Scaling to Long Horizons — Ross Taylor & Chengxi Taylor, General Reasoning

Uh, so this talk is called "Scaling to Long Horizons." My name is Ross, uh, I'm the CEO of GR, we're a London-based reinforcement learning company. Before GR, I was the reasoning lead at Meta AI, working on Llamas, Galactica, lots of other models back in the day.

I'm joined by Chengxi, uh, co-founder and president of GR, and yeah, today we're going to talk about algorithms, environments, compute, all the things you need to do to get agents scaling, uh, to kind of longer tasks. So we're going to have two parts to this talk today.

I'm going to first of all start with a personal perspective about, you know, the early days, the golden age of language modeling, between like maybe 2020 and 2023. I'll talk about, like I said, all those models and some of our early reinforcement learning efforts for LLMs.

And then Chengxi is going to talk about, you know, what's ahead, you know, what the next frontiers. And yeah, that's going to be a really interesting talk with a lot of alpha, so I'd encourage you to stick around for that.

So, the journey so far. So my journey started here. Um, so this was the Papers with Code team. I'm sure many of you used Papers with Code back in the day. So we were a London-based startup, 2019. Uh, we were acquired by Meta later that year, and then we had a crazy transition within Meta to do research.

Um, so we did, like I said, Galactica. Then after ChatGPT came out, we started the post-training for LLaMA 2, LLaMA 3. So all the great work you, uh, saw there was folks in this room, and lots of other interesting stuff that never got published as well, uh, reasoning LLaMA and lots of other things.

Um, so yeah, this, this small team. I like to think like the open weight kind of revolution started in this room. And, you know, it really like hit home this idea to me that kind of small focused teams, like even in like the age of scaling, can do amazing things if people are aligned.

Now for me, uh, things got particularly crazy in 2022. Uh, so let me tell you a story. Um, the media perception is that ChatGPT came out of nowhere, you know, shocked the world, and that's how kind of the modern AI wave started.

But, you know, I have a different personal perspective on this because two weeks before ChatGPT came along, there was another language model called Galactica. So let's talk about Galactica. Galactica and, you know, ChatGPT, you know, they were both, you know, in some respects quite similar.

They were both based on pretty good base models. Galactica itself was a base model, and then ChatGPT was based on GPT-3.5. But there was a clear difference in outcomes. So Galactica at the time shipped with this like base model demo.

And as you guys know now, like base models, they come with a lot of quirks. You know, they hallucinate. You prompt them to do, you know, silly things, they will do silly things. Whereas ChatGPT wasn't just a base model, but had this like crucial reinforcement learning from human feedback pipeline.

And this was the key thing that made LLMs like really products for the first time. So I like to think in a way, kind of way, this is like the first like kind of natural experiment showing you that kind of RL like provides value,right?

Uh, and to my misfortune, it was like a very personal like kind of a natural experiment. You know, Galactica blew up. Uh, but that's like a good like kind of lesson there. A good base model is not enough.

So I took that lesson quite early on. So like I said, RLHF made LLMs products. They were the thing that kind of made LLMs cross the Rubicon into something that wasn't just a toy, but used by now billions of people.

But you didn't have to wait until ChatGPT to see this. Like even at the time, like InstructGPT in 2022 had this pretty stunning results. Like a 1 billion parameter model with RLHF was outperforming 175 billion models. So two orders of magnitude fewer parameters, but getting better results.

So that was astonishing. So if you were paying attention closely, you know, maybe we should have been as well, but we were focused on a, you know, a bloody base model, which is hard work in 2022. But that shows you how, you know, important even, you know, basic RL is.

And the Galactica demo itself, I mean, it set off a storm. So ancient history now, but we put out a demo. We let people play around with it. Um, we thought it was kind of cool. And at the time, people got scared.

So it was like, you know, like I said, you prompt it on like a research paper on Dyson spheres or a report on the benefits of eating crushed glass, and people were like, "Oh my God." Um, so that was the state of things in 2022.

And yeah, I'll be honest, the Meta association didn't, you know, help us either. Um, and yeah, the tragic story in a way was a lot of the, you know, novel work was maybe overshadowed. But, you know, the paradox of this whole thing is that at the time, Galactica was actually a bloody good model.

Um, like it outperformed PaLM, Chinchilla, GPT-3.5 with a lot less compute in scientific domains. It will stay at the art. So again, that also shows you how powerful RL is. You can have a SOTA-based model, but that is not enough.

Um, so here you see on like a map, so it was kind of beating Chinchilla, uh, kind of latex equations, you know, science, you know, it was getting around 68% compared to GPT-3.5, 49%, so crushing there. And chain of thought as well.

So, you know, PaLM at the time, which was a Google brain model, 540 billion. You know, it's 30 billion in Galactica, was getting, you know, 36 versus 19%. So double the performance, order of magnitude less results. So that again reinforces base good base models, not enough.

But it introduced some key ideas, which I think are very important. I mean, Galactica was the first LLM to really crack data efficiency. 105 billion, uh, token corpus compared to, you know, trillion, uh, tokens in Chinchilla. And it was really contrarian at the time because, you know, at the time everyone was like, "Okay, we just need more tokens."

And Galactica said, "No, high quality, you know, curated data sets really matter." And that was a real driver of those results you just saw. And it was also like the first major LLM to really crack multi-epoch training. It sounds ridiculous now, but at the time the consensus was you don't do more than an epoch.

Um, but this kind of rule of thumb, you may have heard of it, like four epochs of repeated data. That was formalized later, but Galactica was the first like real empirical result for that. Now, perhaps more importantly, there was this idea of thinking tokens.

And some of you might remember this, but it was like really quite buried within the paper. So around this time, there were like different ideas for reasoning. There was chain of thought, which was one idea, which is where you prompt for like kind of like the steps.

There were scratch pads where you just like put like very light numerical kind of intermediate steps. But Galactica was really this first idea which said, "No, this is an internal working memory process. This is an internal thinking. It should be inside these tags, and you should spend inference compute before you get to an answer."

Right? So these are all like quite like prescient ideas. But, you know, Galactica came and went, blew up, and then, you know, we were kind of tasked as a team to kind of spin up the post-training effort for Llama.

But I had like a personal obsession, which was like reasoning. And I had like a really simple idea at the time, which is, what if we applied kind of reinforcement learning pressure to this like kind of thinking tags as work?

Like what if we just optimized the thing in between the thinking? And if that sounds familiar, then this is like kind of what DeepSeek at 01 ended up doing two years later. But there's a key difference. So at the time, we only had LLaMA 2 base models, terrible mathematics corpus, terrible results on math.

And, you know, the context window, you know, we're all like, you know, very context rich now, 1 million tokens. You know, back in the day, it was, uh, 4,000, which wasn't too fun. But we still had a recipe at the time, and this was unpublished, but it was really good within Meta.

So our recipe was this. Number one, continue pre-training on LLaMA 2 towards mathematics and science data. So that's the first thing. LLaMA 2, ship math corpus, so let's fix that. Number two, PPO with verifiable rewards. But notice, this isn't GRPO,right?

So we had the time, had a strong outcome reward model to initialize the value model. That was a key thing. Lots of data on value models at the time. And internally at the time, this kind of recipe led to state-of-the-art results on math and reasoning.

So we were kind of like, "Wow, this is like really shows the power of having theright objective." But the really fascinating thing is like we had great results, but we didn't have like inference time scaling. We didn't have this reflective behavior that became like the hallmark of R1 and O1, you know, butt weights, you know, backtracking, all this kind of stuff.

So it begs like the question, like why? Why didn't we have that moment? And we got an answer around like two years later. So there's a couple of like things going on here, but essentially better base models were the thing that really got RL cooking.

And when DeepSeek came out, I was kind of shocked at the time. I was like, "Holy shit, we just like tried the same thing. We didn't have this. What's going on here?" And in a weird kind of way, the real lesson was it was just like the bitter lesson, like the most purest form of bitter lesson possible.

Like better base models, more RL compute, bigger context windows, and that's all you need for this kind of emergent behavior. It also like says something like quite important about the sociology of like research because the fact that OpenAI had this model, you know, GPT-4 level model before anyone else, it allowed them to see further,right?

So that's a really interesting point. Like the, the age of scaling means that if you have certain prerequisites in place, you become smarter, you see further, you see more ideas. So a really interesting point. So this was me in 2024.

I was a very sad panda, uh, defeated by ChatGPT in 01. But I wasn't deterred, um, so I wanted to seek the next wave. And I still was like convinced that kind of reasoning hadn't been solved. So we started GR, uh, to take on truly like big tasks.

And with that in mind, we're going to hand over to Chengxi. He's going to talk about what we're kind of thinking about now.

Ross, Ross, Ross. Give me the.

Hi everyone. I'm Chengxi Taylor, co-founder and president General Reasoning. I'm going to share what it takes to scale to long horizon.

First, I want to make it clear. Long horizon task is not just an engineering problem. It is a mindset. If we want to solve humanity's biggest problems, such as cure cancer, solve Millennium's price problem, or going to Mars, we have to be patient.

It takes time. And if we want AI to move us towards that level of impact, we have to think about long horizon.

But here's the first problem. We have a scarce context window. If you take Ferman's last theorem as an example, what it takes for the mathematician was over 10 years' time of reading paper, writing thoughts in the scratch pad, or taking a walk to generate creative ideas.

If we convert to token, that's probably tens of billions or even hundreds of billions. But where we are now, just 1 million token context window. So one solution is use compaction. So what it essentially does is generate a token until the end of the context window, summarize, and then on top of that, generate more tokens.

And the beauty of applying RL in this situation is kind of like kill two birds with one stone. You apply RL to the compaction and also the task. But here's the problem. With long horizon, there are three issues.

The first is the gradient variance scales with the length. And the second is a sparse reward. And you have this, uh, credit assignment problem. And finally, there's also variable length of the trajectory that adds to the problem of optimization.

So to solve this issue, we can apply critics, which is the value model. And value model can reduce the variance and also have a couple of advantages, such as on the trajectory level, it fits compaction very well and also encourage the batch diversity.

And also I'll talk later on bootstrapping, basically get signal before the end of the episode. But the downside for this is that, um, it's more complicated than GRPO and basically you have to train another value model alongside with the policy model.

And there are some tools to help with the context limitations, such as file system tools, which essentially like a scratch pad for AI to write the, uh, reasoning thought. And self-search tools, which allows agent to search over the previous trajectory.

And then you have archive tools. In the case like auto-research, you can build upon your previous result. But we have to be careful. In other scenarios, you don't want AI to cheat by just to grab the previous answer without thinking.

So how good are the current model on those long horizon tasks? In the general reasoning, we construct a benchmark called Kelly Bench, where we're actually featuring the front page of the Financial Times kind of caught us off guard how much the mainstream have interest in this.

So basically what we did is that we allowed the agents to build machine learning models to, um, trade in the football matches over one year horizon. In this case, it's Premier League if you're interested in football. And we're so fascinated by this because there's real money to be made.

And if it was successful, it could make billions. There's a whole industry on sports betting. And unlike things like a cargo competition, this has a real world implication. But here's the result. As you can see, we gave all the frontier models 100K to start.

All of them lost money. Sad. And that captured the public's imagination. Oh, AI is not as great as they thought. And why are models so bad at long horizon? First, I believe now the AI industry is a little bit too biased towards coding and procedure task.

What I mean is that the current task is most formulated like do this and fix that. Normally that limits the solution like one or two. There isn't just too much space for creativity. And second of all, not enough focus on open-ended task.

We're living the real world with a lot of complexity, uncertainty, and that's not fully captured by the current benchmark. And also, there isn't enough simulation of the real world. We're living the world. There are other players. Like in today's conference room, there are other real people who have a different thought, different games than you have in your mind.

That's the complexity that's not fully captured.

And another thing I want to talk about is the long horizon impact on compute. We know that GPUs are scarce and precious resources. And in this case of a long horizon reasoning, you have to be careful about how to optimize your use between training and inference.

And pipeline RL is a quite popular technique nowadays. So basically, it's a trade-off between off policy and the GPU utilization. So traditionally, you let the inference run towards the end and then you start to train the model. But in the case of long horizon, you have to wait until the inference finish.

What the pipeline RL does is that you let the sequences be generated and you start to train the model while there's still more, uh, sequences being generated. And you see this created off policy. But from our experience, normally off policy up to eight steps is okay.

So essentially, we made a trade-off between the off policy and the GPU utilization. But here comes the issue. As we the long horizon indicates, sometimes the inference will take weeks or even more. In that case, inevitably, it will go beyond the constraint of eight steps of off policy.

So your GPU have just sit there idle and wait for it to finish. And if you don't want to wait, as I mentioned before, applying the value model allows you to bootstrap. What it means is that before the end of the episode, you generate expectation.

It's like a dopamine in human brain. And that allows you to train the model. But here's another trade-off. While you utilize the GPU fully, you introduce the value model bias. So there's always a bigger trade-off in those solutions.

And I want to also mention that in the long horizon, infrastructure is important, especially for the environment. And open reward was a product. It's a platform by General Reasoning. If you're interested, you can check it out, openreward.ai. So it's a place where it hosts over 350 environments and with a single API endpoint.

And we use this for our internal RL and also some frontier labs and new labs are using this.

So to summarize both Ross and my speech, it has been a long journey, as long horizon indicate. We as a team have seen the paradigms in AI reasoning, um, pre-training and agents in the past few years. But looking ahead, what makes us really excited is the long horizon.

And it requires us to think, have a new thinking on the algorithm, environments, and compute. There are a lot of challenges and trade-offs, but we find it's really exciting to take on this journey because, as I mentioned in the very beginning, long horizon is not just an engineering problem.

It is a mindset. And if we really are ambitious to solve humanity's biggest problems, this is the journey for everyone. And that's also the mission for General Reasoning. So if you are interested, follow us, General Reasoning. We're a London-based AI research company.

Thank you.