This has been a super-fun project. Can modern LLMs/LRMs learnto play simple video-games from scratch? Can we outperform existing models of these tasks when predicting fMRI data?
In the words of @botoscsabi.bsky.social , "the short answer is yes and yes".
If you want the long answer, see below :-)
Laurence Hunt
Jeee 🐦⬛
I am very proud of our joint effort with @sreejan.bsky.social on the project "Reason to Play"
LRMs show human-like rule discovery, and their hidden states predict human brain activity during gameplay 10x better than previous methods
Interactive demo + paper:
botcs.github.io/reason-to-pl...
32 fMRI-scanned humans and 8 frontier open weight LLMs play ARC-AGI like games with no rules given. The reasoning models match the human learning trajectories and their hidden states predict human bra...
