Computer Use at the Edge of the Statistical Precipice — Pierluca D'Oro, Programma Labs

0:01 [music] Hi everyone, I am Pugadoro and I'm the founder at Programmabs and today I'm going to talk about computer use agency evaluation and most of the work um and

0:24 the details about it are in a in a paper with this title and I did this work while Meta super intelligent labs with the collaborators you see on this slide. And so to start I want to introduce this

0:39 type of uh agent is a weird type of agent that I call a replay agent. So imagine we run this process um we run our frontier a frontier model a good one uh on a benchmark we like and then for

0:52 every task we collect uh a successful trace or a successful trajectory and we have a recorded tape of this type. So the actions might be tapping uh typing scrolling and we record this and then we

1:08 do this for all the tasks in the benchmark and we com sort of compile this into a replay agent that just when the tasks arrive replace that sequence of actions

1:20 blindly. So if you do this for common benchmark with hundreds of tasks, this is going to be a script that is like less than a megabyte. And this is a completely valid like type of agent that

1:31 you can evaluate on the benchmark. And if you try to evaluate this uh this kind of agent on standard benchmarks such as OSW word or mobile word, you will see that the success rate of this agent

1:43 compared to the uh frontier model from which the agent was extracted is actually uh the same or even better. So this is a kind of a weird but maybe trivial phenomenon but I would argue

1:57 that we shouldn't accept this kind of blind scripts beating the frontier models and the trick here or like the reason for why this happens is that um determinism of of the existing of most

2:10 existing benchmarks and so uh if the benchmark is static is deterministic then it is somehow gameable by this sort of strategy and it goes even deeper than this like if you look at one of the

2:23 metrics that people have been using in the past for evaluating computer use agents um and it's pass at K this metric is defined as you know the probability of at least uh one of K attempts

2:35 succeeding um but if you look into like the details of how this metric works on a deterministic environment you will see that is literally um and we prove it like formally in the paper uh like

2:49 evaluating the success rate of the replay agent that I've shown to you. Um so so basically if that replay agent felt weird uh to you also pass at K on computer use tasks should somehow feel

3:03 weird to you. Um or in other words pass K is sort of a metrification of that uh of that exploit of the replay agent. So these are two problems specific problems but they point at two general classes of

3:19 problems in in Kua benchmarks and uh these problems are around environments. So building environments and evaluation building good metrics to know if your agent is good or not. uh and in

3:32 particular we we want to have environments that don't have exploitable structure and we want to have metrics that are not fragile or they are not based on fragile statistics. Uh and so

3:43 I'm going to talk about the both of the aspects now. So let's talk about building principle environments first. So the first aspect um that I uh that I worked on while uh you know working on

3:57 environments is to try to design a set of principles that could be guiding principle when building environments so that build like robust environments and trustworthy environments. Uh and so if

4:10 you think about the problem that I was describing with uh replay agents, uh the first thing that you could think about as a solution not to have a replay agent to like hack your benchmark is uh to

4:21 have your benchmark to be multiffactorial. So that means uh varying um generating variation for your benchmark. So having stockasticity into the benchmark and for computer use

4:33 environment that means uh varying stuff like uh data or appearance or or simply the initial state. But if you do that you want to make sure that all the combinations that you generate are

4:46 valid. Uh and so you want to have as a design principle in your environment also a system for checking and verifying that everything is working as as intended for every combination. And of

4:59 course you want uh the usual uh the usual things for your environment. So you want your environment to be sandboxed and you want uh your environment to support like verifiers.

5:10 So privileged information and you want of course your environment to be realistic. So if it's a reproduction of a real system, you want uh that reproduction to to be faithful so that

5:21 the score that you get out is um is a good one. Um and so if you sort them out uh you can remember these sort of principles as the prism principles for for environment design. And we tried

5:35 that method to build uh a benchmark that would be satisfying all of these principles. And if you look at existing benchmarks, some of them do some things uh in a good way. Some of uh some others

5:46 do other things in a good way. But there is no unified benchmark that sort of matches all of these boxes. And we built one that is called DGword. So the way DG word uh in practice was built is as a

5:59 set of like mobile apps um for for Android devices. So it's 15 apps on spanning different domains uh with 387 uh verified scenarios and uh a number of configurations. So uh these

6:15 configurations they are in a large number of 3.2 million. Uh but the important thing is that they are verified. Um and indeed uh the axis are the ones that I was mentioning before.

6:27 So you can imagine for each one of the tasks you can vary um things like the the instance. So what is the exact amount of money that you're sending for instance or the data profile like which

6:38 kind of contacts or emails you have in the data for your uh for your task or like the theme or the starting screen. So do you start from the login page or do you start from another valid page? So

6:51 if you do the math, even if you start from a relatively low number of uh base cases for each one of these variables, you end up having many many combinations. So you you can get to like

7:03 millions of combinations and if you scale this up, you can get to easily to billions of of combinations. Um and all of these uh you know different axis can be can be manipulated by coding agents

7:16 because in the end they are like forms of software. So you can have a coding agents to generate different instances different themes and such. Um, so you might think maybe it's easy to build an

7:27 environment. You just generate as much software as you can uh with a coding agent and then you have like a diverse environment. But it's a little bit trickier than that. And indeed coding

7:38 agents can generate a lot of software, but a lot of software is not the same as an effective cool environment. And the reason for this is that you need to verify the correctness of of your uh

7:51 combination, right? Um and so the the the key to scale these up is to have a verification strategy for the variations of of your tasks. Um and so the kind of verification strategy to follow is is

8:06 this one. So you you can generate many configs all the combinations uh of the different factors that I that I've explained before h and you can then have a system that rejects the broken ones

8:18 the ones that are not valid and just keeps the the valid configs and so in the case of dig uh we did this by building a system that looks a little bit like a compiler and that works in

8:31 the in the following way. So you start from a parameterized task template. Um and so uh this might look like something like this. So you have send a certain amount or certain recipient and then you

8:44 have a verifier that corresponds to that template and then you have mock data uh for that task. So data that you need for that specific task to happen. uh and then we have a system that is like the

8:56 dig compiler that takes all of these and given a base case of data base case of UI state puts all of these together and creates like a valid configuration and so you can build systems like this in

9:11 which the the main craft is good software engineering uh to make sure that actually the combinations that you have are both diverse and valid. And so if you build a benchmark in this

9:24 way um and you follow the the prism principles that I was talking about before you actually can have a benchmark that is not hackable in that way. Uh and so if you do the same process we did

9:35 before you evaluate your frontier model and then you evaluate the corresponding replay agent uh you will see that the replay agent doesn't get a lot of performance. Um it gets a little bit of

9:46 performance that is probably what you want. Sometimes some tasks maybe are repeatable by nature, but on average you you shouldn't expect a replay agent to have good performance on the benchmark.

9:59 Once you build like this diverse um combinations, you also can do other things like um measuring the um robustness of frontier models on different over different axis of

10:12 variation. So the axis of variation uh I described before are here represented there. Um and you can see that in the worst case frontier models are pretty bad actually at um being robust to these

10:26 variations. And so for instance if you have a model uh that seems to be good at a given task, you would expect that if you just very you know which screen the task is starting from or like what is

10:39 the team of the app, the model should pretty much have the same performance. But this is actually not the case for most frontier models. And so if you have infrastructure like this, you can

10:48 actually measure measure that and like tailor your expectation uh about this kind of uh robustness. So this was about the first aspect that was building an environment that

11:01 supports um diversity uh and that is robust enough to evaluate models. But the second aspect is as important as the first one is to measure uncertainty honestly. Uh so once you have all of

11:16 this variation, how do you handle like computing um the real performance of your agent and basically there are two sources of of stoasticity of variation and uh they are not exactly the same but

11:30 they are equally important and so the one that we usually think about is the one about the actions right and so you you run your model multiple times in many cases you can have even quite

11:43 different trajectories out of it because the action at each step would be different. uh but if you have a benchmark like the one that I've described with multiple combinations

11:52 with multiple variations then also the variability from the environment becomes important and we want to capture that because that is what we are going to find uh in the real world and so we need

12:04 a methodology that captures uh both of these types of variation and uh in the paper there are the details but but basically we build a meth we built a methodology that can

12:16 accurately um capture these two types of variation taking into account the structure of the benchmark and so if you start like in in practice it is useful to use this concept of coverage. Uh so

12:30 when you when you compute a confidence interval uh basically you have some confidence that the performance of the model is inside of that range. Um and so you would expect that a 95% confidence

12:43 interval would say that you know 95% of the time the performance of the model is on that range. But if you only use rollouts, so you only use the base case and what people would use normally

12:55 actually in realistic cases you have something like 17% on 20% coverage. And that means that basically you only 20% of the time you uh guess the right performance of the agent which can be

13:08 pretty bad. But like if you take it into into account the hierarchy and you use the proper way of computing confidence intervals, you can get to the to the full confidence interval and be 90 95%

13:19 accurate. And so if this seems quite abstract uh you know in practice that means that if you want to make a decision about which models to deploy maybe you have model A model B and you

13:32 do an eval for for those two models um you can have cases in which the confidence intervals seem really really small uh and so you make a decision based on those small confidence

13:45 intervals but they are actually overconfident and so uh this was the wrong decision. So the the orange bars are the real performance here. Uh so you make this decision and if a mistake is

13:57 pretty costly for you and you have many tasks like if you have one million tasks and there is a 4% mismatch in performance for real in the models and each mistake is like $20 uh like $12 on

14:11 average. Uh it can cost you like hundreds of thousands of dollars uh in a single month. So it can be like super costly as a mistake just for a confidence interval being overconfident.

14:24 But if you have like a reliable way of computing the confidence interval, the method would tell to you um I'm not confident enough to make an informed decision. And so you can choose like to

14:34 spend more money to spend more time on evaluating models and avoid the co costly mistake. So we don't want to like elude ourselves with like wrong confidence intervals because there's uh

14:45 you know money on the table essentially. And so this is sort of a final checklist of uh the things that I've uh that I've discussed so far. Uh so again to recap some of the things that are important in

15:01 building on benchmark are about the environment and some other things are about the metrics. So on the environment you can follow uh the principles uh that I described before like prism

15:13 principles. So some of these things are rather common but some things like um varying initial state across runs they are pretty rare across existing benchmarks but they are very important

15:25 and so I would suggest you to try to incorporate these into your evals and and things about the metrics. Uh you can of course read the paper for the details but essentially um it's very important

15:39 to avoid uh replayability as something that uh you can have in your benchmark and also to focus on having accurate confidence intervals. Uh so respecting the benchmark structure and trying to

15:51 avoid um underestimating the uncertainty overall. And so I've heard many times sentences like this. this benchmark can be gamed but everybody's still using it or like

16:03 there is no error bar but I don't see people using them. So these are sort of uh things that we can think when we don't have enough time but actually a non-rigorous benchmark is misleading. Um

16:16 you know it can be misleading for the field because everybody could be seeking you know um maximizing a score on a benchmark that maybe is not capturing what we care about but especially it can

16:27 be misleading for uh you know your own decisions. And so if you um are deluding yourself on thinking that a score is like confident and that is confidently telling that your model is good actually

16:41 you are going to pay for those mistakes and so I think it's very good uh usually to be honest with with yourself and to try to be rigorous in the evaluations that you have. Um as the last slide uh I

16:55 just started uh this company programmer and we are building uh the best infrastructure for ku enable verification uh and so we are hiring if you're interested or want to chat and

17:06 this is our website. Thank you very much. [applause] >> [music]