Reasoning with LLMs (UdS; Summer 2025/26)
Seminar on reasoning with large language models, covering inference, learning, and analysis.
Instructor: Yuekun Yao
Term: Summer
Course Description
Large language models are now capable of a wide range of human-like tasks, but reasoning, planning, and systematic generalization remain major open problems. This seminar surveys recent work on reasoning with LLMs, including prompting strategies, verification, learning-based approaches, and analyses of what these models are actually doing.
Prerequisites
Students should have a solid background in NLP and machine learning. Familiarity with language models, transformer architectures, in-context learning, and supervised fine-tuning is assumed.
Registration
Please contact me by email if you are interested in presenting a paper. Students should indicate their top paper preferences from the schedule readings.
Course Format
The seminar is discussion-driven. Each week we focus on one or more recent papers. Students are expected to come prepared to discuss the central ideas, assumptions, limitations, and open questions.
Each participant leads one session during the semester. Session leaders should:
- present the paper’s main goal, method, and contributions;
- raise discussion points about strengths, weaknesses, and implications;
- engage actively with questions from the group.
Evaluation
For students taking the seminar for 4 credits:
- Presentation: 60%
- Participation in discussion: 40%
For students taking the seminar for 7 credits:
- Presentation: 30%
- Participation in discussion: 20%
- Term paper: 50%
Term Paper
Students taking the 7-credit version may write either a survey paper or a replication paper. The expected paper length is at most 8 pages in ACL, ICLR, or NeurIPS format, excluding references.
Contact
Please email ykyao.cs@gmail.com with questions about the seminar.
Schedule
| Week | Date | Topic | Materials |
|---|---|---|---|
| 1 | 2025-04-15 | Course logistics | |
| 2 | 2025-04-22 | Overview | |
| 3 | 2025-04-29 | Chain-of-thought prompting | |
| 4 | 2025-05-06 | Rationale exploration | |
| 5 | 2025-05-13 | Task decomposition | |
| 6 | 2025-05-13 | Chain-of-thought reasoning without prompting | |
| 7 | 2025-05-20 | Verification | |
| 8 | 2025-05-27 | Step-by-step verification | |
| 9 | 2025-06-03 | Supervised fine-tuning | |
| 10 | 2025-06-10 | Bootstrapping | |
| 11 | 2025-06-17 | Reinforcement tuning | |
| 12 | 2025-06-24 | Test-time scaling | |
| 13 | 2025-07-01 | Latent reasoning | |
| 14 | 2025-07-08 | Multi-hop reasoning analysis | |
| 15 | 2025-07-15 | Reasoning or reciting? |