Reward over 48 steps, three seeds
Mean reward per step for three seeds of the same GRPO run on Qwen3-0.6B and GSM8K. Each step jumps around, and the trend goes up for all three seeds.
0.5
0.6
0.7
0.8
0.9
0
12
24
36
47
step
mean reward of the step
seed 1234
seed 2
seed 3
Gain, last 8 steps against the first 8: +0.122, +0.083, +0.151.