New training paradigm: instead of just predicting tokens, models reason about each prediction using RL The model thinks through context, considers alternatives, then makes a prediction. 14B model matches 32B baseline, though training costs are significantly higher.

→ View original post on X — @askalphaxiv