LLMs for Engineering: Teaching Models to Design High Powered Rockets
a year ago
- #LLMs
- #Rocketry Design
- #Reinforcement Learning
- LLMs are explored for high-powered rocketry design using RocketBench, a benchmark connecting LLMs to rocket simulations.
- Two design tasks tested: target altitude optimization and precision landing challenges.
- State-of-the-art LLMs show strong baseline engineering knowledge but struggle to iterate on designs with simulation results, plateauing below human performance.
- A 7B parameter model enhanced with reinforcement learning (RL) outperforms both SoTA foundation models and human experts.
- RL-trained LLMs can serve as effective tools for complex engineering optimization, extending beyond software development.