1The lecture map
Four acts, seven chapters. First the rules: what you deliver and how it is graded. Then the longest part, choosing a topic from eight tracks and a handful of fresh ideas. Then what good looks like, from past students' projects. And finally writing, because clarity is worth as many points as soundness. Click a chapter to watch it.
Chapter by chapter what each chapter covers, and the question it leaves for the next
- IRulesWhat you deliver, when, and how it is graded
- 1
The Final Project: Deliverables, Deadlines and Grading
Your first mini research project, worth 55% of the grade: teams of up to three with a concrete contribution statement, a poster session on Dec. 18 (file in by 17:00 on Dec. 17 for free printing), and a six-page report on Dec. 25. Writing clarity (15%) counts as much as technical soundness (15%).
→ So what should you actually build? - IIChooseEight tracks, fresh ideas, and how to pick
- 2
Choosing a Track: Agents and Training Your Own Model
Pick one of 77 briefs in eight tracks, or propose your own LLM-related topic, and follow six rules of thumb. Track A studies the agent harness at a matched budget (a new skill, an audited AI-written survey); Track B trains models, from nanoGPT speedruns to vertical models such as HuatuoGPT, AceGPT and Soundwave.
→ What if you would rather build something people interact with, or measure the models? - 3
Interactive Systems, Evaluation and Datasets
Track C makes interaction the variable: LLM NPCs on a virtual campus, tools for education, picture books and open tutorials, judged on fidelity and transfer rather than engagement. Track D builds benchmarks (a dataset with a scoring rule), physician-style rubrics, datasets and RAG systems whose citations can be checked.
→ What about projects that need no GPU at all — or real users? - 4
Society, Efficiency, Competitions and Startups
Track E argues a falsifiable position and tests it, for example with a matched-persona audit; Track F measures speed with fresh-input correctness checks; Track G turns a contest into a controlled comparison; Track H runs a startup prototype judged on retention and cost, with two student startups from our campus.
→ Beyond the tracks, where else can ideas come from? - 5
Fresh Ideas, Hospital Deployments and Picking a Topic
Seeds to brainstorm with (a robot bar, an AI seller, simulated patients, AI + VR, text-to-3D printing, text-to-game, an incubator simulator), agents deployed in the CUHK-Shenzhen hospital, and how to choose: your expertise, your background, real needs around you, and above all your interest. The teaching team is there to help, not only to judge.
→ What does a really good project look like? - IIILearn from the bestWhat strong projects look like
- 6
Outstanding Projects from Past Terms
A valuable question, solid evidence and clear communication. MedGen links paper, code and data; ConsumerSim brings a question from economics; FutureSelf reports a result against its own hypothesis; TripGenie gives an honest number with a proper baseline; and four posters run from a local-LLM safety audit to an analysis of in-context learning.
→ How do you write it up so the reader sees it too? - IVWriteShow it so the reader sees it too
- 7
Writing Your Paper: Tips That Save Points
A paper should read like a film with no 尿点: go from abstract to concrete, keep the straight path, motivate the paper and every component, use plain words, cut empty ones, define every term, be specific, spell out the logic, keep a consistent style — and related work is not a shopping list. Then a wrap-up and Q&A.
2Get started
Project guide & tracks
- Eight tracks, 77 briefs, each with a question, comparison, metric and four-week plan
- The track finder and the compute filter
- The research playbook and a copyable proposal template
- Schedule & grading and the FAQ
Writing tips & samples
- Ten writing tips from this lecture, with examples
- Report structure and the arXiv + Hugging Face + GitHub habit
- Poster samples & the 4-minute pitch, past papers
- A checklist before you submit
Report, poster and grading
- Overleaf report template (copy the project, max. 6 pages plus references)
- Poster template (slide 9)
- Grading policy: poster 10, report 45, bonuses and the ethics penalty
- TAs' office hours and research areas on the course page (slide 68)
A fair comparison, in code
- Freeze and fingerprint a held-out test split
- Compare a baseline and one change at a matched tuning budget
- Seeds, a paired bootstrap interval and an exact sign test
- Check a judge against gold labels; write traces and a booktabs table
.ipynb · CPU, no API keyYour first week
The closing advice of the lecture (slide 98), plus the rules of thumb from the project page (slide 18):
- Form your team this week, and agree on who does what; the report needs a concrete contribution statement (slide 5).
- Skim the project page tonight and pick two candidate briefs, or write down your own LLM-related question (slide 16).
- Talk to a TA whose background fits your topic, with a concrete question (slides 67–68).
- Freeze the task, a held-out split and the primary metric, and have a working baseline by the end of week one.
- One change — test one idea at a time.
- Matched budget — your method and the baseline get the same compute, tokens or model calls.
- Held-out split — never tune on your test data.
- Independent checker — don't let the model grade its own work.
- Seeds — run more than once.
- Replayable traces — log everything so someone can check it.
3Concept index
Key ideas of the lecture. Each link opens the video at the slide where the idea is explained.
4What's new in 2026
The Fall 2026 tracks add several things that were not there last year. Six to know before you pick a topic:
Skills are a project
An Agent Skill is a folder of instructions, sometimes with scripts, that an agent loads only when it needs it. New this year: build one for a real recurring task and show that it measurably helps (brief A9).
AI can write a passable paper
A paper generated end to end by Sakana AI's AI Scientist-v2 passed peer review at an ICLR 2025 workshop (scores 6, 7, 6) — reporting a negative result — and was withdrawn after review. Use AI tools, describe how, and own every claim.
Rubrics for open-ended answers
When there is no single gold answer, experts write what a good answer must contain. HealthBench has 5,000 health conversations and 48,562 rubric criteria from 262 physicians; the same rubrics can become RL rewards (brief D8).
Speedruns as a research lab
The community nanoGPT speedrun trains a 124M-parameter GPT-2-style model to 3.28 validation loss on FineWeb as fast as possible on 8×H100. Try one architecture change at matched time and plot the curve.
Agents at the doctor's desk
OpenClaw-Med runs on a small local machine in the CUHK-Shenzhen hospital: it drafts a reference from the record, then suggests diagnosis and treatment once results are in. The doctor stays in control at every step.
5Quotes & reading
"If you do not work on an important problem, it's unlikely you'll do important work."
Chapters 2–5: pick a question worth four weeks — and one you can actually attack.
"The first principle is that you must not fool yourself — and you are the easiest person to fool."
Why the rules of thumb exist: matched budgets, held-out splits and independent checkers.
"Never use a long word where a short one will do."
"Utilize" or "use"? Just say "use" (slide 89).
"一篇好的论文读起来，应该是像看一个流畅的电影，没有尿点。"
A good paper reads like a smooth film: no moment where the reader can safely leave.
Further reading
- The course's own project guide and writing tips & samples — read these first
- Vered Shwartz: Tips for writing NLP papers — the source of the example on slide 85
- Simon Peyton Jones: How to write a great research paper (talk, slides and video) · Richard Hamming: You and Your Research (transcript)
- Dror et al.: The Hitchhiker's Guide to Testing Statistical Significance in NLP (ACL 2018) — for the comparison in your report
- AutoSurvey (NeurIPS 2024) · HealthBench (2025) · Hofmann et al.: AI generates covertly racist decisions about people based on their dialect (Nature, 2024)