Frontier 01
Connect time series with language.
Generate, edit, and forecast time series with language and real-world context.
View sequence researchMachine Learning & Data Intelligence · ShanghaiTech University
Trustworthy AI
for the real world.
We build multimodal foundation models that observe real signals, reason about possible futures, and remain understandable.
Our big picture
Intelligence should meet the world where it is: continuous, multimodal, uncertain, and consequential.
MLDI develops learning systems from the structure of real data outward. We model irregular signals, build adaptive foundation models, turn models into tool-using agents, and open their internal mechanisms to inspection.
These are not separate themes. They are layers of one research program: understand the data, generalize the model, ground the decision, and make the mechanism legible.
Research atlas
One program, four connected frontiers
Multimodal models, grounded reasoning, and mechanisms we can inspect. Select a direction to explore the original research.
Research impact
Advancing capability.
Building evidence for trust.
Open forecasting models that people use. Reasoning whose steps can be evaluated. Internal mechanisms that can be explained and tested through intervention.
cumulative model downloads
Kairos is available in three compact, open forecasting checkpoints.
Hugging Face · 3 official Kairos models · all time · 27 Sep 2026 Explore the models ↗fewer parameters
Kairos-small uses 23M parameters versus Sundial’s 128M, with lower normalized MASE in the reported GIFT-Eval comparison.
MASE 0.748 vs 0.750 · lower is better · paper comparison Read the benchmark setting ↗AIME24 accuracy
CRM reaches 43.3%, compared with PURE’s 26.6%, when verifiable rewards are disabled in the reported RL experiment.
Qwen2.5-Math-7B · Pass@1 · same evaluation setting Read the reasoning experiment ↗Capability needs a meaningful test.
Our agent and evaluation research received two paper awards. Together, these works ask how models solve a task—and whether the result answers the actual request.
Outstanding Paper · EACL 2024
MLCopilot
Use experience from previous experiments to solve new machine-learning tasks.
Paper and award record ↗Best Paper · IEEE VIS 2024
VisEval
Check whether a generated visualization answers the question, beyond whether the code runs.
Paper and award record ↗Selected research
Inside the papers
Original figures and demonstrations, with the contribution and evidence alongside. Open any figure to inspect it at full size.
The people behind the research
MLDI at ShanghaiTech
The Machine Learning and Data Intelligence Group is part of the School of Information Science and Technology at ShanghaiTech University, led by Prof. Kan Ren.
We work across multimodal time series, foundation models, agentic reasoning, and mechanistic interpretability, connecting fundamental methods with healthcare, finance, science, and other real-world settings.
Principal investigator
Prof. Kan Ren
Assistant Professor, Research Fellow, and PhD Supervisor at ShanghaiTech SIST. Previously Senior Researcher at Microsoft Research Asia.
- 40+papers in leading ML and AI venues
- 2best / outstanding paper awards
- OpenSeqML · Qlib · AutoRL · PhysioPro
Complete index
All research work
0 works shown
Open research