Course project

Course project

All students will write a mini-paper as a final project—empirical, theoretical, or both—using the NeurIPS conference format.

Working in groups

You may work solo or in groups of up to three (special permission required to exceed).
Effort scales with team size (2× for two authors; 3× for three).
Include a contributions paragraph naming each author’s concrete contributions.
Paper length is 3 + n pages (where n = group size), excluding references and the contributions paragraph.

Milestones

Date	Deliverable	Points
Oct 12	Project matching spreadsheet	3 points
Oct 31	1–2 page progress report	7 points
Dec 4	In-class presentation	10 points
Dec 12	Final report	20 points

Project matching spreadsheet (3 pts)

Each student will add information about topics they find interesting to a project matching spreadsheet. The course staff won’t match students, but the goal is that this process will help students find others with similar interests.

Progress report (7 pts)

Each group will submit a short progress report of 1-2 pages. Describe your project and partial progress. Your writeup should include answers (implicitly or explicitly) to the following simplified version of the Heilmeier Catechism:

What are you trying to do?
How is it done today, and what are the limitations?
What’s new in your idea, and why might it succeed?
If successful, what difference will it make?
What’s the first thing you will try or have already tried to test the idea? Include any initial progress you’ve made so far.

Final project grading: Course relevance

The final project is graded out of 30 points (10 for the in-class presentation and 20 for the final report). Naturally, the project must be clearly about AI for algorithmic reasoning & optimization. Projects not related to the course will receive 0 of the 30 final project points.

In-scope examples include:

Algorithmic reasoning as a lens to understand ML models (aligning with the lectures from 10/7–10/16)
ML for graph/constraint/NP-hard problems (aligning with lectures 10/21–10/30)
Learning to formalize optimization problems (aligning with lecture 11/6)
ML to guide/configure solvers (aligning with lectures 11/11, 11/13)
Theoretical guarantees for learned algorithms/heuristics (aligning with lectures 11/18, 11/20)

You will receive feedback about whether your project is in-scope when you turn in your progress report and will have the opportunity to turn in a revision if not.

Final paper: Novelty & insight (10 pts)

Originality (3 pts)

Examples include:

New algorithm or heuristic
New theoretical result
New problem formulation/evaluation/analysis protocol or benchmark
New combination of known elements that yields a surprising result

Score	Description
3	Clear novelty with strong supporting evidence.
2	Straight-forward extension of an existing idea.
1	Minor variation/ablation; weak distinction from prior work.
0	No identifiable novelty.

Depth of analysis (3 pts)

You should include a coherent analysis that explains how and why your idea works or fails. Tools to illustrate depth include:

Ablations isolating methodological components
Fair baselines/metrics (e.g., both classical and learned methods)
Generalization tests (e.g., out-of-distribution instances)
Thorough proofs
Counterexamples/failure modes (e.g., why a hypothesized theorem doesn’t hold) (Not all depth tools will be relevant to all projects.)

Score	Description
3	Relevant depth tools executed well.
2	Some depth but gaps (e.g., missing ablation or weak baselines).
1	Superficial analysis; results reported with little interpretation.
0	No analysis beyond raw outcomes.

Potential impact (2 pts)

It should be clear who benefits from your contributions and how. Impact can come in many different forms, including:

Methodolgical: E.g., improves solution quality/gap, runtime, etc.
Theoretical: E.g., clarifies when/why learned components help or fail.

Score	Description
2	Clear, well-argued value to a broader community.
1	Some value but narrow or under-substantiated.
0	Unclear who benefits or how.

Future work (2 pts)

Include specific, realistic next steps.

Score	Description
2	Specific 2–3 next steps.
1	Plausible but high-level ideas with limited detail.
0	Vague or no future work.

Final paper: Writing & completeness (10 pts)

Any hallucinated reference will result in 0 out of 10 points for this section. “Hallucinated” includes any citation that cannot be verified or misrepresents bibliographic details (incorrect title, author list, venue, or year). Citing arXiv is fine, but if the paper appeared in a conference/journal, it’s preferable to cite the conference/journal version.

Structure & formatting (2 pts)

Your report should comply with the required structure.

Score	Description
2	NeurIPS format; within 3 + n pages (references/contributions excluded); clear sectioning; specific contributions paragraph (if applicable).
1	Slightly under length; contributions paragraph vague.
0	Major non-compliance with formatting instructions.

Problem formulation (2 pts)

It should be clear exactly what problem you solved and under what assumptions.

Score	Description
2	Precise definitions, assumptions, objectives, and evaluation setup (if applicable).
1	Mostly clear but some missing assumptions/ambiguity.
0	Unclear scope; it’s not obvious what is solved or proved.

The report should have accurate positioning and proper attribution.

Score	Description
2	Accurate/sufficient citations; clear where prior methods fall short.
1	Coverage OK but thin; shortcomings of prior work unclear.
0	Missing or very weak context.

Readability (2 pts)

The writing, notation, and figures should be clear.

Score	Description
2	Clear narrative; consistent notation; legible, labeled figures/tables referenced in text.
1	Understandable but with jargon jumps/undefined symbols or hard-to-read figures.
0	Disorganized; illegible figures; math without explanation.

Reproducibility (2 pts)

The reader should be able to re-run your experiments or verify your proofs. Examples include:

Empirical: runnable code + configs (public repo or share privately with the course staff’s GitHub handles vitercik and zzyunzhi); data sources/creation, seeds, splits, metrics, hardware, hyperparams; clear how to reproduce each table/figure.
Theoretical: complete proofs (appendix allowed); assumptions stated; auxiliary lemmas included or cited clearly (e.g., Lemma 3.5 by Vitercik et al., ‘25).

Score	Description
2	Readers can reproduce or check results/proofs.
1	Partially reproducible (e.g., missing seeds/configs/proof details).
0	Not reproducible; key details missing.

Final presentation (10 pts)

Problem & context (3 pts)

It should be clear exactly what problem you solved.

Score	Description
3	Plain-English problem statement with sufficient context for the audience to follow.
2	Mostly clear; minor jargon.
1	Heavy jargon in describing the problem.
0	Totally unclear what the talk is about or why it belongs here.

Method (3 pts)

Your method should be clearly explained.

Score	Description
3	Key definitions/assumptions/steps clearly described; notation defined on first use.
2	Understandable but skips a few definitions.
1	Heavy jargon in describing the method; key steps left implicit.
0	Method/evidence largely unexplained.

Takeaways of results (2 pts)

The “so-what” should be clear, as should the potential impact.

Score	Description
2	Clear “so-what” and interpretation of the results.
1	Numbers/theorems with little interpretation.
0	No takeaways beyond raw results.

Delivery (2 pts)

Emphasize slide readability, pacing, and timing.

Score	Description
2	Clean slides (large fonts, consistent notation); one idea per slide; confident delivery (not script-reading); on time (± a few seconds).
1	Minor readability/pacing issues.
0	Illegible figures/text; script-reading; unable to answer basic questions; major pacing issues.