Incoming AP @Tsinghua_Uni, Researcher @AlibabaGroup, BS @UofIllinois, PhD @Yale
New Haven
Joined August 2021
- Tweets404
- Following662
- Followers649
- Likes1.9K
Pinned Tweet
35B total. 3B active. Built for serious agentic work. Excited to share Occamy-1.0, our team's first open-source agentic model. Occamy delivers strong results across long-horizon co-work, automation, business, tool-use,while sitting near the low-cost knee of the Pareto frontier.
Tianyu Liu retweeted
What if AI could discover better ways to discover molecules?
Meet MolRSI: an AI scientist that recursively improves its own search policy, with results validated in the wet lab.
๐งช Antibiotic discovery: 15.9% hit rate vs 8.7% for an expert-designed non-agentic workflow.
๐งฌ De novo pMHC binder design: 8 selective binder hits across 2 targets vs 0 for Claude Science.
Paper: arxiv.org/abs/2610.39714
Project page: yikunpku.github.io/MolRSI/
How it works ๐
Tianyu Liu retweeted
Introducing Hyra 2.0, the next evolution of Hunyuan Research Agent. ๐ก
From solving problems to building reusable tools, models, and workflows that accelerate future discoveries.
Read our perspective on the next chapter of AutoResearch:
hy.tencent.com/research/hyraโฆ
Tianyu Liu retweeted
Scientific research runs on software โ Jupyter, RStudio, ImageJ, specialized analysis tools. An AI agent that wants to do real science must not just understand the science; it must plan and execute correctly inside these tools, step by step. A misconfigured option, an unsaved file, a suboptimal intermediate result โ any mistake cascades. Can current AI agents actually use scientific software?
A team led by Tsinghua University (with Yale and other institutions) presents OSWorld-Science, a computer-use agent benchmark for real scientific research.
OSWorld-Science contains 146 expert-designed tasks spanning 7 scientific domains and 14 scientific software configurations, evaluated with 12 vision-language models under a unified prompt and interaction protocol. The result is sobering: the 12 models average only 34.3% task score (measuring completion degree, not just success rate). Even the leader, Claude Fable 5.1, reaches 73.7% โ leaving substantial room for improvement. The benchmark adapts OSWorld's multimodal PromptAgent for scientific software, with consistent step budgets and response limits for fair comparison. Led by Assistant Professor Tianyu Liu (Tsinghua) and Professor Lei Liu (Yale), the project includes open code, dataset, paper, and a project website.
Result: OSWorld-Science is the first benchmark for AI agents using real scientific software โ 146 expert tasks across 7 domains and 14 tools, 12 VLMs averaging only 34.3% (best: Claude Fable 5.1 at 73.7%) โ from Tsinghua and Yale, with open code, dataset, paper, and project site.
OSWorld-Science: A Benchmark for Computer-Use Agents in Real Scientific Software
Paper: discoailab.github.io/osworldโฆ
Project: discoailab.github.io/osworldโฆ
Code: github.com/DiscoAILab/OSWorlโฆ
Dataset: huggingface.co/datasets/SciAโฆ
Our report: jiqizhixin.com/articles/2026โฆ
๐ฌ #PapersAccepted by Jiqizhixin
Tianyu Liu retweeted
I'm excited to see so many recent initiatives on recursive self-improvement. I finally find some time to write a short blog for my thoughts! From well-defined to semi-well-defined problems, exciting steps forward towards AI scientists!
One problem I realize I ran accidentally into from 2024, how do we accumulate experience with LLMs for solving problems better.
And more later how this can help us build AI scientists for those I call "semi-well-defined problems".
Tianyu Liu retweeted
Hi, Iโm Yongchao Chen, an Assistant Professor at Tsinghua Universityโs College of AI and founder of ApexIntelligence.Weโve raised $50M in angel-stage funding last month.
My background: a bachelorโs from USTC, a PhD through joint training at Harvard and MIT, and research at Google DeepMind, Microsoft, and IBM.
My focus: self-evolving foundation models, robotics, and AI for Science.
The question driving my work: What if models could identify problems, form hypotheses, design experiments, validate results, and continuously improve through feedbackโall on their own?
Thatโs what Iโm working toward: enabling models to evolve themselves.
Here, Iโll share research insights, thoughts on recursive self-improvement, and lessons from building ApexIntelligence.
Researchers, founders, investors, and engineers: letโs exchange ideas and explore opportunities to work together.
Follow along, connect with me, and drop your questions in the replies. Ask me anything!
Tianyu Liu retweeted
๐๐ ๐ฒ๐ฑ๐ถ๐ฐ๐ฎ๐น ๐๐ ๐๐ต๐ผ๐๐น๐ฑ ๐ฏ๐ฒ ๐ฒ๐๐ฎ๐น๐๐ฎ๐๐ฒ๐ฑ ๐ถ๐ป ๐๐ต๐ฒ ๐น๐ฎ๐ป๐ด๐๐ฎ๐ด๐ฒ๐, ๐ฐ๐ผ๐บ๐บ๐๐ป๐ถ๐๐ถ๐ฒ๐ ๐ฎ๐ป๐ฑ ๐ต๐ฒ๐ฎ๐น๐๐ต ๐๐๐๐๐ฒ๐บ๐ ๐๐ต๐ฒ๐ฟ๐ฒ ๐ถ๐ ๐๐ถ๐น๐น ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐ฏ๐ฒ ๐๐๐ฒ๐ฑ.
Delighted to announce the launch of #MAGIC (๐ ๐ฒ๐ฑ๐ถ๐ฐ๐ฎ๐น ๐๐ ๐๐น๐ผ๐ฏ๐ฎ๐น ๐๐ป๐ถ๐๐ถ๐ฎ๐๐ถ๐๐ฒ ๐ฎ๐ป๐ฑ ๐๐ผ๐น๐น๐ฎ๐ฏ๐ผ๐ฟ๐ฎ๐๐ถ๐ผ๐ป) introduced in our new @NaturePortfolio @NatureMedicine correspondence.
๐MAGIC paper: nature.com/articles/s41591-0โฆ
๐MAGIC website: magic-ai.health/
MAGIC brings together clinicians, AI researchers, policymakers, industry stakeholders and local experts to address a critical question: does strong benchmark performance translate into safe, useful support across diverse healthcare settings?
Our initial priorities are:
โ
A global health LLM evaluation benchmark addressing multilingual coverage, regional disease burden and sociocultural adaptability.
โ
A global deployment reporting standard clarifying intended use, target users, healthcare settings, and performance and safety across populations.
Thank you to our colleagues #worldwide for helping launch the network. ๐ข๐๐ฟ ๐ด๐ผ๐ฎ๐น ๐ถ๐ ๐๐ผ ๐๐ป๐ฑ๐ฒ๐ฟ๐๐๐ฎ๐ป๐ฑ ๐๐ต๐ฒ๐ฟ๐ฒ ๐บ๐ฒ๐ฑ๐ถ๐ฐ๐ฎ๐น ๐๐ ๐๐ผ๐ฟ๐ธ๐, ๐ณ๐ผ๐ฟ ๐๐ต๐ผ๐บ, ๐ฎ๐ป๐ฑ ๐๐ป๐ฑ๐ฒ๐ฟ ๐๐ต๐ฎ๐ ๐ฐ๐ผ๐ป๐ฑ๐ถ๐๐ถ๐ผ๐ป๐.
#MAGIC #MedicalAI #AI #LLM #GlobalHealth #ResponsibleAI #HealthEquity #DukeNUS @dukenus
Tianyu Liu retweeted
I built Oh My Mouse for macOS:
โข Side buttons switch desktops
โข Reverse/smooth scrolling
โข Per-app mappings
โข Double-click Chrome tabs to close
Native Swift. Source-only beta; build it yourself. Mixed MIT/MMF licensing. Feedback welcome!
github.com/yuchenwang3/oh-myโฆ
๐ฎ What are the limits of agentsโ ability to solve scientific problems? Can they use research software as effectively as scientists?
๐งช Introducing ๐ข๐ฆ๐ช๐ผ๐ฟ๐น๐ฑ-๐ฆ๐ฐ๐ถ๐ฒ๐ป๐ฐ๐ฒ: a Computer-Use Agent / GUI Agent Benchmark designed for real-world scientific research scenarios.
10/๐ฅฐ Many thanks to our fantastic collaborators: Dingyuan Dai, Heli Qi, Lei Liu, Yinxi Li, Baiding Chen, Zijun Dou, @SteveZeng7, Qi Kang, Oliver Kunyang Sun, Eric Wang, Bo Zhou, Haixin Wang, Yufan Du, Shi Bo, Ruihan Lin, Mengqi Yuan, Dunjie Lu,
@StevenDillmann , Yiming Shi, Tina Su, Amy Xin, Minghao Liu, Xi Wang, Xu Huang, Ge Zhang, Pengyu Nie, Zhen Yang, @jietang , Juanzi Li, Weihao Xuan!
Tianyu Liu retweeted
Occamy-1.0
A cost-efficient 35B model built for long-horizon co-work agents, sitting at the low-cost knee of the Pareto frontier while staying competitive with much larger systems.
better to name them as BME problems lmao
In light of the progress in mathematics, we at Edison Scientific and FutureHouse have assembled a set of Millennium Problems for Biology. They are chosen to be very hard to solve but very easy to validate in a simple laboratory environment. Any of these, if solved, would mark a major advance in biotechnology, and most of them would contribute materially towards curing disease. These are, in some sense, the โlast reasonable evalโ for AI in biology. This was work primarily by @MichaelaThinks and myself, with contributions from many others.
Short descriptions below. The full descriptions of the problems with acceptance criteria are at the Bio Millennium Problems website, linked in the next post. Share more if you have ideas. If they meet our criteria, weโll add them to our list (with attribution and permission).
