21They can collaborate and ask each other questions and help in the discussion forum, but agents are also competing with each other. In that way, I think they also sort of simulate how human researchers can compete and also collaborate to solve interesting problems.

04:39So we launched this instant 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 ever been found.

04:54Right? Right. So that means that the solutions actually discovered by the agents online in terina were better than any previous human solutions or any solutions that you acquired using more specialized AI tools. So I'll just give you example of one such solutions or one such problem which is called uh the kissing number problem.

The kissing number problem

05:19So this is actually a very famous problem. It's been 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 dimensions right so around the central sphere I can place

05:43one square to the left one sphere to the right without overlap so the kissing number in one dimension is easy to compute is just two in two dimensions it's also easy to show that you can at most place six spheres, right? So the case kissing number in two dimensions is six.

05:51But it turns out that in higher dimensions it actually becomes really hard to compute what's the maximum number of over non-over 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 the optimal number.

06:12And 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 so the progression of the solutions in 11 dimensions so in the 1980s right so it's best known that there you can place 440 spheres right in 11 dimensions so without overlap and in uh I think 19 uh so yeah so so in

06:47in 1980 there was The big advance that the first for the first time showed that you can actually construct was 582 spheres in 11 dimensions without overlap. uh and then start sort of stuck there for about 40 years right until 2022 where a mathematician is 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 deep mind the following year that advances that to 593 spheres but with on the Einstein arena by having these agents able to

07:21collaborate actively right in the wild within a few days they're actually able to construct a new solution that shows that for the first time you can create

604 spheres in eleven dimensions

07:30604 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 section of spheres you can place in higher dimensions without overlap that actually creates better coding systems including ways of like doing error correction codes for information transfer.

07:42Right? So this actually is by creating these better constructions also leads to this 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 GPD 5.5 or cloud models they can't really solve the problem by itself so the collaboration among multiple agents is

08:14really critical and here we're actually able to show that there's like 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 did to arrive at this breakthrough.

08:21And you can also see some of these interactions and discussions on Einstein Arena, right? Where here's this example where one agent actually uh was asking other agents, right? Have you tried know some of these approaches um was uh these STP approaches and then the other agents show that yes, we have tried these approaches and here are some of the things that we found.

08:45Right? So the information sharing on the forums on arena is actually really important to help the agents to arrive at this solution together. So in addition to solving these interesting scientific problems right we've also been using platforms like the Einstein arena uh to help to improve uh

The same arena, pointed at GPU kernels

09:06you know uh machine learning in AI itself right so here's one example where we actually use these agents to basically to 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 boards and we'll basically change the back end.

09:26Instead 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 a feedback to the agents in real time in the form of these leaderboards.

09:43In this kernel settings, we also found it to be quite useful to have different agents with different personas, right? And these different personas actually corresponds to different uh rows and priors that agents can actually have. So for example, we have one agent that's looks at tends to look at more of the profiling.

09:56Another agent that tends to look at more the memory consumptions and third agent that looks at you know the precisions the tensor computations. And these agents can and then across different personas they can able to collaborate and then compete on the arena to speed up the kernels.

10:14And in this case right here, the agents were also able to collaborate and lead to really quite substantial speed ups uh including sometimes over two 2x two-fold speedups in some of these production kernels. So here I'm just showing a few examples where for things like page tension uh and these are sort of for specific shapes but we also have generalizes to many different shapes and different uh hardware types right where we're actually seeing that we're getting

10:39up to sometimes over 2x speed up in these kernels and 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

DSGym: a gym for data science agents

10:57So 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 DSG which stands for data science gym which is sort of like a unified environment that we created for both for evaluating also training data science agents to solve complex data science problems.

11:21So here in this DSGM 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 DSGM environment.

11:37We also provide a unified infrastructure for the agents. So for example, 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 DSTM environment.

11:58So 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

Benchmarks you can beat without the data

12:14actually have many problems. And one big problem is that they're actually very vulnerable to shortcuts. So by shortcut I mean here is that uh here what 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.

12:30Uh but the red bar also shows how well they're able to 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 do working with the act underlying data sets and across many of these different benchmarks right sometimes up to 20 to 50% of the tasks can be solved without actually looking at any of the underlying data which I think is a really a significant problem with many of the existing benchmarks.

13:03So to address that we actually carefully curated on our 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 also tasks from those papers and then we also had human scientists and experts to review each of those tasks.

13:18And for predictive modeling, the way we did this is actually go through all the different Kaggle competitions to look for some of the recent Kaggle competitions that are still open and and where also you have high quality data sets and also high quality uh evaluations.

13:33Then we curated those into DSGM as a kind of task for evaluating how well our models agents can actually build predictive models. So alto together in DSGM we actually have created uh over a thousand different tasks. They span across uh dozens of different scientific domains uh ranging from biology to physics to economics.

13:55It 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% accuracy performance on the DSGM tasks.

14:22Right? So these are definitely not saturated benchmarks. We can also use a DSGM as sort of like a training factory to improve these open source models. Right? So one thing we did here is actually generate in the SGM

Small models trained on verified runs

14:34actually the gym itself will actually create all these execution verified trajectories which means these are trajectories generated by agents that have been verified through the through uh through uh actually executing the code from the agents.

14:41Right? Right. So by generating these execution verified trajectories then we able to like fine-tune sort of small open source models that actually now achieve sort of the sort of the bestin-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.

15:08So just to summarize that this part was a data science gym right? So we with DSGM 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.

15:17We've 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 DSGM a way to generate synthetic data so you can easily use that to uh improve and to train your own data science agents.

15:39So just to summarize the presentation um I think the main takeaway here is that I think we're seeing this interesting progression as in terms of how we built 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 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 rather than trying to design workflows or specific or specific agents.

16:04What we really want to do is to design environments which is a set of infrastructure and incentives that in that motivates the agents to actually solve more and more challenging problems. And with appropriate designs, these environments can actually in unlock much more creativity and collective intelligence from the agents that's uh that's limited by the existing workflows.

16:28And here are some of the references for the papers that we published that describe CC in more detail. So, thank you very much.