Translating short-form Python exercises to other programming languages using diverse prompting strategies

Area of Science:

• Computational Biology
• Bioinformatics
• Scientific Computing

Background:

• Life scientists increasingly use programming for data analysis, reproducibility, and collaboration.
• Python offers simplicity, while C++ and Rust provide efficiency for complex computations.
• Translating code between languages is labor-intensive, motivating the exploration of automated solutions.

Purpose of the Study:

• To investigate the efficacy of large language models (LLMs) for semi-automated code translation.
• To assess LLM performance in translating Python programming exercises into C++, Rust, Julia, and JavaScript.

Main Methods:

• Translated 559 short-form Python programming exercises into C++, Rust, Julia, and JavaScript using GPT-4.
• Employed three prompting strategies: instructions only, code only, and a combination of both.
• Compared the output of translated code against the original Python code for accuracy.

Main Results:

• LLM-based code translation demonstrated high success rates across all target languages.
• Rust achieved the highest translation success rate (99.5%), followed by JavaScript (98.9%), C++ (97.9%), and Julia (95.0%).
• Prompting strategy significantly impacted translation success, with combined approaches often proving most effective.

Conclusions:

• LLMs are effective tools for translating small-scale programming exercises between languages, reducing manual effort.
• The study provides valuable, freely available code translations to support scientific education and research.
• Automated code translation using LLMs shows promise for streamlining workflows in computational science.