Haohan Wang

LLMs are like calculators of this era, eventually, when people get used to its power and the hype gets washed off, it's a tool.

Social media has truly warped science consumption. On one hand it has massively democratized it. On the other hand, it's become extremely easy to use hype to simply amplify things that are simply not true. 1/

📰 Thanks to 404 Media for covering our recent paper. 🗣️ Imagine sitting through a long, tedious speech—overstructured, filled with jargon, and clearly incoherent. As a human, you stop listening. But a language model cannot. It must process the entire input and respond. 🧪 Our…

You’re watching a few rounds of poker games. ♠️♠️♠️♠️ The cards look normal — but the outcomes don’t.♦️ No one explains the rules. You just see hands play out. -- Can you figure out what’s going on? 🎯 That’s the setup, for LLMs. Recently, there is heated discussions on…

Thanks for sharing! The world needs to treat research papers as research papers, not as crystal balls.

📊🤖 Vision-language models like GPT-4V, 4o, and even the latest O3 still struggles with shortcuts⚠️ when reasoning over scientific charts 📉📈. The shortcuts are no longer about easy tasks like “is this a dog?”—those are solved. But when asked deeper questions like “how much…

🚨 We’re hiring a postdoc! Join us to push the frontier of AI and machine learning in genomics—aiming to uncover the genetic basis of complex human disorders. Please help RT! 🙏. For more information, please visit: haohanwang.github.io/postdoc_hiring…

Crosswords were once our quiet battles—fought over breakfast tables, in classrooms, on long train rides—testing not just what we knew, but how we thought. For generations, these puzzles have embodied the essence of human intelligence: language, memory, creativity, and cultural…

Great pleasure to be part of this great piece of work, aiming to define what agent is, look back to its history, and project its future, with summary of a wide range of relevant works. Hopefully it will be considered as one of the building blocks of agent research in the…

Sharing an idea that has haunted me for nearly a decade. 🧩 It started when I compiled the development history of deep learning. I noticed something odd: certain matrices suffer from nullspaces—directions they simply cannot represent. But deep learning models, built from layers…

writing a survey could be a lightweight alternative; the main reason I always recommend graduate students to start their journey by writing a survey

Thanks for hosting it, will present some of our recent works, featured by long-term efforts GenoTex arxiv.org/abs/2406.15341, recent IMPROVE arxiv.org/abs/2502.18530 and several works yet to come. Hopefully it will be a kickstart of our upcoming series of agentic scientist works.

Congratulations!