Distillation, curricula and tool-augmented reasoning
Transfer reasoning from a strong model to a small one by supervised distillation of verified traces, rejection sampling, curriculum and data selection, the role of a short SFT stage before RL, and giving models calculators, code interpreters and search; with failure modes and evaluation cautions.
18 min·Verified 2026-11Pro+50 XP
RL (GRPO) improves a model by trial and reward, which is expensive. Two cheaper routes matter just as much in practice: learn from a better model, and give the model tools.
Pro content
Keep reading with Inference Academy Pro
Unlock the rest of this lesson, plus every advanced lesson, hands-on lab, the larger playground models and completion certificates.