Reinforcement learning

DeepSeek-R1 made reasoning-model training a more inspectable question

The January 2025 paper and model releases offered concrete methods to investigate without making evaluation optional.