#Rl

DeepSeek-R1 Architecture: How Large-Scale Reinforcement Learning Solved Reasoning

A deep dive into DeepSeek-R1's zero-SFT pure RL training paradigm, cold-start data generation, and Multi-Head Latent Attention that...

The Rise of Reasoning Models: How Test-Time Compute (o1, o3, R1) Replaces Pure Scaling

Why pre-training scaling laws hit a wall of diminishing returns, and how test-time compute search graphs unlocked new...