Evolving Graph Structured Programs for Circuit Generation with Large Language Models

This repository contains the official implementation of our manuscript Evolving Graph Structured Programs for Circuit Generation with Large Language Models.

Installation

ABC Installation

We provide a precompiled abc executable for Linux as a compressed archive in the GitHub Releases of this repository. Download the archive from the Releases page and extract the abc binary into the root directory of this project. Due to potential compatibility issues across different operating systems and hardware platforms, users on non-Linux platforms are advised to compile the source code manually. For detailed compilation instructions and the official source code, please refer to the upstream repository: berkeley-abc/abc.

For details on how abc is used in this work, please refer to our paper.

Python Environment

This project has no strict Python version requirements. Python 3.12 is used in our environments. The key packages and their exact versions used during development and testing are listed below:

Package	Version
torch	2.10.0
numpy	2.4.2
tensorboard	2.20.0
boolean.py	5.0
requests	2.32.5
joblib	1.5.3

The complete list of dependencies is provided in requirements.txt. Install all required packages via:

pip install -r requirements.txt

Running the Project

Prerequisites

Before running the project, ensure the following conditions are met:

1. ABC Executable: Ensure the abc executable is present in the project's root directory and is runnable. For non-Linux platforms, compile it from berkeley-abc/abc.
2. LLM Credentials: Configure your LLM API credentials in runEoh4AIG.py:
   • Set llm_api_endpoint to your LLM API endpoint.
   • Set llm_api_key to your valid API key.
3. Python Dependencies: Install all required packages as described in the Python Environment section.

Quick Start

Run the core script with the following sample configuration:

python runEoh4AIG.py \
  --truth_file_path ./example/circuit.truth \
  --ec_pop_size 10 \
  --ec_n_pop 10 \
  --llm_model gpt-3.5-turbo \
  --ec_operators e1 e2 m1 m2 \
  --sample_num 1000 \
  --local_search

Parameter Reference

Parameter	Type	Default	Description
--truth_file_path	str	./example/circuit.truth	Path to the input truth table file.
--ec_pop_size	int	3	Number of individuals in each evolutionary population.
--ec_n_pop	int	5	Number of evolutionary populations (generations).
--ec_operators	list	['e1', 'm1']	EOH operators to apply (supports e1, e2, m1, m2).
--llm_model	str	gpt-3.5-turbo	LLM model identifier (e.g., gpt-4o, gpt-3.5-turbo).
--use_fx	flag	False	Enable FX optimization.
--local_search	flag	False	Enable local search optimization.
--random_generated	flag	False	Generate random initial programs.
--shannon_decomposition	flag	False	Apply Shannon decomposition to the AIG.
--LLM_generation_initial	flag	False	Use LLM to generate the initial AIG programs.
--legalized_parent	flag	False	Use legalized parent prompts during generation.

For the full list of available parameters, refer to runEoh4AIG.py.

Output

All experiment results — including the optimal AIG program and TensorBoard logs — are saved to a timestamped subdirectory under ./outputs/.

Additional Benchmarks

The repository includes a minimal example circuit in ./example/circuit.truth for quick testing. For more comprehensive experimentation, we provide a collection of benchmark circuits from IWLS in our releases page:

• Download benchmark.tar.gz from the releases page
• Extract to the project root directory:

  tar -xzf benchmark.tar.gz

• Use benchmark circuits by updating --truth_file_path (e.g., ./benchmark/iwls2024/all/ex29.truth)