briscoooe commited on
Commit
65c605b
·
verified ·
1 Parent(s): 62fd5e6

card: final results from the 2026-09-24 run

Browse files
Files changed (1) hide show
  1. README.md +17 -8
README.md CHANGED
@@ -15,7 +15,7 @@ A 33.6M-parameter network that guesses how many moves a Rubik's Cube is from
15
  solved, plus a search that uses the guess to find the moves. It was trained from
16
  nothing: no solver, no example solutions, no human data.
17
 
18
- - Solves **96%** of fully scrambled cubes (25 cubes, 20-move scrambles)
19
  - Solves **every** cube scrambled with 10 moves or fewer
20
  - Built as a control for [RubiksBench](https://brianbriscoe.dev/lab/rubiksbench),
21
  where language models must write the whole answer in one reply. On the same
@@ -80,13 +80,22 @@ full scrambles from 16% to 96%.
80
  Canonical RubiksBench scrambles (the same cubes every language model on the
81
  board is tested on), 25 cubes per scramble length:
82
 
83
- | Scramble length | Solved (beam 1024) |
84
- |---|---|
85
- | 3 | 100% |
86
- | 6 | 100% |
87
- | 10 | 100% |
88
- | 15 | 100% |
89
- | 20 | 96% |
 
 
 
 
 
 
 
 
 
90
 
91
  Past about 15 moves, every scramble is fully mixed, so 25, 50 or 100 moves are
92
  no harder than 20. A full scramble takes about 80 to 90 seconds on one RTX A6000.
 
15
  solved, plus a search that uses the guess to find the moves. It was trained from
16
  nothing: no solver, no example solutions, no human data.
17
 
18
+ - Solves **95%** of fully scrambled cubes (142 of 150, scrambles of 20 to 100 moves)
19
  - Solves **every** cube scrambled with 10 moves or fewer
20
  - Built as a control for [RubiksBench](https://brianbriscoe.dev/lab/rubiksbench),
21
  where language models must write the whole answer in one reply. On the same
 
80
  Canonical RubiksBench scrambles (the same cubes every language model on the
81
  board is tested on), 25 cubes per scramble length:
82
 
83
+ | Scramble length | Solved (beam 1024) | Average answer |
84
+ |---|---|---|
85
+ | 3 | 25/25 | 3.0 moves |
86
+ | 6 | 25/25 | 6.0 |
87
+ | 10 | 25/25 | 10.0 |
88
+ | 15 | 25/25 | 20.4 |
89
+ | 20 | 24/25 | 22.0 |
90
+ | 25 | 24/25 | 21.9 |
91
+ | 30 | 25/25 | 21.6 |
92
+ | 40 | 23/25 | 23.0 |
93
+ | 50 | 23/25 | 21.9 |
94
+ | 100 | 23/25 | 21.5 |
95
+ | **all** | **242/250 (96.8%)** | |
96
+
97
+ Measured 2026-09-24 on one RTX A6000. Every cube and answer is replayable at
98
+ https://brianbriscoe.dev/lab/tiny-cube-solver.
99
 
100
  Past about 15 moves, every scramble is fully mixed, so 25, 50 or 100 moves are
101
  no harder than 20. A full scramble takes about 80 to 90 seconds on one RTX A6000.