Frontier AI models are shaped by choices across an entire technical stack: architecture, data, optimization, post-training, inference-time computation, tools, memory, and evaluation. This advanced course studies how modern systems are built and how researchers determine whether they work. We will read primary research papers and technical reports, with an emphasis on recent work.

The course has three themes:

  1. Architectures: the modern decoder stack, mixture-of-experts models, efficient attention, long-context methods, state-space and hybrid models, memory, and multimodal integration.
  2. Training and post-training: data and scaling, optimization, open training recipes, supervised fine-tuning, preference learning, distillation, reinforcement learning with verifiable rewards, reasoning models, and inference-time scaling.
  3. Agents and harnesses: tool use, retrieval and memory, planning, coding agents, computer-use agents, agent evaluation, reinforcement learning for agents, reliability, and safety.

The format combines instructor lectures with paper presentations and class discussion. The instructor will open each theme with one or two lectures that establish shared concepts. Each designated paper session will then feature two teams of four students, with each team presenting one paper. The course also plans to include three or four guest lectures. A final team project gives students an opportunity to investigate a focused question through a controlled experiment, replication and extension, evaluation audit, or carefully evaluated system.

See the schedule for the evolving paper list and the logistics page for the overall course structure and policies.

Prerequisites

Prior exposure to deep learning and natural language processing is assumed. Either CPSC 4770, CPSC 4520, or CPSC 4800, CPSC 4830, or permission of the instructor.

Course information

  • Instructor: Arman Cohan
  • Meetings: Monday and Wednesday, 2:35 to 3:50 p.m.
  • Announcements and course materials: Canvas and email
  • Questions and course logistics: Ed Discussion
  • Readings: Publicly available papers and technical reports. There is no required textbook.

Frontier AI develops quickly, and this is the first offering of the course in the current format. The reading list, order of topics, guest schedule, and some implementation details may change as the semester develops. Students will receive reasonable notice if there are substantive changes.