TextGrad: Automatic ''Differentiation'' via Text
TextGrad is a Python package that provides a simple interface to implement LLM-“gradients” pipelines for text optimization!
About TextGrad
If you know PyTorch, you basically already know how to use TextGrad! It mirrors the syntax and abstraction of PyTorch.
TextGrad can achieve many tasks, including:
- LeetCodeHard best score
- GPQA sota
- Designs new molecules
- Improves treatments
TextGrad flexibly optimizes any system of agents + tools. Many such systems involve blackboxes hard to tune by standard gradients, but easy to optimize via text gradient. The "gradients" here are natural language feedback that are easy to interpret!
Start using TextGradCheckout this great intro video from code AI:
Successful Applications
Here are examples of successful implementations that demonstrate the effectiveness and versatility of TextGrad across various applications.
Coding
We optimize solutions to difficult coding problems from LeetCode, where we boost the performance of gpt-4o and best existing method by 20% relevant performance gain.
Problem solving
We optimize solutions to complex scientific questions to improve the zero-shot performance of GPT-4o. For instance, in Google-Proof Question Answering bench-mark, we improve the zero-shot accuracy from 51% to 55% by refining the solutions at test-time.
Reasoning
We optimize prompts to improve the LLM performance, where we push the performance of GPT-3.5 close to GPT-4 in several reasoning tasks.
Chemistry
We design new small molecules with desirable druglikeness and in silico binding affinity to drug targets.
Medicine
We optimize radiation treatment plans for prostate cancer patients to achieve desirable target dosage and reduce side effects.
TextGrad Team
Mert Yuksekgonul
Stanford University
Federico Bianchi
Stanford University
Joseph Boen
Stanford University
