Share of rubric items satisfied (%) on 100 held-out validation questions, one full agent run per question. Solid: Essential items; dashed: All rubric items. Hover a point for its value.
Figure 5: Each stage starts from the previous stage's chosen checkpoint (circled). Best-teacher SFT peaks at epoch 3, which starts the RL stages. Online GRPO shows the 26-step run whose final checkpoint is Samaya Sterling.