Rethinking Environments for Long-Horizon Work — Rayan Garg, Theta Software

AI Engineer · 21 min · 221 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:12It's uh great to see all of you here today.
00:14We're super excited to talk about one of our favorite topics uh here at Theta.
00:18Um before we get started, we just want to introduce ourselves.
00:21Um so, hi, I'm a co-founder and CTO at Data Software.
00:25Hi, I'm Ryan.
00:26I'm a co-founder and CEO at Thata Software.
00:28Prior to this, I was previously a founding engineer at Deep Silken where we did research into turnary models.
00:34Awesome.
00:35So, I can get us started with the topic today.
00:37Um, we're going to be talking about oral environments uh within the context of long horizon tasks.
00:41And I think the most important thing for us to start with at the beginning is just talk about the trends and what long horizon actually means.
00:48So, you know, we all know that the horizon at which AI agents can work autonomously is accelerating really fast.
00:54Um this is just some of the metrics that you can look at to see how this progress is really accelerating.
00:59Um but I think it's really important to actually define what the time horizon here actually means.
01:03Uh we've gotten this data from one of the most common benchmarks out there that you've probably heard of for time horizons.
01:08I'm sure you've seen in your Twitter feed all the time.
01:10It's um comes from meter.
01:13Uh and meter kind of has one response or answer to this really important question about how do we actually define long horizon?
01:20Um, it's really important to understand because what we consider long horizon a year ago probably isn't really long horizon in our definition today.
01:27And what's long horizon today probably won't be long horizon in a year or two.
01:30Um, and I think that gets to our first point which is that long horizon is really kind of a scalar metric.
01:35Uh, it's useful for kind of measuring relative tasks like one task might be more long than another, but it's really hard to define into kind of a binary category of this task is long and this task is not.
01:46Uh, especially as the kind of scope changes over time.
01:49So I think the first way we can talk about defining this is how meter kind of looks at it which is human horizon meaning can we use humans as a benchmark of oh this task takes humans a certain amount of time so if AI agents can do that then they've reached this certain critical
02:04uh level of kind of a time horizon uh and the way meter kind of does it is they have thresholds for the tasks they care about.
02:10So, you know, they have a 50% threshold, meaning, you know, if a certain model reaches a 16- hour threshold on this benchmark, that means it can achieve tasks with a 50% success rate that take a human 16 hours.
02:21And, you know, there's a really rigorous methodology of how they actually measure how did it take human 16 hours, but we will kind of avoid some of those details.
02:28Um, the other way we usually think about what long horizon actually means is not with the reference of humans, but instead the reference of models.
02:36Uh so some of the relevant model units we usually care about are things like tokens.
02:40How many tokens are consumed in a trajectory, how many steps it took, how many tool calls it kind of takes.
02:45Um and these can be really noisy, right?
02:47Because I'm sure you guys have used different models.
02:49Um like you know a lot of the codeex models are seen as more token efficient than some of the cloud models.
02:54Um and it's a pretty noisy estimate for a couple reasons.
02:57One is that which model you're using like I just said and different harnesses you care about have a pretty big impact on how many tokens are actually consumed on a task.
03:04Right?
03:04So, um, this can be pretty hard to interpret when you're not holding variables constant.
03:09Um, you know, if a task takes GPT model 500,000 tokens, that doesn't really tell you a lot about what that task would look like for cloud models until you actually run on those cloud models.
03:18But despite it being a pretty noisy metric, it's actually really useful uh and important for us to understand because, you know, the amount of tokens that are consumed tells us a lot about how difficult a task actually is for an AI agent to kind of tackle it autonomously, right?
03:33uh we have to deal with things like compaction over long horizons.
03:36Um you know they don't really stay coherent over enough steps or trajectory uh length that you kind of achieve.
03:41So even though it's kind of a noisy metric uh it can be really useful when you know if we look at what a GPT 5.5 model can do uh now and then you use the same model generation and kind of see oh now can actually achieve a a million trajectory based on an increased context window or improved
03:55compaction endpoint.
03:56That tells us a lot about how autonomous AI agents can actually go for long periods of time in that sense.
04:02Um, and it really defines for us what the technical frontier actually means uh for models right now.
04:07Maybe not really human adjacent.
04:09It's really hard to say how many tokens a task takes for a human because we don't really think in tokens, but still very useful in that kind of sense.
04:15Um, so these are two different approaches we can think about.
04:18But what's actually the right way to think about this?
04:20Uh, the answer is that we probably want to think about all of these.
04:23And if we just look at one of these metrics in isolation, it's probably not a great way of measuring things.
04:28So, I went through some of the weaknesses with measuring with like model specific metrics like tokens and and steps, but there's also a lot of weaknesses in the other approach of kind of relying on humans.
04:38Um, you know, what's long horizon for a human isn't necessarily that difficult for a model depending on what the actual task you care about is.
04:45You know, there's a lot of tasks that are really tedious and time inensive.
04:48Maybe like, you know, some financial analyst has to go into like an Excel file and fix a bunch of formatting issues throughout uh the the task.
04:55Maybe they're changing like the theming of like the colors in the in the actual file, right?
05:00That might be really tedious for a human to take.
05:02It might take them like days to do that if it's a really big Excel file, but for a model, it can maybe write a Python script or find some other cool trick to do that really quickly.
05:09And that's not really hard for it to do, but you never really expect a financial expert to do that because, you know, they most of them don't really know how to write these Python scripts.
05:16Um, so I think that's one thing to note.
05:19And the other that I kind of briefly touched upon before is that the methodology of how we actually measure this has a really big impact.
05:25And if you, you know, someone is out there saying, "Hey, we have some tasks or environments that are 16 hours long on average, someone else is 20 hours."
05:31That's really hard to compare across people because there's so many different things in the methodology that really impact uh kind of what that actually means.
05:37It can mean, you know, the quality of the experts you're using.
05:40Some more experienced experts might actually be way more efficient at doing a certain type of financial or coding task, whatever it kind of is.
05:46Um and I think this becomes really really important as we start shifting towards uh kind of the frontier of even human capabilities.
05:53So you know as this meter talks about this but as you shift towards more long resin tasks and tasks that only the top 10% the top 1% top.1% of humans can really do these estimates start to get really really noisy and it's something that we really have to consider.
06:06Uh and I think you know the way agents work is kind of developing in its own separate path and there are a lot of different bottlenecks and different things that AI agents are better at than even the way humans work.
06:16And with that in mind uh you know as these paths kind of diverge of how humans do work and what their limitations are and what agents do and what their limitations are.
06:23Uh it's really important to kind of keep both these metrics in mind because they kind of tell and paint different pictures uh of of what's actually relevant and and you don't really get the whole picture by just looking at one in that sense.
06:36Yep.
06:36So now the question becomes, how do you measure model capabilities?
06:40And this is a really important question because fundamentally long horizon tasks aren't the only thing we care about.
06:46This is the larger question that we want to think about every time we're trying to create tasks, create environments to train our models.
06:53And so the first way we can think about this is environment complexity.
06:57And specifically environment complexity related to tool coordination, right?
07:01So how many tools or external dependency does the agent have to coordinate?
07:05How many tools or external dependencies does the agent have to move information across?
07:09So if we start off kind of thinking about what the world looked like before a long horizon task uh you know world, we'll notice that there was you know a low complexity world where the agent maybe had to read one file or one set of files in a code base and that's kind of what a task entailed.
07:24But now we can see increasingly as these tasks become more long horizon what is important to define for measuring model capabilities is okay the the agent should be using a ton of different tools like graphana for observability
07:35to parse logs or GitHub for CI/CD or AWS cloudatch or reading and writing to a database and we're going to notice that as we sort of start to have these agents and these environments
07:48use many tools that we also start to think about environment complexity in regards to state changes which is effectively the degree to which the environment changes throughout the task.
07:57And so fundamentally the way we want to think about this is right all long horizon tasks aren't equal.
08:03So for example, one task can you know maybe be made by artificially long horizon by chaining together unrelated independent tasks.
08:11However, that doesn't actually tell us or meaningfully measure the model capabilities.
08:17Instead, a key component of this is actually being able to have the earlier decisions in the in the environment influence the later decisions.
08:25And this comes back to how the agents are asked to interact with the tools, how these tools change the state of the environment, etc. So, we can look at an a concrete example for this.
08:34One example where you'll see paralyzable complexity, which is effectively not involving a lot of state changes, is when you can maybe have an agent analyzing a large code base and then the agent needs to spawn off multiple sub aents and it can very easily paralyze this, right?
08:46it can look at a lot of the different files in parallel, come back to the to to the master agent and then kind of wrap this all up, right?
08:52But meanwhile, if we look at sequential complexity, we'll see if you have to use a dashboard or logs, a bad early query or a misread can cascade into these downstream steps that really start to have major consequences later on, right?
09:05It's all dependent on how you use those tools and how the state of the environment changed.
09:09So the third area that we also need to consider for measuring model capabilities is ambiguity, right?
09:15And ambiguity is defined as the information you give the agent and the environment when starting the task.
09:20So this could be the instructions, this could be the artifacts, etc. And increasingly as these agents work with more artifacts at the start, right?
09:28We want to have them mirror the work that humans really do.
09:30And the work that humans really do has a lot to deal with ambiguity, right?
09:35They always are are don't have the most complete information and they want to let exploration happen.
09:40And so we believe that to measure model capabilities, we need to test the model's ability to explore and explore throughout the environment as well and explore these artifacts similar to how a human would.
09:50Now the trade-off with this, right, is that if you are going to have ambiguity in the materials you give, there's a lot more possible paths that the agent could take.
09:59There's a lot more ways the agent could be right.
10:01And that means that standardized evaluation gets much, much harder.
10:06Awesome.
10:06So I'm going to talk about one of the hardest things there are to build environments and one of the most complex things to really think about where there's a lot of nuance which is the verifier in the environment.
10:16How do we actually know that the work the agent did was correct and give it some reward signal during the training process.
10:20So I think there's a few challenges here.
10:22Um you know I think just to give a high level overview um you know tasks are getting more complex the environments are getting more complex the trajectories are getting longer
10:30and we've shifted a lot from you know a lot of the early RL that we were doing in in recent times was really in hard verifiable domains and that's why we saw these gains in in math and kind of uh like data structure style coding problems but what's happened over time is now we really care about a
10:45bunch of economically valuable work in software file domains is is a way to put it where uh you know we can't just run a Python script or run test cases is or or write a proof to really see whether or not the output was correct or whether or not the environment was changed correctly.
10:58We have to start using other techniques and the main way we're really going to use that is kind of introduce a judge model or critic model as some people put it and they kind of can add a lot of nuance to how we actually look at a few things here.
11:10Um you know very critical for how we actually determine correctness and assign reward.
11:14Um they'll look at two things mainly.
11:15One is usually either the state final state of the environment and kind of how it was impacted.
11:20Um and the other is looking at the trajectory of how the model that you're actually training made kind of changes to the the state of the environment as well and and what kind of correctness look like there.
11:29Um so you know why do we actually use uh judges and usually rubrics um as a technique?
11:37I think it's really important to understand before we can even understand how to use them properly which you know I think there's a few reasons.
11:41One is that like I said for these software fellow domains there's like an entire class of problems that are really important and a lot of the problems we care about that really you can't really write a deter deterministic
11:50verifier for um they would be really impractical brittle or just downright impossible depending on what the problem setup really is.
11:56Um and you know I think the other thing also is that like I said we're going to look at the trajectory uh and you know not all solutions are really created equal and not all paths of those solutions are equal either.
12:07uh you know the worst case of a bad solution we can get is some reward hacking that happens.
12:11Lots of different types of reward hacking can happen depending on the setup or the task you kind of care about.
12:15You know agent can escape a sandbox maybe see privilege information it shouldn't be seeing about maybe a hidden test suite for like a coding task.
12:22Um this is all behavior that we want to prevent obviously because those are not actually really valid solutions we really care about.
12:28Um, and you know, a lot of this mitigating this is going to require strengthening your verifier and your environment setup, but the judge is really, really important in actually catching this behavior.
12:36And that's an important reason of why we actually look at the trajectory that the agent actually took to get there.
12:41Um, yeah, and I think there's a lot of careful things you want to be doing here.
12:45Uh, one is that there's nuance in how much guidance or uh like explicit rigidness you want to add to the trajectories that the model can actually take.
12:53um if we kind of enforce this too tightly, we collapse the state space of how many actual paths the agent actually explores.
13:00Uh and that can be really bad, especially because I think some of the more simple approaches we've seen with judges early on is, hey, we'll just give it a reference answer or solution or maybe a sample trajectory of what a good solution looks like and then just compare against
13:12uh what the model did and say, hey, does it match up with that?
13:14Uh and that really does not work for these more ambiguous or open-ended tasks because there's so many possible correct solutions.
13:19It's basically impossible to account for every single one.
13:22and we want to check for more robust methods that allow for these different solutions.
13:28So now that we've kind of established why we use judges, uh we want to go through some of the general heristics and kind of principles we think about when we're designing good judges.
13:37Um some of the things that we think about at data.
13:39So you know I think the first important uh consideration to make is that judges are agents too.
13:45Um, you know, so as environments get really complex, oftent times we a consideration we kind of have is like, hey, we have to make sure the harness can kind of scale and and kind of match up with whatever environment
13:55uh you kind of have.
13:56Maybe that means introducing a bunch of new tools and making sure your harness can support those tools really well.
14:00The agent has clear observability over what's happening in the environment.
14:03Um, but I think like we said, the way the judge determines correctness is that it oftent times has to look at the state of the environment itself as well.
14:10So a lot of the harness that you've designed for the agent might also be reused uh for the judge as well.
14:15Um, I think the best way to illustrate this is the example we have here.
14:17Let's say you've defined a task where, you know, there's some deployment failure with the software engineering task of some platform you're deploying and the agent's task is to like sift through the CI/CD logs on GitHub,
14:27look through the cloud cloudatch logs, figure out whatever happened, uh, kind of apply the changes you care about to the codebase and then open a PR and and kind of kick off a redeploy there once the PR is merged.
14:37Um for a lot of for a lot of that if the judge actually wants to verify uh whether or not this is correct besides just like looking at the tool calls agent mate which are usually not very reliable it actually has to also check the GitHub logs it might check the AWS
14:50logs or the GitHub logs after the deployment happened to make sure oh are things actually working properly.
14:54So it's really important that the judge has access to the environment in the same way uh with some important safeguards of course.
15:00One is that we don't want the judge to make an accidental mutation in some way to the environment after the agent is done.
15:04So you want to be very careful about that.
15:06Maybe that means enforcing readon permissions for a lot of this information.
15:09It can't actually kick off a deployment or anything like that.
15:11So those are things to be careful about.
15:12But I think this is really really important especially where there's a lot of open-ended approaches and the only way we can really verify correctness is to actually look at the state itself.
15:20Uh you the answer isn't obvious of whether or not the agent completed the task just from looking at the trajectory.
15:24So I think that's one example where this approach is really really important.
15:27Um I think uh the other thing to be notable of is you know as these environments get more complex the agent trajectories get longer and longer and part of the reason we also need the judge to be an agent is that you can't just use this really basic approach of taking the trajectory and stuffing it in the
15:42context window of the judge and kind of have it be a basic LM call.
15:45Uh these trajectories can get really really long and really really complex.
15:49So we need to do a lot more thoughtful uh processing of the trajectory in some meaningful way.
15:54So you know that might mean we put into some database.
15:57We use sub agents to actually enrich certain information.
16:00Maybe we parse out specific phases that the agent was actually in.
16:03Maybe the beginning part was it going through logs.
16:05The second part was actually writing code.
16:07The third part was actually it checking what happened after that.
16:09These are all different things we want we want to do.
16:11And in that sense we need to make the trajectory itself queryable.
16:14So that might mean enriching of information like I just said or some other metadata we can kind of look at at certain steps.
16:19is really important so the agent can find critical steps um you know like failure points and and verify whether or not those are actually failures.
16:26Making that uh kind of usable for the agent is really really important.
16:32Um I think another important thing to consider is learnability of our environments.
16:36Um you know the most important thing here is just the density of the reward signal and a lot of that comes from your rubric and kind of how the judge is defining that.
16:43So, I think if you be very careful with just uh you know overloading with density in in your rubric, a lot of times, especially for frontier problems that models aren't really capable of yet, judges will really struggle to apply that rubric consistently.
16:54So, there's a lot of QA we kind of need to do to make sure judges are able to apply that information correctly.
16:59Um, you know, there's other learnability factors that we care about and we measure in environments like the distribution of tasks and and and the actual underlying data there is.
17:06Uh and and these are all kind of things we think about for learnability and it's really important otherwise you're just wasting a bunch of compute on on problems where the model can't actually effectively learn.
17:16Um you know these are some emerging rubric judge patterns we've seen.
17:19I'll quickly skim over this.
17:21Um you know oftent times you deterministic verifiers aren't completely dead.
17:24oftentimes we use them in tandem with judges.
17:26Maybe generating an artifact for the judge to actually look over where you're maybe collecting metrics or interesting thing that we could also use is dynamic evaluation time rubrics where we're actually generating
17:36um you know we're maybe giving partial credit where we've baked in some assumptions that the models made and assume they're correct.
17:42It's like grading a test assuming like if you got the first part wrong, let's just assume it's correct, did they get the rest of the part right?
17:47That can be really important as well as well for kind of assigning credit there.
17:53skip over this part.
17:55Um I will let R just close things off with some things about QA for rubrics.
17:58Yep.
17:59So for each rubric we produce, we run a couple different tests.
18:02We won't go into all of them.
18:03Some of them are pretty basic, right?
18:04Gold, no op variance.
18:06These are tests you want to be considering regardless for your verifiers, but I think increasingly, you know, as as you involve AI in the process of even creating rubrics or verifying rubrics or aiding experts, you need to have more and more tests, especially as the tasks become more long horizon.
18:18And so that really touches on the coverage and the expert agreement.
18:21But I think what we wanted to close off with today is why a lot of this stuff matters, right?
18:25We spent a lot of time earlier in this presentation defining what long horizon means.
18:29And a huge reason we did that is because we feel like a lot of the literature and data, a lot of the literature shows that a lot of the data being produced right now and being used to train and evaluate models is actually flawed.
18:39So we present three major benchmarks in the area of finance predominantly.
18:44And so this is GDP valer toolbench and Apex agents.
18:47There's a couple of notable issues here.
18:49First, if you look at the average human hours per task, based on what Meter has defined for a lot of the leading frontier models, a lot of these different average human hours per task fall far below that and so they wouldn't actually be considered long horizon tasks.
19:01The second notable issue here, right, is that we see that these benchmarks are already reasonably saturated and we think this is a downstream effect of the average human hours per task.
19:10So, it's really important to look at the metrics that are being used here.
19:13If you look at, you know, the Apex agents IB section of this benchmark that they put out, pass at one effectively means that for like 57% of cases, the tasks are 100% solved.
19:24That is effectively telling us that like there's a large part of these tasks that models are solving similar to what we've seen already.
19:30But I think a third key important part here is the breath.
19:34For each of these different uh benchmarks, particularly GDP val, they have a very narrow set of Excel tasks that they consider for finance.
19:41And for Apex agents, they're largely focused on IB.
19:44What this means is that a lot of these more important areas for learnability like, you know, credit, debt, risk in the domain of finance don't really get covered.
19:53And then I think lastly, I'll I'll I'll note there the reward signal as Gver mentioned is really important.
19:59And in regards to the reward signal here, we'll we'll notice that there's like really, you know, if you look at if you look at what you need for a rubric, you need very granular,
20:09detailed reward signal.
20:10You need, you know, we we have 20 different subriteria and 10 different subriteria per criteria.
20:17So, I think there's a lot of room that's left when you read these benchmarks into how granular reward signal they're giving, which is really important for being able to go ahead and train your models.
20:26With that, I think I wanted to round off with a couple of stats about the data we produced.
20:30Here we, you know, you can look at some statistics for our finance data.
20:33We can see that the human time to complete one task on average is 15 hours over a 50 task sample set.
20:38Furthermore, it takes models a pretty long time to work through these tasks.
20:41And after all of that, across all the domains we care about within finance, for example, they still struggle significantly.
20:47And so here we provide mean five notably different than, you know, all of these previous scores uh we see here.
20:52So, thanks for for taking the time to talk with us today.
20:56Thank you.