You will need to submit 8 paper probes from the following 10 papers on the schedule: [2], [5], [6], [9], [14], [15], [24], [27], [30], [31]. You can select which 8 papers to submit a probe for. Paper probes will be due by 9:00 AM on the day of the class. You should upload the submissions through the "Assignments" section on Canvas.
If what you wrote could be about a different paper with a few words changed, it doesn't count. This rules out summarizing the method, saying the results are impressive, and restating the paper's own claimed contribution.
I will call on people at random to discuss what they submitted. Expect questions about what you submitted.
1. The most interesting or surprising thing you found. One paragraph containing three things:
What you expected, and why. Before or while reading.
What the paper actually shows. Point to a specific figure, table, equation, or number.
What it implies, or what it leaves you unsure about.
2. Something you still don't understand. Name the specific thing, and say what you tried in order to understand it. "I don't understand the method" is not an answer. "I couldn't work out why X holds when Y, and I tried Z" is (2-3 sentences)
3. A discussion question, and your answer. Your answer should state your position and the argument for it. See the prompts below.
4. A brief note on AI usage. Which tools you used, what you asked them to do, etc (2-3 sentences).
These are starting points, not questions to submit as written. They lack the specifics that make a discussion work.
Are there empirical results that are surprising, or that a regular reader would miss?
Are there weaknesses in the experiments: unfair comparisons, missing ablations, a baseline that isn't matched?
Are there claims in the paper that the experiments don't actually support?
Which previous papers does this borrow from, and what is genuinely new?
Is there a strength or weakness of the approach that the paper itself doesn't discuss?
If you were reviewing this, would you accept or reject it? Name the single specific thing that decides it.
What follow-up would you run, and which specific limitation does it address?