TextGrad (2024)

TextGrad is a framework for automatic differentiation via text, inspired by backpropagation in neural networks. It leverages large language models (LLMs) to provide rich, natural language feedback that serves as textual gradients for optimizing variables in computation graphs. These variables can range from code snippets and prompts to molecular structures and radiotherapy treatment plans. The framework follows PyTorch’s syntax and abstraction, making it flexible and easy to use. It works out-of-the-box for a variety of tasks without requiring users to tune components or prompts, and has demonstrated effectiveness in improving accuracy on Google-Proof Question Answering, optimizing LeetCode-Hard coding solutions, designing druglike molecules, and planning radiation oncology treatments.