This is Michigan AI Laboratory, a vibrant research group that includes faculty, research scientists, graduate students, and postdoctoral fellows working in AI.
Ann Arbor, MI
Joined January 2018
- Tweets9.4K
- Following337
- Followers5.2K
- Likes19.3K
Pinned Tweet
As #AI continues to advance, so must the safeguards that guide its development and use.
Michigan AI's Prof. Michael Wellman is among 405 scientists, incl. @Yoshua_Bengio, who just authored an open letter calling for safeguards for #AI.
Check it out! 👉ai-security-letter.org/
Over 400 researchers and I have signed this open letter calling for the AI and security communities to come together to improve alignment and containment approaches in order to protect our critical infrastructure and digital systems.
ai-security-letter.org/
MichiganAI retweeted
🌟 Is one example enough to pass LLM fairness benchmarks? In our new paper, “One Example Is Enough to Pass Fairness Benchmarks", presented today at #EMNLP2026, we find that a single BBQ example can boost LLM accuracy to over 90% with GRPO or one-shot learning.
The main takeway is that passing a fairness benchmark does not necessarily imply that a model is fair, and we may need to rethink the benchmarks.
Work led by @NaihaoDeng, jointly with @SameeArif, @ShuaichenChang, @Yulongchen1010.
🌐 Project page: lit.eecs.umich.edu/hacking-f…
💻 Code: github.com/MichiganNLP/hacki…
🤖 Models: huggingface.co/collections/M…
MichiganAI retweeted
📢 Join us this Friday for a CSE Distinguished Lecture! Prof. Ramesh Sitaraman (@UMassAmherst) presents "The Story of the Edge: From Content Delivery to Agentic AI."
October 30 | 1670 Beyster Bldg
Details: #EdgeComputing
#AgenticAI #TechTalks
ALT Promotional graphic for the Distinguished Lecture Series in Computer Science and Engineering at the University of Michigan, featuring Ramesh Sitaraman speaking on "The Story of the Edge: From Content Delivery to Agentic AI" on Friday, October 30 at 3:30 pm at 1670 Beyster Building.
MichiganAI retweeted
Do #LLMs rewrite history?
Our research finds that some models can rewrite history in favor of invaders & perpetrators of massacres and political repression.
We are deeply concerned⚠️
Check out our work at #EMMLP2026 on “Detecting Historical Revisionism in LLMs” tomorrow: Hall H, 16:45.
MichiganAI retweeted
🧠 Can models discover the most effective memory representations for robots?
Excited to share our work showing that LLMs can progressively refine what robots remember and how they represent it through iterative optimization grounded in robot demonstrations. 🚀
MichiganAI retweeted
Using a personal AI agent can automate tasks and remove friction from your busy life, but be prepared for the bot to make mistakes, and assume you won’t catch all of them.
“Even in simple tasks like judging whether AI output is wrong, people consistently accept wrong AI outputs even if they could have gotten it right on their own,” said Vera Liao, associate professor of computer science and engineering @UMengineering. She spoke with @voxdotcom on offloading tedious tasks to AI agents.
🔗 Read more: myumi.ch/2jg7E
MichiganAI retweeted
Check out our poster at 16:45 PM at Hall H, #EMNLP2026!
Do #LLMs rewrite history?
Our research finds that some models can rewrite history in favor of invaders & perpetrators of massacres and political repression.
We are deeply concerned⚠️
Check out our work at #EMMLP2026 on “Detecting Historical Revisionism in LLMs” tomorrow: Hall H, 16:45.
MichiganAI retweeted
Excited to present this work today at the 16:45-18:15 oral session on computational social science, hall G7-8. Feel free to stop by! 😊
Excited to share our new preprint:
Whose Norms? Disentangling Cultural and Personal Alignment in Large Language Models
We ask: when cultural norms and personal preferences conflict, whose norms do LLMs follow? 🌍⚖️
Check out our work here:
📄 Paper: arxiv.org/abs/2606.07877
🤗 Dataset: huggingface.co/datasets/Anga…
💻 Code: github.com/MichiganNLP/pact_…
🌐 Website: lit.eecs.umich.edu/pact_cult… (1/n)
MichiganAI retweeted
Excited to present AtlasNLP, led by @_agirlyengineer, at Poster Session D today 11-12:30, Hall H! 🌍 A collaborative effort mapping geographic representation in NLP datasets and where gaps remain.
#EMNLP2026
💡 Making trouble with #AI!
In this week's AI Seminar: TUESDAY Michigan AI Profs. Silviu Pitis & Gabriel Poesia will explore creating problems with AI for reasoning evaluation and verified programming.
📅 Oct. 27 | 4 PM ET
cse.engin.umich.edu/event/ai…
MichiganAI retweeted
📖 Paper: arxiv.org/abs/2602.17433
💻 Code & dataset: github.com/francescoortu/Pre…
🎓 Authors:
@francescortu @JoeunYk05 @psyonp @bschoelkopf, @albecazzaniga @radamihalcea @ZhijingJin
🤝 Work done at @EuroSafeAI @JinesisLab @MPI_IS @ELLISforEurope @UofTCompSci @VectorInst @TorontoSRI @CIFAR_News @JinesisLab @ELLISInst_Tue @michigan_AI @UMichiganAI @UMichCSE
MichiganAI retweeted
If you’re at EMNLP today, feel free to stop by our poster! We’d love to chat and hear your thoughts 😊
📄 Deep and shallow biases in language models
🔖 1152-MAIN
📅 Oct 26
🕓 4:45–6:15 PM
📍 Hall H | Poster Session E (Session 9)
Heading to #EMNLP2026 in Budapest 🇭🇺 to present our paper "Deep and shallow biases in language models"!
Not all LLM biases are equal.
🎲 Ask an LLM for a random number → 42. Reword it → now it says 3.
🦋 Ask for a random butterfly → Monarch. Reword it → still Monarch.
Why do some biases survive rewording and others don't?
🎬 Watch the video below and learn more at deepbias.github.io
🏆 Congratulations to Michigan AI PhD student Wonje Jeung, one of four inaugural James F. Lu World Fellows!
The prestigious fellowship recognizes academic excellence and promising research toward building safer, more reliable generative AI.
#MichiganAI
cse.engin.umich.edu/stories/…
🤖 How do language, space, time & common ground intersect in #robotics?
TOMORROW Michigan AI's Prof. Joyce Y. Chai explores these topics in the Michigan Robotics Seminar Series!
📅 Oct. 27 | 4–5 PM ET
🔗 robotics.umich.edu/events/ro…
MichiganAI retweeted
We had a great #EMNLP2026 lunch event on @ACLMentorship and AI for Good, co-organized with @EuroSafeAI and @NLP4PosImpact 🎉❤️! Talked about PhD applications, future of #NLProc, …
Thx to our great mentors @mohitban47 @mrinmayasachan @iatitov and my wonderful co-organizers @iamdddaryna @AnganaBorah2 @Ruyuan_Wan with years of my friendship! 🤗
MichiganAI retweeted
If you are at #EMNLP2026, check out @yoonjoo_le2 presenting KnowSim in the Tuesday 9-10:30am poster session. Yoonjoo is also on the job market!
We think a lot about making AI more aligned with the user needs of different cognitive profiles. A key dimension is knowledge level. We built a simulator of how people with varied knowledge seek information, grounded in cogsci theories. Excited to see it out at #EMNLP2026 main!
MichiganAI retweeted
Also, a huge shoutout to my amazing labmate An Vo @an_vo12, who’ll be at #EMNLP2026 presenting his work on how model training and prompt input interact to shape a model’s answer preferences! Check it out 👇 arxiv.org/abs/2609.09901
Also shout-out to Jiatong (@jding646536), a senior undergrad at Michigan, whose work on quantization was accepted as a main conference paper at #EMNLP2026! 🎉 Glad to see so many young and talented researchers contributing to the field 😁! Check out his paper: arxiv.org/abs/2608.27875
MichiganAI retweeted
Unfortunately, I can’t make it to #EMNLP2026 this year, but the amazing @Yulongchen1010 and @AnganaBorah2 will be presenting our papers. Please stop by and say hi! Hope everyone has a wonderful time at the conference 😄!
🌍 Michigan AI Lab at #EMNLP2026!
Our faculty and students are presenting papers exploring #AI alignment, reward hacking, fairness, multilingual #LLMs, and more, check them out!
Go Blue! 💛💙
🏆 Warm congratulations, @__runamu__ !
We’re delighted to recognize the work of the "ScreenTutor" team and its exciting real-world potential!
Excited that our work, "ScreenTutor," received the 🏆 Startup-in-Waiting Award at the 2026 Michigan AI Symposium!
ScreenTutor turns unlabeled tutorial videos into training data for computer-use agents. Great to see its real-world potential recognized.
Grateful to my amazing collaborators @ryan_kwon03, Ruijie Chen, @Jaekyeom__Kim, @lajanugen, @tiangeluo, @sungryulls, and @honglaklee, and to the symposium organizers.
Special thanks to @jinyoung__kim and Jiwoo Han for capturing these moments 🙏✨
MichiganAI retweeted
Excited to be in Budapest for #EMNLP2026! 🇭🇺
I’ll be presenting our work: Whose Norms? (Main) and BeliefSim (Pandora workshop). I’m also excited to present collaborative work on geographical representation in NLP (led by @_agirlyengineer) and child safety (led by @sameearif).
