Paper - L2MAC: Large Language Model Automatic Computer for Extensive Code Generation

Free

LLM-based automatic computer for extensive, coherent code and text generation.

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About Paper - L2MAC: Large Language Model Automatic Computer for Extensive Code Generation

L2MAC is the first practical LLM-based general-purpose stored-program automatic computer (von Neumann architecture) framework, implemented as a multi-agent system for generating long and consistent outputs. It overcomes the fixed context window limitation of transformer LLMs by incorporating a memory system with an instruction registry (populated with a prompt program for the task) and a file store (holding final and intermediate outputs). Each instruction is executed by a separate LLM agent, managed by a control unit that performs precise memory reading and writing. L2MAC achieves state-of-the-art performance in generating large codebases for system design tasks and also works for general-purpose extensive text generation, such as writing entire books. The work was published at ICLR 2024.

Key Features

LLM-based multi-agent system implementing a stored-program (von Neumann) architecture

Instruction registry houses a prompt program that defines the task

File store holds final and intermediate outputs

Control unit manages precise memory reading and writing for each agent

Bypasses the fixed context window constraint of transformer LLMs

Generates long and consistent outputs (codebases, books, etc.)

Pros & Cons

Pros

State-of-the-art performance in generating large codebases compared to other coding methods
Overcomes LLM context window limits for long outputs
General-purpose design applicable to various long-form generation tasks

Best For

Generating large codebases for complex system design tasksWriting entire books or lengthy documentsGeneral-purpose extensive text-based generation

FAQ

What is L2MAC?

L2MAC (Large Language Model Automatic Computer) is a practical LLM-based multi-agent framework that implements a von Neumann stored-program architecture to generate long and consistent outputs, such as extensive codebases or entire books.

How does L2MAC overcome LLM context window limits?

L2MAC uses an instruction registry and file store memory system, along with a control unit that allows precise reading and writing, enabling the generation of outputs far longer than the underlying transformer's fixed context window.

What tasks can L2MAC perform?

L2MAC is designed for extensive code generation in system design tasks, writing entire books, and any general-purpose text generation requiring long, coherent output.

Has L2MAC been peer-reviewed?

Yes, L2MAC was published in the Twelfth International Conference on Learning Representations (ICLR) in 2024.