Getting 50% (SoTA) on Arc-AGI with GPT-4o

(ARC Prize co-founder here).

Ryan's work is legitimately interesting and novel "LLM reasoning" research! The core idea:

> get GPT-4o to generate around 8,000 python programs which attempt to implement the transformation, select a program which is right on all the examples (usually there are 3 examples), and then submit the output this function produces when applied to the additional test input(s)

Roughly, he's implemented an outer loop and using 4o to sample reasoning traces/programs from training data and test. Hybrid DL + program synthesis approaches are solutions we'd love to see more of.

A couple important notes:

1. this result is on the public eval set vs private set (ARC Prize $).

2. the current private set SOTA ~35% solution also performed ~50% on the public set. so this new result might be SOTA but hasn't been validated or scrutinized yet.

All said, I do expect verified public set results to flow down to the private set over time. We'll be publishing all the SOTA scores and open source reproductions here once available: https://arcprize.org/leaderboard

EDIT: also, congrats and kudos to Ryan for achieving this and putting the effort in to document and share his approach. we hope to inspire more frontier AI research sharing like this

Also worth nothing that Ryan mentions

> In addition to iterating on the training set, I also did a small amount of iteration on a 100 problem subset of the public test set

and

> it's unfortunate that these sets aren’t IID: it makes iteration harder and more confusing

It’s not unfortunate: generalizing beyond the training distribution is a crucial part of intelligence that ARC is trying to measure! Among other reasons, developing with test-set data is a bad practice in ML because it hides the difficulty this challenge. Even worse, writing about a bunch of tricks that help results on this subset is extending the test-set leakage the blog post's readers. This is why I'm glad the ARC Prize has a truly hidden test set

... and we know that if we really want to nail it we'd better just pay someone else to create 1,000,000 more harder problems for training (without looking at any in test set, of course). i.e. make the training set distribution similar to test set again.

Because the thing we have now is data-hungry. Your brain is pre-trained on other similar challenges as well. What's the point of requiring it to "generalize beyond the training distribution" with so few samples?

Really, I thought LLMs ended this "can we pretrain on in-house prepared private data for ILSVRC" flame war already.

You seem to misunderstand why generalization is important for making claims about intelligent systems. To illustrate this, we could really easily design a system that encodes all the test set questions and their answers, puts them in an enormous hash table, and looks up the correct answer to each challenge when presented with it. This could probably score 100% on ARC if given the entire test set. Would you call this AGI? What if I put it through a transformer as a hashing function?

The mainstream attention LLMs have garnered has added a bunch of noise to the way we talk about machine learning systems, and unfortunately the companies releasing them are partially to blame for this. That doesn't mean we should change the definition of success for various benchmarks to better suit lay misunderstandings of how this all works

First, LLMs are not AGI. Never will be. Can we talk now?

> if given the entire test set.

I don't want the entire test set. Or any single one in the test set.

The problem here is ARC challenge deliberately give a training set with different distribution than both the public and the private test set. It's like having only 1+1=2, 3+5=8, 9+9=18 in training set and then 1+9=10, 5*5=25, 16/2=8, (0!+0!+0!+0!)!=24 in test set.

I can see the argument of "giving the easy problems as demonstration of rules and then with 'intelligence' [1] you should be able to get harder ones (i.e. a different distribution)", but I don't believe it's a good way to benchmark current methods, mainly because there are shortcuts. Like I can teach my kids how factorial works and ! means factorial, instead of teaching them how addition works only and make them figure out how multiplication, division and factorial works and what's the notation.

[1] Whatever that means.

Okay I admit I'm confused and think I probably missed a crucial thing here. You're saying the publicly available problem set isn't indicative of the distribution of the test set? If so, I can see why you object to that. Still, it's potentially possible that the test's intention is to demonstrate something like progressive integration of compositionality given an abstract model. A lot of machine learning systems can do well as long as they've seen an example similar to the problem they've been presented, but can't do things like respond to a situation that presents them with a novel composition of two abstractions they seem to have already learned in the way a human can trivially.

Like only having [1+1=2, 4+5=9, 2+10=12] in the training set and [2*5=10, 3/4=.75, 2^8=256] in the test set would be bad, but something like [1+1=2, 3+4*2=11, 5*3=15, 2*7=14, 1+3/5=1.8, 3^3=27] vs [2+4*3=14, 3+3^2+4=16, 2*3/4+2^3/2^4=2] might not be, depending on what they're trying to test

Compositionality of information, especially of abstractions (like rules or models of a phenomenon), is a key criterion in a lot of people's attempts to operationally define "intelligence" (which I agree is overall a nebulous and overloaded concept, but if we're going to make claims about it we need at least a working definition for any particular test we're doing) I could see that meaning that the test set problems need to be "harder" in the sense that presenting compositions of rules in training doesn't preclude memorizing the combinations. But this is just a guess, I'm not involved in ARC and don't know, obviously*