Frontiers in Foundation Models

Course Description: This seminar (Advanced Topics in Foundation Models) explores the frontier of Large Language Models (LLMs) and Multimodal Foundation Models. We move beyond standard autoregressive architectures to examine next-generation paradigms, including hierarchical reasoning, latent reasoning, diffusion-based language modeling, and reinforcement learning for reasoning. The course emphasizes both theoretical understanding and practical safety alignment, drawing heavily on recent breakthroughs in mechanism interpretability, test-time compute scaling, and agentic workflows.

Who Should Enroll?

We welcome students from diverse backgrounds who are interested in learning about SOTA foundation models.

For PhD Students: This is an excellent opportunity to read the latest literature and bring your own research into the course. You are encouraged to align the final project with your thesis or ongoing research topics.
For Graduate & Undergraduate Students: If you are interested in Deep Learning and Generative AI, this course will help you get up to speed with the state-of-the-art.
Technical Background: Familiarity with Deep Learning basics (Python/PyTorch) is recommended to get the most out of the course projects, but we will support students in defining projects that match their skill levels (e.g., surveys, reproductions, or novel research).

Grading & Logistics

📍 Time & Location: Friday, Periods 3 & 4 (12:10 PM - 3:10 PM) in Hill 009

Grading Breakdown

Final Research Project and Presentation: 60%
Paper Presentation & Participation: 40%
- Slides (due 1 day before class): 10%
- Presentation Delivery: 20%
- Daily Participation: 10%

Schedule (Spring 2026)

Week	Topic & Readings
Week 1	Introduction & Interpretation Overview of Foundation Models. Safety and Interpretation.
Week 2	LLM Frontiers DeepSeek-V3.2 Kimi-K2
Week 3	Video Generation Wan (arXiv:2503.20314) Hunyuan Video 1.5
Week 4	Part 1: Guest Lecture (1 hr) Xingyu Fu (Princeton) Topic: MLLM (benchmarks, thinking with images) Part 2: Student Presentation Paper: Lessons from the Trenches on Reproducible Evaluation of Language Models
Week 5	Part 1: Guest Lecture (1 hr) Didac Suris (Meta Super Intelligence Lab) Topic: SAM 3 (Vision Foundation) Part 2: Student Presentation Paper: Dino v3 (Vision Representation)
Week 6	Part 1: Guest Lecture (1 hr) Wenhao (NVIDIA Scientist) Topic: Nvidia, autonomous driving (arXiv) Part 2: Student Presentation Paper: Highly accurate protein structure prediction with AlphaFold
Week 7	Part 1: Guest Lecture (1 hr) Ruoshi Liu (Amazon FAR Scientist) Topic: Foundation Model for Robotics Part 2: Student Presentation Paper: DAPO, the art of scaling RL for LLMs
Week 8	Part 1: Guest Lecture (1 hr) Congyue Deng (MIT) Topic: Action Representation Part 2: Student Presentation Paper: TBD
Week 9	Hierarchical Reasoning Hierarchical Reasoning Model (arXiv) Mixture of Depth (arXiv) Scaling up Test-Time Compute with Latent Reasoning
Week 10	Data & Models Openthought, DataComp-LM, s1
Week 11	Part 1: Guest Lecture (1 hr) Yushi Hu (Meta FAIR) Topic: Multimodal Part 2: Student Presentation Paper: Why Do Multi-Agent LLM Systems Fail?
Week 12	New Architectures Diffusion LM: Llada, dream-7B
Week 13	Reasoning Paradigms Parallel thinking, Latent thinking, Soft thinking
Week 14	Hybrid Reasoning Transfusion, Diffusion Forcing
Week 15	Final Final Project Presentations