Every RS paper follows the same skeleton.
Knowing what each section really does changes how you read.
| Section | Official Purpose | What It's Really Doing |
|---|---|---|
| Abstract | Summarise the paper | Sell you on reading further |
| Introduction | Motivate the problem | Establish the gap they claim to fill |
| Related Work | Survey prior work | Argue why existing work is insufficient |
| Methodology | Present the model | The actual contribution ← slow down here |
| Experiments | Validate the model | Selectively show where it works best |
| Conclusion | Summarise findings | Restate + hint at future work |
Three things always hide in the abstract. Find them in 60 seconds.
What gap or failure in existing work motivates this paper?
What did they propose in one sentence?
What improvement, on what dataset, by how much?
K-RagRec example: Problem = LLM hallucination + noisy RAG · Approach = KG subgraph retrieval with GNN encoder · Claim = 41.6% avg improvement over sub-optimal baseline
It tells you exactly what they're claiming to have done — and you can hold them accountable to it in the experiments section.
You're only looking for 3 things:
To make their contribution look bigger, they may understate what existing methods already do. You'll learn to spot this in L3.
Everything else is scaffolding. This is where the contribution lives.
How do they formally define the task? What are the inputs and outputs?
Usually one central formula defines the model. Find it and decode it.
What loss function? What optimiser? What's being minimised?
No traditional RS baselines like LightGCN or SASRec. All comparisons are within the KG-RAG-for-LLM family only.
| What You See | What It Often Means |
|---|---|
| "Outperforms all baselines" | On these specific datasets, with these hyperparameters |
| "Simple and efficient" | Simpler than the most complex competitor |
| "Novel framework" | We combined two existing ideas |
| Weak baselines only | The strongest competitor wasn't included for a reason |
| Single dataset experiments | Didn't generalise to other datasets |
This isn't cynicism — it's critical reading. Good researchers write this way due to space constraints and competition.
Intro and Related Work are rhetoric. Slow down at Methodology and Experiments.
Authors must defend every bullet. Hold them to it in experiments.
What baselines are missing? What datasets weren't tested? Gaps tell you more than results.
Next: M2 · L2 — The 3-Pass Reading Method