Dashun Wang

🔥Download the Science of Science book for FREE🔥 We're VERY excited to make the whole book freely available: dashunwang.com/book/the-scien… Please help us RT & let everyone know of this resource! #ScienceOfScience

Wonderful new work by Eddie and team! Wow!! 🤩

Another hit from @yian_yin's lab! this time, super timely, important, and scary!!!

Dashun Wang talking 'airplanes for the mind' 🛩️ are heavier than air, but can fly faster than sound—remind you of another technology that "shouldn't" be able to work, but does?

Fascinating - Newton's gravity law turns out to be ... correct! largest scale validation. 3 centuries later, are we still doing Newton's homework for us? science.org/content/articl…

Fantastic analysis from journal submission data. It is also encouraging to see key findings from our recent @ScienceMagazine paper on #AI impact on science — increased productivity, more complex scientific writing, and emerging quality concerns — echoed in this setting.

New paper led by Minsu Park, showing that high impact interdisciplinary papers disproportionately come from deeply disciplinary grants. Check it out!

Check out the paper: pnas.org/doi/10.1073/pn… and the press release: news.northwestern.edu/stories/2026/4… 7/7

Bipartisan-cited science thus highlights both the limits of science in bridging polarization and its potential to provide a shared evidentiary foundation for addressing pressing challenges. Key Q going forward is how to improve the evidentiary basis for policymaking. Can gen AI help? or make it much worse? Stay tuned for more! 6/n

🚨Bipartisan-cited science is rare, unevenly distributed, and disproportionately influential. 🚨 Check out our latest paper in PNAS, led by @zfurnas 🧵1/n

Northwestern just ran a test nobody wanted to run. They took two LLMs with nearly identical MMLU-Pro scores and compared their answers question by question. The models disagreed on up to 66% of items. Same benchmark. Same accuracy. Completely different error profiles. Even among frontier models scoring above 60%, disagreement stayed between 16% and 38%. Two models that each "get 7 out of 10 right" systematically get a different 7 right. Then they did the part that should have ended the conversation. They took published studies in education and political science — real peer-reviewed research — and re-ran the analyses using 8 different LLMs as annotators. In an education study on literacy intervention, the estimated treatment effect varied by 84% depending on which model scored the student essays. 0.19 with one model. 0.35 with another. Same essays. Same rubric. In a political science study on Russian state media, switching the model reversed the sign of the finding. The original study concluded officials were more likely to be credited with good economic news. Two of the models concluded the opposite — that they were more likely to be blamed for bad news. Not "slightly different magnitude." The opposite conclusion. This is what the paper calls a benchmark illusion. MMLU-Pro and GPQA are the currency the entire industry uses to decide which model to deploy. The working assumption is that models with matching scores are interchangeable. They aren't. They're making different mistakes on different questions, and those mistakes happen to average out to the same number on a leaderboard. The villain here isn't any specific model. It's the benchmark itself. A single accuracy score compresses away the exact information that matters for science: which items you got wrong and why. Every paper that says "we used GPT-4 to annotate our dataset" was making an invisible methodological choice. Swap the model, potentially flip the result. The uncomfortable part isn't that LLMs disagree. It's that the tools we built to measure them were never designed to catch it. Paper in comments.

Abstract is due in two days! Join us for an exciting day!

CALL FOR APPLICATIONS | CASBS 2026 Summer Institute in AI Methods for Social Scientists (AIMS) July 26-31 | App deadline: May 8 AIMS brings together scholars from across the social sciences for a week of learning, practice & collaboration DETAILS & APP: bit.ly/4muwstd

@MishaTeplitskiy 💰 or put more positively, the success of science itself.

Ok boys and girls, I'm stuck feeding a newborn, so I guess it's time to read and "live-tweet" this bad boy! How did academia go from a quasi-religious and guild-like occupation to the office job that it is today?

Join us in the beautiful Boulder this summer! Get your abstract in by April 6!!!

Excited to share that our SciSciGPT paper is featured on the cover of Nature Computational Science. Special credit to Hanna Renedo — this is also the first piece coming out of our newly launched CSSI art studio. More to come.

@Afinetheorem Why Shenyang? (My hometown)

I am bringing a small group of innovation economists and AI policy folks to China Apr 12-19. Hangzhou, Beijing, Shenyang. Lots of meetings on AI and "real industry"/robotics - if we should talk to someone great in your network, please email and connect us!