Role-play Seminar

After the opening instructor lectures, the class shifts to student-led presentations and panel discussions, following Alec Jacobson and Colin Raffel’s role-play seminar approach. Each session covers two papers, listed on the Calendar; students take on distinct roles for each, and the instructor moderates a debate and ties the work to its broader context.

Roles

Every paper is covered by four students, each taking a different role from the five below. You choose which role you take when you sign up, so the mix of perspectives changes from paper to paper.

Whatever the role, prepare to talk about the paper for roughly 8 minutes — about ten slides.

🔬 Scientific Peer Reviewer

Treat the paper as unpublished and currently under submission to a top venue in this area, with you assigned as one of its reviewers. Complete a full review following the NeurIPS reviewer guidelines, and commit to a recommendation to accept or reject. A strong review does three things: it summarizes the problem, the contribution, and why it matters; it states strengths and weaknesses plainly, including what most needs improvement; and it checks for fundamental flaws — are the assumptions sound, do the experiments support the claims, is prior work properly credited? In class, walk us through your review and put your questions to the room.

🏺 Archaeologist

This paper was found buried underground in the desert. You are the archaeologist who must work out where it sits in the context of previous and subsequent work. Find and report on one older paper, cited by this one, that substantially influenced it, and one newer paper that cites this one. Then step back: why did this paper happen when and where it did, and did the field take the idea up the way the authors intended, or repurpose it for something else?

📊 Data Analyst

Talk about the datasets, the curation, the benchmarks, and the metrics. Be as specific as possible and hunt for the little details that matter. Don’t expect full answers from the paper — investigate and research. If a method is only ever demonstrated on synthetic two-dimensional examples, say so and explain what that does and does not establish.

👾 Hacker

Talk about how to implement the algorithm. Study the open-source repositories and tell us the status quo: officially released (training, inference, or both?), community replication, work in progress, or nothing at all. Dig into the implementation details that decide whether it works — how the data is batched, how the loss is really computed, how much compute a realistic run takes. Best form of presentation: a Colab notebook.

💡 Academic Researcher

You are a researcher starting a new project in this area. Propose an imaginary follow-up project — not merely one that builds on this paper, but one that is only possible because this paper exists and succeeded. Keyword: imaginary. No idea is a bad idea in brainstorming. Bring in your own background here.

Expectations

• Every student presents around four times during the semester. With 20 papers and four roles for each, there are 80 panel slots to spread across the class; the exact count per student is confirmed once sign-up closes.
• Presenters prepare slides and, where the role calls for it, a short demo.
• All students are expected to come having read the assigned papers and ready to contribute to the discussion.

Logistics

• Sign-up. Claim a paper and a role by adding your name to the shared spreadsheet linked from Brightspace. Each paper takes four sign-ups, one per role. The sheet opens at the start of the semester and the deadline is announced with it.
• Uploading your material. Slides — and a notebook, if your role calls for one — are due by 11 AM on the day of the class. Guidance on where to upload them will be provided.

Grading (30%)

Most of this comes from your four panel appearances. Each is graded on whether you took the role seriously, engaged with the paper in depth, and brought the room something it could not have gotten from reading the abstract.