TokenMix Research Lab · 2026-07-06

Do Not Ask AI to Write Related Work First: Build a Literature Map Instead

Many related work sections do not fail because the author has not read enough papers.

They fail because the papers are introduced one by one without a clear relationship.

This problem becomes even more common when people use AI. A typical prompt looks like this:

Please write a related work section based on the following papers.

The model may produce a fluent paragraph. It may mention every paper. It may even sound academic.

But the result often reads like a list:

• Paper A did this;
• Paper B proposed that;
• Paper C evaluated another method;
• Paper D found a different result.

The section looks complete, but the reader still does not understand:

• why these papers belong together;
• whether they support, extend, contradict, or merely resemble each other;
• what problem remains unsolved;
• why your study is necessary.

A stronger workflow is to ask AI to build a literature map before asking it to write the related work section.

What Is a Literature Map?

A literature map is not just a summary table.

It is a structured representation of relationships among papers.

Instead of asking:

What does each paper say?

A literature map asks:

How do these papers form a research landscape?

A good literature map helps you see:

• which papers study the same problem;
• which papers use similar methods;
• which papers use different datasets, populations, benchmarks, or scenarios;
• which findings are consistent;
• which findings conflict;
• where the real research gap is.

Related work is not a paper list. It is a relationship graph.

Why Direct AI Writing Often Produces Weak Related Work

Directly asking AI to write related work usually creates four problems.

1. The Model Follows the Input Order

If you paste ten papers into a prompt, the model often follows the order you provided.

The output becomes:

Smith et al. studied X.
Wang et al. proposed Y.
Li et al. evaluated Z.

This is closer to an annotated bibliography than a related work section.

A real related work section should be organized by research logic: problem, method, evidence, limitation, and gap.

2. Keyword Similarity Is Treated as Conceptual Similarity

Two papers may share the same keyword but study different problems.

For example, the same keyword may appear in:

• a theoretical paper;
• a system implementation paper;
• an application paper;
• an evaluation paper;
• a background section only.

If you do not separate topic and method first, AI may group papers together simply because they contain similar words.

3. The Gap Sounds Fluent but Remains Vague

AI often writes sentences like:

However, existing studies still have several limitations.

This sentence is not wrong, but it is usually too vague.

A useful research gap should specify what kind of limitation it is:

• an object gap;
• a method gap;
• a data or context gap;
• an evidence gap;
• a measurement gap;
• a validation gap.

If the gap cannot be tied to object, method, data, context, or evidence strength, it may only be a writing formula.

4. The Model May Overstate Relationships

The most dangerous issue in related work is not awkward wording. It is inaccurate relationship judgment.

For example:

• a preliminary result may be written as a field consensus;
• a finding from a narrow scenario may be generalized;
• two incomparable studies may be compared directly;
• a missing paper in your input may be mistaken as a true research gap.

That is why the first AI task should be relationship analysis, not final prose.

The Five Lines of a Literature Map

A practical literature map can be built with five lines:

1. Topic line;
2. Method line;
3. Data / context line;
4. Finding line;
5. Research gap line.

These lines are not extra paperwork. They are the structure behind a strong related work section.

01. Topic Line: What Problem Does Each Paper Study?

The topic line answers:

What is each paper really about, and which papers belong to the same theme?

Do not only look at keywords. Look at the research question.

A useful topic-line table can include:

This step prevents you from putting unrelated papers in the same paragraph just because they share a keyword.

When writing related work, ask:

• Do these papers answer the same question?
• Do they study the same object?
• Do they support the same argument?

If not, they probably should not be forced into one paragraph.

02. Method Line: How Did the Papers Study the Problem?

The method line answers:

What method does each paper use, and can the results be compared?

Depending on your field, methods may include:

• controlled experiments;
• modeling;
• simulation;
• surveys;
• interviews;
• case studies;
• theoretical derivation;
• benchmark evaluation;
• system implementation;
• data analysis.

Method matters because conclusions from different methods may not be directly comparable.

For example:

• a small interview study;
• a large public dataset experiment;
• a simulation study;
• a real-world deployment;
• a theoretical proof.

They may all address the same broad topic, but the evidence they provide is different.

A method-line table can look like this:

This helps you write a synthesis sentence instead of a paper-by-paper list:

Existing studies have approached this problem through three main methodological routes: controlled experiments, simulation-based analysis, and data-driven modeling.

03. Data / Context Line: Are the Research Boundaries the Same?

The data / context line answers:

Under what objects, datasets, conditions, or scenarios were the conclusions obtained?

This is critical in many fields:

• medicine: population, diagnosis, protocol, sample source;
• bioinformatics: dataset, sequencing platform, preprocessing;
• mechanical engineering: operating condition, simulation setting, load, temperature;
• AI and computer science: dataset, benchmark, task definition, metric;
• education and social science: region, school type, sample size, participant group;
• management and communication: platform, industry, time window, event background.

The same method may behave differently in different contexts.

A useful table:

This line helps you write a concrete gap:

Most existing studies were evaluated in controlled or benchmark settings, leaving their performance under real-world operational conditions less understood.

04. Finding Line: Which Conclusions Agree or Conflict?

The finding line answers:

What did the papers find, and how do the findings relate to each other?

A weak related work section says:

• A found X;
• B found Y;
• C found Z.

A stronger section asks:

• Are X and Y consistent?
• Does C challenge A?
• Are the differences caused by method, data, or context?
• Which findings are stable?
• Which findings remain uncertain?

A finding matrix can include:

This is where the related work section gains analytical depth.

05. Research Gap Line: What Exactly Is Missing?

The research gap line answers:

What has existing research not solved, and why does my study matter?

Useful gaps often fall into four categories.

Object Gap

Existing studies cover one object but not yours.

Examples:

• adult populations but not adolescents;
• benchmark data but not real-world data;
• a single device but not multi-device coordination;
• English corpora but not Chinese contexts.

Method Gap

Existing methods are limited for your problem.

Examples:

• weak interpretability;
• poor robustness;
• heavy need for labeled data;
• sensitivity to outliers;
• offline-only setting;
• poor scalability.

Data / Context Gap

Existing studies were conducted in different conditions.

Examples:

• controlled laboratory settings;
• public benchmarks only;
• no complex operating condition;
• no long-term data;
• no cross-platform validation.

Evidence Gap

Existing conclusions are not fully validated.

Examples:

• small sample size;
• lack of control group;
• no ablation study;
• no cross-dataset validation;
• limited metrics;
• no failure-case analysis.

A strong gap should follow this pattern:

Existing studies have addressed X, but remain limited in Y. Therefore, this study focuses on Z.

Prompt: Ask AI to Build a Literature Map

Use this prompt before asking AI to write the related work section.

I will provide a set of papers and my research question.

Do not write the related work section yet.

Your task is to build a literature map that helps me understand the relationships among these papers.

Organize the map along five lines:

1. Topic line:
   - What problem does each paper study?
   - Which papers belong to the same topic group?
   - Which papers only share similar keywords but study different problems?

2. Method line:
   - What method does each paper use?
   - Are the methods alternatives, extensions, complements, or not directly comparable?

3. Data / context line:
   - What data, object, experimental condition, or application context does each paper use?
   - Which papers are close to my research context?
   - Which findings cannot be directly transferred to my study?

4. Finding line:
   - What is the main finding of each paper?
   - Which findings are consistent?
   - Which findings conflict or remain uncertain?

5. Research gap line:
   - What remains insufficient in existing research?
   - Is the gap about object, method, data/context, or evidence strength?
   - Which gaps are truly relevant to my research question?

Finally, output:
- a paper grouping table;
- a gap table;
- a suggested related work structure;
- which papers should be discussed together;
- which papers should only be briefly mentioned.

Requirements:
- Do not invent information not provided;
- Mark uncertain points as "to be verified";
- Do not fabricate citations;
- If the material is insufficient, list missing information first.

My research question:
[Paste your research question]

Paper materials:
[Paste titles, abstracts, notes, or citation information]

From Literature Map to Paragraph Plan

After building the map, do not immediately ask AI for final prose.

Ask for a paragraph plan first.

Below is my literature map.

Do not write the final related work section yet.

Please generate a paragraph plan:

1. The central claim of each paragraph;
2. Which papers should be discussed in each paragraph;
3. The relationship among those papers;
4. How each paragraph should transition to the next;
5. Which paragraph introduces the research gap;
6. Which papers should only be briefly mentioned.

Requirements:
- Organize paragraphs by research logic, not paper order;
- Each paragraph must have an argument;
- The gap must be specific to object, method, data/context, or evidence strength.

Literature map:
[Paste the literature map]

Then Write the Related Work Draft

Only after the map and paragraph plan are clear should you ask AI to draft the section.

Based on the literature map and paragraph plan above, write a related work draft.

Requirements:
1. Do not list papers one by one;
2. Start each paragraph with the research logic, then cite specific papers;
3. Do not invent authors, years, or references;
4. Do not turn uncertain claims into definite conclusions;
5. Make the research gap concrete;
6. The final paragraph should naturally lead to my research question;
7. After the draft, provide a claim-evidence checklist showing which paper supports each key claim.

The claim-evidence checklist is important. AI-generated related work should not be treated as final text until you verify:

• whether every citation exists;
• whether each claim is supported by the cited paper;
• whether relationships among papers are overstated;
• whether the gap really exists;
• whether important classic or contrary studies are missing.

How to Handle Different Numbers of Papers

Fewer Than 5 Papers

Do not expand too early.

Ask whether the papers can even belong to the same paragraph.

Check:

• Are they about the same topic?
• Do they share a method or object?
• Which paper is only background?
• Which paper directly supports your gap?

10-30 Papers

This is the ideal range for a five-line literature map.

Ask AI to classify the papers, then manually check:

• whether important papers are placed in the wrong group;
• whether the gap relies too heavily on a few papers;
• whether an entire research direction is missing.

More Than 50 Papers

Do not paste everything at once.

Use a batch workflow:

Batch papers by topic or search source;
Create small maps;
Merge small maps into a master map;
Manually check representative papers;
Generate a paragraph plan;
Then write the draft.

Large literature sets need structure before writing. Otherwise, AI will compress away important distinctions.

GPT, Claude, Gemini: How to Divide the Work

Do not think of one model as responsible for the entire workflow.

A more practical division is by task:

A common workflow is:

Long-context model → build the literature map
Reasoning model → challenge the gap
Writing model → improve the draft
Verification step → check claim-evidence alignment

The principle remains the same:

Map relationships first. Write prose later.

Copyable Literature Map Template

# Literature Map Template

## My Research Question

[Write one sentence describing the problem my study addresses]

## Literature Scope

- Must include:
- Optional:
- Exclude for now:

## Topic Line

| Paper | Research question | Topic group | Relation to my study |
|---|---|---|---|

## Method Line

| Paper | Method type | Method core | Strength | Limitation |
|---|---|---|---|---|

## Data / Context Line

| Paper | Research object | Data/context | Metric | Difference from my study |
|---|---|---|---|---|

## Finding Line

| Paper | Main finding | Evidence strength | Relation to other papers | Caution |
|---|---|---|---|---|

## Research Gap Line

| Gap type | What existing studies have done | What is missing | Supporting papers | Relation to my study |
|---|---|---|---|---|

## Related Work Paragraph Plan

| Paragraph | Central claim | Papers to discuss | Relationship | Transition |
|---|---|---|---|---|

Conclusion

A strong related work section does not simply include many papers.

It explains:

• how the problem has been studied;
• how different research routes relate to each other;
• what the current evidence supports;
• where the existing boundaries are;
• why your study is necessary.

AI can accelerate this process, but it should not skip the analysis step.

A safer workflow is:

Paper list
→ literature map
→ gap check
→ paragraph plan
→ related work draft
→ claim-evidence check
→ human revision

If your related work is stuck, do not first ask, “How can I make the writing more academic?”

Ask first:

Do I understand the relationships among the papers?

Once the relationships are clear, the writing becomes much easier.