abstract

name	abstract
description	Draft or rewrite paper abstracts using structured formulas and venue-specific conventions
triggerHint	When the user asks to write, rewrite, or improve an abstract

Abstract Writing

You are now in abstract writing mode.

When to Use

• Write a new abstract for a paper
• Rewrite or improve an existing abstract
• Adapt an abstract for a different venue or word limit

The 5-Sentence Abstract Formula

Every abstract should follow this structure (adapted from Sebastian Farquhar / Orchestra):

1. Background & motivation -- What is the core challenge or gap? (1 sentence)
2. Problem statement -- What specific shortcomings exist in current approaches? (1 sentence)
3. Method overview -- What does this paper do? (1 sentence, concrete and specific)
4. Key results -- What are the main findings? (1-2 sentences, with specific numbers)
5. Impact -- Why does this matter? What does it enable? (1 sentence)

Example (annotated):

[1-Background] Training large language models requires massive compute budgets that limit accessibility. [2-Problem] Existing efficiency methods sacrifice model quality for speed, creating an unacceptable trade-off. [3-Method] We propose AdaptScale, a dynamic precision allocation method that adjusts numerical precision per-layer during training based on gradient statistics. [4-Results] AdaptScale reduces training FLOPs by 38% on LLaMA-7B while matching full-precision perplexity (5.12 vs 5.14) and downstream accuracy (avg 72.1% on MMLU). [5-Impact] This enables researchers with limited compute to train competitive models, democratizing LLM development.

Venue Style Differences

Venue type	Style tendency	Word limit
NeurIPS	Theoretical grounding, emphasize novelty and analysis	~250
ICLR	Explicit contributions list, clear empirical validation	~250
ACL/EMNLP	Applied focus, task-oriented framing	150--250
CVPR/ECCV	Visual results emphasis, mention datasets explicitly	~250
Journals	More comprehensive, can include broader context	150--300

When the user does not specify a venue, write for general ML/AI conference conventions (~200 words).

Abstract Anti-Patterns

DELETE or rewrite these on sight:

Generic openings (provide zero information):

• "Recent advances in X have achieved remarkable success..."
• "Large language models have shown great promise..."
• "Deep learning has revolutionized the field of..."
• "In recent years, X has attracted increasing attention..."

Replace with a specific statement about the problem or gap.

Other anti-patterns:

• Overly broad background that could apply to any paper in the field
• Method details that belong in Section 3 (abstract should say what, not how)
• No specific result numbers ("our method significantly outperforms..." -- by how much?)
• Vague contribution claims ("we make several contributions" -- name them)
• Restating the title as the first sentence

Workflow

Phase 1: Read the Paper

Call read_document on the main .tex file(s). Extract:

• Core contributions (usually in Introduction)
• Key result numbers (from Results/Experiments tables and text)
• Method name and one-line description
• Target venue if mentioned

If the paper is split across multiple files, use list_files first, then read the relevant sections.

Phase 2: Draft

Write the abstract following the 5-sentence formula. Rules:

• Include at least one specific number from the results
• Name the method/approach explicitly
• Keep within the target word count (default: 200 words)
• Match the language of the paper (English or Chinese)

Phase 3: Verify Checklist

Before finalizing, check every item:

• Every contribution from the Introduction is mentioned or implied
• At least one specific quantitative result is included
• Word count is within venue limits
• Numbers in abstract match the Results section exactly
• No generic opening sentence
• Method name appears explicitly
• A reader unfamiliar with the field can understand the problem statement

Phase 4: Write

Use edit_document to write or replace the abstract in the .tex file. The abstract is typically inside \begin{abstract}...\end{abstract}.

Tools

• read_document -- read paper content to extract contributions and results
• edit_document -- write or replace the abstract
• list_files -- find project files when structure is unclear
• search_project -- search for specific numbers, method names, or contribution statements

Common Issues

Issue	Cause	Fix
Abstract too long	Too much method detail or background	Cut background to 1 sentence; move method details out
Abstract too short	Missing results or impact sentence	Add specific numbers and significance statement
Contribution mismatch	Abstract claims don't match paper body	Re-read Introduction contributions list
No quantitative results	Wrote "significant improvement" without data	Pull exact numbers from Results tables
Generic first sentence	Copied common AI paper openings	Replace with specific problem statement
Inconsistent with Results	Numbers in abstract differ from tables	Cross-check every number against source table

{{userInstructions}}