Researchers built an AI system to solve 36 complex math problems

Large language models often struggle to create custom code for complex math problems on their own. Scientists from Nanyang Technological University and their partners recently introduced a new framework called LACE to address this issue.

LACE stands for LLM-driven Algorithm Construction via Complementary Evolution. It breaks the process of building a computer program into smaller parts, such as defining input and output requirements and selecting from a library of available tools.

Instead of asking an AI to write a whole program at once, the system uses an iterative process. It generates several candidate algorithms and refines them through a technique called complementary evolution. This ensures the tools work well together to solve diverse problems under strict time limits.

In testing, the team used LACE on 36 classical problems in sectors like logistics, manufacturing, and energy management. The system earned an average score of 0.945. For comparison, traditional prompt-based methods scored only 0.571, while other language model frameworks reached 0.870.

When the researchers threw four completely new problems at the system, LACE solved them successfully. Other models failed to produce any working code for these specific challenges.

Imagine you have a series of complex warehouse routing tasks. Instead of manual trial and error, this framework creates a tailored set of instructions specifically designed for those constraints.

While this framework improves how models handle technical algorithm design, it does not replace the need for the human-led design of the framework itself. The system was tested on specific benchmark problems and its performance on undefined, real-world data outside of these settings remains a question for future research.