I Isaac Liao personal blog

ARC-AGI without pretraining

by Isaac Liao, Albert Gu iliao2345.github.io published 2025-03-05

We solve 20% of ARC-AGI-1 with no pretraining by minimizing the description length during inference time, with the only data involved being the target puzzle itself.

Points

351

HN comments

121

Submissions

1

Cited by

0 commenters

AI & Machine LearningMath & Algorithms2025

Related posts from the same era

Statement on US government directive to suspend access to Fable 5 and Mythos 5

Anthropic20263,158 points2,314 commentscited by 11 details

GPT-4o

OpenAI Blog20243,138 points2,366 commentscited by 16 details

Astra can read and write weights to improve Qwen

RRikiya Takehi20262,960 points899 comments details

HackerTube: YouTube built from videos recommended by HN

karpathy20262,889 points641 comments details

Claude Fable 5

Anthropic20262,626 points2,159 commentscited by 10 details

Claude Code is steganographically marking requests

Thereallo20262,445 points750 commentscited by 3 details