optima-cityu/llm4ad_next

LLM4AD_Next is a research platform that uses AI language models to automatically write and improve algorithms. The idea is to combine two techniques: asking a language model to generate candidate solutions to a problem, then running those solutions through a process inspired by biological evolution, where the better-performing ones are kept and further refined over many cycles. The goal is to reduce the manual work that researchers and engineers normally put into designing algorithms by hand. The platform provides a web interface and a command-line tool. A first-time user can start a conversation with a built-in assistant that asks about the problem they are trying to solve, then generates the necessary code files to set up a run. For users who already have a project, the assistant can inspect what is already there and only generate the missing pieces. Once a run is underway, the system repeatedly asks the language model to suggest improvements to a chosen section of code, evaluates how well each version performs, and keeps the best results. Two additional tools help users decide which part of their code is worth trying to improve. One, called the Evolve-Block Advisor, takes a specific section of code and a stated goal and gives back an assessment of whether evolving that block is likely to help, along with concerns and suggestions. The other, called the Evolve-Block Recommender, looks at an entire project folder and ranks the candidate code blocks most likely to benefit from the evolution process. The platform connects to AI providers including OpenAI, Anthropic, and compatible alternatives, configured through a simple settings file. It also supports distributing large experiments across multiple machines using a framework called Ray. Results are saved automatically so a run can be resumed if it is interrupted. LLM4AD_Next is built in Python and requires version 3.10 or newer. An online demo is available that needs no installation or API key, intended for trying the platform before setting it up locally. The project is licensed under BSD-3-Clause.