Number of Partitions

Qwen3.5-9B · 400 held-out puzzles (rating 800–1199) × 4 samples · best calibration checkpoint

Number of Partitions0102030405060Held-out mean reward1248Number of partitionsGRPO 34.2Base Qwen3.5-9B 7.434.3%50.1%51.9%49.7%33.9%44.0%46.3%48.1%
SporkIts teachers (task-weighted)
View data table
SeriesNumber of partitionsHeld-out mean rewardInterval / statistics
Spork34.3%34.3
Spork50.1%50.1
Spork51.9%51.9
Spork49.7%49.7
Its teachers (task-weighted)33.9%33.9
Its teachers (task-weighted)44.0%44
Its teachers (task-weighted)46.3%46.3
Its teachers (task-weighted)48.1%48.1