Excited to share new works on LLMs, agents and AI for science at #NeurIPS this week!👇 Thanks to my awesome students + collaborators. Look forward to meeting old and new friends in San Diego! Let me know if you want to chat!

🤔Want a principled way to RL your diffusion model? Check Data-regularized Reinforcement Learning (DDRL)! Post-train @nvidia #Cosmos World Foundation models with a million GPU hours! 🤯 Novel formulation ➡️ Theoretically integrates SFT into RL ➡️ Robust to Reward Hacking 🛑

This upcoming week centers on the Federal Reserve’s final policy meeting of the year, with the interest rate decision due on Wednesday. Per CME’s FedWatch, the chance of a 25bps rate cut is now 86.2%. Beyond the rate decision itself, investors will be following Jerome Powell’s

“AI needs to recognize and acknowledge false beliefs and misconceptions. That’s still a big gap in current models, even the most recent ones,” says @StanfordHAI faculty affiliate @james_y_zou on AI's current blind spots: https://t.co/U31R3PAVQm

Can LLMs help us interpret genetic variants?🧬 Check out our #NeurIPS2025 paper on CGBench; @oq_35 is presenting it today!

There's tremendous interest in multi-agent systems now, but how to optimize such system is a big challenge. #Sirius is a powerful framework that enables teams of multiple AI agents to self-improve 💯 Check out @WanjiaZhao1203's #NeurIPS2025 poster + paper!

You haven't actually graduated until you've paid off your student loans.

I am at #NeurIPS2025 from 12/2 to 12/7! Looking forward to meeting old and new friends! Come check out our poster: 🗓️Wed, Dec 3, 11AM – 2PM 📍Exhibit Hall C,D,E #5406 🗓️Fri, Dec 5, 11AM – 2PM 📍Exhibit Hall C,D,E #1712 🗓️Sat, Dec 6, 8:00 AM – 5:00 PM📍Upper Level Ballroom 6A

Can we predict the spatial effects of single-cell perturbations? Excited to share our preprint introducing SpatialProp, which computationally propagates single-cell transcriptomic perturbations across tissues, along with key frameworks for evaluating spatial perturbation models.

🚀 Introducing Agentic Context Engineering (ACE) --- a framework for self-improving language models through continuously evolving contexts (not weights). 📈 High-Performing: +10.6% on agent tasks, +8.6% on finance ⚡ Ultra-Efficient: −86.9% latency, −83.6% dollar cost 🛠

As AI takes on agent roles in critical fields, reasoning failures raise risks. New studies from @james_y_zou of @StanfordMed and @ylqzd2011 of @HKUniversity show how reasoning goes off the rails.

Learning to learn by #LLM feedback, by #LLM feedback🤯 Many AI systems are optimized by LM feedback (eg TextGrad, DSPy, etc)🔃. Our #neurip2025 paper introduces metaTextGrad: a powerful way to optimize all these LM optimizers ➡️ better agents. 🧵

✨ 𝐓𝐡𝐫𝐢𝐥𝐥𝐞𝐝 𝐭𝐨 𝐬𝐡𝐚𝐫𝐞 𝐭𝐡𝐚𝐭 𝐨𝐮𝐫 𝐩𝐚𝐩𝐞𝐫 𝐄𝐱𝐆𝐫𝐚-𝐌𝐞𝐝 [1] 𝐡𝐚𝐬 𝐛𝐞𝐞𝐧 𝐚𝐜𝐜𝐞𝐩𝐭𝐞𝐝 𝐭𝐨 NeurIPS 2025! (one of three other ones accepted this year 🚀 ) ✨ Over the past year, my collaborators and I have been exploring a fundamental limitation of

AI is transforming scientific discovery at incredible speed 🚀 Join us at NeurIPS for our AI4Science Panel with Rafael Gómez-Bombarelli, @jeffclune @james_y_zou as they discuss what the future of AI-enabled science might look like. Link to signup below 👇

3/ Amazing work by @Kevin_GuoweiXu and @mertyuksekgonul leading this project🙇 Check out @Kevin_GuoweiXu's excellent thread for all the goods!

2/ Existing LM optimizers are broad and generic. #metaTextGrad automatically adapts them to specific tasks, greatly improving performance and efficiency. 📰 #NeurIPS2025 paper: https://t.co/DvluDROBdm 🧑‍💻 Code: https://t.co/mtkb3HGmNg 📖 Slides: https://t.co/it94UrzIYf

Learning to learn by #LLM feedback, by #LLM feedback🤯 Many AI systems are optimized by LM feedback (eg TextGrad, DSPy, etc)🔃. Our #neurip2025 paper introduces metaTextGrad: a powerful way to optimize all these LM optimizers ➡️ better agents. 🧵

Accepted at #NeurIPS2025 🎉 Look forward to meeting everyone in SD!

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/M4Wj7TVIy4 🧑‍💻Code: https://t.co/9E0M1VrG35 📚 Slides:

⚡️Solving inequality proofs with LLMs is accepted as a #neurips2025 spotlight paper! Mathematical analysis often involves deriving bounds or inequalities. Here we investigate using LLMs to derive tight bounds.

Causal DAG is a really neat approach to generate high-quality reasoning process reward at scale, without relying on LLM judge! 💯 Great job @WanjiaZhao1203 @AquaHorseM @ShiJingzhe41415 w/ awesome collaborators👏