powstik
Prove
LiveProof-of-Work · Oct–Nov: every verified proof counts. 28 days leftAll outcomes
Skills

AI Research capstone: reproduce and extend one paper

12 weeks · 2 milestones

Reproduce the main result of one published machine learning paper from your own code, add one tested extension, write it up as a short paper and defend it in a recorded talk. The last step of the Varsity Curriculum: AI Research, read by a blind review for Honors.

Part of: AI Research · Proves for: Engineers, Career, PG

Milestone map

Milestone map

2 milestones

Reproduce one paper and extend it

8–12 weeks (12–15 hrs/week)

Pick one published machine learning paper. Reproduce its main result from your own code, report where your numbers match and where they do not, then add one extension of your own (a new dataset, an ablation the authors did not run, or a change to the method) and test it. Write it up as a short paper. This is what the Honors blind review reads.

Proof required

Link the public repo (code, configs, logs), the paper you reproduced, and your write-up: the original claim, your reproduction with numbers, your extension and its result, and the limits.

What gets checked

  • The main result is reproduced from your own code, with numbers side by side with the paper's.
  • The extension is tested with a clear comparison and several seeds.
  • The write-up is honest about what did not reproduce.

Common mistakes

  • Rerunning the authors' code and calling it a reproduction.
  • An extension with no comparison.

Powstik Guide

A public repo and a short paper: reproduction of one published result plus one tested extension.

Steps

  1. Pick a paper whose main result fits your compute.
  2. Write down the claim you will reproduce, with its number.
  3. Reproduce it from your own code.
  4. Design one extension and test it over several seeds.
  5. Write it up: claim, reproduction, extension, limits.
  6. Submit for the blind review.

Template

Paper
Title:
Original claim (with number):
Reproduction (setup, results, gaps):
Extension (question, setup, results):
Limits:
Code and logs:

What gets sent back

  • The authors' code rerun without changes.
  • Numbers in the paper that are not in the logs.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

  • Is the reproduction your own code?
  • Do the numbers in the write-up match the logs in the repo?
  • Is the extension real and tested?
  • Is the write-up honest about gaps?

You'll sign in first, then come straight back here.

Defend it in a 15-minute talk

1 week (5 hrs)

Give a 15-minute talk on your reproduction and extension to at least two people who know the field, then take their questions. Record it.

Proof required

Link the recording of the talk and the questions, and the slides.

What gets checked

  • The talk states the claim, the reproduction, the extension and the limits.
  • Questions are answered from the evidence, not from hope.
  • The recording includes the questions.

Common mistakes

  • Slides with no numbers.
  • Cutting the questions from the recording.

Powstik Guide

A recorded 15-minute talk with questions.

Steps

  1. Plan: claim, reproduction, extension, limits.
  2. Practise once, timed.
  3. Give it to people who know the field and record the questions.

Template

Talk plan
Claim:
Reproduction:
Extension:
Limits:
Likely questions:

What gets sent back

  • No questions recorded.
  • No numbers shown.

Resources

What a verifier looks for

  • Does the talk match the write-up?
  • Were hard questions answered honestly?

You'll sign in first, then come straight back here.

Part of