jal278 / lmx

LMX

This is the official code repo for the paper "Language Model Crossover: Variation through Few-shot Prompting" (https://arxiv.org/abs/2302.12170).

Language Model Crossover (LMX) is a simple yet effective method of generating solution variations using Large Language Models (LLMs). Parent solutions (represented as text) are concatenated and fed into the LLM, which naturally produces variations, which are parsed as children.

This repository contains notebooks and scripts for running LMX in the domains from the paper. From these examples we hope you will see how easy it is to apply LMX to new domains :)

Notebooks and Scripts

• Binary Strings: This notebook applies LMX to the classic evolutionary algorithms domain of evolving strings of 1's and 0's.
• QD Sentiment: This notebook uses LMX, coupled with MAP-Elites, to generate rephrasings that maintain the original meaning as well as possible, while covering the space of underlying sentiment.
• Sodaracers: This script uses LMX to generate a diverse array of locomoting virtual creatures represented as python code.
• Image Generation: This script uses LMX to optimize prompts to image generation models to maxmimize desired image properties.
• Symbolic Regression: This script uses LMX to discover compact mathematical expressions that model datasets.
• EDA Comparison: This notebook compares the distribution of offspring produced by LMX to that of a classical Estimation of Distribution Algorithm (EDA).
• The Effect of Parent Ordering: These scripts and notebook analyze the impact of ordering parents in different ways when feeding them into LMX.

All of these notebooks and scripts use the huggingface transformers API, so the LLM can be easily swapped out for one that's better or more appropriate for your particular domain.

Please let us know if you have any questions. Enjoy!

sample results of QD sentiment optimization

sample results of symbolic regression

sample results of prompt optimization for image generation