🚀 Phase 6 — SOTA RL & Research Trends

Cutting-edge reinforcement learning methods powering modern AI systems.

Algorithms

• GRPO — Generalized Reinforcement Policy Optimization
• DPO — Direct Preference Optimization
• DreamerV3 — Latent world-model-based imagination training
• Decision Transformer — Sequence modeling for RL trajectories
• GFlowNets — Sampling trajectories proportional to reward
• Diffusion Policies — Diffusion-based action generation
• RLHF — Reward modeling + policy optimization with human feedback

Mini Projects

• PPO → GRPO comparison experiment
• DreamerV3 pixel-control demo
• DPO fine-tuning on preference data

📁 Source folder:
06-SOTA-RL