Prepare · Explain · Discuss

Student-led seminars

One focused question, developed over two academic hours.

Each meeting develops one fundamental question through a central paper and selected background or comparison material. Start with the older idea, uncover the main mechanism, and use evidence to assess the explanation.

Before your seminar

  1. Choose a numbered candidate from the topic pool and coordinate one focused question with the instructor. Weeks 10–11 cover mechanistic interpretability; remaining selections and all presenter assignments are TBD.
  2. Read the main paper and the relevant foundation material. Identify the minimum background needed for the mechanism. Use selected recent work to extend or question the explanation.
  3. Choose the shared pre-reading with the instructor and announce it one week before class; follow the policy below.
  4. Prepare one careful derivation, one decisive experiment and one unresolved question. State the assumptions behind each claim. A companion paper can supply a comparison without becoming a second full presentation.
  5. Bring slides or notes with complete citations. Announce the pre-reading one week before class; instructor meetings, slide deadlines and other material-sharing arrangements are TBD.

Shared pre-reading

All students complete one required shared pre-reading before each selected meeting. Presenters choose it in consultation with the instructor and announce it one week before class. Budget 30–45 minutes for one tutorial, a short paper or clearly specified sections of a longer paper.

Presenters include two guiding questions with the reading. Each student arrives with one question or point of confusion. Presenters read the main paper; it is optional for the audience unless selected as the shared pre-reading. Specific pre-readings are TBD.

A suggested structure for two academic hours

The allocation below assumes two 45-minute academic hours, totaling 90 minutes. Adjust the split to the local academic-hour convention.

15 min · Foundations

Introduce the older idea and the minimum mathematical background.

35 min · Mechanism

Explain the main paper’s mechanism, with one derivation worked through carefully.

15 min · Evidence

Present one experiment or ablation that tests the explanation.

25 min · Discussion

Discuss limitations, selected recent work, competing explanations and open questions.

Choosing a manageable scope

Every candidate has one reading bundle and a central question. Build the seminar around its main paper, using the foundations to supply the minimum necessary background and the recent work to extend or question the mechanism.

Develop one derivation and one decisive experiment in detail. A companion paper can support a comparison through selected results or sections. The scope notes identify which claims fit a two-academic-hour meeting.

Topic numbers identify candidates, not assigned weeks. Select the candidate and its question first, then agree on the sections to present with the instructor.

What makes a strong presentation?

Separate exact mathematical connections from connections that depend on assumptions, and distinguish both from empirical interpretations. Explain what a method can implement, what the trained system appears to implement, and how far the evidence supports a broader explanation.

Use a small example before the full model. Define notation, explain each assumption, and make the experiment readable enough for the audience to critique it.

Keep conclusions tied to the setting studied. Recent papers are opportunities for careful discussion; their inclusion does not mean their claims or long-term impact are settled.

Participation & assessment

For all participants

Complete the shared 30–45-minute pre-reading, consider its two guiding questions, and arrive with one question or point of confusion. Specific pre-readings are TBD. Contribute to the discussion.

Assessment

No exam. Grading components, weights, attendance requirements and any written submissions: TBD.

Research practice

Cite papers, figures, code and other sources. Explain which claims are the authors’ and which are your own analysis. Formal academic-integrity and AI-use policies: TBD.

Access & communication

Course forum, accessibility arrangements, contact information and office hours: TBD.