I enjoy reading this line of work, I'm really excited about the solid effort here. I just can't wait for the days where the baseline claim and experiment for this kind of work is sth beyond "Well, it maybe sort of kind appears to correspond to human-understandable stuff"

TextGrad: Automatic "Differentiation" via Text - Backprops textual feedback provided by LLMs to improve individual components of a compound AI system - 51% -> 55% on GPQA and 20% rel. gain in LeetCode-Hard repo: github.com/zou-group/text… abs: arxiv.org/abs/2406.07496

TextGrad Automatic "Differentiation" via Text AI is undergoing a paradigm shift, with breakthroughs achieved by systems orchestrating multiple large language models (LLMs) and other complex components. As a result, developing principled and automated optimization

Was such a fun project exploring several applications of LLM-powered “backprop” and designing a PyTorch-like API for it — with a wonderful team! We appreciate AK and Aran Komatsuzaki for sharing our work 🤍🤍 Will share more soon!! For now check out: Paper:

⚡️This is the most fun project! We built PyTorch-for-text! 🔥 #TextGrad: automated "differentiation" via text to optimize AI systems by backpropagating LLM text feedback. TextGrad + GPT4o: 💻LeetCodeHard best score ❓GPQA sota 🧬Designs new molecules 🩺Improves treatments 🧵

This is the most interesting paper I've read in a while. A calculus of language instead of number. infinitesimal = a criticism integration = editing over criticisms you optimize LLM programs via backprop where each update is expressed linguistically

New Hugging Face Daily Papers feature If you have at least one indexed paper on Hugging Face, you can now directly submit papers to HF daily papers try it here: huggingface.co/papers

Exciting new paper: TextGrad: Automatic “Differentiation” via Text! This is a DSPy-like framework for optimizing prompts in a composite LLM system. However, there is one major difference! In DSPy the idea is (basically): - Forward pass: Each Module (LLM call) generates picks a

⚡️Chain rule for text! The key of #TextGrad is to optimize any #AI #agent system by backpropagating text feedback. Autograd for the age of agents🪄 Check out our new hands-on tutorials: Tutorial github.com/zou-group/text… Paper arxiv.org/abs/2406.07496

Engineer, designer, honorary physician, entrepreneur, and now Prof. Huang — a master of all trades but most importantly, a kind and wonderful person. I envy the students who’ll get to work with Zhi!!

*TextGrad: Automatic “Differentiation” via Text* by Mert Yuksekgonul Federico Bianchi Sheng Liu Zhi Huang James Zou PyTorch syntax for optimizing graphs of LLM calls, where "gradients" and "optimization" are computed by additional LLM instances. arxiv.org/abs/2406.07496

#TextGrad now features multimodal reasoning! 🔬 ScienceQA (multimodal scientific reasoning) - Error rate drops by 20%, achieving the highest zero-shot performance we know of. 📊 MathVista (multimodal math reasoning) - Boosting the score from 63.8% to 66.1% on GPT-4o! Explore

🔥#TextGrad is now multi-modal! TextGrad boosts GPT-4o's visual reasoning ability: 📊MathVista score 63.8➡️66.1 w/ TextGrad 🧬Reduces ScienceQA error rate by 20%. Best reported 0-shot score Tutorial: colab.research.google.com/github/zou-gro… Great work Pan Lu Mert Yuksekgonul + team! Works

Work co-supervised with Hidenori Tanaka, led by Udith Haputhanthri, has now been accepted to ICML Mechanistic Interpretability workshop. TLDR: High-dimensional bifurcations underly skill acquisition in task-trained RNNs Link: openreview.net/forum?id=njmXd… A tweetprint🧵

⚡️#TextGrad reduces hallucination in multimodal LLMs! MMVP 🏆 (multiple choice questions) - TextGrad optimized prompts increase the accuracy of GPT-4v from 71% -> 76%! HQH - Relation📍(open-ended generation) - TextGrad boosts the accuracy of GPT-4o from 77.2% to 82.5%!

I am glad to share one of our recent works, now published in nature! Here, we studied the neural circuits of the placebo effect. I am happy to be part of such an amazing team, thank you Chong, Grégory Scherrer, and Dr. Mark Schnitzer! nature.com/articles/s4158…

We talk a lot about the potential of AI for applications, like AI for Education. But actual progress requires that we hill-climb on realistic, hard tasks. Are there any? 🔽 Bridge, Backtracing, and Teacher Coach are 3 real-world AI for Education datasets that are far from

A while ago I wrote a thread about #TextGrad, which is an alternative prompt optimization method, based on "natural language gradients". Cool! Since we are still waiting for Andrej Karpathy's video reimplementing this from scratch... I thought I had to make my own... So here is the