L2MAC can create near unbounded outputs that align exactly with the user input prompt over very long generation tasks

It achieves strong empirical performance of state-of-the-art generation for large codebase tasks and is in the top 3 for the HumanEval coding global benchmark. As L2MAC can detect invalid code and failing unit tests when generating code and automatically error corrects them.

Internally persists a complete file-store memory that enables LLM agents to read files and write to files, creating a large output over many iterations

It can be instructed to follow an exact prompt program

As it generates the output one part at a time, it enables an LLM with a fixed context token limit to be bypassed

The paper, peer-reviewed and recently accepted and published at ICLR 2024, introduces L2MAC.