just a quick update: we managed to replicate these claims Guan's system hits 25% on ARC-AGI 1 with 50 GPU hours I still couldn't audit the code personally though, but, unless cheating somehow, this approach seems to generalize ARC-AGI instances with relatively little compute
Guan Wang
Guan Wang21.7.2025
🚀Introducing Hierarchical Reasoning Model🧠🤖 Inspired by brain's hierarchical processing, HRM delivers unprecedented reasoning power on complex tasks like ARC-AGI and expert-level Sudoku using just 1k examples, no pretraining or CoT! Unlock next AI breakthrough with neuroscience. 🌟 📄Paper: 💻Code:
and to be clear, until I have time to dig deep into the paper, which I won't until at least Bend2's release, this is still LK-99 v2 to me all that we know is that if we run his repo with this amount of compute, it shows the claimed results. plenty of other ways to cheat
129,03K