Ari Autoresearch
Held-out reward rose from 0.118 to 0.629 across 9 training runs
Best so far
Not evaluated
Failed launch
Ari Autoresearch
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Held-out eval reward
1
2
3
4
5
6
7
8
9
Training run
Baseline 0.118
Baseline
Move list
Variance filter
SFT→GRPO
Longer SFT→GRPO
SFT stage
SFT stage
SFT, cut off
Launch failed