Forty-eight GRPO steps on a GB10
Raw reward and the share of samples cut at the 1,024-token limit for each of 48 plain GRPO steps of Miles with Qwen3-0.6B on GSM8K on one NVIDIA GB10 with the default seed. The mean reward of the first 8 steps is 0.636 and of the last 8 is 0.758. The cut share falls from a mean of 0.33 to 0.08 over the same steps. Dashed lines show the raw rewards of runs with seeds 2 and 3, which rise from means of 0.653 to 0.736 and 0.614 to 0.766.
0
0.25
0.5
0.75
1
0
8
16
24
32
40
47
GRPO step
fraction
raw reward
share of samples cut at 1,024 tokens
seeds 2 and 3