M3 · Lesson 3 — Documenting Findings

Connecting Ideas
Across Papers

Individual notes are observations. Connections between them are understanding.

01
M3 · L3 — Why This Matters

From isolated notes to a knowledge map

The field is a
conversation, not a list

"Every paper responds to prior work. If you can't see the chain of influence, you can't position your own work in it."

Connecting ideas helps you:

  • Trace how an idea evolved over time
  • Identify when papers solve the same problem differently
  • Spot when papers solve different problems similarly
  • Write related work that positions, not just lists
02
M3 · L3 — Research Lineage

Tracing an idea's evolution

MF → NeuMF → LightGCN
→ K-RagRec

2009
MF
Dot product of latent vectors
adds neural layers
→
2017
NeuMF
MLP on top of MF
adds graph structure
→
2020
LightGCN
Simplified graph convolution
adds LLM + KG
→
2025
K-RagRec
GNN + KG RAG for LLM

What this lineage reveals: Each step solves a limitation of the previous. K-RagRec solves LightGCN's lack of semantic knowledge — but inherits LightGCN's graph structure insight. This is how to write a related work section.

03
M3 · L3 — Connection Types

Four types of inter-paper connections

Not all connections
are the same

Evolves

A directly extends B

Same problem, stronger method. LightGCN → K-RagRec. The newer paper explicitly cites and improves upon the older.

Same problem

Different methods, same goal

KAPING and K-RagRec both address LLM knowledge gaps in RS — but via triple retrieval vs. subgraph encoding. Compare their trade-offs.

Same method

Same technique, different domain

G-Retriever applies graph RAG to general QA; K-RagRec applies it to RS. Transferring a method is a valid contribution.

Contradicts

Two papers disagree

One claims LLMs need KGs; another shows text-only RAG is sufficient. Contradictions are the most valuable connections — they define open questions.

04
M3 · L3 — Documenting Connections

Practical method

How to record
a connection

For each paper you finish, ask:

  • Which papers does this directly respond to?
  • Which papers solve a similar problem differently?
  • Which papers does this enable or enable extension of?
  • Does this paper contradict anything I've read?

Write 1–2 sentences per connection. That's enough.

Connection notes — K-RagRec
K-RagRec
evolves
G-Retriever

Applies same graph RAG idea to RS specifically, adds popularity filtering and re-ranking.

K-RagRec
same goal as
KAPING

Both address LLM knowledge gap. K-RagRec uses subgraphs; KAPING uses triples. Subgraphs capture structure better.

K-RagRec
misses
LightGCN

No comparison to classical CF. This is a gap that could invalidate the paper's practical claim.

05
M3 · L3 — From Connections to Research

The payoff

Connections become
your positioning

"In your related work section, every paragraph should end with: 'but none of these address X — which we solve'."

Your connection notes are the raw material. Group them by theme, not chronology.

Bad related work

"MF was proposed in 2009. Then NeuMF in 2017. Then LightGCN in 2020. Then K-RagRec in 2025..."

This is a list. It positions nothing.

Good related work

"Graph-based methods (LightGCN, NGCF) capture collaborative structure but lack semantic knowledge. LLM-based methods (KAPING, G-Retriever) add semantics but ignore structure. We propose X which addresses both."

This positions. It earns its space.

06
M3 Complete — Documenting Findings

Module 3 complete

L1 · Summary Template

One page per paper

Problem → Approach → Key Idea → Strengths → Weaknesses → Relevance. In your own words.

L2 · Literature Matrix

See the field at a glance

Rows = papers, Columns = dimensions. Consistent blanks = your research opportunity.

L3 · Connections

Map the conversation

Evolves / Same problem / Same method / Contradicts. Connections become related work positioning.

Next: M4 · Reproducing Works — from equations to pseudocode to implementation

07
← → arrow keys to navigate
1 / 7