Degree-Thesis Writing Philosophy

“A thesis is not a pile of experiments; it is an academic story with a clear contribution.” - adapted from Neel Nanda

Contents

• Narrative Principle
• Five-Sentence Abstract Formula
• Seven Reader-Expectation Principles
• Precise Wording
• Precision First
• Micro-Writing Techniques
• Section-Specific Guide

Narrative Principle

(Core idea)

A degree thesis tells a story around one explicit research contribution.

Three pillars (must be clear before the introduction ends):

Pillar	Description	Example
What was done	1-3 concrete innovations	“This thesis proposes method X and achieves Z under condition Y”
Why it is credible	Rigorous experimental evidence	Sufficient baseline comparisons and ablation studies
Why it matters	Why readers should care	Connection to a recognized problem in the field

If you cannot state the contribution in one sentence, the thesis is not yet whole.

Five-Sentence Abstract Formula

(Adapted from Sebastian Farquhar, DeepMind)

1. Core result (“This thesis proposes...”, “This thesis demonstrates...”)
2. Why the problem is difficult and important
3. Method, including key terms for searchability
4. Experimental evidence
5. Most important quantitative result

Delete generic openings such as “With the rapid development of deep learning...”

Seven Reader-Expectation Principles

(Adapted from Gopen & Swan)

#	Principle	Rule	Example
1	Subject-verb proximity	Keep subject and predicate close	❌ “The method proposed in this thesis based on..., achieved” -> ✅ “The proposed method achieved...”
2	Stress at the end	Put sentence focus last	✅ “Through attention, accuracy increased by 15%”
3	Old information first	Known information first, new information later	✅ “Based on the analysis above, this thesis proposes...”
4	One idea per paragraph	One point in each paragraph	Split multi-point paragraphs
5	Use verbs for actions	Avoid nominalization	❌ “performed an analysis” -> ✅ “analyzed”
6	Context before display	Explain before presenting a formula	State the meaning before the formula
7	Explicit transitions	Use signals between paragraphs	“therefore,” “however,” “in addition”

Precise Wording

(Adapted from Zachary Lipton)

Be Specific

• ❌ “performance” -> ✅ “accuracy,” “inference latency”
• ❌ “significant” -> ✅ “statistically significant (p < 0.05)” or delete it

Remove Ambiguity

• ❌ “may help improve” -> ✅ “accuracy increased by 3.2%”

Avoid Empty Phrases

Delete: obviously, undoubtedly, as everyone knows, undeniably

Precision First

(Adapted from Jacob Steinhardt, UC Berkeley)

• Terminology consistency: use one term for one concept throughout
• Explicit assumptions: list every assumption before a theorem
• Intuition + rigor: pair intuitive explanation with formal proof
• Define before use: define every symbol and term before first use

Micro-Writing Techniques

(Adapted from Ethan Perez, Anthropic)

• [ ] Avoid vague reference: ❌ “This shows...” -> ✅ “The experiment shows...”
• [ ] Move verbs forward: keep the predicate near the sentence opening
• [ ] Delete fillers: actually, to some extent, quite, relatively, basically
• [ ] Active voice: ❌ “The method was applied to” -> ✅ “This thesis applies the method to”
• [ ] Quantify: ❌ “large improvement” -> ✅ “improved by 12.3%”

Section-Specific Guide

Section	Suggested Length	Core Requirement
Abstract	300-500 Chinese characters	Five-sentence formula; remove generic opening
Introduction	3-5 pages	Background -> problem -> contribution list -> organization
Literature review	5-10 pages	Group by theme; discuss each work's contribution and limitation; no more than 2 parallel citations in one sentence
Methods	As needed	Ensure reproducibility; list every hyperparameter
Experiments	Main body	State which claim each experiment validates; include error analysis
Conclusion and outlook	1-2 pages	Answer research question; actionable future work

Introduction Must Include:

• Clear research-question statement
• 2-4 contributions, each no more than 1-2 lines
• Brief method overview
• Chapter organization
• Citation integration: every cited work must include a brief description of its contribution, method, or limitation; never stack 3 or more citations in one sentence without analysis (see guide.md category 6)

Experiments Must Include:

• Specific claim validated by each experiment
• Error analysis (standard deviation vs standard error)
• Hyperparameter search range
• Computing resources (GPU model, total duration)

Time Allocation

Spend approximately equal time on:

1. Abstract
2. Introduction
3. Figures and tables
4. All other content combined

Reason: most reviewers form a judgment before reading the method chapters.

Reader order: title -> abstract -> introduction -> figures/tables -> perhaps the rest.

Revision Order (Logic -> Sentence -> Vocabulary; Irreversible)

When polishing takes several passes, follow this order and never reverse it:

1. Argument / logic: paragraph order, repeated/missing main ideas, chapter transitions.
2. Sentence structure: split very long sentences, passive to active, move information-dense components forward.
3. Vocabulary / typesetting: AI-frequency terms, number/unit format, terminology/abbreviation consistency.

Why fixed: if wording is polished first (Layer 3), a Layer 1 change may delete or merge that paragraph, wasting the work. Coarse-to-fine revision is several times more efficient.

Sources

Source	Core Contribution
Neel Nanda (Google DeepMind)	Narrative principle
Sebastian Farquhar (DeepMind)	Five-sentence abstract formula
Gopen & Swan	Seven reader-expectation principles
Zachary Lipton	Precise wording
Jacob Steinhardt (UC Berkeley)	Precision
Ethan Perez (Anthropic)	Micro-writing techniques