Scientific Harness
An execution and modification substrate that orchestrates tools, workflows, compute, memory, and verification—and constructs new task-specific capabilities on demand.
夏应策 / Yingce Xia, Ph.D.
Self-evolving systems for scientific discovery.
Associate Professor at Zhongguancun Academy and former Principal Research Manager at Microsoft Research. His research vision is Agentic Science: building verifiable AI systems for scientific discovery and using the principles of scientific inquiry to make AI itself more capable, reliable, and self-improving.
Research Vision
Agentic Science: Self-Evolving Scientific Design
My research vision is to build verifiable, self-evolving systems for scientific discovery. Scientific agents should do more than generate candidates or operate a fixed set of tools: they should plan and execute long-horizon research, learn from trajectories, failures, experimental outcomes, and expert feedback, and improve the models, tools, evaluators, memory, and decision processes that make up the discovery system itself. I ground this work in drug discovery, protein design, and medicine and healthcare, where computational, experimental, and real-world feedback can drive continuous system improvement. The long-term goal is twofold: to use AI to advance science, and to use the methods of science—hypothesis formation, experimentation, verification, and cumulative evidence—to advance AI itself.
Application domains
Drug Discovery
Protein Design
Medicine & Healthcare
Current research focus
Agentic Post-Training
Learning from scientific traces, execution failures, expert corrections, and verified outcomes to improve planning, tool use, recovery, and efficiency.
Multimodal Scientific Reasoning
Extending NatureLM toward evidence-grounded reasoning across text, sequences, structures, and other scientific modalities, with an emphasis on mechanism inference and testable hypotheses.
Towards RSI in Science
A verified self-evolution loop that diagnoses bottlenecks, modifies models, tools, evaluators, memory, or workflows, and retains only improvements that generalize.
Experience
Appointments, education, and leadership
Associate Professor, Zhongguancun Academy
On duty for the AI + Education direction.
Principal Research Manager, Microsoft Research
Led research programs in generative AI, AI for Science, drug discovery, and NLP.
Joint Ph.D. program between University of Science and Technology of China
and Microsoft Research Asia.
Supervisors: Prof. Tie-Yan Liu and Prof. Nenghai Yu.
Microsoft Research Asia PhD Fellowship, 2016.
Research topics: dual learning and machine translation.
Thesis: A Theoretical and Empirical Study of Dual Learning.
B.S., University of Science and Technology of China
School of Information Science and Technology.
Research Contributions
Selected research contributions
- NatureLM The first science foundation model for universal scientific entity generation, translating natural language goals into molecules, proteins, materials, RNA, and text, with wet-lab validation in infectious disease drug discovery and de novo DNA binding protein design. Paper: Nature Language Model: Deciphering the Language of Nature for Scientific Discovery. Technical report, 2025.
- TamGen A target aware molecular generation model developed with GHDDI to accelerate Gates Foundation supported tuberculosis drug discovery. Paper: TamGen: drug design with target-aware molecule generation through a chemical language model. Nature Communications, 2024.
- BioGPT The first GPT style generative pretrained Transformer model specialized for biomedical text generation and mining. Paper: BioGPT: generative pre-trained transformer for biomedical text generation and mining. Briefings in Bioinformatics, 2022.
- Dual Learning A pioneering closed-loop learning paradigm for inverse NLP tasks, later integrated into MSRA's WMT'19 systems with 8 first-place wins and transferred to real-world translation products. Papers: Dual Learning for Machine Translation, NeurIPS 2016; Dual Supervised Learning, ICML 2017; Microsoft Research Asia's Systems for WMT19, ACL WMT 2019.
See the full publication list or visit Google Scholar.
Highlights
Awards and achievements
- 2025 Released Nature Language Model, serving as project lead for a 50-member virtual team across China, the U.S., and the U.K.
- 2024 Developed TamGen and delivered it to GHDDI and additional institutions; designed hit and lead compounds targeting tuberculosis.
- 2023 Released BioGPT, the first AI model to achieve human parity on the PubMedQA benchmark.
- 2019 Contributed to Microsoft Research Asia's WMT'19 systems with 8 first-place finishes; received Excellent PhD Thesis of Chinese Academy of Sciences.
- 2018 Received the Chinese Academy of Sciences President Award; contributed to human parity on Chinese-to-English news translation.
- 2016 Microsoft Research Asia PhD Fellowship.
Community
Teaching and professional service
Professional service
- 2023-2026 Senior Area Chair for NeurIPS ED track 2026; Area Chair for NeurIPS, ICML, and ACL.
- 2020-2023 Senior Program Committee member for AAAI, IJCAI, and AAMAS.
- 2018-2025 Reviewer for Nature Methods, Nature Communications, CVPR, ICCV, ECCV, EMNLP, and TPAMI.
Teaching
- 2023 Chinese Academy of Sciences: advanced machine learning, including deep learning basics, meta learning, generative adversarial networks, and dual learning.
- 2021, 2022 AI School China and Microsoft: advanced machine learning, automatic machine learning, dual learning, and natural language processing.
- 2019, 2021 Tsinghua University: advanced machine learning topics including meta learning, generative adversarial networks, dual learning, and NLP.