Online Self-Weighted Fine-Tuning
Not assessedPlan blockedFindingclaim-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.

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

Reported and observed measurements

pass_at_16

m-osw-k1-aime24-p16

Reported 6.67 percentage_points

Observed — percentage_points

pass_at_16

m-osw-k1-aime25-p16

Reported 6.67 percentage_points

Observed — percentage_points

pass_at_1

m-osw-k1-math500-p1

Reported 22.8 percentage_points

Observed — percentage_points

pass_at_1

m-osw-k1-gpqad-p1

Reported 15.15 percentage_points

Observed — percentage_points

pass_at_16

m-osw-k2-aime24-p16

Reported 6.67 percentage_points

Observed — percentage_points

pass_at_16

m-osw-k2-aime25-p16

Reported 10 percentage_points

Observed — percentage_points

pass_at_1

m-osw-k2-math500-p1

Reported 24.2 percentage_points

Observed — percentage_points

pass_at_1

m-osw-k2-gpqad-p1

Reported 20.2 percentage_points

Observed — percentage_points

pass_at_1

m-osw-k4-gpqad-p1

Reported 20.71 percentage_points

Observed — percentage_points

pass_at_1

m-osw-k8-gpqad-p1

Reported 20.71 percentage_points

Observed — percentage_points

Assessments (0)

No immutable Assessment has been published for this Claim yet.

Runs (0)

No execution has been linked to this Claim yet.