Agentic Evaluations at Scale, For Everybody — Nicholas Kang & Michael Aaron, Google DeepMind

Alright, hi everybody. Um, let me just try to stand straight so I don't have to crouch over. Thank you all for coming. This is our talk on agentic evaluations at scale, for everybody. I hope everyone's in theright room, and if you are, thank you for coming.

We were expecting like 20 people, so this is like way more than what we expected. So, alright, who are we? So I'm Nick, I'm a product manager in Kaggle Benchmarks, and I basically run and build our benchmarks platform alongside a couple of our engineers, and I also focus on our agentic eval solutions.

I'm originally from Singapore, but I live in the San Francisco Bay Area, and so I flew in to do this talk and attend all the great talks out in this conference today.

And hi, I'm Michael. I'm a software engineer on Kaggle. I've been working at Google for about a third of the time that I've been alive, and Kaggle for about half of that. So, uh, yeah, but mostly working on evaluations and benchmarks for Kaggle at the moment.

Alright. Has anyone here heard of Kaggle? Put your hands up if you have. Okay, great. So a lot of people know us for competitions, but we don't just do that. We're the world's largest AI/ML community of 30 plus million users, and we've been working a lot in the Gen AI eval space over the course of the past two years.

And we think there are lots of interesting problems in the industry that not many people are trying to solve, and we feel like we're positioned well to solve them. And we want to share more of the work that we've been doing, and also invite contributions if you want to get involved in the space.

So we have a simple agenda for today. First is AI evals today are kind of broken, and we'll talk about why. And then step two is we're trying to solve it, not saying we're the all-cure and we have everything kind of sorted out.

We'll talk about what we're trying to do, the challenges we're running into, and also maybe, like, it might inspire some of you in terms of how you think you might be able to help contribute to this very important problem that we're trying to solve for.

With that, I'll jump into the first section. So, first problem: evals are scattered, decentralized, and get stale fast. I don't know how many of you have tried to keep track of, like, AI benchmarks, but basically like 10 plus of them drop every single day.

And the best way to find out what they are is go to archive and spend, like, hours scrolling through them, reading every paper. That doesn't make sense. We don't think it makes sense. I can't even do it, even though it's my full-time job.

And I think what happens after that paper gets published is that, you know, you see some of the leaderboards in the papers, and what happens after that? They just get stale. The authors move on to the next best benchmark because they just want to publish lots of papers, no fault of their own, but these leaderboards no longer become relevant as time goes on.

The second issue is that evals aren't always transparent, accessible, and verifiable. Sure, many of us have seen these charts on these model publisher notes when they release a new model, but what's the problem with that? You know, we don't actually know how these benchmarks are set up.

It's a lot of configurations you could use for the models themselves, and also how the benchmark is orchestrated and facilitated. And we don't always know what's actually being tested here. I'll give you one real anecdote, which is that we had published a benchmark with one of these AI labs, and another competing AI lab came to us and said, "Hey, we don't like the results of, you know, this particular benchmark you've published.

Let's run it on our own." And so they ran it, and then they published it with much higher, much better results. And the difference was that they were optimizing it for their model. So they had used compaction that they had provided through their API, and we did it for all the models we ran.

So the results you're seeing don't always reflect the actual state of things, and that's a problem. Number three: big circle represents all the world and its knowledge. Small circles represent AI researchers and technical professionals. They're like something like 30,000 AI researchers, at least that's what Google AI Search told me.

And I think there's like 30 million software engineers, data scientists, technical folks. We expect AI to help most of humanity, but then a very small percentage of people are creating all these evals. And if something's not being evaluated, not being benchmarked, we cannot hill climb on it, we cannot know how good we are at those things.

And what will this lead to? More of these cognitive edges or jaggedness with these models, as we're already seeing now. And that's only going to get more exacerbated over time as we see superhuman intelligence in some areas and just, like, very mediocre performance in other areas.

That's not equitable AI that we want that will benefit all of humanity.

This is a kind of fun anecdote, but not many AI researchers are also wastewater treatment plant engineers. And I give you this example because this is an actual benchmark built by one of our users. He lives in Turkey.

He's been a wastewater plant engineer for 20 years. I don't know what that entails, but it's clearly a very important job because he built this benchmark because he cares. And he recounted a story where, you know, he's been doing this for 20 years.

There have been severe incidents in his country where there was some incident, people didn't follow the safety protocols, and people ended up dying as a result of that. So he built this benchmark to evaluate how AI could help him in his job and help avoid these incidents in the future.

And so this is like a proprietary novel dataset that he's created from his own experience. It doesn't live anywhere else in the web, doesn't live in any of the AI lab-focused area, because that's not something that's economically productive for them at the moment.

So it's very important why we think we should work on open-source contributions to the eval space. Alright, enough of that. Talking about the solutions ahead. We're working on a couple solutions, but it's tough. So I'll very quickly cover the first two at the top, and then I'll hand it over to Michael to kind of deep dive into the remaining two products at the bottom.

So at the top left, we have hackathons. We have a platform that lets anybody host a hackathon, and I'll talk about why I think it's relevant to the problem of evals. Second, on theright-hand side, we have agent exams.

We want to democratize the process of taking evals. We heard a lot about OpenClaw and consumer agents being a big thing, but the problem is that most people don't care about evaluating their own OpenClaw agents, which is kind of crazy.

Number three, at the bottom left, we have Game Arena. That's where we have this evergreen benchmark where models are playing PvP games against each other. So it's an ELO score-type rating, and it's forever hill climbable and unsaturated because you're just fighting against each other, and there must be one winner and one loser.

At the bottomright, we have Benchmark. So it's the product I run, which is basically a platform that enables anybody to build, run, and share evals to the open community. So very quickly, hackathons. Why I think it's important. Hackathons are a great way to channel people's energy and expertise solving a problem.

I think we've seen that, you know, with theright energy investment and time, we can do a lot with very little. I think a great example is that the world galvanized over the past three years to make Gen AI happen.

And this is like a very small form of how we want to help facilitate that process towards solving theright problems in this space. It's important that we put guardrails around the problems that we're trying to solve so that people don't go crazy, but also give them enough space so that they can flourish and their creativity can show.

And the results of everything will be open source for the benefit of everybody and not just a small group of people. But there are some challenges. And actually, before I dive into the challenges, so on the screenshot on theright is a hackathon that we're actually runningright now with the Google DeepMind AGI team.

So Google DeepMind, a couple weeks ago, published a paper on how we can measure the cognitive faculties of AGI. And so we started this hackathon to focus on five particular faculties off the ten. And we want people to build benchmarks in those areas, and we want to give everybody the chance to contribute to AI research and not just a few people.

And also knowing that everyone has something unique to contribute that these AI labs couldn't do so themselves.

But running a hackathon platform isn't all that easy as well. I think these are fairly self-explanatory. Maybe I'll talk about the second and the third one in particular. The second one being that we need to provide them theright tools in order for them to do their best work.

It might sound trivial, but things like, okay, if you have 1,000 participants, everyone's operating globally online, how can we give them the tools like hosting their own datasets to have access to AI models? That's something I never thought about before.

But for a lot of people, you know, coming from a poorer background, they might not have money to pay for these five API keys to access all these state-of-the-art models. And how can we then let them share their work in a way that's understandable through write-ups so others can see, perceive, understand, learn, and build on the work that they've been doing.

And then the third point is that as much as, you know, AI agents and everything you hear about this conference are very good at a lot of things, they're not very good at, like, judging innovation and creativity. So a lot of the work still requires human experts, and even alignment within experts is difficult and not trivial.

And that's something that we have to facilitate as part of this platform too. The second thing that we're working onright now is what we're calling standardized agent exams. I originally called it, like, SATs, standardized agent tests, but, you know, there was a trademark issue so I had to change the name.

Yeah, it's a true story. So how it works is that you just paste a one-line prompt to your agent, and essentially it takes an exam and we return a score for you on a leaderboard that you can compare its performance against.

This was a very experimental MVP that we just launched last week. And I think this is important because if you look at AI evals that people are doing, it's kind of two ends of the spectrum. You have, like, research labs and enterprises, you know, using BrainTrust, using all these state-of-the-art technology to set up to measure their agents and their models.

And then on the other end, you have consumer agents, people who are building OpenClaws and then, you know, filing 1,100 security advisories this morning, as we heard. But most of them aren't actually testing their agents before they're sending them out to the real world, which I think is a huge, huge problem as we've seen and will become even more important.

So, like, one conversation that came up this week was, how can we maybe do more safety-focused exams so you can do a quick baseline of your agent before you send it out into the world to run your inbox, to run your Amazon accounts, and to do stuff for you.

So I talked about the first point. I think the second and third one are quite interesting. On the second one is that when something's accessible, we want to make sure that it's also challenging enough. And so we have this spectrum of if we make something too difficult, people can take the exam, but no one finishes it because it runs for too long, it's too difficult.

But if you make it too easy, then it doesn't give you theright signal for what you want to measure. And then finally, the chart at the bottom shows maybe there is a market for agent consumers. We only launched a week ago, and we have, you know, hundreds, like 500-plus agents already evaluated on our exam without us even really promoting it very much.

So I think that's an interesting insight that we've gleaned from this experience. On theright-hand side, on the screenshots that you see there, so we posted in MOPBook, and then we started seeing these weird kind of spin-off posts about people sharing, agents sharing their exam results, and even, like, an SAE prep course that came up on MOPBook.

So, you know, that's always interesting to see. With that, I'll hand over to Michael.

Yeah, thanks, Nick. Yeah, so I really selfishly wanted to talk about Game Arena and benchmarks with y'all AI engineering and mostly kind of give an overview of how it works, some of the really cool things we've seen it, but mostly I want to talk about the challenges that we've been having with them.

And so please come find me at the DeepMind booth after this to talk about them. I would love, like, some more insight on these kind of things too. So for Game Arena, it's a benchmarking platform. One of the problems with benchmarks is they very quickly get saturated.

We see this with community benchmarks, we see it with AI and, like, researching benchmarks as well. So Game Arena is an approach to help us work against the saturation by just having PvP. And so you can never have saturation because you'll always have one model able to compete against others.

So saturation might just be, for a while, a model is the best. Really quick engineering slide here of, like, how is all this set up and how does it work. So when we're trying to figure out games to put into Game Arena, we want to analyze, like, separate capabilities of AI models, and so we try to pick good, varied games.

So far, the ones we've vested a lot in are Werewolf to, like, play around with what's best at deception, Poker for, like, the randomization, also some of the deception and how good, like, Grok loves to go all-in on Poker.

Less models are a little bit less crazy, are a little bit more conservative. Most interesting, some of the newer generation of models are worse at Poker because they are more risk-averse. And so you just see these personalities start to emerge over time.

And then chess, because anytime you're analyzing ML things, you have to be analyzing chess. And so for the quick just overview of, like, how all this works, we design and iterate on a game, figure out a good game we want to do, make sure models can actually play it, and then spend a lot of time on iterating prompts to try to make sure that our prompts are fair.

And this is all, like, open source and, like, very viewable for, like, what's happening. So if you want to check it out, I put the GitHub link there. Really, all of the things that we've talked about in this presentation so far are live on Kaggle.

A lot of them are open source, so please come look at it, play around, give us feedback. We love it. So after that, we work on building this harness. Mostly we've done OpenShpiel Games so far, which is like an RL framework.

And so, like, we test, like, are these models better at playing at random? Sometimes they are. Usually they are, but, like, sometimes not that much better. Are there any other interesting properties that emerge? Finally, we end up running the simulations.

We use LLM model proxy. This is actually available on Colab, if anybody uses that, to just, like, talk in a consistent way to all of the models that we want to run these games against. And so it runs on top of the Kaggle simulation platform, which was, like, initially an RL platform that we had for Kaggle before LLMs became a few things.

We schedule game runs, use this Bradley Terry pairwise to try to, like, not have too many games we have to run. I'll talk about that in a second. And then finally, we publish the results. We have all these LLM conversations.

We stick them in a dataset that's available on Kaggle. People can check it out and, like, learn things from that. We put it on to benchmarks to show the ELO scores. And then we also have a game visualizer for all these things so you can go and, you know, see Grok go all-in on poker hands.

That's a little demo to the left of that. So now, as promised, some of the really big, like, challenges that we've had with this. And so love to hear all y'all's ideas and, like, different ways to approach this.

You can imagine it gets very expensive very quick. I'm sure you all have seen, you know, your Claude 4.6 builds are something along those lines. So you can imagine that for poker, in order to get statistical significance, we had to run about 400,000 poker hands.

And there's many turns inside of each of those hands. You can imagine what those builds start to look like. And so the Bradley Terry pairing is, like, part of this, but any way that we can get statistical significance without having to run millions of games is great.

And, like, we're always trying to think of new ways to, like, be able to be sure that the models are best at the things that we're claiming they're best at, but, you know, running as quickly as possible. It gets a little boring to just watch, like, LLMs play against each other all the time.

Like, for some of the initial games, it's pretty fun. But, you know, as we're trying to build this out, it might get a little bit repetitive. And so trying to figure out ways to engage Kaggle's community to be able to participate in this process.

And so, you know, even things like, could we have a hackathon that somebody provides a prompt for, for example. And then, like, you know, we would have, like, prompts given by our community as part of the competition to, like, play these games and see who prompts, like, the model the best and, like, climbs on a leaderboard.

And then comparison over time is difficult. Old models disappear, new models come along. Sometimes when you're talking to a model endpoint, if you're not talking to yourself, they're not exactly honest on what model is happening in the background.

So that's always a little bit of a difficulty. But yeah, go check it out. Pretty fun. So yeah, moving on to benchmarks with my limited time here. So what is this not? Is this not, like, a production evaluation platform?

There's plenty of people talking about that over and more if you want to go and run this for your production code. Very cool things. This is much more about community involvement. Anyone can, like, build and run and share evals in a hopefully open and verifiable way.

For time's sake, I'll just talk about this really second, but we basically, it looks very similar to the production evaluations platform where you write some assertions. So, like, this thing worked, you know, what gets wetter is the driest.

Like, you can check, okay, does this contain a towel? We also do LLM judging similar to the production platforms. These all get grouped together in a task. They then get evaluated against a collection of models that the users want to run against.

And then all these tasks get aggregated together in a benchmark, such as the wastewater treatment one that Nick was talking about previously. So Paige, who just presented in this room before this, actually made this nice little task for us.

It was parsing an SVG from XKCD and is like, can you recreate this SVG? And so you can kind of see the code for this is over on the left. One of the models is a little bit outdated, so Sonnet 4 created this nice reproduction below it.

And then Paige created a number of assertions. So, like, you know, can it generate an SVG at all? Does it have the correct text and some other checks? And then an easy way to compare these things side by side.

So yeah, some of the challenges with this of that, like, inspiration and incentivization are hard. If you want to, like, it's not hard for a production evaluations platform because you're, you know, shipping a thing to consumers. They care about, like, does this model work or not.

But to just inspire people in the community to create benchmarks that, like, other people find interesting. We've had good luck with hackathons and, like, you know, Kaggle has points and, like, medal system and things like that. So we have some things baked in the platform to help inspiration and incentivization, but it just takes a lot of work to write a good evaluation.

For agentic benchmark execution, like, oh, yeah. So when we started this, people were really interested about analyzing just models. As we've moved more on to what are agents doing, it gets really hard to figure out, like, what we're actually testing against.

So I pulled out this little thing from MorphLLM paper, like, blog post that they published on March 16th that basically called out that, you know, against SWE-Bench Pro, the six frontier models are within a couple of percentage points of each other.

Definitely, like, go check out this blog post. I haven't specifically verified it, but it doesn't seem likely to me. The thing that really matters a lot for coding performance is what harness is it running inside of with, like, a 22% difference depending on the harness.

And so that can get, you know, really tricky of, like, are you testing the harness? Are you checking the model? Things like actual ambiguity under test. And then, again, fast release and deprecation cycles of models, it gets a little bit tricky to figure out what we're or, like, to be able to do comparisons over time.

So yeah, I think that's about time. But as I mentioned, we'll all be at the DeepMind booth for the in-between times, or you can email either me or Nick here. But thank you so much.