LLM4AD
An open-source Python platform that uses large language models to automatically design and optimize algorithms, with unified interfaces across methods, tasks and models plus a graphical interface for running experiments.
LLM4AD is an open-source Python platform for automatic algorithm design using large language models. It provides unified interfaces across search methods, tasks and language models, so a user can combine different algorithm-design methods and LLMs without rewriting integration code for each combination.
It includes documentation and worked examples to help users test, build, deploy and research their own algorithm-design applications, along with a graphical interface for selecting methods and tasks and viewing convergence and the current best algorithm. Runs can be resumed, evaluation can run across multiple processes with a timeout and process isolation for safety, and results can be logged locally or to Weights & Biases and TensorBoard. Although developed for optimization tasks, the authors describe it as general enough for other areas such as machine learning, scientific discovery, game theory and engineering design.
The project reports results including new best-known solutions on large-scale vehicle routing benchmarks and a circle-packing world record produced by one of its methods.