Four training stages

Each stage starts from the best checkpoint of the one before.

01 · Imitation
Synthetic SFT
Learn the tools, gather the evidence and call for answer generation
→
02 · Imitation
Best-teacher SFT
Learn which research path fits which question
→
03 · RL
Offline GRPO
Learn what separates good research from weak
→
04 · RL
Online GRPO
Learn from its own mistakes

Figure 3: Two imitation stages, then two reinforcement-learning stages, first on the teachers' traces and then on the model's own attempts.