Top Tweets for #TextGrad
LLMs can be an "Expert-In-The-Loop" for physics problems that require simulation. Here's a minimal example using #TextGrad (Automatic “Differentiation” via Text) solving a simple physics problem iteratively: https://t.co/6h4x9eUshz
@SimAI4Science #TesseractHackathon
(8/8) Finally, my gratitude to the #TextGrad project (https://t.co/YJLBKWA9Rv) is beyond words — without the foundation it provided, this research simply wouldn’t exist.
(1/8) Existing LLM optimizers such as #TextGrad, #DSPy, and #LangChain are often too broad, which can make them inefficient at times. Also, it's difficult to combine the strengths of different optimizers.
Introducing #metaTextGrad🌟: a meta-optimization framework built on #TextGrad , designed to improve existing LLM optimizers by aligning them more closely with specific tasks.
📰 NeurIPS 2025 paper: https://t.co/a32seWX4t6
🧑💻Code: https://t.co/YrKIxwOhmy
📚 Slides: https://t.co/IItVarRxGw
🚀 I’m thrilled to announce that #textgrad has been published in @Nature today! It’s been an incredible journey working with the TextGrad team, I am grateful for the wonderful collaboration within the Zou Group. @james_y_zou. 🙌
#Nature #AI #LLMs #AgenticAI
🚀 Thrilled to share that #textgrad is published in @Nature today! 🎉
It’s been an incredible journey working with the amazing TextGrad team and the Zou Group @james_y_zou. 🙌
✨ What is TextGrad?
A groundbreaking framework that automates optimization of LLMs and compound systems using insights from "textual gradients."
🔍 Check it out for more details!
📄 Paper: https://t.co/B4RCEDL7WP
💻 Code: https://t.co/R6TPUDY2wL
📚 Docs: https://t.co/uLqIRNc5Mp
🎥 Video by Discover AI: https://t.co/EfwGUvRjcZ
#Nature #AI #LLMs #AgenticAI
💡The key idea of #textgrad is to optimize by backpropagating textual gradients produced by #LLM.
Paper: https://t.co/CjBpSxcnpn
Code: https://t.co/nCGqp15kJJ
Amazing job by @mertyuksekgonul leading this project w/ fantastic collaborators @federicobianchy Joseph Boen @ShengLiu_ @lupantech @guestrin @ZhiHuangPhD 👏
🚀 The Future is Multi-LLM-based AI Systems🚀
In the upcoming multi-LLM systems, there’s a BIG question on the horizon:
𝗛𝗼𝘄 𝗱𝗼 𝘄𝗲 𝗘𝗩𝗔𝗟𝗨𝗔𝗧𝗘 𝘁𝗵𝗲𝘀𝗲 𝗔𝗜 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗱𝘂𝗿𝗶𝗻𝗴 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲?
(as in #TextGrad @mertyuksekgonul @james_y_zou)
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 @karpathy's video reimplementing this from scratch... I thought I had to make my own...
So here is the just 300 lines of code! a lot of which is (ironically) prompts: https://t.co/SeL9Ssekhv
This screenshot is real code using tiny-text-grad!
- The equality_loss function asks an LLM judge whether the answer is correct, or provide feedback otherwise.
- The loss.backwards() call distributes the feedback through the call-graph--using more LLM calls!
- And optimizer.step() updates all the parameters (prompts) using a "gradient step" in the direction of the feedbacks received.
The MultihopModel is based on #dspy's Simplified Baleen: https://t.co/0tXiSCYbeu and allows the LLM to perform multiple Wikipedia calls as it builds up context.
This creates an interesting call graph, which is visualized here:
The loss node is furthest to the right, and all the nodes that it depends on are to the left.
In the text-backprop step each node accumulates feedback in a list, rather than with a sum as in "real" backprop.
That's all!
Except... Does it work?
Well... It's definitely better than bad prompts and no tuning. But sometimes the optimization "explodes" just like normal gradient descent.
I'm not 100% convinced this is the way to go, over few-shot optimization or just giving the LLM the complete call graph directly, and asking it to optimize it.
⚡️#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%!
🚀 #TextGrad is advancing multimodal reasoning and reducing hallucinations! Join us in contributing to TextGrad, an innovative framework that automatically optimizes foundation models via natural language gradients!
Check it out here: https://t.co/WpwqCKvpn0! 🌟
⚡️#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%!
⚡️#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%!
⚡️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 🧵
Great presentation by @james_y_zou on mixture of agents and text grad at hackathon organized by @SambaNovaAI, @togethercompute, @NumbersStnAI at @agihouse_org . Multiple agents collaborating to achieve a task is becoming an increasingly important aspect to develop useful enterprise products! By combining multiple agents, they were able to beat #GPT4 on AlpacaEval leaderboard.
These systems require switching between multiple agents and running them at extremely fast speed. This is where SambaNova’s SN40L chip really shines. 100s of models on a single node each at 1000s tokens per second. Try our apis at -
1.https://t.co/RPrWNbxA4j
2.https://t.co/u6bOwITBG2
🔥#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: https://t.co/9NGJJtsQf8 Great work @lupantech @mertyuksekgonul + team! Works w/ any VLM. Check out @lupantech's 🧵for more examples! https://t.co/2lQb4Fch6J
#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 more:
💻 Code: https://t.co/WpwqCKvpn0
📄 Doc: https://t.co/Wwb5XTNTkF
🌐 Project: https://t.co/3n8OfPYM91 🧵
⚡️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 🧵
The Top ML Papers of the Week (June 17 - June 23):
- TextGrad
- PlanRAG
- Claude 3.5 Sonnet
- DeepSeek-Coder-V2
- Mitigating Memorization in LLMs
- Tree Search for Language Model Agents
...
Last Seen Hashtags on Sotwe
لب_العسل
pakistanislut
Seen from Netherlands
daddysissues
Seen from Mexico
Ukraine
Seen from United States
dizzlelife
Seen from United States
สวิงกิ้งสาวใหญ่
Seen from Thailand
DiplomadoATuRitmo
Seen from Brazil
momson #nolimit
Seen from Turkey
kayserieskort
Seen from Turkey
DumbassHeadline
Seen from Ukraine
Trends for you
Most Popular Users
Elon Musk 
@elonmusk
240.1M followers
Barack Obama
@barackobama
119.3M followers
Donald J. Trump
@realdonaldtrump
111.6M followers
Cristiano Ronaldo
@cristiano
108.9M followers
Narendra Modi
@narendramodi
107M followers
Rihanna
@rihanna
97.3M followers
NASA
@nasa
92.1M followers
Justin Bieber
@justinbieber
90.6M followers
KATY PERRY
@katyperry
86.8M followers
Taylor Swift
@taylorswift13
80.6M followers
Lady Gaga
@ladygaga
72.1M followers
Kim Kardashian
@kimkardashian
69.4M followers
YouTube
@youtube
68.6M followers
Virat Kohli
@imvkohli
68.5M followers
Bill Gates
@billgates
63.4M followers
The Ellen Show
@theellenshow
62.5M followers
CNN
@cnn
61.9M followers
Neymar Jr
@neymarjr
61M followers
X
@x
60.9M followers
CNN Breaking News
@cnnbrk
59.9M followers