M2 · Lesson 3 — Reading Papers Effectively

Critical Questions
to Ask Every Paper

Reading is passive. Questioning is what makes you a researcher.

01
M2 · L3 — Why This Matters

The difference between reading and researching

Papers are arguments,
not facts

"Every paper is trying to convince you of something. Your job is to decide whether to be convinced."

Critical questions help you:

  • Spot hidden assumptions in formulations
  • Identify unfair comparisons in experiments
  • Find gaps that become your research opportunities
  • Write better reviews at venues
02
M2 · L3 — The Framework

Ask these every time

5 questions for
every paper

What is the actual problem?

Not what they say — what do they actually solve? Are these the same thing?

What assumptions are hidden?

Every formulation buries assumptions. What are they assuming is true?

Is the baseline comparison fair?

Are hyperparameters tuned equally? Are strong recent baselines missing?

Would this generalise?

Does it work on different datasets? Different domains? Cold-start?

What does the ablation reveal?

Remove each component — what actually drives the performance gain?

Your own question

Add one specific to your research area.

03
M2 · L3 — Question 1

Question 1

What is the
actual problem?

Papers often claim to solve one problem while actually solving a narrower or different one. Look for the gap.

K-RagRec example

Claimed: Addresses LLM hallucination and lack of domain knowledge in RS.

Actual: Addresses this only when a KG exists for the domain — which is not always available in practice.

Ask yourself: If I removed the KG, what problem is left? Is the problem actually the KG retrieval design, not LLM hallucination itself?

04
M2 · L3 — Question 2

Question 2

What assumptions
are hidden?

Authors rarely announce their assumptions. You have to find them by reading the formulation carefully.

Assumption type 1

Data assumptions

Does the method assume explicit ratings? Dense interactions? A specific popularity distribution?

Assumption type 2

Model assumptions

Independence of ratings? Linear relationships? That the KG is complete and accurate?

Assumption type 3

Evaluation assumptions

Random negative sampling? Leave-one-out? These choices dramatically inflate reported numbers.

K-RagRec hidden assumption

The popularity threshold p=50% assumes that popularity follows a clean binary split. In reality, popularity is a continuous power-law distribution — the threshold is a design choice presented as fact.

05
M2 · L3 — Question 3

Question 3 — the most commonly violated

Is the comparison fair?

  • Were all baselines tuned with equal effort?
  • Are the strongest recent baselines included?
  • Are baseline implementations from their original papers or reimplemented (possibly weaker)?
  • Is the same train/test split used for all methods?

K-RagRec: fair or not?

Baselines included
KG-Text, KAPING, G-retriever, GraphToken
Fair
Traditional RS baselines
LightGCN, SASRec, BPR — none included
Missing
Negative sampling
19 random negatives — too easy
Weak
06
M2 · L3 — Questions 4 & 5

Questions 4 and 5

Generalisation &
Ablation

Question 4 — Generalisation

Would this work elsewhere?

Ask: different domain? Different language? No KG available? Cold-start users?

K-RagRec: They do test generalisation (Table 6) — MovieLens→Amazon Book zero-shot. Good. But still only movie/book domains with KG coverage.

Question 5 — Ablation

What drives the gain?

When they remove the GNN Encoder, accuracy drops 37–45%. That's the biggest single component.

Implication: The graph encoding is doing most of the work — not the retrieval strategy itself.

💡 Research insight: If encoding matters most, could you get similar results with better text encoding instead of a KG?

07
M2 · L3 — Practice

How to use these questions

Make it a habit,
not a checklist

The 5 questions aren't meant to be run mechanically. They're lenses — use the ones that apply.

  • For any new paper: Q1 and Q3 always apply
  • For papers in your area: Q2 and Q5 matter most
  • For papers you're reviewing: all 5, systematically
  • For papers you want to extend: Q5 reveals the gap
"The best research ideas come from taking a paper's ablation study seriously."

Remove component X → performance drops → X is important → can you do X better? That's a paper.

08
M2 · L3 — Key Takeaways

What to remember

01

Papers are arguments

Decide whether to be convinced. Every result has a context that limits it.

02

Missing baselines > weak baselines

The most revealing thing about an experiment is often what was left out.

03

Ablations reveal gaps

The component with the biggest drop is often the best target for your extension idea.

Next: M2 · L4 — Reading Experimental Sections

09
← → arrow keys to navigate