When Should Code Models Switch from NTP Pretraining to RL?

Researcher Brian Ryhuang proposes a PhD-worthy scaling law question on code data: training repos with next-token prediction optimizes perplexity but overfits, while RL can absorb unlimited compute, raising the question of when code models should switch from NTP to RL.

2026-10-15 ~ 2026-10-15 · 2 related posts

• Open scaling-law questions: NTP on a code repo vs RL on its tasks — which is more compute-efficient? — brianryhuang · 2026-10-15
• When Should Coding Models Switch from NTP to RL? NTP Overfits Fast, RL Scales — brianryhuang · 2026-10-15