Rules, Logistic, Grading

This will update through September.

---

1) Philosophy of the Class

Updates to Grading

This past July, a UCLA‑led team (Huang & Yang) demonstrated that the publicly available Gemini 2.5 Pro (later GPT-5 and Grok) could attain a gold‑level score at IMO with structured prompting and verification—it solved 5 of the 6 problems.
We believe LLM are good tools both for students and for researchers and we expect and encourage the students to use them.
This, however, raises questions on the utility and effectiveness of grading student based on problem sets which are far more standard than IMO problems.

We decided to remove the problem sets and to introduce an oral presentation. Unlike the previous years, we are thus grading only attendance and the projects through their presentation and their deliverables (see below).

• We will devote more time to the projects with respect to the last year.
• We will grade a few early deliverables of the project in October.
• The goal of the presentation is to understand to what extent you own the project you performed and to what extent are you aware of its positioning in SLT and of the related topics SLT.

Expectations

This class is aimed at PhD students. We believe our role is to provide resources to learn, resources to research, and help them to our best towards this goal. We are happy (and expect) to help substantially with the projects during in person office hours.

We will devote more time to the projects with respect to the last year. The students will have little more than 2 months for the project (see below).
We expect a number of these projects to be submitted to ICML2026 (end of Jan) or COLT2026 (beginning of Feb).

Prerequisites

Part II is designed for students with a good background in ML.

We will make extensive use of basic notions of calculus, linear algebra and probability. The essentials are covered in class and in the math camp material. We will introduce a few concepts in functional/convex analysis and optimization. Note that this is an advanced graduate course and some exposure on introductory Machine Learning concepts or courses is expected: for course 6 students prerequisites are 6.041 and 18.06 and (6.036 or 6.401 or 6.867).

---

2) Timeline and Deliverables

Sep 26 — Groups & Proposals.

Task: Fill the Google Form indicating:

• Your group (1 or 2 people).
• The intent of working on either:
  • 3 projects from the official list, or
  • 2 projects from the official list + 1 self-proposed project.
• If you self-propose a project, attach a pdf project proposal of ~0.5/1 pages.

Oct 10 — Literature Reviews & Implications

Task: For each of the 3 projects you indicated (it is a group submission), hand in a pdf (written in latex) of 3–4 pages:

• 2-3 Pages of literature review — motivations, prior attempts, related empirical and mathematical work (not exhaustive, but substantial and in detail).
• 1 final page of implications — at least 3 detailed consequences for ML theory and practice.

Oct 17 — Plan

Task: If you want to pick the problem you proposed you need to schedule a meeting with TP and PB and get our approval by this date. Hand in 1-2 pages:

• Communicate succintly:
  (i) which problem you picked,
  (ii) what you plan to achieve,
  (iii) how you plan to achieve it (what to prove and/or what to plot),
  (iv) a realistic timeline/milestones (split in 5 weeks).

Nov 3 — Initial Checkpoint

Task: Hand in a short commentary (1-2 pages) containing initial results (first plots or a sketch of proof).

First 2 weeks of November — Discussion

Task: Show up at office hour to discuss your work.

Nov 25 and Dec 2–4 — Presentation

Task: 8-minutes sharp presentation per group.

• Upload up to 10 slides (of content) by the end of the day of the presentation.
• Discuss: motivations, related work, open question, results, implications.

Dec 10 — Final Paper

Task: Submit the final paper and a link to public code repository.

• Length (main text): min 5 pages, expected ~8 pages, max 9 pages. Followed by references and Appendix.
• Appendix: allowed for proofs, ablations, additional observations.
• Code: a python notebook runnable on google colab has to be povided with a runnable small experiment.

---

3) Grading (10 Participation, 90 Project, 5 bonus)

Outline:

• Up to 10 points for active paticipation in the class. We want to see you in class and actively engaging with the project.
• Up to 90 points (possibility of +5 bonus) for the project-related activities (see here for details).

Classes are conducted in-person.

The deadlines are on Fri and are meant to be end of the day (11:59pm local time).

Project Grading:

Delay penalty: -3 points for every deadline missed after the first one, -3 points if delay is more than 3 days (only applicable to Sep 26, Oct 10, Oct 17, or Nov 3).

Conflict penalty: −15 points to the whole group if the project overlaps with current or past research or coursework of a member.

---

Sep 26 — Form filled: (no grade)

Oct 10 — Literature reviews & implications: (up to 15 points)

• Each of the three documents are graded from 1 to 5.

Oct 17 — Plan: (no grade, 5 possible bonus points)

• +3 bonus points if you pick a problem from the list.
• +2 bonus points if it is particularly well planned.

Oct 31 — Initial Checkpoint: (up to 10 points)

• The purpose is make sure you have started early enough and are on track.
• Ideally, 15 points means we believe you fruitfully worked 9h prioritizing well.

First 2 weeks of November — Discussion: (no grade, but show up)

Nov 25 and Dec 2–4 — Presentation: (up to 25 points)

• 25 points project presentation question answering
• Penalty if the presentation goes overtime: −5 points.

Dec 10 — Final paper: (up to 40 points)

Graded for:

• Execution of the idea (10 points): Completeness of proofs; reasonable assumptions; tightness of conditions; counterexamples as needed; justification of experimental setting; strong baselines, ablations, uncertainty reporting, compute noted; conclusions match evidence.
• Positioning and Related Work (10 points): Precisely situates contributions vs. strongest prior art; explains why prior methods don’t already solve it.
• Clarity (5 points): quality of the organization; succintness of the text; quality of graphics; readability of the mathematics.
• Novelty (5 points): New problem/technique/proof idea, or a sharp recombination yielding genuinely new insight/bounds/behavior.
• Limitations (5 points): Candid limits (failure modes, data/compute, assumptions); proposes mitigations/tests.
• Implications and Significance (5 points): Considers impacts/ethics when relevant; Spells out in detail non-trivial consequences to both theory and practice of machine learning; Presents experiments hinting at implications where relevant. Significance to SLT of deep learning or DL theory in general.

---

4) Others

Classes are conducted in-person.