International Conference on Machine Learning

Um, because a lot of people talk about knowledge being quite situated, and what they're meaning in that case is that it's only coherent if you, if you respect the constraints, and sometimes it's not possible to break the constraints.

And, um, another thing you spoke about, and this is also related to your 2024, um, ICML paper, which was, uh, open-endedness is, is, um, I, I think necessary or required for edge.

Yes.

Yes.

Um, I hope that that will persist, in fact, because I think that will lead us to, um, different kinds of creativity from different kinds of constraints being broken.

But then, of course, what you need if you're gonna have that kind of approach is you need a translation layer.

Um, and so this is why, why when we talk about the, um, the definition of open-endedness in the paper that I wrote with Michael, um, uh, at ICML a couple of years ago, um, we talk about the idea of a, an open-ended system to an observer having to produce artifacts that are both novel and learnable.

And it's really that learnable piece which is to this translation point.

And then I think that environment creates like great work.

And it was something about that like ethos and, um like farming that organism that can just like keep innovating is it's is this happened i think this has happened i would say back in the day in other labs right the bell labs and other places so it's like it's really hard to recreate and it does need like an all like a big industry like at least like back in the day there was uh multiple companies that had money to fund this and google definitely had the money to fund it and had the right people at the top to kind of think about the impact of these in an altruistic way in some sense, right? Because it was like publication, right? At some point, Google had like the most number of submissions for NeurIPS or ICML or in the top three.

So it was one of the biggest institutions, including Meta as well and other institutions

as well.

We're trying to find sort of the most glaring issues that we can and that we can tackle with the resources we have.

I want to talk about ICML and what brought you to ICML.

So the paper, for someone who doesn't know, well, first of all, what's the title of the paper? And then the second question is, well, like, what is fine tuning for people who might not know?

Sure.

And I think that's more generally true across algorithms for machine learning.

It could go back to NeurIPS from a decade ago or ICML a decade ago and, you know, or two decades ago.

There's lots of papers there which are, you know, oh, this is a better learning algorithm, doesn't get used, doesn't get adopted.

And why is that? Well, I think often you have weak baselines effectively.

Once we d- once we realize that's what this musing is, a couple major problems arise.

One, there's a growing body of research, including a brand-new paper at ICML this year and a well-known paper from NeurIPS from twenty twenty-three and many others as well as far as I know, that have, uh, established the fact that the reasoning that these reasoning models output can be performative, that it, that it can be unrelated to how it actually gener- how and why it generated the response it did.

And this makes sense because the way you tune a reasoning model to reason is you reward it for giving correct reasoning for questions.

They give it questions where they know the answer and, um, and they have examples of correct reasoning.

Again, it's kind of the key, the Einstein test of, you know, when these things can do stuff with a corpus that's lobotomized, you know, post 1905 or 1911, as the case may be.

And it's a position paper in, uh, ICML, uh, 2026, which Roman probably knows what that means, uh, by Tom Zahavi, and it's called "Position: LLMs Can't Jump".

And there's a famous movie called White Men Can't Jump with Woody Harrelson and Wesley Snipes, and it was about, you know, it was called...

Basically, white men aren't good at basketball and, and, and it was kind of a funny comedy and, and drama, you know, coupled together.

I don't know which ones are, like, necessarily the most popular.

There's, like, the PyTorch conference, the, uh, MLSys, like, ICML, like, all the big AI conferences.

Like, I think those ones are still, like, you know, super important and relevant to go to.

And, like, you do also have to keep in mind that, like, the...