research-paper-writing

AI-Assisted Academic Paper Writing: Best Practices

A practical reference for researchers using LLMs to draft, refine, and polish scientific papers. Last updated: 2026-03

---

Table of Contents

1. Overview & Philosophy
2. Workflow Architecture
3. Prompt Engineering by Section
4. Academic Writing Style Guide
5. Common Pitfalls & Mitigations
6. Tool Ecosystem
7. Section-Specific Strategies
8. Prompt Template Library

---

1. Overview & Philosophy

The Right Mental Model

AI is a writing accelerator, not a writing replacement. The researcher owns:

• The research question and hypothesis
• Experimental design and data
• Interpretation of results
• Intellectual contribution claims

AI assists with:

• Drafting prose from structured notes
• Improving clarity and flow
• Suggesting transitions and hedging language
• Reformatting and restructuring sections

Core Principle: Input Quality Determines Output Quality

The single biggest lever is what you feed the model. Bullet-point notes → mediocre draft. Structured outline with key claims, evidence, and citations → publication-ready prose.

Bad input:

Write the introduction for my paper about transformers.

Good input:

Write an introduction for a NeurIPS paper. Context:
- Problem: Transformer attention is O(n²) — bottleneck for long sequences
- Gap: Existing sparse attention methods sacrifice accuracy for speed
- Our contribution: FlashLinear — linear attention with no accuracy loss on GLUE/LongBench
- Tone: confident but measured, cite [Vaswani 2017] and [Dao 2022]
- Length: ~300 words, 3 paragraphs (hook → gap → contribution)

---

2. Workflow Architecture

Phase 1: Pre-Writing (Research → Structure)

Raw research notes
      ↓
Outline with claims + evidence per section
      ↓
Key citations identified (Semantic Scholar / Elicit)
      ↓
Section-by-section bullet points