4B RL-Trained Model Outperforms GPT-6 Astra in Physics Simulation

A 4B parameter model trained via Reinforcement Learning outperformed GPT-6 Astra and other frontier models in a high-wind physics simulator. While frontier models struggled with the 0.2s decision latency requirement, the specialized RL model maintained flight stability in 33 m/s winds.