A quick comment: the o3 and o3-mini announcements each have two significantly different scores, one ⇐ 10%, the other >= 25%. Our own eval of o3-mini (high) got a score of 11% (it’s on Epoch’s Benchmarking Hub). We don’t actually know what the higher scores mean, could be some combination of extreme compute, tool use, scaffolding, majority vote, etc., but we’re pretty sure there is no publicly accessible way to get that level of performance out of the model, and certainly not performance capable of “crushing IMO problems.”
I do have the reasoning traces from the high-scoring o3-mini run. They’re extremely long, and one of the ways it leverages the higher resources is to engage in an internal dialogue where it does a pretty good job of catching its own errors/hallucinations and backtracking until it finds a path to a solution it’s confident in. I’m still writing up my analysis of the traces and surveying the authors for their opinions on the traces, and will also update e.g. my IMO predictions with what I’ve learned.
Thanks a lot for the answer, I put in an edit linking to it. I think it’s a very interesting update that the models get significantly better at catching and correcting their mistakes in OpenAI’s scaffold with longer inference time. I am surprised by this, given how much it feels like the models can’t distinguish its plausible fake reasoning from good proofs at all. But I assume there is still a small signal in the right direction, and that can be amplified if the model think the question through a lot of times (and does something like a majority voting within its chain of thought?). I think this is an interesting update towards the viability of inference time scaling.
I think many of my other points still stand however: I still don’t know how capable I should expect the internally scaffolded model to be given that it got 32% on FrontierMath, and I would much rather have them report results on the IMO or a similar competition, than on a benchmark I can’t see and whose difficulty I can’t easily assess.
Yes, the privacy constraints make the implications of these improvements less legible to the public. We have multiple plans for how to disseminate info within this constraint, such as publishing author survey comments regarding the reasoning traces and our competition at the end of the month to establish a sort of human baseline.
Still, I don’t know that the privacy of FrontierMath is worth all the roundabout efforts we must engage in to explain it. For future projects, I would be interested in other approaches to balancing preventing models from training on public discussion of problems vs being able to clearly show the world what the models are tackling. Maybe it would be feasible to do IMO-style releases? “Here’s 30 new problems we collected this month. We will immediately test all the models and then make the problems public.”
I like the idea of IMO-style releases, always collecting new problems, testing the AIs on them, then releasing to the public. What do you think, how important it is to only have problems with numerical solutions? If you can test the AIs on problems with proofs, then there are already many competitions that regularly release high-quality problems. (I’m shilling KöMaL again as one that’s especially close to my heart, but there are many good monthly competitions around the world.) I think if we instruct the AI to present its solution in one page at the end, then it’s not that hard to get an experience competition grader to read the solution and give it scores according to the normal competitions scores, so the result won’t be much less objective than if it was only numerical solutions. If you want to stick to problems with numerical solutions, I’m worried that you will have a hard time regularly assembling high-quality numerical problems again and again, and even if the problems are released publicly, people will have a harder time evaluating them than if they actually came from a competition where we can compare to the natural human baseline of the competing students.
A quick comment: the o3 and o3-mini announcements each have two significantly different scores, one ⇐ 10%, the other >= 25%. Our own eval of o3-mini (high) got a score of 11% (it’s on Epoch’s Benchmarking Hub). We don’t actually know what the higher scores mean, could be some combination of extreme compute, tool use, scaffolding, majority vote, etc., but we’re pretty sure there is no publicly accessible way to get that level of performance out of the model, and certainly not performance capable of “crushing IMO problems.”
I do have the reasoning traces from the high-scoring o3-mini run. They’re extremely long, and one of the ways it leverages the higher resources is to engage in an internal dialogue where it does a pretty good job of catching its own errors/hallucinations and backtracking until it finds a path to a solution it’s confident in. I’m still writing up my analysis of the traces and surveying the authors for their opinions on the traces, and will also update e.g. my IMO predictions with what I’ve learned.
Thanks a lot for the answer, I put in an edit linking to it. I think it’s a very interesting update that the models get significantly better at catching and correcting their mistakes in OpenAI’s scaffold with longer inference time. I am surprised by this, given how much it feels like the models can’t distinguish its plausible fake reasoning from good proofs at all. But I assume there is still a small signal in the right direction, and that can be amplified if the model think the question through a lot of times (and does something like a majority voting within its chain of thought?). I think this is an interesting update towards the viability of inference time scaling.
I think many of my other points still stand however: I still don’t know how capable I should expect the internally scaffolded model to be given that it got 32% on FrontierMath, and I would much rather have them report results on the IMO or a similar competition, than on a benchmark I can’t see and whose difficulty I can’t easily assess.
Yes, the privacy constraints make the implications of these improvements less legible to the public. We have multiple plans for how to disseminate info within this constraint, such as publishing author survey comments regarding the reasoning traces and our competition at the end of the month to establish a sort of human baseline.
Still, I don’t know that the privacy of FrontierMath is worth all the roundabout efforts we must engage in to explain it. For future projects, I would be interested in other approaches to balancing preventing models from training on public discussion of problems vs being able to clearly show the world what the models are tackling. Maybe it would be feasible to do IMO-style releases? “Here’s 30 new problems we collected this month. We will immediately test all the models and then make the problems public.”
I like the idea of IMO-style releases, always collecting new problems, testing the AIs on them, then releasing to the public. What do you think, how important it is to only have problems with numerical solutions? If you can test the AIs on problems with proofs, then there are already many competitions that regularly release high-quality problems. (I’m shilling KöMaL again as one that’s especially close to my heart, but there are many good monthly competitions around the world.) I think if we instruct the AI to present its solution in one page at the end, then it’s not that hard to get an experience competition grader to read the solution and give it scores according to the normal competitions scores, so the result won’t be much less objective than if it was only numerical solutions. If you want to stick to problems with numerical solutions, I’m worried that you will have a hard time regularly assembling high-quality numerical problems again and again, and even if the problems are released publicly, people will have a harder time evaluating them than if they actually came from a competition where we can compare to the natural human baseline of the competing students.