Events

W375 Westgate Building
10:00AM

With rapid advances in machine learning (ML), data-driven methods have become a powerful tool for decision-making. Yet in real-world deployment, we repeatedly face three bottlenecks spanning all stages of the ML pipeline: the observational scarcity gap (data), the policy synthesis gap (modeling/learning), and the human–AI alignment gap (deployment). My research develops generative-AI methods to close these gaps. In this talk, I will show: (1) how flow matching addresses the observational scarcity and policy synthesis gaps by selectively reusing shifted offline data to improve RL sample efficiency and synthesize expressive policies under combinatorial constraints, and (2) how LLM agents bridge the alignment gap by incorporating expert guidance directly into planning, yielding optimization strategies consistent with stakeholder priorities. Throughout, I will highlight collaborations with field partners that translate these algorithmic ideas into real-world impact—from environmental health, including conservation planning in African national parks, to human health, including work with the WHO on optimizing HIV testing policies in South Africa. I will close with a broader vision for generative AI that scales real-world decision-making—and how it can directly benefit society.

Additional Information:

Lingkai Kong is a Postdoctoral Fellow at the Harvard John A. Paulson School of Engineering and Applied Sciences. He earned his Ph.D. in Computational Science and Engineering from the Georgia Institute of Technology. His research advances generative AI by integrating it with optimization and reinforcement learning to address high-stakes decision-making challenges in public health and sustainability. Dedicated to bridging theory and practice, Lingkai collaborates closely with field partners to translate algorithmic innovations into tangible social impact. His work has been published in top-tier venues such as ICML, NeurIPS, and ICLR, and he has delivered tutorials at major data science conferences like KDD. He is also a recipient of the Otto & Jenny Krauss Fellowship.

Details...

W375 Westgate Building
10:00AM

ABSTRACT As AI models become increasingly powerful and integrated into society, their "black box" nature poses significant risks and distrust. In this talk, I will introduce principled methods to scientifically understand AI systems and improve their trustworthiness across the full spectrum, from training processes to model mechanisms to data features. First, I will discuss techniques to trace model behavior back to specific training updates, enabling training data assessment, model auditing, and credit assignment. Next, I will address the need for revealing the internal mechanisms of Large Language Models (LLMs), particularly regarding their knowledge representation and safety during post-training. Finally, I will introduce a structure-aware framework for examining how data features drive AI decisions, allowing non-AI experts to interpret and effectively use AI in healthcare and science applications. The presentation concludes with a look toward the future: controlling advanced agentic AI systems and creating frameworks where AI empowers humans.

Additional Information:

BIO

Shichang Zhang is a postdoctoral fellow at the Digital Data Design Institute at Harvard University. He earned his Ph.D. in Computer Science from the University of California, Los Angeles (UCLA), an M.S. in Statistics from Stanford University, and a B.A. in Statistics from the University of California, Berkeley (UCB). His research focuses on developing principled methods to understand and improve the trustworthiness of AI systems, with applications in high-stakes domains such as science and healthcare. His work has been published in leading venues, including NeurIPS, ICML, ICLR, ACL, WWW, and ISR, and was highlighted in a Nature News Feature for its educational impact. He delivered a comprehensive tutorial on Explainable AI at NeurIPS 2025 and has industry experience at Amazon Web Services (AWS) and Snap Research. He is a recipient of the J.P. Morgan AI Ph.D. Fellowship, the Amazon Ph.D. Fellowship, and the NeurIPS Outstanding Paper Award, and has received multiple Outstanding Reviewer Awards (ICML 2022; KDD 2023, 2025).

Details...

Hammond Building, Room 220
1:25 – 2:35 p.m.

Details...

W375 Westgate Building
10:00AM

Given the fast development of technology and automation systems, researchers are generating tons of data from scientific experiments to facilitate discoveries, especially in the areas of biology and medicine. However, large-scale biomedical data make it challenging for biologists or medical researchers to more effectively utilize these data with robust and powerful methods for novel and reproducible discoveries. Therefore, I pioneered the research for building Knowledge-Enhanced AI Co-Scientist System to overcome these challenges. First, I will show the importance of having a fair evaluation framework to rank the contributions of foundation models (FMs) and AI Scientists for biological data analysis and demonstrate the importance of prior knowledge. Second, I will show the strength of leveraging knowledge to gain insights in analyzing multi-modal data for clinical and biomedical applications by building AI Bioinformaticians and biologists. Finally, I will discuss our progress of building an AI Bioengineer for accelerating scientific discoveries which facilitate generation tasks in multiple areas to enclose this robust and extensive system, as well as some future directions I would like to explore.

Additional Information:

Tianyu is a PhD candidate from Yale University working on artificial intelligence (AI) and machine learning (ML) methods for accelerating scientific research. Previously, Tianyu obtained bachelor’s degrees with the highest honors from a joint program between Zhejiang University and UIUC. Having a good understanding for both machine learning concepts and biological knowledge, Tianyu has led multiple research projects based on multimodal AI Agents and Large Reasoning Models for computational biology and chemistry research. He has published several papers in the high-impacted science journals (Nature and Cell series) and ML conferences (NeurIPS, ICML, etc.). The model he designed has also been adopted by some well-known companies such as Genentech. He also served as reviewers and organizers for key conferences in this field. His research is supported by fundings from NIH, NSF, Google, and OpenAI. 

Details...