Haofei Yu
@haofeiyu44
Followers
273
Following
458
Media
11
Statuses
99
CS PhD student @UofIllinois | previously CS undergrad @ZJU_China, MS @LTIatCMU | ex-intern @Apple @TencentGlobal @MITIBMLab
Champaign, IL
Joined April 2022
🚀 Excited about automatic research? What if we can combine graphs and LLMs to simulate our interconnected human research community?. ✨ Check our latest paper ResearchTown: Simulator of Human Research Community (. #AI #LLM #AutoResearch #MultiAgent #Graph.
4
14
68
RT @JiaxunZhang6: ⚠️ Rogue AI scientists? 🛡️ SafeScientist rejects unsafe prompts for ethical discoveries. Check out paper ➡️ ( https://t….
0
7
0
RT @xwzliuzijia: 💥Time-R1 is here! Can a 3B LLM truly grasp time? 🤔 YES! . Excited to share our new work, Time-R1: Towards Comprehensive Te….
0
3
0
RT @youjiaxuan: 🤯NeurIPS 2025 might break records as the most submitted-to academic conference ever. One of our submission IDs is already ~….
0
18
0
🌐 Join our growing community passionate about auto-research: 🔗 Connect, collaborate, and innovate together. Let's redefine the future of research automation!.
0
0
1
📦As a python package, tiny-scientist targets minimizing users' efforts to start. Try with two lines of code:.from tiny_scientist import TinyScientist.scientist = TinyScientist(model="gpt-4o").
1
0
0
Want to augment the pipeline?.Add your own tools—like:.🎨 a figure drawer.🔎 a code searcher.📊 a paper searcher.Every tool can be plugged in at any stage of research to enhance your agent's abilities.
1
0
0
At the heart of tiny-scientist are four intelligent agents:.🧠 Thinker – for brainstorming and ideation.💻 Coder – writes and refines code.🖋️ Writer – drafts your paper.🔍 Reviewer – critiques and improves it.Each plays a role in your auto-research pipeline.
1
0
0
Tiny-scientist is designed with 3 core modules. All of them are modularized and separated for better development. 📦 formatter – handles complex LaTeX formatting.🧠 core functions – simulate research workflow.🛠️ tools – pluggable extensions to boost capabilities
1
0
0
🧪 Want an AI-generated paper draft in just 1 minute? Or dreaming of building auto-research apps but frustrated with setups?.Meet tiny-scientist, a minimal package to start AI-powered research:.👉 pip install tiny-scientist.🔗 #AIAgent #pythonpackages.
3
10
30
RT @youjiaxuan: Attended an #ICLR workshop on self-improving LLMs. Asked what might be my best audience question, to renowned panelists Yos….
0
12
0
RT @youjiaxuan: 🚀Congrats to U Lab member @Kunlun_Zhu for leading MultiAgentBench—the first comprehensive benchmark evaluating both collabo….
0
1
0
Congrats @KiwiJWY.
SGLang Powers lm-eval-harness, The Gold Standard in LLM Eval! . Harness is a highly fair and authoritative framework in LLM evaluation. We’ve previously used it internally for calibration. Now, thanks to the efforts of Jinwei Yao, Jin Pan, Xiaotong Jiang, and Qiujiang Chen,
0
0
1
🎉 Our GitHub repo ( has reached 100 stars! 🌟. 🚀 Thanks for the support! Feel free to explore, use, and share your feedback to help us create a better platform for automatic research in one step:. `pip install research-town`.
🚀 Excited about automatic research? What if we can combine graphs and LLMs to simulate our interconnected human research community?. ✨ Check our latest paper ResearchTown: Simulator of Human Research Community (. #AI #LLM #AutoResearch #MultiAgent #Graph.
0
0
5
🙌 Huge thanks to my incredible collaborators: Zhaochen Hong, @Zirui_Cheng_, @Kunlun_Zhu, Keyang Xuan, @KiwiJWY, Tao Feng. 💡 Special thanks to our amazing ULab advisor, @youjiaxuan, for invaluable guidance!.
0
0
0
Thread(11/11).🌍 Our vision goes beyond a paper—we built an open-source platform for every developer to create auto-research workflows with ease. ⛽️ Join us to build the open-source auto-research platform together.
1
0
1
Thread(10/11).🤝 One key strength of ResearchTown is its ability to simulate cross-domain collaboration—e.g., researchers from NLP and criminology working together. 🌟 This leads to the generation of underexplored ideas, highlighting the power of multi-agent simulation.
1
0
0
Thread(9/11).📈 Experiments reveal that performance improves as the number of research agents increases.
1
0
1
Thread(8/11).🧾 Experimental results show that ResearchTown achieves realistic simulation performance on both collaborative paper writing and review writing tasks.
1
0
0
Thread(7/11).🧪 Based on node masking evaluation, we introduce a novel research benchmark: ResearchBench. 📄 It includes 1,000 paper writing tasks and 200 review writing tasks, designed to test simulation algorithms' ability to reproduce real-world papers and reviews.
1
0
0
Thread(6/11).🌐 Using a graph-based simulation not only unifies the definition of diverse research activities but also enables scalable evaluation via node masking prediction. 🤖 It allows the evaluation of research simulations using similarity scores without humans.
1
0
0