Show HN: I reimplemented TRM – trains on Sudoku 6×6 in minutes on an A10
github.comHi folks!
Tiny Recursive Models reach impressive results on ARC AGI. I implemented a version from scratch, with ease of experimentation in mind:
- cleaner config: hydra, uv, lightning - smaller datasets for faster iteration (Sudoku 6x6 and 9x9) - introduction, in-code video
All important implementation details have been carefully kept. The results of the paper are reproducible (Sudoku Extreme, Maze Hard).
Feedback/contributions welcome. Is it possible to measure which rules are more useful in each case? When my older daughter was a kid, she used to play Sudoku 6x6. I notices she was trying the rules in a weird order, that was different from mine that played 9x9. After playing a few 6x6 I notice some of the rules were more useful in 6x6 than in 9x9 and she was right to try them in a different order.