This research transforms LLMs into powerful scientific optimizers through an innovative edited-LLM framework that overcomes traditional prompt-based limitations.

• Introduces bi-level optimization where an LLM proposes solutions while an external simulator provides objective feedback
• Develops a dynamic editing mechanism that augments the LLM's context with previous optimization attempts
• Demonstrates superior performance across engineering problems compared to standard prompt-based methods
• Creates a more robust optimization approach less sensitive to prompt variations

For engineering applications, this framework represents a significant advancement in leveraging AI for complex optimization tasks without requiring specialized domain knowledge in prompt engineering.

Exploiting Edited Large Language Models as General Scientific Optimizers