Online Self-Weighted Fine-Tuning
尚未评估计划受阻研究发现claim-rollout-ablation-k

In rollout count ablation on Qwen3-0.6B-Base, increasing K from 1 to 2 yields a significant performance jump (from 15.15% to 20.20% on GPQA-Diamond, and from 6.67% to 10.00% on AIME 25 Pass@16), while further increasing K to 4 or 8 yields negligible marginal benefits.

来源:source-paper:Table 2, Section 4.3, page 6

报告指标与观测值

pass_at_16

m-osw-k1-aime24-p16

论文报告 6.67 percentage_points

实际观测 — percentage_points

pass_at_16

m-osw-k1-aime25-p16

论文报告 6.67 percentage_points

实际观测 — percentage_points

pass_at_1

m-osw-k1-math500-p1

论文报告 22.8 percentage_points

实际观测 — percentage_points

pass_at_1

m-osw-k1-gpqad-p1

论文报告 15.15 percentage_points

实际观测 — percentage_points

pass_at_16

m-osw-k2-aime24-p16

论文报告 6.67 percentage_points

实际观测 — percentage_points

pass_at_16

m-osw-k2-aime25-p16

论文报告 10 percentage_points

实际观测 — percentage_points

pass_at_1

m-osw-k2-math500-p1

论文报告 24.2 percentage_points

实际观测 — percentage_points

pass_at_1

m-osw-k2-gpqad-p1

论文报告 20.2 percentage_points

实际观测 — percentage_points

pass_at_1

m-osw-k4-gpqad-p1

论文报告 20.71 percentage_points

实际观测 — percentage_points

pass_at_1

m-osw-k8-gpqad-p1

论文报告 20.71 percentage_points

实际观测 — percentage_points

Assessment(0)

这条 Claim 暂无已发布的不可变 Assessment。

关联运行(0)

这条 Claim 暂无关联执行。