LW - Getting 50% (SoTA) on ARC-AGI with GPT-4o by ryan greenblatt

Link to original article

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Getting 50% (SoTA) on ARC-AGI with GPT-4o, published by ryan greenblatt on June 17, 2024 on LessWrong.
I recently got to 50%[1] accuracy on the public test set for ARC-AGI by having GPT-4o generate a huge number of Python implementations of the transformation rule (around 8,000 per problem) and then selecting among these implementations based on correctness of the Python programs on the examples (if this is confusing, go here)[2]. I use a variety of additional approaches and tweaks which overall substantially improve the performance of my method relative to just sampling 8,000 programs.
[This post is on a pretty different topic than the usual posts I make about AI safety.]
The additional approaches and tweaks are:
I use few-shot prompts which perform meticulous step-by-step reasoning.
I have GPT-4o try to revise some of the implementations after seeing what they actually output on the provided examples.
I do some feature engineering, providing the model with considerably better grid representations than the naive approach of just providing images. (See below for details on what a "grid" in ARC-AGI is.)
I used specialized few-shot prompts for the two main buckets of ARC-AGI problems (cases where the grid size changes vs doesn't).
The prior state of the art on this dataset was 34% accuracy, so this is a significant improvement.[3]
On a held-out subset of the train set, where humans get 85% accuracy, my solution gets 72% accuracy.[4] (The train set is significantly easier than the test set as noted here.)
Additional increases of runtime compute would further improve performance (and there are clear scaling laws), but this is left as an exercise to the reader.
In this post:
I describe my method;
I analyze what limits its performance and make predictions about what is needed to reach human performance;
I comment on what it means for claims that François Chollet makes about LLMs. Given that current LLMs can perform decently well on ARC-AGI, do claims like "LLMs like Gemini or ChatGPT [don't work] because they're basically frozen at inference time. They're not actually learning anything." make sense? (This quote is from here.)
Thanks to Fabien Roger and Buck Shlegeris for a bit of help with this project and with writing this post.
What is ARC-AGI?
ARC-AGI is a dataset built to evaluate the general reasoning abilities of AIs. It consists of visual problems like the below, where there are input-output examples which are grids of colored cells. The task is to guess the transformation from input to output and then fill out the missing grid. Here is an example from the tutorial:
This one is easy, and it's easy to get GPT-4o to solve it. But the tasks from the public test set are much harder; they're often non-trivial for (typical) humans. There is a reported MTurk human baseline for the train distribution of 85%, but no human baseline for the public test set which is known to be significantly more difficult.
Here are representative problems from the test set[5], and whether my GPT-4o-based solution gets them correct or not.
Problem 1:
Problem 2:
Problem 3:
My method
The main idea behind my solution is very simple: 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). I show GPT-4o the problem as images and in various ascii representations.
My approach is similar in spirit to the approach applied in AlphaCode in which a model generates millions of completions attempting to solve a programming problem and then aggregates over them to determine what to submit.
Actually getting to 50% with this main idea took me about 6 days of work. This work includes construct...