Assistant professor @iSchoolUI at UIUC, affiliated at CS and IGB. Previously @CarnegieMellon (across CBD, LTI, MLD). Trustworthy AI. & Computational Biology
Champaign, Illinois
Joined November 2012
- Tweets401
- Following774
- Followers1.6K
- Likes1.6K
Yesterday I spoke at @WMdatascience Data Science. Huge thanks to @jindong_wang92 for hosting me, and for the sharp questions!
It was also heelo's first public release, after nearly a year of baking.
Short version: AI agents reach almost any tool, yet still fail at the last mile 🧵
Why "heelo"? It comes from heluo, 何罗之鱼, a creature in the ancient Classic of Mountains and Seas: one head, ten bodies. Most likely an octopus.
An octopus thinks with its arms as much as its head. That's the idea: intelligence spread across every resource.
AI talks; heelo delivers.
Out in the world for the first time. Try it: heelo.dev
Thanks again to @jd92wang and @williamandmary Data Science for the warm welcome 🙏 Joint work with my amazing students and collaborators
Ask a language model for a metaphor about time and you will probably get a river.
Ask again, or ask another model, and you get a river again.
Our new paper asks how to make LLM answers more diverse in substance, not just wording: search for the answer that is missing.
The honest part: the ablation is mixed. Searching for distance alone cut similarity slightly more, and random directions passed the novelty check more often, so the filtering and prompting stages do a lot of the work.
There is no human evaluation and no temperature baseline yet. Turning a well-chosen gap into a better answer, not just a different one, is the open problem.
We also tried it for product positioning in AI search, building on GEO by @PranjalAggarw16 and colleagues: being cited and being differentiated turned out not to be the same thing.
Paper: arxiv.org/abs/2610.20162
Blog: dream.ischool.illinois.edu/b…
Led by Xucheng Yu, with @Haibo_Jin97, Xiaoqi Wang, Runze Kong.
#EMNLP2026 is happening, check out our findings paper on multi-agent systems on transcriptomics data.
A sumary blog of the work and follow-up papers can be found here:
dream.ischool.illinois.edu/b…
Thanks for the report, it's quite impressive we got there within a very short time of launching.
@TopCitedAI helps boost your company / project's visibility online and especially through major chatbots like ChatGPT, Claude, Gemini, Grok etc.
If you have built something and worry about how to get more users, let's talk!
TopCited AI, founded by #iSchoolUI students and alumni with Assistant Professor @HaohanWang, placed fourth in the 2026 Grainger Engineering Tech Startup Challenge! The startup was 1 of 7 finalists selected from 100+ applications across 35 universities.
▶️ bit.ly/4if1kMw
ALT From left to right: MSIM student Ethen Huang, PhD student Haibo Jin, Haojing Luo, and Assistant Professor Haohan Wang.
Just want to put a random note here for my students that they will no longer be evaluated based on papers accepted at top conferences for graduation.
In fact, it's never that case, but a public note might be helpful for them now.
Pleasing (AI) reviewers will only draw you to mediocre.
Most multi-agent LLM systems make agents talk in text, so every agent re-reads the whole history.
What if they handed each other their internal memory, the KV-cache, instead?
Four works from our lab give that memory four jobs: communicate, learn, score and check.
Caveats: agents must share one model; AIME has 30 problems, so one problem is 3.3 points; two of the four papers are preprints.
For speeding up shared context at the systems level, see KVCOMM by @HankYe1 and colleagues.
Agent Primitives: arxiv.org/abs/2602.03695
DiffMAS: arxiv.org/abs/2604.21794
KV-PRM: arxiv.org/abs/2607.09153
ReLAT: arxiv.org/abs/2606.06252
Blog: dream.ischool.illinois.edu/b…
Led by Haibo Jin, Ye Yu, Peng Kuang and Xiaopeng Yuan, with @Yushun_Dong, @LijunYu0 and others.
